DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Applicant’s amendments and arguments filed 10/24/2025 have been entered but are not persuasive.
Claims 1-5, 7-17, 21 and 22 are pending in this application. Claims 3, 4, and 11-14 stand withdrawn from consideration as being drawn to nonelected species.
Claims 1, 2, 5, 7-10, 15-17, 21 and 22 are under examination to the extent the read on the elected species of: a) neural networks, and b) the two specific biomarker assays are hemolytic activity, and erythrolytic activity. The election of species was made 2/28/2022 without traverse.
Rejections maintained
Claim Interpretation
The claims in this application are given their broadest reasonable interpretation (BRI) using the plain meaning of the claim language in light of the specification as it would be understood by one of ordinary skill in the art.
The term “similarity criterion” amended into claim 1 does not have clear basis in the specification, however, the specification does disclose several differing types of similarity scoring, or similarity measures which appear to meet the plain meaning of this limitation. (Tanimoto similarity, FCD scoring, internal/external diversity scoring, KL divergence score, isomer capability score, nearest neighbor score, fragment similarity score, scaffold similarity score, SMARTS score “and the like” [0118].)
This indicates that the BRI of “similarity criterion” can be met by any one or more of the listed types of similarity, or distance calculations.
The term “rediscovery criterion” amended into claim 1 does not have clear basis in the specification, however these also appear to be similarity scores. The specification states: “[0117] The goal-directed benchmark may include generating a score for an ability of the creator module 151 to generate candidate drug compounds similar to a real drug compound, a score for an ability of the creator module 151 to rediscover the potential viability of previously-known drug compounds (e.g., using a drug which is prescribed for certain conditions for a new condition or disease), and the like.”
Goal directed benchmark is defined as “[0114-0116] …The goal-direct benchmark may evaluate whether the creator module 151 generates a best possible candidate drug compound to satisfy a pre-defined goal (e.g., activity level in a design space). A resulting benchmark score may be calculated as a weighted average of the candidate drug compound scores. In some embodiments, the candidate drug compounds with the best benchmark scores may be assigned a larger weight… For example, the resulting benchmark score may be a combination of the top-1, top-10, and top-100 scores, in which the resulting benchmark score is determined by the following relationship: [algorithm and explanation of variable s].”
“[0127] In some embodiments, a goal-directed benchmark may include determining a rediscovery score for the creator module 151. In some embodiments, certain real drug compounds may be removed from the training dataset and the creator module 151 may be retrained using the modified training set lacking the removed real drug compounds. If the creator module 151 is able to generate ("rediscover") a candidate drug compound that is identical or substantially similar to the removed real drug compounds, then a high rediscovery score may be assigned. Such a technique may be used to validate the creator module 151 is effectively trained and/or tuned.”
This indicates the BRI of “rediscovery criterion” is a “high” similarity score with respect to a “real drug compound.”
The new limitation “responsive to…” is a contingent or conditional limitation.
MPEP 2111: “The broadest reasonable interpretation of a method (or process) claim having contingent limitations requires only those steps that must be performed and does not include steps that are not required to be performed because the condition(s) precedent are not met. For example, assume a method claim requires step A if a first condition happens and step B if a second condition happens. If the claimed invention may be practiced without either the first or second condition happening, then neither step A or B is required by the broadest reasonable interpretation of the claim. If the claimed invention requires the first condition to occur, then the broadest reasonable interpretation of the claim requires step A. If the claimed invention requires both the first and second conditions to occur, then the broadest reasonable interpretation of the claim requires both steps A and B.”
Generative Artificial Intelligence (Generative AI) is a subset of the mathematic discipline of Machine Learning (ML). ML is the use of statistical algorithms to learn from data, find patterns and make predictions of decisions without being explicitly programmed for each specific task, allowing systems to improve performance.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1, 2, 5, 7-10, 15-17, 21 and 22 is/ are/ remain rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea of mental steps, mathematic concepts, organizing human activity, or a natural law without significantly more.
The claims have been significantly amended. The steps have been numbered below solely to facilitate the analysis.
Applicant is directed to MPEP 2106 and the Federal Register notice (FR89, no 137 (7/17/2024) p 58128-58138) for the most current and complete guidelines in the analysis of patent- eligible subject matter. The current MPEP is the primary source for the USPTO’s patent eligibility guidance.
With respect to step (1): YES. The claims are drawn to statutory categories: Processes.
With respect to step (2A) (1): YES the claims recite an abstract idea. The claims recite an abstract idea, law of nature and/or natural phenomenon. The claims explicitly recite elements that, individually and in combination, constitute one or more judicial exceptions (JE). (MPEP 2106.04).
Mathematic concepts, Mental Processes or Elements in Addition (EIA) in the claim(s) include:
For claim 1:
The preamble: “A method for pre-clinical validation of an effectiveness of a candidate drug compound comprising:”
The preamble indicates that the method is intended to judge whether a candidate drug meets a validation condition “of an effectiveness” in a pre-clinical setting.
No particular category of drug is set forth.
No particular effect to be judged is set forth.
No particular clinical relevance (such as disease, symptom management, treatment, or prevention) is set forth.
This encompasses all possible types of small molecules, macromolecules, chemicals, small molecules, peptides, proteins, complex macromolecules, lipids, polymers, drugs, pharmaceuticals, or substances which may be applied in a “clinical” situation.
The preamble does not limit any aspect of the ensuing method steps.
Step 1) “removing a real drug compound from a training dataset to create a modified training dataset;”
Step 1) now recites “removing a real drug compound” from a “training dataset.” The nature of the real drug compound is not set forth. No initial structure, framework, class of drug or type of drug is provided. It encompasses any known possible compound, composition, small molecule, macromolecules, chemicals, small molecules, peptides, proteins, complex macromolecules, lipids, polymers, drugs, pharmaceuticals, or substances which may be applied in a “clinical” situation.
The clinical relevance, activity, property, effectiveness, or aspect of the “real drug compound” to be assessed is not set forth.
The initial training set is not provided within the bounds of the claim nor is it described as having any particular attributes, activities, or structures. The initial training set has no particular members, parameters, data values, chemical structures or “real drug compounds.” The training set is not specifically linked to the “clinical effectiveness” of the preamble.
The effect which is to be analyzed is not identified in any way.
The plain meaning of “removing” a compound from a training dataset” is to delete, mask, remove or alter the dataset to prevent the “real drug compound” from further analysis.
This limitation falls within the abstract idea grouping of Mental processes: the observation of a “real drug compound” within a dataset and removing it from a non-specific “training dataset” which is a mental process of data modification. (MPEP 2106.04(a)(2), subsection III)
Step 2) “retraining one or more trained machine learning models using the modified training dataset;”
New step 2) describes “retraining one or more trained machine learning models…”
There is no initially trained ML model present in claim 1.
There is no indication what the initially trained model was intended to carry out, nor how it was trained.
The ML is completely generic as to what is being modeled, what the specific type of ML process is performed, or how any particular training data affects any given result.
The purpose of re-training the model is completely unclear.
The structure of the ML is undefined.
The purpose of the modeling is not set forth.
How the retraining affects the model with respect to the structure of the ML, or the output, is not set forth.
The plain meaning of “training”, in the discipline of machine learning, is the iterative application of data values to a set of algorithms, which change the weight of various parameters based on the goal of the model. Training is performed to teach the algorithm to recognize patterns, adjust its internal parameters, and make more accurate predictions or decisions.
This is a mathematic concept which encompasses mathematic relationships, calculations, and formulae or algorithms. (MPEP 2106.04(a)(2) subsection I.)
Step 3) “generating a first candidate drug compound using the one or more retrained machine learning models;”
New step 3) recites using the “retrained” ML model to “generate a first candidate drug compound.” How the retrained model acts to generate any candidate is not set forth.
How the retrained model generates a candidate with respect to the “real drug” of step 1 is not set forth.
The retrained model is not generating a modification of the “real drug compound” of the first step, and the retrained model neither uses nor provides any particular framework, structure, physical or chemical property, or any desired effect. The retrained ML model does not specifically modify a previously provided structure or candidate.
The first candidate is not required to have any particular physical or chemical structure, property, function, effect, or clinical relevance. This encompasses all possible types of molecules, macromolecules, chemicals, small molecules, peptides, proteins, drugs, pharmaceuticals, or substances which may be applied in a “clinical” situation.
The generation of a new first candidate is completely without limitation, detail, or information as to how the retrained ML acts to make such a candidate.
The plain meaning of “generating a first candidate drug compound,” in the context of machine learning, is the creation of a structure (or the creation of a mathematic/ numerical/ graphical representation of a substance) that has a desired property, function, activity, effect or action. No particular property, function, activity, effect or action is provided by this limitation, or the claim overall.
This limitation falls within the category of Mental processes: thinking up candidate drug compound(s) based on the output of the retrained model, and using mental processes of comparison and judgement to determine whether the candidate should proceed. (MPEP 2106.04(a)(2), subsection III)
Step 4) “determining whether the first candidate drug compound satisfies at least one similarity criterion with respect to the real drug compound that was removed;”
New step 4) recites determining whether the first candidate compound meets or satisfies “at least one similarity criterion” with respect to the “real drug compound” of step 1.
The aspect of the “real drug compound” which is to be used to assess the similarity is not identified. No particular property, structure, function, activity or effect is specifically compared between the first candidate and the “real drug compound” using any particular similarity criteria.
The similarity criterion is not specified.
How any particular candidate can “satisfy” the similarity criterion is not specified.
This first candidate is not synthesized, or physically tested for any particular property.
The plain meaning of “similarity criterion” in the context of machine learning involves comparing and analyzing similarities between two elements, whether it is a structural similarity, a functional similarity, a similar property, a similar effect, et al. The assessment of similarity in machine learning encompasses a wide variety of scoring methodologies depending on the aspect to be compared.
A reading of the specification indicates these are mathematic distance -related scoring mechanisms, such as Tanimoto similarity, FCD scoring, internal/external diversity scoring, KL divergence score, isomer capability score, nearest neighbor score, fragment similarity score, scaffold similarity score, SMARTS score “and the like” [0118].
This limitation falls within the abstract idea category of mathematic concepts: assessing the similarity by determining the statistical differences between two elements. Each score is obtained by applying the data to a differing algorithm, equation or formula and calculating the results. (MPEP 2106.04(a)(2) Subsection I.)
Determining whether the calculated value “satisfies at least one similarity criterion” falls within the abstract idea grouping of Mental Processes: comparing the score to an aspect of the “real drug” compound, and making a judgement as to whether they are sufficiently similar. (MPEP 2106.04(a)(2) subsection III)
Step 5) “responsive to determining that the first candidate drug compound satisfies the at least one similarity criterion, determining that the one or more trained machine learning models satisfy at least one rediscovery training criterion;”
New step 5) recites, in a conditional format, determining that the [re-]trained ML models satisfy at least one rediscovery training criterion.
The rediscovery training criterion is not defined.
How the model is assessed to determine whether such a criterion is satisfied is not set forth.
The aspect of the model to be assessed for this criterion is not set forth.
The specification was probed for the meaning of this limitation. The sole description of this limitation occurs in [0127]:
“If the creator module 151 is able to generate ("rediscover") a candidate drug compound that is identical or substantially similar to the removed real drug compounds, then a high rediscovery score may be assigned. Such a technique may be used to validate the creator module 151 is effectively trained and/or tuned.”
The “creator module 151” is not clearly the same as the “retrained ML model” of this limitation.
This does not set forth how the ML is to be assessed to make this determination, as there is no clear comparison of the ML pre and post retraining, or the outputs of the trained or retrained ML models.
The rediscovery limitation does not require that the first candidate drug generated needs to be “identical or substantially similar to the removed real drug compounds” as indicated by the specification. Only a single undefined similarity criterion must be met.
No further action is taken with the first candidate drug compound, or compound information.
This limitation falls within the abstract idea groupings of Mathematic Concepts and Mental processes. The Mathematic Concept is the scoring of the “rediscovery training criterion” (calculation of a similarity) or the assignment of a similarity value. (MPEP 2106.04(a)(2) subsection I). The Mental Processes include comparing the similarity scores, comparing the models, and making a judgement as to whether the [re-]trained ML model meets the “rediscovery training criterion.” (MPEP 2106.04(a)(2) subsection III.).
Step 6) “generating a second candidate drug compound at least one candidate drug compound(s) using generative artificial intelligence,
wherein the one or more retrained machine learning models use causal inference to remove one or more potential candidate drug compounds from being considered as the second candidate drug compound to be selected, thereby forming a set of candidate drug compounds,
wherein using causal inference comprises using one or more counterfactuals, each comprising determining one or more alternative scenarios based on at least one of one or more past actions, occurrences, results, regressions, regression analyses, correlations, or some combination thereof;”
Step 6) now “uses generative AI” to generate a second candidate drug compound. The AI uses “one or more trained ML models” that use causal inference to remove one or more compounds from consideration as the second candidate drug compound.
No limitations to the drug compound type, class, structure, or information are provided.
The second candidate drug does not have any similarity to the first candidate of steps 3-5, or the initial “real drug compound” from step 1).
The basis for generating and selecting the drug to be tested is not set forth specifically.
There is no specific property, function, action, effect, or activity that the second candidate is intended have. This encompasses all possible types of compounds, compositions, macromolecules, chemicals, small molecules, peptides, proteins, complex macromolecules, lipids, polymers, drugs, pharmaceuticals, or substances which may be applied in a “clinical” situation.
There are no specific limitations as to the chemical space to be explored by the generative AI which might provide information as to the generated second candidate and its intended properties.
The claim does not provide any details about how the Generative AI operates or how the selection is made, and the plain meaning of “selected” encompasses mental observations or evaluations, e.g., a researcher’s mental identification of a property-related correlation in a data set.
No particular type of “generative AI” algorithm is set forth in this limitation. The generative AI has no particular structure, layers, generators, discriminators, or other elements. There are no limits to of the AI, other than including the ML that has a functional description, to delete a member of a group, or prevent the selection of a candidate. No positive or negative example data or ground truth data is provided.
How the “causal” inference or relationship is defined within the trained/ retrained ML is not provided. The plain meaning of “use causal inference” is to make a mental judgement as to whether a candidate drug should remain in the candidate group. The “use causal inference to remove one or more potential drug compounds from being considered” has no limits as to how or why a compound is selected to be deleted or masked.
No particular counterfactual is provided. The new limitation merely describes what counterfactuals are, in general, without providing any specific to generation and screening of candidate drug compounds.
No error, validation, or similarity comparison are specifically recited.
The “trained ML” (of this step) selects and deletes/ masks data.
The “determining alternative scenarios” based on a broad list of widely varying characteristics fails to set forth what the alternative scenarios are particularly to describe.
No particular scenario, or alternative scenario is provided related to the generation of the second candidate drug compound.
The second candidate drug compound does not need to be “identical or substantially similar to the removed real drug compounds.” How the second candidate relates to the first candidate, or to the “real drug” is completely absent.
This limitation falls within the abstract idea grouping of Mental process: thinking up candidate drug compound(s) based on the output of the retrained model, and using mental processes of comparison and judgement to determine whether the second candidate should proceed. The second candidate has no desired properties, parameters, effectiveness, or other specificity. (MPEP 2106.04(a)(2), subsection III)
Step 7: “administering to a proxy organism the second candidate drug compound from the set of candidate drug compounds;”
A step of applying the second candidate drug to a proxy organism. A routine laboratory step wherein a candidate drug is applied to a cell culture, or other proxy organism.
The second candidate drug is not synthesized physically, or obtained or retrieved or otherwise physically available for application to a proxy system. The generated second candidate compound is not the generation of a known or available compound, thus, there is no expectation that it exists in the physical world.
There are no limits on the second candidate drug, how it is administered, the assays performed or the biomarkers assessed.
There are no limits on the proxy organism. The proxy organism is not specifically related to any particular “effectiveness,” physical or chemical property, structure, function or parameter of the “real drug” or any other candidate. This encompasses any type of proxy organism testing, for any purpose by any means. The proxy organism is not linked to any particular test, activity, assay or function. It has no particular clinical effect-related elements, properties, activities, or functions. The proxy organism has not been modified to carry out any particular known process, assay, or activity. The proxy organism does not comprise any particular, or even generic categories of biomarkers, or biomarker systems.
All aspects of applying a second candidate drug to the proxy organism are generic.
This limitation is an element in addition to the JE: a step of necessary data gathering, recited at a high level of generality, and is extra solution activity. MPEP 2106.05(g).
Step 8: “8) receiving, at a processing device, a signal that comprises at least two wavelengths that are each associated with a respective biomarker, wherein, when the second candidate drug compound is administered, at least two assays configured to reveal the respective biomarkers are performed on the proxy organism;”
Receiving a signal that comprises at least two wavelengths. This is a step of data acquisition, acquiring a signal resulting from carrying out at least two assays, representing at least two biomarkers.
The assays are undefined and encompass any assay scientifically available.
The biomarkers are undefined and encompass any known or measurable biomarker.
The particular proxy organism is undefined and encompass any known proxy system.
The property or function, effect or activity to be detected is not identified.
The assays are not linked to any aspect of the “real drug” or candidate drugs. The assays are not linked to any “pre-clinical” or clinically relevant activity.
The assays are not linked to any particular property, structure, function, effect, or parameter related to the generated candidates.
This limitation is an element in addition to the JE related to necessary data gathering, recited at a high level of generality, and is extra solution activity (MPEP 2106.05(g).)
Step 9: “analyzing, by the processing device, the signal to obtain the at least two wavelengths from the signal, wherein obtaining the at least two wavelengths includes separating or decoupling the signal into the at least two wavelengths, and wherein the at least two wavelengths indicate whether each of the respective biomarkers is present in the signal received at the processing device; and”
Step 9 recites analyzing the received signal, to obtain the at least two wavelengths.
The claim does not limit how the analysis (evaluation) is performed, and there is nothing about a detected fluorescence itself that would limit how it can be analyzed. The analysis begins by decoupling or separating the two wavelengths, and if the two wavelengths are detected, each respective biomarker is present.
No limits of the analysis beyond decouple and observe are present. The separation and observation of the signals can be performed mentally by observation.
No ramifications for the second candidate drug are provided based on the results of the analysis (i.e. if A is positive, make the candidate, but if A is negative, the candidate is not made).
No particular parameters of the assay, the desired properties, or the fluorescence are analyzed to provide a final result.
No details as to any “preclinical validation” are provided. (i.e. if AB < threshold, it will not have the desired property and would not be useful in the trial; OR if AB = toxicity it should not be used; OR IF AB< 0.05 the compound is likely to have clinical utility in a particular trial.)
No rationale is provided as to whether any tested candidate should or should not be synthesized.
No judgement of the result with respect to the goal of the method (pre-clinical validation of “effectiveness”) is carried out.
The biomarkers detected are not specifically linked to the “real drug” or the candidates of any prior step.
How the biomarker relates to any activity, structure, function, property, effect or ability of the second candidate is not set forth.
The plain meaning of “analyzing” encompasses evaluating information, which in this claim is limited to evaluating detected fluorescent signals that result from testing the second candidate in the proxy organism.
This limitation falls within the abstract idea groupings of Mental process: observation of two wavelengths of light, and judgement as to whether the related biomarker is present, or not. (MPEP 2106.04(a)(2) subsection III.)
Step 10: “10) causing, via the processing device, the second candidate drug compound to be synthesized.”
Step 10) directs routine synthesis of the tested, yet undescribed compound, regardless of the results of the analysis.
There are no particular synthesis steps, particular formulations, or limitations to any part of the synthesis process.
It is unclear how the processing device causes the synthesis. No computer directed, robotic or otherwise processor -related instructions are provided for this step.
The second candidate is completely undefined and has no particular properties, effects, effectiveness or function. It is not compared to any previous candidate for any particular property, effect, effectiveness or function prior to synthesis.
There is no clear link between the second candidate and any preclinical validation of “effectiveness” as set forth in the preamble.
The candidate is synthesized whether or not the biomarkers are judged to be present in the previous step.
This is an element in addition to the JE, directing synthesis by any means of a completely generic “second candidate” having no particular structure, property, activity, effect, effectiveness or action. This is extra-solution activity. MPEP 2106.05(g).
Therefore, the claims explicitly recite elements that, individually and in combination, constitute one or more judicial exceptions (JE). Step 2A-1: YES, the claims recite one or more judicial exceptions.
With respect to step 2A (2): NO the identified JE are not integrated into a practical application. The claims were examined further to determine whether they integrated any JE into a practical application (MPEP 2106.04(d)). The claimed additional elements are analyzed alone, or in combination to determine if the JE is integrated into a practical application (MPEP 2106.05(a-c, e, f, g and h)).
Claim(s) 1, 5, 10, 15, 17, 21-22 recite the additional non-abstract element(s) of data gathering, or a description of the data gathered.
In claim 1, steps 7- 8 and 10 as numbered by the examiner represent routine data gathering. The data gathering steps are generically recited, having no particular steps related to the application of the undefined, undescribed second candidate to any particular proxy system for the purpose of performing any assays or identifying any biomarkers.
With respect to step 7, the second candidate drug compound is not produced, or compounded into a form suitable for application to any particular proxy organism. No particular dosages, timelines, or route of administration is required.
The proxy organism itself has no particular characteristics, structure, cells, or biomarkers. The proxy organism has no connection to any effect, or effectiveness of any aspect of any of the “real drug compounds” or any of the candidates of the claim.
No particular steps related to performing a particular controlled test, utilizing control and candidate substances or data are provided.
With respect to step 8, the receipt of the signals does not set forth what two assays were performed. The receipt of the signals does not identify the biomarkers being assayed.
There are no positive active steps requiring the assays be performed at all, be it together or separately on the proxy organism. There are no positive active steps requiring any control testing.
Any technically feasible assay is encompassed, therefor this step does not meaningfully limit or apply any particular aspect of the identified JE.
With respect to step 10, the production of the second candidate is always carried out, whether or not the analysis might indicate that the second candidate might have a desired property, function, structure, effect or effectiveness. The synthesis has no specific structure to be synthesized. The synthesis has no specific synthesis steps. The production of the second candidate is not subsequently tested to see if the second candidate actually has the activity. This does not meaningfully apply any aspect of the identified JE.
As previously pointed out, the identified dependent claims reciting aspects of data gathering, alone or in combination with claim 1, do not integrate any identified JE into a practical application.
Data gathering steps are not an abstract idea, they are extra-solution activity, as they collect the data needed to carry out the JE. The data gathering does not impose any meaningful limitation on the JE, or how the JE is performed. The additional limitation (data gathering) must have more than a nominal or insignificant relationship to the identified judicial exception. (MPEP 2106.04/.05, citing Intellectual Ventures LLC v. Symantec Corp, McRO, TLI communications, OIP Techs. Inc. v. Amason.com Inc., Electric Power Group LLC v. Alstrom S.A.).
Claim(s) 1 recite(s) the additional non-abstract element (EIA) of a general-purpose computer system or parts thereof.
The “processing device” of claim 1 has no specific attributes beyond carrying out certain routine computer-based processes (receipt of signals, decoupling signals and analyzing signals), or “causing” the synthesis of the second candidate. A general-purpose computer does not integrate any identified JE into a practical application.
The newly added steps (1-5) of claim 1 are not limited to computer-implemented processes, or the use of a computing device, or processing device.
The EIA do not provide any details of how specific structures of the computer elements are used to implement the JE. The claims require nothing more than a general-purpose computer to perform the functions that constitute the judicial exceptions. The computer elements of the claims do not provide improvements to the functioning of the computer itself (as in DDR Holdings, LLC v. Hotels.com LP); they do not provide improvements to any other technology or technical field (as in Diamond v. Diehr); nor do they utilize a particular machine (as in Eibel Process Co. v. Minn. & Ont. Paper Co.). Hence, these are mere instructions to apply the JE using a computer, and therefore the claim does not recite integrate that JE into a practical application.
Dependent claim(s) 2, 7-9, 17 recite(s) an abstract limitation to the JE reciting additional mathematic concepts, or mental processes. Additional abstract limitations cannot provide a practical application of the JE as they are a part of that JE.
In combination, the limitations of data gathering, for the purpose of carrying out the JE, using a general-purpose computer merely provide extra-solution activity, and fail to integrate the JE into a practical application.
With respect to step 2B: NO the claims do not provide a specific inventive concept. The claims recite a JE, do not integrate that JE into a practical application, and thus are probed for a specific inventive concept. The judicial exception alone cannot provide that inventive concept or practical application (MPEP 2106.05). The additional elements were considered individually and in combination to determine if they provide significantly more than the judicial exception. (MPEP 2106.05.A i-vi).
With respect to claim(s) 1, 5, 10, 15, 17, 21-22: The limitation(s) identified above as non-abstract elements (EIA) related to data gathering do not rise to the level of significantly more than the judicial exception.
Stratton (2019; of record) discloses applying candidate drugs to proxy organisms, carrying
out two assays and detecting the fluorescence.
Benz (2016) discloses applying candidate drugs to proxy organisms, carrying out two assays
and detecting the fluorescence. Benz applies machine learning to the results of the assays to make a
prediction about responsiveness to a treatment.
Bedoya (2020) discloses applying candidate drugs to proxy organisms, carrying out assays
and detecting fluorescence.
Das (2021) discloses applying candidate drugs to proxy organisms, carrying out assays and
detecting fluorescence, and directing synthesis. The methods of Das particularly test compounds
which were generated by AI processes.
These elements meet the BRI of the identified data gathering limitations. As such, the prior art recognizes that this data gathering element was routine, well understood and conventional in the art (as in Alice Corp., CyberSource v. Retail Decisions, Parker v. Flook).
Activities such as data gathering do not improve the functioning of the computer itself, or comprise an improvement to any other technical field. The limitations do not require or set forth a particular machine, they do not effect a transformation of matter, nor do they provide an unconventional step (citing McRO and Trading Technologies Int’l v. IBG). Data gathering steps constitute a general link to a technological environment. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception are insufficient to provide significantly more (as discussed in Alice Corp.,).
With respect to claim(s) 1: the limitations identified above as non-abstract elements (EIA) related to general-purpose computer systems do not rise to the level of significantly more than the judicial exception. Claim 1 recites a “processing device.”
Stratton (2019) provides processing devices meeting the limitations of claim 1.
Oskooei (2020) provides processing devices meeting the limitations of claim 1, and particularly those that comprise trained AI elements, for the same purpose.
Oono (2017) provides processing devices meeting the limitations of claim 1, particularly for
drug design, comprising generative ML.
Isayev (2020) provides processing devices meeting the limitations of claim 1, particularly for
drug design, using generative AI.
Each of Stratton, Oskooei, Oono and Isayev disclose computer systems or computing elements which meet the BRI of the claimed “processing device.”
As such, the prior art recognizes that these computing elements were routine, well understood and conventional in the art.
These elements do not improve the functioning of the computer itself, or comprise an improvement to any other technical field (Trading Technologies Int’l v IBG, TLI Communications). They do not require or set forth a particular machine (Ultramercial v. Hulu, LLC., Alice Corp. Pty. Ltd v. CLS Bank Int’l), they do not effect a transformation of matter, nor do they provide an unconventional step. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception are insufficient to provide significantly more (as discussed in Alice Corp., CyberSource v. Retail Decisions, Parker v. Flook, Versata Development Group v. SAP America).
Dependent claim(s) 2, 7-9, 17 each recite a limitation requiring additional mathematic concepts or mental processes. Additional abstract limitations cannot provide significantly more than the JE as they are a part of that JE (MPEP 2106.05).
In combination, the data gathering steps providing the information required to be acted upon by the JE, performed in a generic computer or generic computing environment fail to rise to the level of significantly more than that JE. The data gathering steps provide the data for the JE, which is carried out by the general-purpose computers. No non-routine step or element has clearly been identified.
The claims have all been examined to identify the presence of one or more judicial exceptions. Each additional limitation in the claims has been addressed, alone and in combination, to determine whether the additional limitations integrate the judicial exception into a practical application. Each additional limitation in the claims has been addressed, alone and in combination, to determine whether those additional limitations provide an inventive concept which provides significantly more than those exceptions. For these reasons, the claims, when the limitations are considered individually and as a whole, are rejected under 35 USC § 101 as being directed to non-statutory subject matter.
Applicant’s arguments:
Applicant’s arguments have been considered but are not persuasive.
It appears Applicant is attempting to model the claims presented in the recent Desjardin decision, without providing any of the necessary details in the claim, including structures, training steps, modeling processes, or how the retraining affects the ML or generative AI in any way.
The ML of the rejected claims here has no particular initial structure, which is changed in particular ways with the initial training data. There is no step of the initial training, nor what the initial training required or what data was used in that initial training.
The ML of the rejected claims has no particular structure after the initial training.
The processes being modeled by the initial untrained ML, or the initially trained ML are not identified.
There is no link between any initial model, any initial training set and any particular line of inquiry which must be addressed by the model.
The initially trained model does not provide any particular output.
The “modified” training set only differs by the masking or removal of one “real drug compound” from the unprovided, undescribed initial training set. The remaining members of the modified set are completely undescribed and unlimited in any fashion.
The claim fails to set forth how the modification of the training set has ANY effect on the initially trained model to produce the “retrained ML model.”
The retraining is not for any particular purpose. The retraining is not to identify any particular property, structure, action, activity, effect or effectiveness of anything.
The retraining does not result in any particular changes to the initially trained model.
There are no outputs related to the retraining of the model such that they could be judged in comparison to the initially trained model outputs.
There is no comparison of “generated candidates” between the initially trained model and the retrained model such that any improvement based on the retraining could be recognized.
The retrained model does not comprise any specific “generator” elements with which to generate any new (or first) candidates. The retrained model does not comprise any specific “discriminator” elements by which the retrained model could assess any generated candidates for a likelihood of having any desired property.
The determination of satisfying the similarity criterion is not carried out by the retrained ML.
There are no consequences for the first candidate if it does not meet the similarity criterion. No specific aspect of the first candidate or the real drug reference is to be assessed. No specific type of similarity analysis is applied.
The determination that the retrained ML model meets a “rediscovery training criterion” has no specific aspects. There is no consequence if the retrained model does not meet the criterion.
The retrained ML model is separate from the “generative AI” of step 6.
The Generative AI of step 6 is a separate, distinct ML model, described only in results-based language that fails to set forth how the causal inference particularly works with the data at hand (meeting a similarity criterion and an indication that the other ML meets a retraining criterion) to generate a second candidate. This candidate is completely unrelated to the first candidate and the initial “real drug compound. The ML is completely unrelated to the ML of steps 1-5.
The generative AI of step 6 does not comprise any “generative elements” which may act to generate a candidate structure based on desired results, or any previously gathered data.
The generative AI of step 6 does not comprise any “discriminators” which could act to assess a generated candidate structure for whether it may possess or exhibit a particular property.
The “causal inference” of the ML is used to REMOVE structures from a candidate pool, not to GENERATE any new structure. HOW the inference selects any compound for removal from consideration is not set forth.
The “counterfactuals” used in the “causal inference” do not set forth any particular counterfactual to be used in understanding why the AI generated (or removed) a structure, or what its properties may encompass. The counterfactual limitation provides a definition of what a counterfactual could be, but not what it IS, or SHOULD be for this method.
The “counterfactuals” do not provide any particular scenarios, or alternate scenarios of any kind, and are “based on” a long list of widely disparate ideas, results, actions, or correlations. No specific alternatives are provided.
The generative AI of step 6 is not modifying a previously provided drug structure with previously identified characteristics.
The AI is not generating a compound intended to have any particular property, function, effect, effectiveness or activity. It is entirely unrelated to the previously generated candidate (by the first ML) and is entirely unrelated to the initial “real drug candidate”. It does not use the “retrained ML” of step 2) to generate the compound.
The training, the re-training and the ML model are all recited at such a high level of generality that the recitation merely informs the user of a “field of use” of the invention: the broad discipline of machine learning. (MPEP 2106.05(h).)
The generative AI, comprising the separate ML, and the generation of the second candidate compound are all recited at such a high level of generality that the recitation merely informs the user of a field of use of the invention: the broad discipline of AI and ML processes. (MPEP 2106.05(h).)
The second drug compound (not a known compound, or a modification of a known compound) is not synthesized prior to application to the proxy organism.
As noted at length above, all aspects of the proxy organism, applying the candidate to the proxy organism, the assays performed, and the biomarkers identified are completely generic, without any particularity. The proxy organism, the assays, and the biomarkers have no link to any aspect of the idea of “pre-clinical validation of an effectiveness of a candidate drug” as no particular drug, assay, biomarker, effectiveness or effect is identified.
The proxy organism does not need to identify any aspect of a biological response that may have been similar to the results using the “real drug compound.” It is completely unrelated.
The biomarkers are not identified as being indicative of a particular clinical, biological, chemical or physical property of any kind.
The assays are not identified as being indicative of any particular clinical, biological, chemical or physical process which may be related to a “clinical” effectiveness, as set forth in the preamble.
The proxy system, the applying of the candidate, the receipt of the signals, the analysis of the signals and the direction of synthesis are all recited at such a high level of generality that the recitation merely informs the user of a field of use of the invention: the broad discipline of AI and ML processes. (MPEP 2106.05(h).)
MPEP 2106.05(f) provides the following considerations for determining whether a claim
simply recites a judicial exception with the words "apply it" (or an equivalent), such as mere instructions to implement an abstract idea on a computer:
(1) whether the claim recites only the idea of a solution or outcome i.e., the claim fails to recite details of how a solution to a problem is accomplished; The claims fail to recite any details as to how any particular compound is to be generated, analyzed, produced, tested and the test results analyzed for any particular clinical effectiveness. At each stage the elements are broad, generic or completely undefined.
(2) whether the claim invokes computers or other machinery merely as a tool to perform an existing process; The “processing device” is a tool to deconvolve wavelengths of light and identify the presence or absence of those wavelengths. The processing device directs that a compound be synthesized by any means. The processing device does not comprise any ML or AI structures. The processing device is not changed by the process. The processing device is a convenient tool for observing, receiving and deconvolving signals.
(3) the particularity or generality of the application of the judicial exception. Steps 1-7 are not carried out using the “processing device.” None of the AI or ML structures exist on the “processing device.” Carrying out the JE of steps 1-7 does not require the “processing device” of step 8. "analyzing... the signal to obtain the at least two wavelengths from the signal...' is performed using "a processing device". The processing device is used to generally apply the abstract idea without placing any limits on how the device functions. Rather, these limitations recite the outcome of observing the two wavelengths, to determine whether each respective biomarker is present. There are no limits placed on the processing device, and any method of observing the two wavelengths is encompassed. This outcome invokes computers or other machinery as a tool to perform an existing process (detecting wavelengths of light). The outcome is generally stated, with no limits to the analysis, and no rationale as to how a tested compound should be selected for synthesis. The processing device causes the synthesis of the second candidate whether or not it has any particular property, or whether or not any wavelength was detected.
MPEP: 2106.05: “The recitation of claim limitations that attempt to cover any solution to an identified problem with no restriction on how the result is accomplished and no description of the mechanism for accomplishing the result, does not integrate a judicial exception into a practical application or provide significantly more because this type of recitation is equivalent to the words "apply it". See Electric Power Group, LLC v. Alstom, S.A., 830 F.3d 1350, 1356, 119 USPQ2d 1739, 1743-44 (Fed. Cir. 2016); Intellectual Ventures I v. Symantec, 838 F.3d 1307, 1327, 120 USPQ2d 1353, 1366 (Fed. Cir. 2016); Internet Patents Corp. v. Active Network, Inc., 790 F.3d 1343, 1348, 115 USPQ2d 1414, 1417 (Fed. Cir. 2015). In contrast, claiming a particular solution to a problem or a particular way to achieve a desired outcome may integrate the judicial exception into a practical application or provide significantly more. See Electric Power, 830 F.3d at 1356, 119 USPQ2d at 1743.”
Applicant fails to explain sufficiently and persuasively how the rejected claims analogous to the cited decisions as set forth in the response.
With respect to the production of a candidate, as previously indicated, at length, there is no particular candidate to synthesize. The second candidate is completely without any hint of structure, composition, physical or chemical properties, functions, or effects. This is not akin to any of the considerations provided in Vanda Pharmaceuticals, or any other cited decision. This is a field of use limitation: the second candidate has to have the capability to be synthesized. It is not the direction of the production of a particular compound, or even class of compound. The second candidate could be nearly any substance, including metabolites, vitamins, chemical elements, small molecules, chemicals, peptides, polypeptides, proteins, lipids, other macromolecules and other complex structures. This is a general idea that the compound should or could be synthesized, but is not the specific integration of any JE into a practical application and it fails to provide significantly more than the JE.
As previously pointed out, the claims fail to successfully model any other decision or cited case where claims have been determined to be patent-eligible, such as McRo, Enfish, BASCOM, Finjan, Visual Memory or Data Engine Tech. LLC. (MPEP 2106.05(a).)
In McRO, it was not the mere presence of unconventional rules that led to patent eligibility. In McRO, "[t]he claimed improvement was to how the physical display operated (to produce better quality images)." SAP Am. v. InvestPic, LLC, 898 F.3d 1161, 1167 (Fed. Cir. 2018). The claims in McRO recited a step of applying the data sets generated using the specific claimed rules to a specific target: a sequence of animated characters to produce lip synchronization and facial expression control of those animated characters. McRO, 837 F.3d at 1308. Thus, the claims were directed to an improvement in computer animation and used rules to automate a subjective task of humans to create a sequence of synchronized, animated characters. Id. at 1314--15.
In the claims at issue here, there is no such application of specifically claimed rules to produce an improved technological result. There are no specific rules guiding the generation of any particular candidate. There are no specific rules regarding the analysis of the data from the proxy organism experiments. There are no rules identifying when a candidate should, or should not be synthesized. The claims lack specific rules of any kind.
The disputed claims in Enfish were patent-eligible because they were "directed to a specific improvement to the way computers operate, embodied in [a] self-referential table." Enfish, 822 F.3d at 1336. The court found that the "plain focus of the claims" there was on an improvement to computer functionality itself-a self-referential table for a computer database, designed to improve the way a computer carries out its basic functions of storing and retrieving data- not on a task for which a computer is used in its ordinary capacity. Id. at 1335-36. The court noted that the specification identified additional benefits conferred by the self-referential table (e.g., increased flexibility, faster search times, and smaller memory requirements), which further supported the court's conclusion that the claims were directed to an improvement of an existing technology. Id. at 1337 (citation omitted).
Here, in contrast, the purported data structures of the ML, the trained ML, the re-trained ML and the AI, which comprises an additional ML are not confined to being carried out on a computer. Therefore, these aspects of the claim cannot be relied upon to provide the improvement to computer functionality.
Applicant fails to explain sufficiently and persuasively how claim 1 is directed to an improvement in the way computers operate analogous to the case in Enfish. The examiner finds instead that Applicant's claim merely uses a generic “processing device” that operates in its normal, expected manner: analyzing data.
Similarly, in BASCOM, the claims recited a "specific method of filtering Internet content" requiring "the installation of a filtering tool at a specific location, remote from the end-users, with customizable filtering features specific to each end user." BASCOM, 827 F.3d at 1345--46, 1350. The installation of a filtering tool at a specific location, remote from the end users, with customizable filtering features specific to each end user, provided an inventive concept in that it gave the filtering tool both the benefits of a filter on a local computer and the benefits of a filter on the ISP server. Id. at 1350. Applicant fails to explain sufficiently and persuasively how the rejected claims are analogous to BASCOM. The examiner finds no analogous non-conventional, non generic arrangement of known, conventional physical elements within a computer system. That is the claim does not recite any physical element positioned in an unconventional manner in the system.
In Finjan, the courts found that a computer-implemented method that generates a security profile that identifies both hostile and potentially hostile operations, and can protect the user against both previously unknown viruses and "obfuscated code," was an improvement over traditional virus scanning. (Finjan Inc. v. Blue Coat Systems, 879 F.3d 1299, 1304, 125 USPQ2d 1282, 1286 (Fed. Cir. 2018)) In contrast, in the rejected claims, no elements are present dealing with online security, or virus protection, or code are present. No element of the computer, or processing device is improved.
In Visual Memory, the courts found that particular recited details of a memory system with
programmable operational characteristics, configurable based on the type of processor, provided an
improvement in the computer itself (Visual Memory, LLC v. NVIDIA Corp., 867 F.3d 1253, 1259-60,
123 USPQ2d 1712, 1717 (Fed. Cir. 2017)). In contrast, no particular, configurable aspect of the computer is specifically set forth in the rejected claims. The ML and AI are not limited to being carried out in a computer. The “processing device” is generic and acts to analyze data. The “processing device” is not improved.
In Data Engine Tech, LLC, the courts found that a specific interface and implementation for
navigating complex three-dimensional spreadsheets using techniques unique to computers provided
an improvement to the computer itself. (Data Engine Techs., LLC v. Google LLC, 906 F.3d 999, 1009,
128 USPQ2d 1381, 1387 (Fed. Cir. 2018)). In contrast, the rejected claims do not use any techniques
unique to computers to carry out the JE. The ML and AI are not limited to being carried out in a computer. The “processing device” is generic and acts to analyze data. The “processing device” is not improved.
Claims that recite performing information analysis (e.g., statistical analyses and correlating similarity information), as well as the collection and manipulation of information related to such analysis, have been determined by our reviewing court to be an abstract concept that is not patent eligible. See SAP, 898 F.3d, 1165, 1167, 1168 (Claims reciting "[a] method for providing statistical analysis" (id. at 1165) were determined to be "directed to an abstract idea" (id. at 1168)); see also Content Extraction & Transmission LLC v. Wells Fargo Bank, Nat'l Ass 'n, 776 F.3d 1343, 1345, 1347 (Fed. Cir. 2014) (finding the "claims generally recite ... extracting data ... [and] recognizing specific information from the extracted data" and that the "claims are drawn to the basic concept of data recognition").
"As many cases make clear, even if a process of collecting and analyzing information is limited to particular content or a particular source, that limitation does not make the collection and analysis other than abstract." SAP, 898 F.3d at 1168 (internal quotation marks omitted)).
With respect to the identified additional elements, and their analysis individually, and in combination, and of the claim as a whole, these additional elements have been addressed multiple times over the history of this application.
Whether examined individually or in combination, the EIA in the claims beyond the JE, i.e.,
data collection and a “processing device” which may include routine laboratory equipment, are
insufficient to establish that the abstract ideas recited are integrated into a practical application nor
do the EIA provide significantly more than the JE.
MPEP 2106.04: “Limitations the courts have found indicative that an additional element (or combination of elements) may have integrated the exception into a practical application include:
• An improvement in the functioning of a computer, or an improvement to other technology or technical field, as discussed in MPEP §§ 2106.04(d)(1) and 2106.05(a);
• Applying or using a judicial exception to effect a particular treatment or prophylaxis for a disease or medical condition, as discussed in MPEP § 2106.04(d)(2);
• Implementing a judicial exception with, or using a judicial exception in conjunction with, a particular machine or manufacture that is integral to the claim, as discussed in MPEP § 2106.05(b);
• Effecting a transformation or reduction of a particular article to a different state or thing, as discussed in MPEP § 2106.05(c); and
• Applying or using the judicial exception in some other meaningful way beyond generally linking the use of the judicial exception to a particular technological environment, such that the claim as a whole is more than a drafting effort designed to monopolize the exception, as discussed in MPEP § 2106.05(e).
The courts have also identified limitations that did not integrate a judicial exception into a practical application:
• Merely reciting the words "apply it" (or an equivalent) with the judicial exception, or merely including instructions to implement an abstract idea on a computer, or merely using a computer as a tool to perform an abstract idea, as discussed in MPEP § 2106.05(f);
• Adding insignificant extra-solution activity to the judicial exception, as discussed in MPEP § 2106.05(g); and
• Generally linking the use of a judicial exception to a particular technological environment or field of use, as discussed in MPEP § 2106.05(h).”
The rejected claims do not provide an improvement in the function of the computer itself.
The rejected claims do not provide an improvement to other technology or technical field.
The rejected claims do not provide or effect a particular treatment or prophylaxis for a disease or medical condition.
The rejected claims do not implement the JE in conjunction with a particular machine or manufacture integral to the claim.
The rejected claims do not effect a transformation or reduction of a particular article to a different state or thing.
The rejected claims merely link the use of the JE to technological environments of the general use of machine learning, and the use of proxy organism systems.
As set forth above, the recitations of the rejected claims are equivalent to the words “apply it”, and use the processing device as a tool to analyze data.
The rejected claims add insignificant, completely generic steps to the proxy organisms test system. These represent necessary data gathering and extra-solution activity.
These elements were reconsidered under step 2B, and added the consideration as to whether the EIA represented routine, well understood and conventional limitations.
(MPEP 2106.05: ii. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception, e.g., a claim to an abstract idea requiring no more than a generic computer to perform generic computer functions that are well-understood, routine and conventional activities previously known to the industry, as discussed in Alice Corp., 573 U.S. at 225, 110 USPQ2d at 1984 (see MPEP § 2106.05(d));”)
The EIA identified as data gathering, and the use of the “processing device”, were each shown, to be specified at a high level of generality, and by citation of prior art references to be well-understood, routine, and conventional limitations in bioinformatics.
An important consideration in determining whether a claim improves technology is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome, as opposed to merely claiming the idea of a solution or outcome. (McRO, 837 F.3d at 1314-15, 120 USPQ2d at 1102-03; DDR Holdings, 773 F.3d at 1259, 113 USPQ2d at 1107.) Here there is no particular solution, or a particular way to achieve any desired outcome. The claims recite the ideas of training, retraining and use of AI, without any particular indication of what processes are being modeled, or how the model acts on any data to achieve any result.
The examiner has repeatedly pointed out specific ML, specific training sets, specific AI elements, and specific processes disclosed in the specification which, if incorporated into the independent claim would improve the analysis under this statute.
Further, with respect to the arguments regarding the alleged improvement, it is unclear that the independent claims recite all the necessary and sufficient steps required to achieve that improvement. MPEP 2106.05(a): “An important consideration in determining whether a claim improves technology is the extent to which the claim covers a particular solution to a problem or a particular way to achieve a desired outcome, as opposed to merely claiming the idea of a solution or outcome. McRO, 837 F.3d at 1314-15, 120 USPQ2d at 1102- 03; DDR Holdings, 773F.3d at 1259, 113 USPQ2d at 1107.”
The MPEP sets forth that “if the examiner concludes the disclosed invention does not improve technology, the burden shifts to applicant to provide persuasive arguments supported by any necessary evidence to demonstrate that one of ordinary skill in the art would understand that the disclosed invention improves technology. Any such evidence submitted under 37 CFR 1.132 must establish what the specification would convey to one of ordinary skill in the art and cannot be used to supplement the specification.” Applicant’s arguments cannot take the place of evidence.
The examiner Strongly suggests Applicant review the following differing ideas:
1) Applicant’s patent 11,049,590 contains claims to a system for generating candidate compounds, which clearly set forth the structure of the system, the nature of the candidate drug, biomedical activities to be assessed, the ML, the training data, how the ML act on the training data, how the AI acts, the structure of the AI including generators, discriminators, descriptor modules, benchmark analyses, changes in the processing device based on certain processed information, specific feature comparisons, and the generation of additional candidates, and testing these candidates for specific biomedical activities.
Were this system to be specifically and completely included in claim 1, this would SIGNIFICANTLY improve the analysis under this statute, depending on the form of the amendment, and any amendments made to improve the proxy organism testing process.
2) Applicant’s patent 11,403,316 has patented claims to method of presenting a graphical user interface for the analysis of protein sequences, wherein the protein sequences are each associated with a biomedical activity or biochemical activity. The trained ML model of these claims generate protein sequences, and use causal inference to execute a plurality of alternative scenarios to filter the superset of generated protein sequences, and to generate new sequences in a design space. The training required embedding/ encoding known sequences with known activities and were to be included in the superset acted on by the ML. Further, the GUI, in a second screen space, presented specific information about the generated sequences, the generated parameters, any flag or label information, whether the sequence is previously known, or novel, and other information. Dependent claims to specific forms of output of data, or the addition of differing types of data are provided.
Should the claims be amended to specifically and completely add these method steps, followed by the proxy organism testing, the analysis under this statute would be significantly improved, depending on the form of the amendment, and any amendments to improve the proxy organism testing process.
3) Applicant’s patent 11,462,304 has patented claims to methods of generating candidate drug compounds. The process utilizes AI engines in a computing environment, and use knowledge graphs comprising information about candidate drug compounds, including structure, activity and semantic information. The AI analysis the shape of the multidimensional representation, slices the representation, and decodes the dimensions present in the slice, which then are analyzed to determine an effectiveness of a biomedical feature of the slice. Claims to attention message passing neural networks are provided. Benchmarking is claimed. Validation steps are claimed. Specific scoring and ranking is carried out. Particular attributes of the AI are enumerated.
Should the claims be amended to specifically and completely add these method steps, followed by the proxy organism testing, the analysis under this statute would be significantly improved, depending on the form of the amendment, and any amendments to improve the proxy organism testing process.
4) Applicant’s patent 12,462,902 contains claims to computer implemented methods for using an AI engine to generate candidate drug compounds. Multidimensional representations of protein drug compounds are obtained, which include specific structure, function and activity information. The AI translates the representation into encodings. The encodings are acted upon by the AI in a particular way to generate particular sets of encodings. Specific structures of the AI (autoencoders) act on the encoded information in a particular way for a particular purpose. Candidate drugs are generated based on the autoencoded vectors, using a specific module of the AI engine. The AI engine uses a particular decoding mechanism to represent structure, activity or sematic information. An effectiveness of a biomedical feature of the candidate drug compound is determine by the AI, using a specific generator/discriminator pairing. Benchmarking is claimed. Tuning parameters of the creator module (re-training) is claimed. Specific attributes of the AI are claimed. Steps of validation are performed.
Should the claims be amended to specifically and completely add these method steps, followed by the proxy organism testing, the analysis under this statute would be significantly improved, depending on the form of the amendment, and any amendments to improve the proxy organism testing process.
5) Applicant’s patents 11,424,008 and 12,087,404 contain claims to methods of generating a design space of a peptide for an application, where the peptide has biomedical, or biochemical activities. Peptide sequences are obtained, updated within the design space, and a solution space comprises a target of updated sequences each having the updated respective plurality of activities. These are used in a ML model to process the solution space, to identify a candidate that represents a sequence having at least one level of activity that exceeds a threshold, performing validation of these compounds and generate a second candidate. The second candidate is specifically compared for the aspect of the activity. Additional ML models are claimed. Updating the ML in response to a query is claimed. Particular statistical analyses are claimed. Anti-infectives are a specific category of drug that is generated by the claims. Specific information is claimed with relation to the peptide/ protein sequences, including information for synthesis, structure, stability, solubility, clinical data, market data et al. Claims to directing synthesis of the candidate are present. Specific claims assessing computer metrics are set forth.
Applicant’s patents 11,436,246 and 11,848,076 have claims directed to/ related to the method of the ‘008 and ‘404 patents in a specific GUI of a therapeutic tool.
(The ‘008 / ‘404 set forth methods of generating a design space of a peptide for an application, where the peptide has biomedical, or biochemical activities. Peptide sequences are obtained, updated within the design space, and a solution space comprises a target of updated sequences each having the updated respective plurality of activities. These are used in a ML model to process the solution space, to identify a candidate that represents a sequence having at least one level of activity that exceeds a threshold, performing validation of these compounds and generate a second candidate. The second candidate is specifically compared for the aspect of the activity. Additional ML models are claimed. Updating the ML in response to a query is claimed. Particular statistical analyses are claimed. Anti-infectives are a specific category of drug that is generated by the claims. Specific information is claimed with relation to the peptide/ protein sequences, including information for synthesis, structure, stability, solubility, clinical data, market data et al. Claims to directing synthesis of the candidate are present. Specific claims assessing computer metrics are set forth.)
The GUI have particular design areas, specific graphical elements, particular screen areas, solution spaces, and act in response to the selection of a specific sequence to generate a candidate drug, an interaction, an activity, a drug, a gene, a pathway or combinations thereof.
Applicant’s patent 11,967,400 contains claims that combine the GUI design space for protein sequences, ML models, query parameters, specific graphical elements, second GUI screens, the specific display of certain information, and the analysis of the generated sequences to determine whether they are novel to the model, and comprise certain specifically listed types of information. Specific trails can be selected. Specific biasness intelligence screens are produced and acted upon by the AI. Specific generated sequences are produced. The ML can be adjusted. Biomedical information is incorporated. Direction of the synthesis of the candidate is claimed.
6) Applicant has patents to methods and platforms for directing synthesis of peptides, or polypeptides, based on spectral profiling, known recipes for synthesis of peptides, specific coupling reactions, and control elements. (11,587,643, 11,512,345, 12,006,541) Particular ML having particular layers, information, parameters and outputs are claimed. The recipe comprises specific information regarding reagents, chemistry, temperatures, solvents, protection groups, anchors, linkers, catalysts etc. The ML can generate a sequence based on a desired drug activity level in a therapeutic domain. Particular reaction chambers, detectors and devices are claimed. Particular training data is applied to the ML.
As this Final Rejection now closes prosecution after the third RCE, Applicant is strongly encouraged to either amend the claims consonant with the requirements of this statute and the examiner’s suggestions, or decide whether the issue(s) should be decided by a higher authority.
New Grounds of Rejection
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1, 2, 5, 7-10, 15-17, 21, 22 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
The metes and bounds of claim 1 as amended are entirely unclear. Claim 1 fails to particularly point out and distinctly claim the subject matter of the invention. As set out in detail above, the newly amended claims introduce multiple instances of indefiniteness as to 1) the “effectiveness” under consideration, 2) the nature and members of the initial, unprovided, undescribed initial training set. 3) the nature of the “real drug compound”, 4) the structure of the untrained ML, 5) the structure of the initially trained ML, 6) the structure, nature or members of the modified training set, 7) the actual process being modeled by any of the ML, 8) the output of the initially trained ML, 9) how applying the completely undescribed “modified” training set affects the initially trained ML, 10) how the retraining affects the structure or the output of the initially trained ML, 11) the purpose of the retraining, 12) how the retrained ML generates any candidate, 13) what happens if either criterion is not met by the generated first candidate, or the retrained ML, 14) the structure of the AI, which comprises a SEPARATE ML, described in functional language, 15) the purpose of the AI with respect to the processes being modeled, 16) how the AI acts to generate the second candidate, 17) how the second candidate is related to any other candidate, real drug, or “effectiveness”, 18) the nature of the proxy organism, and how it is intended to “validate” any “effectiveness”, 19) the assays carried out, 20) the biomarkers to measure, and how they are related to any “effectiveness”, specifics as to the analysis and the consequences of the analysis, and the “causing” of the synthesis of the second candidate.
While breadth of a limitation is not the same as indefiniteness, the complete lack of specificity in these issues raises the issue as to what particular ML, AI, training data, clinical or biological effect, proxy system and analysis steps should actually be carried out to achieve Applicant’s Invention. It is entirely unclear what the boundaries of each element are, individually and as a whole, and the scopes of each are entirely unclear. MPEP 2173:
“The essential inquiry pertaining to this requirement is whether the claims set out and circumscribe a particular subject matter with a reasonable degree of clarity and particularity. "As the statutory language of ‘particular[ity]' and 'distinct[ness]' indicates, claims are required to be cast in clear—as opposed to ambiguous, vague, indefinite—terms. It is the claims that notify the public of what is within the protections of the patent, and what is not." Packard, 751 F.3d at 1313, 110 USPQ2d at 1788.”
The metes and bounds of claim 1 are unclear. As set forth above, the newly added limitations to the initially trained ML fail to particularly point out and distinctly claim the nature of that ML with respect to the structure of the ML, the purpose of the ML, the initial training set and training of the ML, and what the output of the initially trained ML is intended to be. The ML of the rejected claims (steps 1-5) here has no particular initial structure, which could then be changed in particular ways with the application of initial training data. There is no step of the initial training, nor any indication of what the initial training required or what data was used in that initial training. The only hint is that the initial training set apparently included at least one “real drug compound” but that provides no information as to the other members of the training set, what the ML is intended to predict or output, or why retraining might be required. The ML of steps 1-5 has no particular structure after the initial training. There is no link between any initial model, any initial training set and any particular line of inquiry which must be addressed by the model. The initially trained model does not provide any particular output. The “modified” training set only differs by the masking or removal of one completely generic “real drug compound” from the unprovided, undescribed initial training set. The remaining members of the modified set are completely undescribed and unlimited in any fashion. The claim fails to set forth how the modification of the training set has ANY effect on the initially trained model to produce the “retrained ML model.” The retraining is not for any particular purpose. The retraining is not to identify any particular property, structure, action, activity, effect or effectiveness of anything. The retraining does not result in any particular changes to the initially trained model. There are no outputs related to the retraining of the model such that they could be judged in comparison to the initially trained model outputs. There is no comparison of “generated candidates” between the initially trained model and the retrained model such that any improvement based on the retraining could be recognized.
The metes and bounds of the retrained model are entirely unclear, in terms of the structure of the ML, the process carried out, and how any new candidate is to be generated. The retrained model does not comprise any specific “generator” elements with which to generate any new (or first) candidates. The retrained model does not comprise any specific “discriminator” elements by which the retrained model could assess any generated candidates for a likelihood of having any desired property.
The metes and bounds of the “similarity criterion” limitation are unclear. The determination of satisfying the similarity criterion by the first candidate, generated by the retrained ML, is not clearly carried out by the retrained ML. It is entirely unclear what aspect of the various models and processes are to be carried out by the retrained ML. There are no consequences for the first candidate if it does not meet the similarity criterion. There are no directions for modifying the structure, retraining the model, tuning the model, or any direct comparisons between a specific property of the first candidate and the generic “real drug compound.” No specific aspect of the candidate or the real drug reference is to be assessed. No specific type of similarity analysis is applied. It is entirely unclear how the satisfaction of this criteria affects any subsequent step in the method.
The metes and bounds of the “rediscovery training criterion” are entirely unclear. This step does not appear to be carried out by the retrained ML, but is assessing the ML in some other system, on some undefined characteristic. The determination that the retrained ML model meets a “rediscovery training criterion” has no specific aspects. There is no consequence if the retrained model does not meet the criterion. There is no direction to tune, re-train or modify the retrained model if it does not meet the criterion. The retrained ML is not used in any subsequent step, and thus the purpose of the retraining is unclear. The retrained ML is not specifically a part of the next steps (6-10), and it is entirely unclear how the satisfaction of this criteria affects any subsequent step in the method. The retrained ML model is separate from the “generative AI” of step 6.
The metes and bounds of the generative AI and the steps performed by that AI are entirely unclear. The Generative AI of step 6 is a separate, distinct ML model, described only in results-based language that fails to set forth how the causal inference particularly works with the data at hand (meeting a similarity criterion and an indication that the other, retrained ML meets a retraining criterion) to generate a second candidate. This second candidate is completely unrelated to the first candidate and the initial “real drug compound. The ML of step 6 is completely unrelated to the ML of steps 1-5. The generative AI of step 6 does not comprise any “generative elements” which may act to generate a candidate structure based on desired results, or any previously gathered data. The generative AI of step 6 does not comprise any “discriminators” which could act to assess a generated candidate structure for whether it may possess or exhibit a particular property.
The “causal inference” of the ML is used to REMOVE structures from a candidate pool, not to GENERATE any new structure. HOW the inference selects any compound for removal from consideration is entirely unclear. The “counterfactuals” used in the “causal inference” do not set forth any particular counterfactual to be used in understanding why the computer generated the structure, or what its properties may encompass. The counterfactual limitation provides a definition of what a counterfactual could be, but not what it IS, or SHOULD be for this method. The “counterfactuals” do not provide any particular scenarios, or alternate scenarios of any kind, and are “based on” a long list of widely disparate ideas, results, actions, or correlations. No specific alternatives are provided. The generative AI of step 6 is not modifying a previously provided structure with previously identified characteristics. The AI is not generating a compound intended to have any particular property, function, effect, effectiveness or activity. It is entirely unrelated to the previously generated candidate (by the first ML) and is entirely unrelated to the initial “real drug candidate”. It does not use the “retrained ML” of step 2) to generate the compound.
The metes and bounds of the application of the second candidate to the proxy organism are completely unclear. The second candidate compound (not a known compound, or a specific modification of a known compound) is not synthesized prior to application to the proxy organism. As the candidate was “newly” generated by the distinct AI of step 6, there is no assumption that can be made as to whether the second candidate was known or existed, such that it could be applied to a physical proxy system. It is not compared to a database of known, available compounds such that one would recognize that the second candidate existed in the physical world in a form to be applied to the proxy organism. The proxy organism, the assays, and the biomarkers have no link to any aspect of the idea of “pre-clinical validation of an effectiveness of a candidate drug” as no particular drug, assay, biomarker, effectiveness or effect is identified.
The proxy organism does not need to identify any aspect of a biological response that may have been similar to the results using the “real drug compound.” It is completely unrelated. The biomarkers are not identified as being indicative of a particular clinical, biological, chemical or physical property of any kind. The assays are not identified as being indicative of any particular clinical, biological, chemical or physical process which may be related to a “clinical” effectiveness, as set forth in the preamble. It is entirely unclear what Applicant considers their actual invention with respect to the methods of the claims, versus those disclosed in the specification.
While claims are read in light of the specification, limitations from the specification are not to be read into the claims.
MPEP 2111, referring to In re Prater: “The court explained that "reading a claim in light of the specification, to thereby interpret limitations explicitly recited in the claim, is a quite different thing from ‘reading limitations of the specification into a claim,’ to thereby narrow the scope of the claim by implicitly adding disclosed limitations which have no express basis in the claim." The court found that applicant was advocating the latter, i.e., the impermissible importation of subject matter from the specification into the claim.”
MPEP 2111: "Though understanding the claim language may be aided by explanations contained in the written description, it is important not to import into a claim limitations that are not part of the claim. For example, a particular embodiment appearing in the written description may not be read into a claim when the claim language is broader than the embodiment." Superguide Corp. v. DirecTV Enterprises, Inc., 358 F.3d 870, 875, 69 USPQ2d 1865, 1868 (Fed. Cir. 2004).”
The dependent claims fail to remedy these issues, and all claims are indefinite.
Rejections Maintained
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
Claim(s) 1, 2, 5, 7-10, 15-17, 21-22 remain rejected under 35 U.S.C. 103 as being unpatentable over Oono in view of Munoz-Delgado and Stratton.
The elected invention comprises: The use of NN and generative AI to generate and select anti-infective compounds which are then tested in a proxy organism for the elected assays of hemolytic activity and erythrocytic activity. Applicant’s earliest priority date is 8/24/2020.
The claims have been heavily amended, however, the amendments are taught by one or more of the cited references.
Oono, K. et al. Generative machine learning systems for drug design. US 2017/0161635 A1, 6/8/2017.
Munoz-Delgado, Device and method for the generation of synthetic data in generative networks. US 12,242,957 B2, published 3/4/2025, with priority to at least 10/28/2019.
Stratton et al. (2009) In vitro biocompatibility studies of antibacterial quaternary polymers. Biomacromolecules, vol 10, p 2550-2555.
Oono is directed to: “[0006] In a first aspect, the methods and systems described herein relate to a computer system for generation of chemical compound representations. The system may comprise a probabilistic autoencoder. The probabilistic autoencoder may comprise a probabilistic encoder configured to encode chemical compound fingerprints as latent variables; a probabilistic decoder configured to decode latent representations and generate random variables over values of fingerprint elements; and/or one or more sampling modules configured to sample from a latent variable or a random variable. The system maybe trained by feeding it chemical compound fingerprints and training labels associated with the chemical compound fingerprints and generating reconstructions of chemical compound fingerprints, wherein the system's training is constrained by the reconstruction error. The reconstruction error may comprise the negative likelihood that an encoded chemical compound representation is drawn from the random variable generated by the probabilistic decoder. The system may be trained to optimize, for example to minimize, the reconstruction error. In some embodiments, the training is constrained by a loss function comprising the reconstruction error and a regularization error. The probabilistic autoencoder may be trained to learn to approximate an encoding distribution. The regularization error may comprise a penalty associated with the complexity of the encoding distribution. The training may comprise minimizing the loss function. In some embodiments, the training labels comprise one or more label elements having predetermined values.”
Munoz-Delgado is directed to: Generative AI process of generating synthetic data using a generator and a discriminator GAN, or VAE (variational autoencoder) that outputs a classification score for the generated data. The GAN/ VAE uses counterfactuals to find meaningful explanations for the results of the classification model.
Stratton is directed to proxy organism testing systems for testing the effects of a candidate drug in a relevant proxy system, (for hemolytic activity and erythrocytic activity) carrying out certain assays with fluorescent readouts at two different wavelengths, and analyzing the results.
Claim 1 is now directed to methods of retraining an initially trained ML, generating a first candidate, assessing the retraining, then generating and testing a second candidate drug compound, using generative AI and trained ML, as well as a proxy system.
Oono et al. disclosed methods of generating candidate drug compounds, using generative AI, and trained ML models.
Oono provides more than one ML model which has been pre-trained, meeting the limitation of the initially trained ML model of claim 1 at [0054]. The initially trained ML can be used to generate candidate compounds in various forms. The initially trained ML of Oono is semi-supervised, in that not all the training data for the pre-training (initial training) is labeled. [0059] The initially trained ML can be a neural network of any number of layers, classifiers, generators and discriminators [0060]. The pre-trained ML of Oono is then trained again (retrained) using labeled chemical compound representations, that are different from the initial set of data used in the pre-training. [0060]. This meets the BRI of removing one compound from the training set to provide a modified training set of claim 1. This pre-trained and then trained ML model can be assessed to determine whether the output has improved by meeting a similarity criterion, and by determining a retraining criterion. Oono discloses the comparison by calculating certain distance functions which fall within the BRI of both the similarity criterion and the retraining criterion. [0065, 0076-77, 0097-0098, 0104-0106, 0114, 0121, 0125-0126, Example 22, et al.] Once the required similarities are obtained, and the ML is validated, Oono then generates additional compounds using the trained (retrained) ML/ AI model.
Oono provides probabilistic autoencoders meeting the BRI of the generative AI and trained NN models of claim 1. See Fig 1-2 for the depiction of the autoencoder and the multi-component generative model with and without a predictor. Fig 5A and its description in the text depict the encoder and decoder. Fig 6 depicts the variational autoencoder [0032-0036, 0039+]. The probabilistic autoencoders of Oono generated chemical compound representations, based on provision of a seed compound and related training data. Oono also generated chemical compound representations ab initio, that is, without a seed structure or framework (See Fig 11). The variational autoencoders of Oono use penalties and rewards to determine what subset of generated chemical compound representations should be synthesized for testing/ production.
The details of the generative model are set forth in Oono beginning at [0048].
“[0049] Generative models, according to the methods and systems of the invention, can be used to randomly generate observable-data values given values of one or more hidden parameters. Generative models can be used for modeling data directly (i.e., modeling chemical compound observations drawn from a probability density function) or as an intermediate step to forming a conditional probability density function. Examples of generative models include, but are not limited to probabilistic autoencoders, variational autoencoders, Gaussian mixture models, hidden Markov models, and restricted Boltzmann machines. Generative models described in further detail elsewhere herein typically specify a joint probability distribution over chemical compound representations, i.e., fingerprints, and labels associated with the compounds.”
“[0054] A trained autoencoder, such as a trained probabilistic or variational autoencoder, may be used to generate or simulate observable-data values by sampling from the modeled joint probability distribution to generate a latent representation and by decoding this latent representation to reconstruct an input data point. In one embodiment, the weights of the autoencoder are adjusted during training by an optimization method. In one embodiment, the weights are adjusted by using backpropagation in conjunction with gradient descent to optimize, for example to minimize, the loss function. In one embodiment, individual layers of the autoencoder may be pre-trained and the weights of the entire autoencoder are fine-tuned together.”
“[0056] In various embodiments, generative models described herein comprise probabilistic autoencoders with multiple components. For example, a generative model may have one or more of an encoder, decoder, sampling module, and optional predictor (FIGS. 2A-2B). The encoder may be used to encode representations of chemical compounds, e.g., fingerprints, as an output of a different form, e.g. a latent variable.”
Oono provides trained ML models to classify the output of the autoencoder, and select a subset of candidate compounds for further testing based on the predicted presence of the desired property.
“[0060] The predictor may comprise a machine learning classification model. In some embodiments, the predictor is a deep neural network with two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen, fourteen, fifteen, sixteen, or more layers. In some embodiments, the predictor is a random forest classifier. In some embodiments, the predictor is trained with a training data set comprising chemical compound representations and their associated labels. In some embodiments, the predictor may have been trained previously, on a set of chemical compound representations and their associated labels that is different from the training data set used to train the generative model.”
Oono does not use the term “counterfactuals” to describe the predictor, however it is the predictor, (the trained ML) that is used to make decisions as to what candidate drug to select from a pool of candidate molecules to test for their predicted activity. Oono minimizes loss functions of the generated set of compounds in comparison to known drugs or compounds to optimize the structures generated by the generator and select those with the desired predicted activity. The predictor of Oono meets the BRI of the counterfactuals as claimed.
Oono sets forth in [0064] that “The compounds associated with the generated representations may be added to a compound database, used in computational screening methods, and/or synthesized and tested in assays.”
This is a direct suggestion to synthesize the candidates predicted to have (or classified as having) the desired property for testing, validation of the presence of the predicted property, or for production.
In the same field of using generative AI or GAN to generate new synthetic data meeting particular parameters, Munoz-Delgado disclosed GAN for generating new synthetic data based on the input data, and disclosed the use of counterfactuals to provide meaningful explanations for the results of the classification model. Described by Munoz-Delgado is an “optimization procedure that includes the computation of gradients for the generator based on the response of the classifier (discriminator) to outputs of the generator. The optimization may for example be based on the optimization of an objective function, e.g. on the maximization or minimization of an objective function, and comprise one or more neural networks.” [Columns 1 and 2].
From the Background [col 1-2]:
“The research for deep generative models such as Generative Adversarial Nets (GANs) or Variational Autoencoders (VAE) has exploded over the last few years. These usually unsupervised models are showing very promising results, being nowadays already able to reproduce natural looking images at high resolution and at sufficient quality to even fool human observers.
The future potential uses for deep generative models are considerable, ranging from anomaly detection, high quality synthetic image generation to providing explainability tools for datasets and models. This can for example enable the creation of corner-case datasets to test automotive perception systems and sensors or the creation of meaningful inputs, i.e., inputs that have significance for a human being, that test the decision boundaries of a given model.
However, these potential benefits come at the cost of a harder and more complex training compared to the standard classification models. For example, during the training of a GAN, several problems may occur such as training instability, mode collapse, high variance gradients, vanishing generator gradients, etc.
Within the context of explainable AI (artificial intelligence), understanding how the generator of a generative model has structured its latent space is crucial to understanding what it has really learned.
Typically, the latent space of a generator is high-dimensional, so that an exhaustive search or analysis is intractable. In recent years, several methods for sampling and visualizing the latent spaces of generative models, such as the ones described in the paper “Sampling Generative Networks” by Tom White, published in CoRR abs/1609.04468, 2016 (available at https:// arxiv.org/ abs/ 1609.04468 and incorporated herein by reference) have been developed to search through the latent space, producing visualizations of the latent codes along the way which illustrate the relationships that have been learned by the generator.
This search for explanations, i.e., the search through the latent space of the generator, may be conditioned on the response (output) of a classification (discriminative) model, as a way of making sure that the path taken by the generator through the latent space produces realistic and meaningful instances (samples). This enables a better understanding of how well the generator models the transitions between semantic concepts.”
Beginning at Col 11, Munoz-Delgado discussed why it is important to explain and understand what a trained generator of a GAN has learned:
“As mentioned earlier, the generator of a trained GAN may also be used in the field of explainable AI to increase the understanding of the outputs of deep neural networks, in particular of classification models.
For example, a generator of a GAN may be used to try to find meaningful explanations for the outputs of a classification model, wherein the generator of the GAN is trained on the same dataset as the classification model.
The trained generator of the GAN is then used together with the classification model to try to find meaningful explanations for the outputs of the classification model.
In particular, in order to find these meaningful explanations, a counterfactual to a synthetic data sample is generated using the trained generator of the GAN.
A counterfactual data sample is a data sample which, according to the classification model, has a different classification and/or a significantly different classification score (e.g. the classification score drops below a predetermined threshold) than a corresponding (synthetic) data sample.
It should be noted that in the following reference will be made to images for data samples, however, this does not exclude that other types of data samples may be used and/or generated.” [col 11].
The counterfactuals of Munoz-Delgado meet the BRI of the counterfactuals in claim 1, as they correlate to how a human operator would look at the synthetic data to classify it. [ col 11].
Col 11: “The generated counterfactual should ideally have a small amount of meaningful (pixel) differences, i.e., where the (pixel) differences correlate to parts of the image that represent semantic objects or parts, in other words correlate to parts of the image that a human operator would also likely look at when classifying the objects. Thus, to a human operator, the meaningful differences should plausibly justify the change in classification and/or change in classification score.
Generating a counterfactual for a synthetic image may in particular be achieved by use of a perturbation mask m, wherein this mask indicates perturbations that disturb the classification of the synthetic image by the classification model. This perturbation mask m is applied to the latent space representations of the generated synthetic image. This will cause the generator to produce a new synthetic image which the classification model classifies differently. In other words, the mask m causes the generator to generate a counterfactual to a corresponding synthetic image.
It should be noted that the new generated synthetic image does not necessarily have to be a counterfactual and/or be classified differently; it may be classified the same but with less certainty, it may be close to the decision boundary of the classification model, etc.
Finding a suitable (small), ideally the smallest, perturbation mask m which significantly affects the classification of a synthetic image is useful to understand the properties of the classification model, in particular to find critical decision boundaries of the classification model, i.e. non-smooth, typically overfitted decision boundaries where the predictions are neither uniform nor predictable. In others words, it leads to a better understanding of how sensible the classification model is to changes in its input.
This investigation of the operation of the classification model may then be further developed and quantified into validation and verifications metrics, in particular into an evaluation of the reliability of the classification model.”
The counterfactuals of Munoz-Delgado would have been able to have been applied to the input data of Oono, as both take predicted properties of synthetic representations of compounds, and attempt to optimize them based on desired activity. The counterfactuals of Munoz Delgado in this situation would have had the predictable result of an optimization of an objective function (min/ max) using a trained neural network, and would have optimized the subset of generated structures for the desired property (here, anti-infective candidates). The counterfactuals of Munoz-Delgado would have provided increased interpretability of the generated structures, and an increased understanding of the latent space covered by the generative model.
Neither Oono nor Munoz-Delgado carry out testing of the generated candidate drug in a proxy organism system.
In the same field of testing candidate drug compounds, Stratton discloses testing candidate drug compounds in a proxy organism system, which can reveal biomarkers, wherein the assay is a hemolysis assay or a Caco-2 epithelial cell culture proxy system. The hemolysis assay provides a readout of hemolytic concentration (HC50) which is the minimum concentration of compound at which 50% of the red blood cells (erythrocytes) are lysed.
Stratton notes that:
“In vitro cell viability assays offer specific advantages over RBC hemolysis assays as a measure of biocompatibility… erythrocytes lack a nucleus, the ability to reproduce, and the ability to synthesize protein. Alternatively, in vitro cell cultures provide living, metabolizing cells undertaking a wide array of cellular functions... In vitro cell viability testing directly exposes living cells to a material, providing for an alternative and in some ways advantageous method of exploring biocompatibility.” (introduction).
The proxy organism of Stratton is the Caco-2 human epithelial cell culture system, which is a model system for human epithelial tissues. Stratton discloses the proxy organism cell culture protocol at pages 2551-2552, “Cell Culture.” Upon the application of a candidate drug compound to the live Caco-2 cells, the cells can be subjected to LIVE/DEAD assay experiments which reveal biomarkers, by detecting wavelengths associated with each biomarker.
“To measure the viability of our cells after exposure to the polymers we chose to use the LIVE/DEAD Viability/Cytotoxicity Kit for Mammalian Cells (Invitrogen), one of the most common assays for the determination of cell viability or cytotoxicity. The LIVE/ DEAD assay kit uses two fluorescent dyes: a green calcein AM dye that is cleaved by ubiquitous intracellular esterase activity after being taken up by living cells, and an ethidium bromide homodimer-1 (EthD1) dye that binds to nucleic acids but is unable to penetrate intact, healthy cell membranes.”
Figure 2 illustrates the cell culture system after exposure to the compound, wherein green cells represent cells which uptake and metabolize the calcein AM and are living cells, while red cells are the nuclei of dead or dying cells with disrupted membranes. The biomarkers are the metabolization of the Calcein AM, and cell lysis. Observing and correlating the green wavelength of light and the red wavelength of light meet the newly added limitation to separating the light signals and observing the presence of each wavelength.
In KSR Int 'l v. Teleflex, the Supreme Court, in rejecting the rigid application of the teaching, suggestion, and motivation test by the Federal Circuit, indicated that “The principles underlying [earlier] cases are instructive when the question is whether a patent claiming the combination of elements of prior art is obvious. When a work is available in one field of endeavor, design incentives and other market forces can prompt variations of it, either in the same field or a different one. If a person of ordinary skill can implement a predictable variation, § 103 likely bars its patentability.” KSR Int'l v. Teleflex lnc., 127 S. Ct. 1727, 1740 (2007).
Applying the KSR standard of obviousness to Oono, Munoz-Delgado and Stratton, the examiner concluded, that the combination of the pre-training, retraining using a modified training set, generating a candidate, and satisfying certain criteria, as well as subsequent GAN generation of candidate drug compounds, utilizing trained machine learning models which take into account compound information and target tissue information, using counterfactuals related to past actions, occurrences, results, regressions, regression analyses or correlations, as disclosed by Oono and Munoz-Delgado, with the proxy organism testing system of Stratton, represents a combination of known elements which yield the predictable result of a workflow which generates compounds likely to have a particular desired activity, which can be assessed in a well characterized proxy organism system, with simple fluorescent detection. The use of the generative processes as taught by Oono and Munoz-Delgado would have achieved in the predictable result of a pool of candidate compounds structures likely to have had a desired property, as the counterfactuals provided by Munoz-Delgado are applicable to data describing chemicals to be tested in a given proxy system. Both Oono and Munoz-Delgado provided specific details as to the structure of the GAN, the reward or optimization routines, and the output of synthetic data meeting the initial requirements. Generating each selected candidate compound for either testing within the proxy organism system, or for production would have been well within the skill of the art in biochemistry. The use of the standard proxy organism system to test the candidate compound as taught by Stratton in this combination would further serve to achieve the predictable result of fluorescent readout data which easily measures cell viability and the ability of a compound to disrupt cell membranes, both critical for the analysis of a drug in development for the treatment of a living organism. Once the suitable pool of candidate was generated by the generative AI process, one of skill in the art of bioinformatics would have looked for an established proxy organism system in which to test each candidate drug compound and identify whether the compound was likely to be hemolytic or erythrocytic. One would have looked for an established proxy system such as the LIVE/DEAD assay of Stratton as it has the simple readout of fluorescence at two wavelengths, using standard laboratory equipment including an oscillator. One of skill would have had a reasonable expectation of success in this combination, as once the candidate drugs are generated and synthesized, their application to a test system is fully within the skill of ordinary skill in molecular biology, and use standard cell culture, and fluorescence detection methods. The further production of the selected and tested candidate drug would have been well within the skill of the ordinary artisan, as the candidate had been produced at least once to be used in the testing. Thus, the invention would have been prima facie obvious to one of ordinary skill in the art at the time of filing, absent evidence to the contrary.
Such a combination is merely a "predictable use of prior art elements according to their established functions." KSR Int’l 7, 127 S. Ct. at 1740.
With respect to claim 2, Oono discloses generating a pool of compounds to be tested and included compounds which meet a particular parameter.
With respect to claim 5, the tests of Stratton are those testing anti-infective properties. Compounds which have a high rate of cell lysis in the LIVE/DEAD assay are anti-infective, in that the cells are killed.
With respect to claims 7-8, fluorescent signals are acquired and processed routinely by Stratton. (Cell viability assays).
With respect to claims 9-10, the LIVE/DEAD and MTT assays of Stratton are assays which test erythrolytic activity. Standard hemolytic assays as disclosed by Stratton test for hemolysis activity. These are membrane interaction assays, as the LIVE/DEAD assays and the MTT assays test for membrane integrity.
With respect to claim 15, the red and green fluorescent dyes have two specific wavelengths, representing two different biomarkers, in Stratton.
With respect to claim 16, fluorescence is the signal received by the processing device of Stratton.
With respect to claims 21-22, Oono discloses ranking the generated compounds, which can be transmitted as desired. Munoz-Delgado discloses ranking the generated compounds according to their optimized parameters and transmitting them as desired.
Claim(s) 1, 2, 5, 7-10, 15-17, 21-22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Isayev in view of Munoz-Delgado and Stratton.
The elected invention comprises: The use of NN and generative AI to generate and select anti-infective compounds which are then tested in a proxy organism for the elected assays of hemolytic activity and erythrocytic activity. Applicant’s earliest priority date is 8/24/2020.
Isayev et al. Methods, systems and non-transitory computer readable media for automated design of molecules with desired properties using artificial intelligence. US 2020/0168302 A1, published 5/28/2020, with priority to at least 7/20/2018.
Munoz-Delgado, Device and method for the generation of synthetic data in generative networks. US 12,242,957 B2, published 3/4/2025, with priority to at least 10/28/2019.
Stratton et al. (2009) In vitro biocompatibility studies of antibacterial quaternary polymers. Biomacromolecules, vol 10, p 2550-2555.
Isayev is directed to: “computational methods, systems and non-transitory computer readable media for de-novo drug discovery, which is based on deep learning and reinforcement learning techniques. The subject matter described herein allows generating chemical compounds with desired properties. Two deep neural networks—generative and predictive, represent the general workflow. The process of training consists of two stages. During the first stage, both models are trained separately with supervised learning algorithms, and during the second stage, models are trained jointly with reinforcement learning approach.”
Munoz-Delgado is directed to: Generative AI process of generating synthetic data using a generator and a discriminator GAN, or VAE (variational autoencoder) that outputs a classification score for the generated data. The GAN/ VAE uses counterfactuals to find meaningful explanations for the results of the classification model.
Stratton is directed to proxy organism testing systems for testing the effects of a candidate drug in a relevant proxy system, (for hemolytic activity and erythrocytic activity) carrying out certain assays with fluorescent readouts at two different wavelengths, and analyzing the results.
Claim 1 is now directed to methods of training and retraining ML models for generating and testing candidate drug compounds, using generative AI and trained ML, as well as a proxy system.
With respect to claim 1:
Isayev provides more than one ML model which has been pre-trained, meeting the limitation of the initially trained ML model of claim 1 at [0070]. The initially trained ML can be used to generate candidate compounds in various forms. The initially trained ML of Isayev is semi-supervised, in that not all the training data for the pre-training (initial training) is optimized for a particular property. [0070] The initially trained ML can be a neural network of any number of layers, classifiers, generators and discriminators [0070]. The pre-trained ML of Isayev is then trained again (retrained) using labeled chemical compound representations, that are different from the initial set of data used in the pre-training. [0072-0073, Fig 13]. This meets the BRI of removing one compound from the training set to provide a modified training set of claim 1. This pre-trained and then trained ML model can be assessed to determine whether the output has improved by meeting a similarity criterion, and by determining a retraining criterion. Isayev discloses the comparison by calculating certain distance functions which fall within the BRI of both the similarity criterion and the retraining criterion. [0050-0053 similarity, 0054-0059 retraining] Once the required similarities are obtained, and the ML is validated, Isayev then generates additional compounds using the trained (retrained) ML/ AI model.
Isayev et al. disclosed methods of generating candidate drug compounds, using generative AI, and trained ML models in Fig 1, and
“[0008] … method based on deep reinforcement learning (RL) for generating chemical compounds with desired physical, chemical or bioactivity properties. Reinforcement learning (RL) is a subset of artificial intelligence, which is used to solve dynamic decision problems. RL involves analyzing possible actions, estimating the statistical relationship between the actions and their possible outcomes and then determining a treatment regime that attempts to find the most desirable outcome based on the analysis.”
Isayev provides generative AI meeting the BRI of the limitation, throughout. Fig 2 discloses a generative stack augmented RNN, with the training elements. Sample results of the generative process of structure generation are shown in Fig 3. The trained NN is described beginning at [0029] and the particular trained NN /ML at [0034]. Isayev applies their process to candidate drug design, with a focus on estimation and prediction of physical chemical or bioactivity properties of synthetic or generated drug molecules. The elected “anti-infective properties” are met by this limitation. A data element used by Isayev in the training is a predicted IC50 level, which is defined by Isayev as “the concentration of a drug that is required to inhibit 50% of a specific biological target in vitro.” This is a direct suggestion to utilize assays which have an IC50 parameter, in proxy (in vitro) organism systems, to validate the generated compound. Other data points collected and used by Isayev include Melting Temperature of the generated compound, (Tm) and Log P value, an n-octanol / water partition value, to “better mimic the requirements of drug-likeliness instead of property minimization.”
Isayev does not use counterfactuals to make decisions as to what candidate drug to select from a pool of candidate molecules to test for their predicted activity, but uses reward functions to optimize the structures generated by the generator and select those with the desired predicted activity. The reward function is an optimization (a minimization or maximization) of an objective function: exponential functions of pIC50.
“[0046] The reward function in both cases (min and max) was defined as exponential functions of pIC50 (see FIG. 11). The results of optimization are demonstrated in FIG. 11, graph (c). With minimization, the mean of predicted pIC050 distribution was shifted by about one unit. However, distribution is heavily tailed, and 24% of molecules are predicted to have practically no activity (pIC50<=4). In the maximization strategy, properties of generated molecules were more tightly distributed bet. In both cases models virtually synthesized known and novel compounds based on one scaffold as well as suggested new scaffolds.”
Isayev provided a database of novel chemical entities with optimized desired properties, that are “possible to synthesize” based on their process [0062]. This is a direct suggestion to synthesize the candidates predicted to have (or classified as having) the desired property for testing, validation of the presence of the predicted property, or production.
In the same field of using generative AI or GAN to generate new synthetic data meeting particular parameters, Munoz-Delgado disclosed GAN for generating new synthetic data based on the input data, and disclosed the use of counterfactuals to provide meaningful explanations for the results of the classification model. Described by Munoz-Delgado is an “optimization procedure that includes the computation of gradients for the generator based on the response of the classifier (discriminator) to outputs of the generator. The optimization may for example be based on the optimization of an objective function, e.g. on the maximization or minimization of an objective function, and comprise one or more neural networks.” [Columns 1 and 2].
From the Background [col 1-2]:
“The research for deep generative models such as Generative Adversarial Nets (GANs) or Variational Autoencoders (VAE) has exploded over the last few years. These usually unsupervised models are showing very promising results, being nowadays already able to reproduce natural looking images at high resolution and at sufficient quality to even fool human observers.
The future potential uses for deep generative models are considerable, ranging from anomaly detection, high quality synthetic image generation to providing explainability tools for datasets and models. This can for example enable the creation of corner-case datasets to test automotive perception systems and sensors or the creation of meaningful inputs, i.e., inputs that have significance for a human being, that test the decision boundaries of a given model.
However, these potential benefits come at the cost of a harder and more complex training compared to the standard classification models. For example, during the training of a GAN, several problems may occur such as training instability, mode collapse, high variance gradients, vanishing generator gradients, etc.
Within the context of explainable AI (artificial intelligence), understanding how the generator of a generative model has structured its latent space is crucial to understanding what it has really learned.
Typically, the latent space of a generator is high-dimensional, so that an exhaustive search or analysis is intractable. In recent years, several methods for sampling and visualizing the latent spaces of generative models, such as the ones described in the paper “Sampling Generative Networks” by Tom White, published in CoRR abs/1609.04468, 2016 (available at https:// arxiv.org/ abs/ 1609.04468 and incorporated herein by reference) have been developed to search through the latent space, producing visualizations of the latent codes along the way which illustrate the relationships that have been learned by the generator.
This search for explanations, i.e., the search through the latent space of the generator, may be conditioned on the response (output) of a classification (discriminative) model, as a way of making sure that the path taken by the generator through the latent space produces realistic and meaningful instances (samples). This enables a better understanding of how well the generator models the transitions between semantic concepts.”
Beginning at Col 11, Munoz-Delgado discussed why it is important to explain and understand what a trained generator of a GAN has learned:
“As mentioned earlier, the generator of a trained GAN may also be used in the field of explainable AI to increase the understanding of the outputs of deep neural networks, in particular of classification models.
For example, a generator of a GAN may be used to try to find meaningful explanations for the outputs of a classification model, wherein the generator of the GAN is trained on the same dataset as the classification model.
The trained generator of the GAN is then used together with the classification model to try to find meaningful explanations for the outputs of the classification model.
In particular, in order to find these meaningful explanations, a counterfactual to a synthetic data sample is generated using the trained generator of the GAN.
A counterfactual data sample is a data sample which, according to the classification model, has a different classification and/or a significantly different classification score (e.g. the classification score drops below a predetermined threshold) than a corresponding (synthetic) data sample.
It should be noted that in the following reference will be made to images for data samples, however, this does not exclude that other types of data samples may be used and/or generated.” [col 11].
The counterfactuals of Munoz-Delgado meet the BRI of the counterfactuals in claim 1, as they correlate to how a human operator would look at the synthetic data to classify it. [ col 11].
Col 11: “The generated counterfactual should ideally have a small amount of meaningful (pixel) differences, i.e., where the (pixel) differences correlate to parts of the image that represent semantic objects or parts, in other words correlate to parts of the image that a human operator would also likely look at when classifying the objects. Thus, to a human operator, the meaningful differences should plausibly justify the change in classification and/or change in classification score.
Generating a counterfactual for a synthetic image may in particular be achieved by use of a perturbation mask m, wherein this mask indicates perturbations that disturb the classification of the synthetic image by the classification model. This perturbation mask m is applied to the latent space representations of the generated synthetic image. This will cause the generator to produce a new synthetic image which the classification model classifies differently. In other words, the mask m causes the generator to generate a counterfactual to a corresponding synthetic image.
It should be noted that the new generated synthetic image does not necessarily have to be a counterfactual and/or be classified differently; it may be classified the same but with less certainty, it may be close to the decision boundary of the classification model, etc.
Finding a suitable (small), ideally the smallest, perturbation mask m which significantly affects the classification of a synthetic image is useful to understand the properties of the classification model, in particular to find critical decision boundaries of the classification model, i.e. non-smooth, typically overfitted decision boundaries where the predictions are neither uniform nor predictable. In others words, it leads to a better understanding of how sensible the classification model is to changes in its input.
This investigation of the operation of the classification model may then be further developed and quantified into validation and verifications metrics, in particular into an evaluation of the reliability of the classification model.”
The counterfactuals of Munoz-Delgado could have been applied to the input data of Isayev, as both take an objective function, and perturb it to either extreme. The counterfactuals of Munoz Delgado in this situation would have had the predictable result of an optimization of an objective function (min/ max) using a trained neural network, and would have optimized the subset of generated structures for the desired property (here, anti-infective candidates). The counterfactuals of Munoz-Delgado would have provided increased interpretability of the generated structures, and an increased understanding of the latent space covered by the generative model.
Neither Isayev nor Munoz-Delgado carry out testing of the generated candidate drug in a proxy organism system. However, Isayev provided useful characteristics to study in a proxy system, including the IC50 value and Munoz-Delgado indicated that generated synthetic datasets should be validated in an appropriate model.
In the same field of testing candidate drug compounds, Stratton discloses testing candidate drug compounds in a proxy organism system, which can reveal biomarkers, wherein the assay is a hemolysis assay or a Caco-2 epithelial cell culture proxy system. The hemolysis assay provides a readout of hemolytic concentration (HC50) which is the minimum concentration of compound at which 50% of the red blood cells (erythrocytes) are lysed.
Stratton notes that:
“In vitro cell viability assays offer specific advantages over RBC hemolysis assays as a measure of biocompatibility… erythrocytes lack a nucleus, the ability to reproduce, and the ability to synthesize protein. Alternatively, in vitro cell cultures provide living, metabolizing cells undertaking a wide array of cellular functions... In vitro cell viability testing directly exposes living cells to a material, providing for an alternative and in some ways advantageous method of exploring biocompatibility.” (introduction).
The proxy organism of Stratton is the Caco-2 human epithelial cell culture system, which is a model system for human epithelial tissues. Stratton discloses the proxy organism cell culture protocol at pages 2551-2552, “Cell Culture.” Upon the application of a candidate drug compound to the live Caco-2 cells, the cells can be subjected to LIVE/DEAD assay experiments which reveal biomarkers, by detecting wavelengths associated with each biomarker.
“To measure the viability of our cells after exposure to the polymers we chose to use the LIVE/DEAD Viability/Cytotoxicity Kit for Mammalian Cells (Invitrogen), one of the most common assays for the determination of cell viability or cytotoxicity. The LIVE/ DEAD assay kit uses two fluorescent dyes: a green calcein AM dye that is cleaved by ubiquitous intracellular esterase activity after being taken up by living cells, and an ethidium bromide homodimer-1 (EthD1) dye that binds to nucleic acids but is unable to penetrate intact, healthy cell membranes.”
Figure 2 illustrates the cell culture system after exposure to the compound, wherein green cells represent cells which uptake and metabolize the calcein AM and are living cells, while red cells are the nuclei of dead or dying cells with disrupted membranes. The biomarkers are the metabolization of the Calcein AM, and cell lysis. Observing and correlating the green wavelength of light and the red wavelength of light meet the newly added limitation to separating the light signals and observing the presence of each wavelength.
In KSR Int 'l v. Teleflex, the Supreme Court, in rejecting the rigid application of the teaching, suggestion, and motivation test by the Federal Circuit, indicated that “The principles underlying [earlier] cases are instructive when the question is whether a patent claiming the combination of elements of prior art is obvious. When a work is available in one field of endeavor, design incentives and other market forces can prompt variations of it, either in the same field or a different one. If a person of ordinary skill can implement a predictable variation, § 103 likely bars its patentability.” KSR Int'l v. Teleflex lnc., 127 S. Ct. 1727, 1740 (2007).
Applying the KSR standard of obviousness to Isayev, Munoz-Delgado and Stratton, the examiner concluded, that the combination of the pre-training, retraining using a modified training set, generating a candidate, and satisfying certain criteria, as well as subsequent GAN generation of candidate drug compounds, utilizing trained machine learning models which take into account compound information and target tissue information, using counterfactuals related to past actions, occurrences, results, regressions, regression analyses or correlations, as disclosed by Isayev and Munoz-Delgado, the GAN generation of candidate drug compounds, utilizing trained machine learning models which take into account compound information and target tissue information, using counterfactuals related to past actions, occurrences, results, regressions, regression analyses or correlations, with the proxy organism testing system of Stratton, represents a combination of known elements which yield the predictable result of a workflow which generates compounds likely to have a particular desired activity, which can be assessed in a well characterized proxy organism system, with simple fluorescent detection. The use of the generative processes as taught by Isayev and Munoz-Delgado would have achieved in the predictable result of a pool of candidate compounds structures likely to have had a desired property, as the counterfactuals provided by Munoz-Delgado are applicable to data describing chemicals to be tested in a given proxy system. Both Isayev and Munoz-Delgado provided specific details as to the structure of the GAN, the reward or optimization routines, and the output of synthetic data meeting the initial requirements. Generating each selected candidate compound for either testing within the proxy organism system, or for production would have been well within the skill of the art in biochemistry. The use of the standard proxy organism system to test the candidate compound as taught by Stratton in this combination would further serve to achieve the predictable result of fluorescent readout data which easily measures cell viability and the ability of a compound to disrupt cell membranes, both critical for the analysis of a drug in development for the treatment of a living organism. Once the suitable pool of candidate was generated by the generative AI process, one of skill in the art of bioinformatics would have looked for an established proxy organism system in which to test each candidate drug compound and identify whether the compound was likely to be hemolytic or erythrolytic. One would have looked for an established proxy system such as the LIVE/DEAD assay of Stratton as it has the simple readout of fluorescence at two wavelengths, using standard laboratory equipment including an oscillator. One of skill would have had a reasonable expectation of success in this combination, as once the candidate drugs are generated and synthesized, their application to a test system is fully within the skill of ordinary skill in molecular biology, and use standard cell culture, and fluorescence detection methods. The further production of the selected and tested candidate drug would have been well within the skill of the ordinary artisan, as the candidate had been produced at least once to be used in the testing. Thus, the invention would have been prima facie obvious to one of ordinary skill in the art at the time of filing, absent evidence to the contrary.
Such a combination is merely a "predictable use of prior art elements according to their established functions." KSR Int’l 7, 127 S. Ct. at 1740.
With respect to claim 2, Isayev discloses generating a pool of compounds to be tested and included compounds which meet a particular parameter.
With respect to claim 5, the tests of Stratton are those testing anti-infective properties. Compounds which have a high rate of cell lysis in the LIVE/DEAD assay are anti-infective, in that the cells are killed.
With respect to claims 7-8, fluorescent signals are acquired and processed routinely by Stratton. (Cell viability assays).
With respect to claims 9-10, the LIVE/DEAD and MTT assays of Stratton are assays which test erythrolytic activity. Standard hemolytic assays as disclosed by Stratton test for hemolysis activity. These are membrane interaction assays, as the LIVE/DEAD assays and the MTT assays test for membrane integrity.
With respect to claim 15, the red and green fluorescent dyes have two specific wavelengths, representing two different biomarkers, in Stratton.
With respect to claim 16, fluorescence is the signal received by the processing device of Stratton.
With respect to claims 21-22, Isayev discloses ranking the generated compounds by reward, which can be transmitted as desired. Munoz-Delgado discloses ranking the generated compounds according to their optimized parameters and transmitting them as desired.
Applicant’s arguments:
Applicant’s arguments that the cited references do not teach the new limitations have been considered, but are not persuasive, as the new limitations are taught by Oono/ Isayev in each rejection as pointed out above.
New Grounds of Rejection
Claim Rejections - 35 USC § 103
Claim(s) 1-2, 5, 7-10, 15-17, 21-22 is/are rejected under 35 U.S.C. 103 as being unpatentable over Oskooei, Yang (2018), Munoz-Delgado in view of Karabasz (2018).
Oskooei et al. Drug compound identification for target tissue cells. US 2020/0365238 A1, (11/19/2020, filed 11/15/2019).
Yang (2018) Concepts of artificial intelligence for computer-assisted drug discovery. Chemical Reviews, vol 119, p10520-10594.
Munoz-Delgado, Device and method for the generation of synthetic data in generative networks. US 12,242,957 B2, published 3/4/2025, with priority to at least 10/28/2019.
Karabasz, A. et al. (2018) In vitro toxicity studies of biodegradable polyelectrolyte capsules. International Journal of Nanomedicine, p5159-5172.
Claim 1 is now directed to methods of retraining an initially trained ML, generating a first candidate, assessing the retraining, then generating and testing a second candidate drug compound, using generative AI and trained ML, as well as a proxy system.
Oskooei et al. disclosed methods of generating candidate drug compounds, using generative AI, and trained ML models.
Oskooei provides more than one ML model which has been pre-trained, meeting the limitation of the initially trained ML model of claim 1 at [abstract, 0006, 0035, 0046-0047]. The initially trained ML can be used to generate candidate compounds in various forms. The initially trained ML of Oskooei is semi-supervised, in that not all the training data for the pre-training (initial training) is labeled. [0034-0035] The initially trained ML can be a neural network of any number of layers, classifiers, generators and discriminators [0035, Figures]. The pre-trained ML of Oskooei is then trained again (retrained) using updated parameter weightings, that are different from the initial set of data used in the pre-training. [0034, 0041-0051]. This pre-trained and then trained ML model can be assessed to determine whether the output has improved by meeting a similarity criterion, and by determining a retraining criterion. Oskooei discloses the comparison by calculating certain distance functions which fall within the BRI of both the similarity criterion and the retraining criterion. [0045, 0049] Once the required similarities are obtained, and the ML is validated, Oskooei then generates additional compounds using the trained (retrained) ML/ AI model.
Oskooei does not specifically remove one real drug compound from the training set, however this “leave one out” analysis was well known and used in the training of ML and AI networks, as shown by Yang.
Yang discusses the use of semi-supervised ML models in the generation of candidate drug compounds, beginning at page 7.
“2.1.1.3. Semisupervised Learning. Semisupervised learning (Figure 3A) sits at a junction between supervised and unsupervised learning approaches71 and can be useful when many input data (X) are available, with only relatively few labeled samples (Y). Many real-world drug discovery problems fall into this area. Semisupervised learning can maximize the use of unlabeled data to either modify or reprioritize hypotheses obtained from limited labeled data alone. This feat is usually achieved by (i) using a supervised learning algorithm to train a model with the available labeled data; (ii) applying this trained model to predict labels for the unlabeled data; and (iii) retraining the model with those pseudo-labeled samples and the labeled data. In this way, distributions of the original labeled samples are used for model building and for potentially increasing its predictive power with little additional real-world, e.g., practical experimental, cost.
There are many semisupervised learning methods, and they make different assumptions about the relationship between the unlabeled data distribution and the function learned from the labeled set. Some often-used methods include: self-training, which assumes that a learner’s high confidence predictions are correct;72 cotraining, with the idea to include different “views” of the objects to be predicted;73 transductive support vector machines,74 which implement the idea of transductive learning by including unlabeled data in the computation of the margin; and graph-based methods,75 which rely on the idea of building a graph on the labeled and unlabeled data where instances connected by heavily weighted edges tend to be assigned the same label. Such assumptions are equivalent to prior domain knowledge, and the success of semisupervised learning depends to a large extent on choosing a priori assumptions (constraints, hypotheses) that fit the underlying problem
structure well.”
After retraining, Yang notes that the retrained model must be validated and tested to determine its performance on the desired task. (p19)
“One should first determine whether the training set performance is acceptable. If the training set performance in unacceptable, even with model regularization (such as dropout226) disabled, then it is likely that the architecture chosen could benefit from some tuning or reconstruction. In general, it should be possible to learn the training data with even a poorly optimized architecture, so it is best to focus efforts here rather than immediately going in search of more data. If, after tuning or trying another architecture (where appropriate), one still does not see improvement, it might be time to consider the quality of the training data, in terms of noise and errors, and check if the training distribution is balanced and the input-output correspondence is logical. 270−272 A further recommended test is to retrain the model with scrambled Y-data as a sanity check, where poor
performance of the trained model on the scrambled data is desirable to demonstrate that the model is learning meaningful relationships in the unscrambled case.273”
Yang reviews known NN which were used to generate SMILES representations of molecules beginning at p 29, and Fig 24. This meets the BRI of removing at least one real drug compound to generate a modified training set, followed by retraining, and evaluating the trained model.
“Segler et al.478 demonstrated that RNNs trained on the SMILES representations of molecules can both learn the grammar required to generate valid SMILES and generate candidate molecules with similar properties to those of template compounds but with differing scaffolds. The de novo drug design cycle of this method adopts transfer learning (Figure 24), in which an RNN model is first trained on a large set of molecules and then further retrained with a small set of active molecules to bias the sampled molecules toward a given template set. Their retrospective results showed that their de novo RNN model could reproduce 28% of 1240 known active compounds against Plasmodium falciparum, without having seen the compounds in the initial training, having utilized a roughly equivalent number for fine-tuning. For Staphylococcus aureus, the corresponding figures were 14% of 6051 test molecules, having trained on 1000.”
Oskooei et al. provide generative AI, (generative adversarial networks) which take an initial compound molecule which can be a real drug, and generates multiple additional candidate molecules that may have the desired activity.
Beginning at [0034] Oskooei discloses methods of utilizing ML models based on neural networks, which take initial compound information, and target tissue (proxy organism) information, to generate additional compounds for testing. The trained ML are trained in an iterative process of signal propagation and weight-update operations until a convergence condition is reached. [0034-0035] The trained ML model is then used for inference in the design of additional compounds [0036-0039, 0045-46, et al. Fig 2, Fig 5]. Once a candidate structure has been identified which may have the desired property it can be synthesized for testing. The property desired for the compound can be an efficacy value, such as an IC50, gene expression data, microarray data, MS spectra, SNV or CAN values, or any biomolecular measurements representing a target cell sample [0038]. Each property, such as an IC50 level, is related to a proxy organism / proxy test system. Generation of compounds and validation for the property is suggested by Oskooei to reduce the testing of compounds unlikely to have the desired property.
In the same field of using generative AI or GAN to generate new synthetic data meeting particular parameters, Munoz-Delgado disclosed GAN for generating new synthetic data based on the input data, and disclosed the use of counterfactuals to provide meaningful explanations for the results of the classification model. Described by Munoz-Delgado is an “optimization procedure that includes the computation of gradients for the generator based on the response of the classifier (discriminator) to outputs of the generator. The optimization may for example be based on the optimization of an objective function, e.g. on the maximization or minimization of an objective function, and comprise one or more neural networks.” [Columns 1 and 2].
From the Background [col 1-2]:
“The research for deep generative models such as Generative Adversarial Nets (GANs) or Variational Autoencoders (VAE) has exploded over the last few years. These usually unsupervised models are showing very promising results, being nowadays already able to reproduce natural looking images at high resolution and at sufficient quality to even fool human observers.
The future potential uses for deep generative models are considerable, ranging from anomaly detection, high quality synthetic image generation to providing explainability tools for datasets and models. This can for example enable the creation of corner-case datasets to test automotive perception systems and sensors or the creation of meaningful inputs, i.e., inputs that have significance for a human being, that test the decision boundaries of a given model.
However, these potential benefits come at the cost of a harder and more complex training compared to the standard classification models. For example, during the training of a GAN, several problems may occur such as training instability, mode collapse, high variance gradients, vanishing generator gradients, etc.
Within the context of explainable AI (artificial intelligence), understanding how the generator of a generative model has structured its latent space is crucial to understanding what it has really learned.
Typically, the latent space of a generator is high-dimensional, so that an exhaustive search or analysis is intractable. In recent years, several methods for sampling and visualizing the latent spaces of generative models, such as the ones described in the paper “Sampling Generative Networks” by Tom White, published in CoRR abs/1609.04468, 2016 (available at https:// arxiv.org/ abs/ 1609.04468 and incorporated herein by reference) have been developed to search through the latent space, producing visualizations of the latent codes along the way which illustrate the relationships that have been learned by the generator.
This search for explanations, i.e., the search through the latent space of the generator, may be conditioned on the response (output) of a classification (discriminative) model, as a way of making sure that the path taken by the generator through the latent space produces realistic and meaningful instances (samples). This enables a better understanding of how well the generator models the transitions between semantic concepts.”
Beginning at Col 11, Munoz-Delgado discussed why it is important to explain and understand what a trained generator of a GAN has learned:
“As mentioned earlier, the generator of a trained GAN may also be used in the field of explainable AI to increase the understanding of the outputs of deep neural networks, in particular of classification models.
For example, a generator of a GAN may be used to try to find meaningful explanations for the outputs of a classification model, wherein the generator of the GAN is trained on the same dataset as the classification model.
The trained generator of the GAN is then used together with the classification model to try to find meaningful explanations for the outputs of the classification model.
In particular, in order to find these meaningful explanations, a counterfactual to a synthetic data sample is generated using the trained generator of the GAN.
A counterfactual data sample is a data sample which, according to the classification model, has a different classification and/or a significantly different classification score (e.g. the classification score drops below a predetermined threshold) than a corresponding (synthetic) data sample.
It should be noted that in the following reference will be made to images for data samples, however, this does not exclude that other types of data samples may be used and/or generated.” [col 11].
The counterfactuals of Munoz-Delgado meet the BRI of the counterfactuals in claim 1, as they correlate to how a human operator would look at the synthetic data to classify it. [ col 11].
Col 11: “The generated counterfactual should ideally have a small amount of meaningful (pixel) differences, i.e., where the (pixel) differences correlate to parts of the image that represent semantic objects or parts, in other words correlate to parts of the image that a human operator would also likely look at when classifying the objects. Thus, to a human operator, the meaningful differences should plausibly justify the change in classification and/or change in classification score.
Generating a counterfactual for a synthetic image may in particular be achieved by use of a perturbation mask m, wherein this mask indicates perturbations that disturb the classification of the synthetic image by the classification model. This perturbation mask m is applied to the latent space representations of the generated synthetic image. This will cause the generator to produce a new synthetic image which the classification model classifies differently. In other words, the mask m causes the generator to generate a counterfactual to a corresponding synthetic image.
It should be noted that the new generated synthetic image does not necessarily have to be a counterfactual and/or be classified differently; it may be classified the same but with less certainty, it may be close to the decision boundary of the classification model, etc.
Finding a suitable (small), ideally the smallest, perturbation mask m which significantly affects the classification of a synthetic image is useful to understand the properties of the classification model, in particular to find critical decision boundaries of the classification model, i.e. non-smooth, typically overfitted decision boundaries where the predictions are neither uniform nor predictable. In others words, it leads to a better understanding of how sensible the classification model is to changes in its input.
This investigation of the operation of the classification model may then be further developed and quantified into validation and verifications metrics, in particular into an evaluation of the reliability of the classification model.”
The counterfactuals of Munoz-Delgado could have been applied to the input data of Oksooei and Yang, as both take predicted properties of synthetic representations of compounds, and attempt to optimize them based on desired activity. The counterfactuals of Munoz Delgado in this situation would have had the predictable result of an optimization of an objective function (min/ max) using a trained neural network, and would have optimized the subset of generated structures for the desired property (here, anti-infective candidates). The counterfactuals of Munoz-Delgado would have provided increased interpretability of the generated structures, and an increased understanding of the latent space covered by the generative model.
Neither Oskooei, Yang, nor Munoz-Delgado carry out testing of the generated candidate drug in a proxy organism system.
In the field of testing candidate drug compounds for efficacy, safety, or suitability, Karabasz provides multiple test strategies to analyze candidate compounds. As set forth in the abstract, Karabasz discloses hemolysis assays, viability tests, biomarker expression, nitric oxide production, intracellular reactive oxygen species levels, detection of antioxidant enzyme activity, and analysis of DNA damage.
The compounds of Karabasz (different layered differently charged nanocapsules) were tested for toxicity, individually. Proxy organism systems included the use of human erythrocytes, human PBMC’s, human hepatoma HepG2 cell lines, mouse brain endothelial cells and mouse macrophage cell lines. For the hemolysis assay, the compounds were applied to the human erythrocyte cultures, and after treatment, the concentration of the released hemoglobin was determined by spectrophotometer at 540nm. For the cytotoxicity assay, human PBMC’s were exposed to the compound, and viability was determined by the trypan dye exclusion test. For the MTT assay, mouse brain endothelial cells and human hepG2 cells were cultured and exposed to the compound. Cell viability was determined by measuring absorbance at 545nm by spectrophotometry. Nitric Oxide production was determined by culturing P388D1 and MBE cells, which were exposed to the compound. Nitrite levels were measured using Griess reagent, and the absorbance was measured at 545nm using spectrophotometry.
VCAM-1 biomarker expression was measured by culturing MBE cells with the compound, followed by staining with fluorescent tagged antibodies, and analyzed by flow cytometry for fluorescence.
Reactive oxygen species were detected by culturing HepG2 cells with the compound, performing the DCFH-DA assay, and measuring fluorescent area with an excitation of 470nm and an emission of 535 nm by spectroscopy. SOD activity was measured by culturing HepG2 cells with the compound, then stained using a staining solution followed by gel analysis. DNA damage was measured by the Comet assay, which tested three different cell types which were incubated with the compound, then DNA damage was induced by Etopside. After lysis, the DNA from each cell sample was analyzed by electrophoresis.
Observing and correlating the 545nm wavelength of light and the 535nm wavelength of light meet the newly added limitation to separating the light signals and observing the presence of each wavelength, correlated to each respective biomarker.
In KSR Int 'l v. Teleflex, the Supreme Court, in rejecting the rigid application of the teaching, suggestion, and motivation test by the Federal Circuit, indicated that “The principles underlying [earlier] cases are instructive when the question is whether a patent claiming the combination of elements of prior art is obvious. When a work is available in one field of endeavor, design incentives and other market forces can prompt variations of it, either in the same field or a different one. If a person of ordinary skill can implement a predictable variation, § 103 likely bars its patentability.” KSR Int'l v. Teleflex lnc., 127 S. Ct. 1727, 1740 (2007).
Applying the KSR standard of obviousness to Oskooei, Yang, Munoz-Delgado and Karabasz, the examiner concludes, that the combination of the pre-training, retraining using a modified training set, generating a candidate, and satisfying certain criteria, as well as subsequent GAN generation of candidate drug compounds, utilizing trained machine learning models which take into account compound information and target tissue information, using counterfactuals related to past actions, occurrences, results, regressions, regression analyses or correlations, as disclosed by Oskooei, Yang and Munoz-Delgado, the GAN generation of candidate drug compounds, utilizing trained machine learning models which take into account compound information and target tissue information, using counterfactuals related to past actions, occurrences, results, regressions, regression analyses or correlations, with the proxy organism testing system of Stratton, represents a combination of known elements which yield the predictable result of a workflow which generates compounds likely to have a particular desired activity, which can be assessed in a well characterized proxy organism system, with simple fluorescent detection other signal generation.
The use of the generative processes comprising pre-training, retraining and model analysis as taught by Oskooei, Yang, and Munoz-Delgado would have achieved a pool of candidate compounds structures likely to have had a desired property, because of the information used in the training of the ML and the use of the counterfactuals. Generating each compound was well within the skill of the art in biochemistry. The use of the standard proxy organism systems to test the candidate compound as taught by Karabasz in this combination would further serve to achieve the predictable result of data which easily measures cell viability, biomarker expression, DNA damage and the ability of a compound to disrupt cell membranes, all critical for the analysis of a drug in development for the treatment of a living organism.
Once the suitable pool of candidate was generated by the generative AI process, one would have looked for established proxy organism systems to test each candidate drug compound and identify whether the compound was likely to be hemolytic or erythrolytic, or express a given biomarker. The well-known proxy assay systems including the proxy organisms of HepG2 cells, PBMC, mouse brain endothelial cells et al meet those requirements. Karabasz demonstrates testing for a variety of properties which can be measured by spectroscopy, flow cytometry, and other routine processes. One of skill would have had a reasonable expectation of success in this combination, as once the candidate drugs are generated and synthesized, their application to a test system is fully within the skill of ordinary skill in molecular biology, and use standard cell culture, and fluorescence detection methods.
The final synthesis of the tested compound would have been well within the level of one of ordinary skill in the art.
Thus, the invention would have been prima facie obvious to one of ordinary skill in the art at the time of filing, absent evidence to the contrary.
Such a combination is merely a "predictable use of prior art elements according to their established functions." KSR Int’l 7, 127 S. Ct. at 1740.
With respect to claim 2, Ooskoei discloses generating a pool of compounds to be tested and included compounds which meet a particular parameter.
With respect to claim 5, the tests of Stratton are those testing anti-infective properties. Compounds which have a high rate of cell lysis in the LIVE/DEAD assay are anti-infective, in that the cells are killed.
With respect to claims 7-8, fluorescent signals are acquired and processed routinely by Stratton. (Cell viability assays).
With respect to claims 9-10, the LIVE/DEAD and MTT assays of Stratton are assays which test erythrolytic activity. Standard hemolytic assays as disclosed by Stratton test for hemolysis activity. These are membrane interaction assays, as the LIVE/DEAD assays and the MTT assays test for membrane integrity.
With respect to claim 15, the red and green fluorescent dyes have two specific wavelengths, representing two different biomarkers, in Stratton.
With respect to claim 16, fluorescence is the signal received by the processing device of Stratton.
With respect to claims 21-22, Munoz-Delgado discloses ranking the generated compounds according to their optimized parameters and transmitting them as desired.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MARY K ZEMAN whose telephone number is 5712720723. The examiner can normally be reached on 8am-2pm M-F. Email may be sent to mary.zeman@uspto.gov if the appropriate permissions have been filed.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Larry Riggs can be reached on 571 270-3062. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/MARY K ZEMAN/ Primary Examiner, Art Unit 1686