DETAILED ACTION
Notice of Pre-AIA or AIA Status
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Status
2. Claims 10-25 are cancelled.
Claims 1-9 and 26-36 are currently pending and under exam herein.
Claims 1-9 and 26-36 are rejected.
Priority
3. This application is a National Stage Application under 35 U.S.C. 371. The claimed benefit of International Application No. PCT/EP2021/082707, filed 23 November 2021, which
claims the benefit of the filing date of U.S. Provisional Patent Application Serial No. 63/118,919 filed on 28 November 2020 is acknowledged. In this action, all claims are examined as though they had an effective filing date of 28 November 2020. In future actions, the effective filing date of one or more claims may change, due to amendments to the claims, or further analysis of the disclosure(s) of the priority application(s).
Information Disclosure Statement
4. The information disclosure statements (IDSs) submitted on 08 November 2023, 04 September 2025 and 06 February 2026 are being considered by the examiner.
Drawings
5. The drawings submitted 27 April 2023 are accepted by the examiner.
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.
6. Claims 4-9 and 29-34 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.
Claims 4-5, 7-9, 29-30 and 32-34 recite the limitation "generating the combined input". There is insufficient antecedent basis for this limitation in the claims because “a combined input” has not been introduced in the claims or the claim from which they depend. Dependent claims 6 and 31 are similarly rejected as they do not resolve the indefiniteness issue. For the purpose of examination, claims 4-5 and 7-9 will be considered to depend on claim 3 and claims 29-30 and 32-34 will be considered to depend on claim 28, which each recite ‘a combined input’.
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.
7. Claims 1-9 and 26-36 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 2A, Prong 1
In accordance with MPEP § 2106, claims found to recite statutory subject matter (Step 1: YES) are then analyzed to determine if the claims recite any concepts that equate to an abstract idea, law of nature or natural phenomenon (Step 2A, Prong 1). In the instant application, the claims recite the following limitations that equate to an abstract idea:
Claims 1, 26 and 35 recite: processing the network input for the first iteration to generate structure parameters for the first iteration that define an initial predicted structure for the protein
Claims 1, 26 and 35 recite: generating, from (i) the structure parameters generated at a preceding iteration that precedes the subsequent iteration in the sequence, (ii) one or intermediate outputs generated by the protein structure prediction neural network while generating the structure parameters at the last iteration, or (iii) both, features for the subsequent iteration
Claims 1, 26 and 35 recite: processing the features and the network input for the subsequent iteration to generate structure parameters for the subsequent iteration that define another predicted structure for the protein
Claims 3 and 28 recite: wherein processing the features and the network input for the subsequent iteration using the protein structure prediction neural network to generate structure parameters for the subsequent iteration that define a predicted structure for the protein comprises: generating a combined input from the features and the network input for the subsequent iteration
Claims 3 and 28 recite: processing the combined input to generate the structure parameters for the subsequent iteration
Claims 4 and 29 recite: at each iteration, repeatedly update the initial pair embeddings while generating the structure parameters for the iteration
Claims 4 and 29 recite: generating the features comprises generating, from updated pair embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of pair embeddings
Claims 4 and 29 recite: generating the combined input comprises combining the transformed set of pair embeddings and the initial pair embeddings for the subsequent iteration
Claims 5 and 30 recite: at each iteration, the structure prediction neural network is configured to generate one or more sets of single embeddings that each include a respective single embedding for each amino acid the protein
Claims 5 and 30 recite: while generating the structure parameters for the iteration
Claims 5 and 30 recite: generating the features comprises generating, from one of the sets of single embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of single embeddings
Claims 5 and 30 recite: generating the combined input comprises combining the transformed set of single embeddings and the initial MSA representation for the subsequent iteration
Claims 6 and 31 recite: the method of claim 5, wherein combining the transformed set of single embeddings and the initial MSA representation for the subsequent iteration comprises adding the transformed set of single embeddings to a first row of the initial MSA representation
Claims 7 and 32 recite: generating the features comprises: generating, from the predicted 3-D spatial locations for the amino acids specified by the structure parameters at the preceding iteration, a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein
Claims 7 and 32 recite: generating, from the distance map, a transformed distance map that has a same dimensionality as the initial pair embeddings
Claims 7 and 32 recite: generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration
Claims 7 and 32 recite: at each iteration, the structure parameters specify, for each amino acid, a predicted 3-D spatial location of a specified atom in the amino acid in the structure of the protein
Claims 8 and 33 recite: at each iteration, the structure parameters specify a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein
Claims 8 and 33 recite: generating the features comprises: generating, from the distance map specified by the structure parameters at the preceding iteration, a transformed distance map that has a same dimensionality as the initial pair embeddings
Claims 8 and 33 recite: generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration.
Claims 9 and 34 recite: generating the combined input comprises modifying the initial pair embeddings for the subsequent iteration by adding the protein and the structure prediction defined by the embeddings at the preceding iteration to the set of one or more template sequences and the corresponding known structures
The limitations regarding: ‘generating structure parameters’, ‘generating features’, ‘generating a combined input’, ‘update the initial pair embeddings’, ‘generating a transformed set of pair embeddings’, ‘combining embeddings’, ‘generate one or more sets of single embeddings’, ‘generating a transformed set of single embeddings’, ‘combining single embeddings and the initial MSA representation’, ‘adding the transformed set of single embeddings to a first row of the initial MSA representation’, ‘generating a distance map’, ‘generating a transformed distance map’, ‘combining the transformed distance map and the initial pair embeddings’, ‘generating a transformed distance map’, ‘combining the transformed distance map and the initial pair embeddings’ and ‘adding the protein and the structure prediction to the set of one or more template sequences and the corresponding known structures’ are verbal equivalents that describe mathematical calculations such as representing data as vectors and matrices and performing simple operations with them including adding, averaging and/or concatenating vectors, scaling numbers, calculating Euclidian distance between atoms using an equation, measuring dihedral angles between planes and calculating attention with an equation. These mathematical calculations are so simple that they could be performed in the human mind or with pen and paper Therefore, these limitations fall under the "Mathematical concepts" and "Mental processes" groupings of abstract ideas.
While some of the mathematical concepts are performed using a protein structure prediction neural network, including the limitations: ‘generating structure parameters’, ‘updating the initial pair embeddings’, ‘generating one or more sets of single embeddings that each include a respective single embedding for each amino acid in the protein’, the claims do not provide any details about how the neural network operates or how the embeddings are generated. The plain meaning of ‘generating an embedding’ encompasses math because vector representations of data are generated using mathematical equations and calculations such as matrix multiplication, cosine similarity or neural network weights. Furthermore, the steps of generating the embeddings and manipulating the data as generically disclosed could also practically be performed in the human mind, thus they also qualify as judicial exceptions under the mental process grouping.
While claims 26-36 recite performing some aspects of the analysis with a processor, there are no additional limitations that indicate that this processor requires anything other than carrying out the recited mental process or mathematical concept in a generic computer environment. Merely reciting that a mental process is being performed in a generic computer environment does not preclude the steps from being performed practically in the human mind or with pen and paper as claimed. If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components, then if falls within the "mental processes" grouping of abstract ideas.
As such, claims 1-36 recite an abstract idea (Step 2A, Prong 1: YES).
Step 2A, Prong 2
Claims found to recite a judicial exception under Step 2A, Prong 1 are then further analyzed to determine if the claims as a whole integrate the recited judicial exception into a practical application or not (Step 2A, Prong 2). This judicial exception is not integrated into a practical application because the claims do not recite an additional element that reflects an improvement to technology or applies or uses the recited judicial exception in some other meaningful way. Rather, the instant claims recite additional elements that amount to mere instructions to implement the abstract idea in a generic computing environment or insignificant extra-solution activity. Specifically, the claims recite the following additional elements:
Claims 1, 26 and 35 recite: a method performed by one or more computers for predicting a structure of a protein comprising one or more chains, wherein each chain comprises a sequence of amino acids, the method comprising: at a first iteration of a sequence of iterations that comprises the first iteration followed by one or more subsequent iterations: obtaining a network input for the first iteration that characterizes the protein
Claims 1, 26 and 35 recite: at each subsequent iteration in the sequence of iterations: obtaining a network input for the subsequent iteration that characterizes the protein
Claims 1, 26 and 35 recite: using a protein structure prediction neural network
Claims 1, 26 and 35 recite: determining a final predicted structure for the protein from the structure parameters for the last iteration in the sequence
Claims 2, 27 and 36 recite: wherein the network input for each iteration in the sequence of iterations is the same network input
Claims 3 and 28 recite: using a protein structure prediction neural network
Claims 4 and 29 recite: wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein
Claims 5 and 30 recite: the method of claim 1, wherein: the respective network input for each iteration comprises an initial multiple sequence alignment (MSA) representation that represents a respective MSA corresponding to each chain in the protein
Claims 7 and 32 recite: the method of claim 1, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein
Claims 8 and 33 recite: the method of claim 1, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein
Claims 9 and 34 recite: the method of claim 1, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein that is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences
Claim 26 recites: a system comprising: one or more computers
Claim 26 recites: one or more storage devices communicatively coupled to the one or more computers
Claim 26 recites: wherein the one or more storage devices store instructions
Claim 35 recites: one or more non-transitory computer storage media
Claim 35 recites: storing instructions
Claim 36 recites: the non-transitory computer storage media
Claim 36 recites wherein the network input for each iteration in the sequence of iterations is the same network input
The limitations for ‘obtaining a network input’ merely serve to gather data that is used an input for the judicial exception. Therefore, these limitations are mere data gathering activities. As set forth in MPEP 2106.05(g), mere data gathering activity has been identified by the courts as insignificant extra-solution activity that does not provide a practical application. Dependent claims 2, 4, 7-9, 27, 29, 32-34, and 36, directed to either the format or type of embedding of the input data, or the source or content of the input data, further limit the data gathering activities, but don’t change their position as data gathering activities.
The limitations directed to ‘determining a final predicted structure’ in claims 1, 26 and 35 are recited generically with no details provided in the specification or the claims as to how the structure parameters output from the neural network are used to determine the final predicted structure. Therefore, these limitations amount to mere instructions to apply the judicial exception in a generic way and do not integrate the exception into a practical application (MPEP 2106.05(f)).
There are no limitations that indicate that the computers, ‘computer readable medium for storing instructions’ or ‘using a protein structure prediction neural network’ requires anything other than a generic computing system. As such, these limitations equate to mere instructions to implement the abstract idea on a generic computer that the courts have stated does not render an abstract idea eligible in Alice Corp., 573 U.S. at 223, 110 USPQ2d at 1983. See also 573 U.S. at 224, 110 USPQ2d at 1984.
The limitations of claims 1, 3-5, 26, 28-30 and 35, directed to ‘using a neural network to generate structure parameters’, and claims 1, 26 and 35 directed to ‘determining a final predicted structure’ fail to integrate the exception into a practical application (for example by improving either a computer or protein structure-prediction technology).
The above recited additional elements do not provide a practical application of the recited judicial exception. As such, claims 1-9 and 26-36 are directed to an abstract idea (Step 2A, Prong 2: NO).
Step 2B
Claims found to be directed to a judicial exception are then further evaluated to determine if the claims recite an inventive concept that provides significantly more than the judicial exception itself (Step 2B).
The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception because the claims recite additional elements that equate to mere instructions to apply the recited exception in a generic computing environment or well-understood, and conventional activity.
Regarding limitations directed to obtaining a network input, as set forth in MPEP section 2106.05(g), the courts have decided that limitations that merely add an insignificant extra-solution activity, do not amount to an inventive concept, particularly when the activities are well-understood and conventional. Parker v. Flook, 437 U.S. 584, 588-89, 198 USPQ 193, 196 (1978). As set forth in MPEP section 2106.05(d), the courts have recognized that limitations directed to data gathering that are claimed as insignificant extra-solution activity are routine, well understood and conventional. Mayo Collaborative servs. V. Prometheus Labs., Inc., 566 U.S. at 79, 101 USPQ2d at 1968.
The limitations of claims 1 and 26-36, pertaining to the computers and computer readable storage medium used to execute the method, are directed to performing judicial exceptions with a generic computing system on a generic computer. These limitations are not sufficient to amount to significantly more than the judicial exception because, as set forth in the MPEP section 2106.05(d)(II)), using a generic computing environment or generic computer to perform the judicial exception, has been deemed well-understood, routine and conventional activity including receiving or transmitting data over a network (Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362), performing repetitive calculations (Bancorp Services v. Sun Life, 687 F.3d 1266, 1278, 103 USPQ2d 1425, 1433 (Fed. Cir. 2012)), and storing and retrieving information in memory (Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334, 115 USPQ2d 1681, 1701 (Fed. Cir. 2015)). 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 above, the limitations in claims 1, 26 and 35 directed to ‘determining a final predicted structure for the protein from the structure parameters for the last iteration in the sequence’ are recited at a high level with no details provided in the specification or the claims of how the final predicted structure is determined, thus they amount to generic instructions to apply the judicial exception.
Additionally, the limitations of claims 1, 3-5, 26, 28-30 and 35, pertaining to using a neural network to predict protein structure, and limitations of claims 1, 26 and 35 directed to determining a final predicted structure from parameters, were well-understood, routine and conventional at the time of the effective filing date as evidenced by Kuhlman et al. (Nature Reviews Molecular Cell Biol., 2019, Vol. 20, p. 681-698; IDS 08 November 2023).
Kuhlman et al. discloses that methods and systems to determine a final predicted protein structure from template-free structure predictions was available in multiple software packages including PSIPRED and Rosetta, in 2019. Specifically, these software packages included: 1) starting with local sequence alignments, 2) using local sequence alignment data to predict local structures, 3) use local sequence data to predict residue contacts and residue-residue distance constraints, and 4) use the predictions to generate a final predicted 3D structure (Fig. 2).
Kuhlman et al. further discloses that many neural network approaches to recognize patterns in the sequence and structures of proteins were used in the process of predicting protein structures by 2019 including DeepContact, RaptorX-Contact, DeepCov and Triplet Res (p. 686, col. 2, para. 3 – p. 688, col. 1, para.1)
The additional elements do not comprise an inventive concept when considered individually or as an ordered combination that transforms the claimed judicial exception into a patent-eligible application of the judicial exception. Therefore, the claims do not amount to significantly more than the judicial exception itself (Step 2B: No). As such, claims 1-9 and 26-36 are not patent eligible.
Claim Rejections - 35 USC § 102
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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
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.
8. Claims 1-3, 5-6, 26-28, 30-31 and 35-36 are rejected under 35 U.S.C. 102(a)(1) and 102(a)(2) as being unpatentable over Senior et al. (WO2020058174A1; publication date 26 March 2020; 11/08/2023 IDS document). The italicized text corresponds to the instant claim limitations.
Regarding claims 1, 26 and 35, Senior et al. discloses a system implemented as computer programs on one or more computers in one or more locations that perform protein tertiary structure prediction and protein domain segmentation. Senior et al. further teaches a method performed by one or more data processing apparatus for determining a final predicted structure of a given protein. Senior et al. further discloses that the structure prediction systems disclosed can be used to predict the structure of protein complexes in which a group of multiple proteins that may be non-covalently linked fold together into a global structure (e.g. protein dimers comprising more than one chain). (para. 0005-0006; para. 0129; a method performed by one or more computers for predicting a structure of a protein comprising one or more chains, wherein each chain comprises a sequence of amino acids).
Regarding claims 1, 26 and 35, Senior et al. discloses a method wherein the process of predicting the protein structure is iterative, and at the first iteration, generating a predicted structure of the given protein may comprise obtaining initial values of the plurality of structure parameters defining the predicted structure (network input) and updating the initial values of the parameters. Senior et al. further discloses that the network input can be components of the sequence of amino acid residues, or data from multiple sequence alignments (para. 0006; para. 0030; Fig. 1 and 9; the method comprising: at a first iteration of a sequence of iterations that comprises the first iteration followed by one or more subsequent iterations: obtaining a network input for the first iteration that characterizes the protein).
Pertaining to claims 1, 26 and 35, Senior et al. discloses processing the network input for the first iteration using a neural network configured to process the current values of the structure parameters or a representation of the sequence of amino acids of the given protein or both to generate a predicted structure of the given protein defined by values of a plurality of structure parameters (para. 0006; processing the network input for the first iteration using a protein structure prediction neural network to generate structure parameters for the first iteration that define an initial predicted structure for the protein).
Regarding claims 1, 26 and 35, Senior et al. discloses that at each update iteration, the neural network processes the current values of the structure parameters, a representation of the sequence of amino acids of the given protein, or both (para. 0006; at each subsequent iteration in the sequence of iterations: obtaining a network input for the subsequent iteration that characterizes the protein).
With respect to claims 1, 26 and 35, the claim is interpreted to mean that features are generated for the next iteration by using structure parameters and/or intermediate model outputs generated at a preceding iteration. Senior et al. disclose that for each iteration, the neural network processes the current values of the structure parameters (which, in a round subsequent to the first, is the updated initial values of the structure parameters). Senior et al. further discloses that the predicted structure of the given protein is defined by values of a plurality of structure parameters (features). Senior et al. further discloses that in some cases the distance prediction neural network (which characterizes the distances between pairs of amino acids) can be configured to generate outputs characterizing torsion angles between amino acids in each of the first position sand second positions in the amino acid sequence corresponding to a distance map crop. In another example, outputs characterizing estimated secondary structures like beta sheets or alpha helices are generated, which may improve the predictions from the distance prediction neural network (para. 0006; para. 0252-0253; generating, from (i) the structure parameters generated at a preceding iteration that precedes the subsequent iteration in the sequence, (ii) one or intermediate outputs generated by the protein structure prediction neural network while generating the structure parameters at the last iteration, or (iii) both, features for the subsequent iteration).
Regarding claims 1, 26 and 35, Senior et al. discloses that for each update iteration, the neural network processes the current values of the structure parameters, a representation of the sequence of amino acids of the given protein, or both, to generate a predicted structure of the given protein (para. 0006; processing the features and the network input for the subsequent iteration using the protein structure prediction neural network to generate structure parameters for the subsequent iteration that define another predicted structure for the protein).
Regarding claims 1, 26 and 35, Senior et al. discloses that the method may further comprise determining the predicted structure of the given protein to be defined by the current values of the plurality of structure parameters after a final update iteration of the plurality of update iterations (para. 0008; determining a final predicted structure for the protein from the structure parameters for the last iteration in the sequence).
Pertaining to claim 26, Senior et al. discloses that their protein structure prediction method can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus (para. 0391; a system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform [the claimed method]).
Regarding claim 35, Senior et al. discloses that their protein structure prediction method can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus (para. 0391; one or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform [the claimed method]).
Pertaining to claims 2, 27 and 36, according to the specification, network input is interpreted to include multiple sequence alignments of the protein of interest with other proteins. Senior et al. disclose that the network input can include a representation of the amino acid sequence, alignment features corresponding to the amino acid sequence, or both. Senior et al. further discloses that the alignment features are derived from a multiple sequence alignment (MSA) of amino acid sequence. Because MSA sequence data are from databases and not from the iterative model output, these data are inherently the same with each iteration of the modeling. Senior et al. describes other examples of where network inputs are fixed: 1) the geometry neural network is trained on a static set of training data which remains fixed throughout training (because variation is provided b randomly perturbing the predicted structures of the proteins) and 2) the value neural network is trained on a static set of training data which remains fixed throughout training (and variation is provided by computing predictions for a large number of sequences for different proteins (para. 0187; 0213-0215; 0296; the method of claim 1/system of claim 26, wherein the network input for each iteration in the sequence of iterations is the same network input).
Regarding claims 3 and 28, Senior et al. discloses that the system uses one or more of (i) a representation of the amino acid sequence, (ii) the structure parameters defining the predicted structure, (iii) the alignment features, and (iv) the structure distance map, to generate respective inputs for the geometry neural network, the value neural network, and the distance prediction system. Senior et al. further discloses that in some cases, the representation of the amino acid sequence and the structure parameters defining the predicted structure are represented as one-dimensional (1D) features (i.e., features that are represented as linear sequences), while some of the alignment features and the structure distance map are represented as two-dimensional features (2D) (i.e., features that are represented as matrices). To generate the input, the system may combine the 1D and 2D features by broadcasting and concatenating the 1D features along the rows and the columns of the matrix representation of the 2D features (para. 0205; the method of claim 1/system of claim 26, wherein processing the features and the network input for the subsequent iteration using the protein structure prediction neural network to generate structure parameters for the subsequent iteration that define a predicted structure for the protein comprises: generating a combined input from the features and the network input for the subsequent iteration, processing the combined input using the protein structure prediction neural network to generate the structure parameters for the subsequent iteration).
Regarding claims 5 and 30, single embedding is interpreted to mean transforming vector representations of data by multiplying them by an embedding matrix. Senior et al. describes combining features (sequence features and multiple sequence alignment (MSA) features of an amino acid sequence) and grouping them together into a single 2-D matrix, which is processed by the generative neural network. Senior et al. further discloses that he embedding neural network may include one or more convolutional residual blocks followed by a mean pooling layer that outputs the conditioning vector. The conditioning vector is then passed into a l-D convolutional long short-term memory (LSTM) convolutional decoder subnetwork (para. 0380; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises an initial multiple sequence alignment (MSA) representation that represents a respective MSA corresponding to each chain in the protein at each iteration, the structure prediction neural network is configured to generate one or more sets of single embeddings that each include a respective single embedding for each amino acid the protein while generating the structure parameters for the iteration generating the features comprises generating, from one of the sets of single embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of single embeddings generating the combined input comprises combining the transformed set of single embeddings and the initial MSA representation for the subsequent iteration).
Regarding claims 6 and 31, this is interpreted to mean that MSA representations contain the target sequence in the first row with homologous sequences aligned below it. In a combining step, adding the transformed single embeddings to the first row of the MSA representation enforces feature alignment of the structural properties derived from the single embeddings into the position corresponding to the target sequence in the MSA. Senior et al. disclose that the system uses one or more of: (i) a representation of the amino acid sequence, (ii) the structure parameters defining the predicted structure, (iii) the alignment features, and (iv) the structure distance map, to generate respective inputs for the geometry neural network, the value neural network, and the distance prediction system. Senior et al. further discloses that in some cases, the representation of the amino acid sequence and the structure parameters defining the predicted structure are represented as one-dimensional (1D) features (i.e., features that are represented as linear sequences), while some of the alignment features and the structure distance map are represented as two-dimensional features (2D) (i.e., features that are represented as matrices). To generate the input, the system may combine the 1D and 2D features by broadcasting and concatenating the 1D features along the rows and the columns of the matrix representation of the 2D features (para. 0205; the method of claim 5/system of claim 30, wherein combining the transformed set of single embeddings and the initial MSA representation for the subsequent iteration comprises adding the transformed set of single embeddings to a first row of the initial MSA representation).
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
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.
9. Claims 4, 9, 29 and 34 are rejected under 35 U.S.C. 103 as being unpatentable over Senior et al. (WO2020058174A1; 11/08/2023 IDS document), as applied to claims 1-3, 5-6, 26-28, 30-31 and 35-36 above, in view of Ju et al. (BioRxiv 7 October 2020, p. 1-9), Jorgensen et al. (arXiv:1806.03146; 2018, p. 1-10) and Adhikari et al. (Bioinformatics, Vol. 34, 2018, p. 1466-1472). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-3, 5-6, 26-28 , 30-31 and 35-36 have been taught by Senior et al. above.
Pertaining to claims 4, 9, 29 and 34, Senior et al. is silent to: the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein, at each iteration, the structure prediction neural network is configured to repeatedly update the initial pair embeddings while generating the structure parameters for the iteration, generating the features comprises generating, from updated pair embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of pair embeddings; and generating the combined input comprises combining the transformed set of pair embeddings and the initial pair embeddings for the subsequent iteration (claims 4 and 29); and the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein that is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences; and generating the combined input comprises modifying the initial pair embeddings for the subsequent iteration by adding the protein and the structure prediction defined by the embeddings at the preceding iteration to the set of one or more template sequences and the corresponding known structures (claims 9 and 34). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Ju et al. and Jorgensen et al. and Adhikari et al.
Pertaining to claims 4 and 29, pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1 (or system of claim 26) wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein at each iteration).
Pertaining to claims 4 and 29, Jorgensen et al. teaches a graph neural network method for predicting properties of molecules that has an edge update network that iteratively refines an initial bond embedding alongside node representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom. Therefore Jorgensen et al. teaches a method of updating edges (relationships between nodes) (p. 2, para. 3; abstract; Fig. 1; the structure prediction neural network is configured to repeatedly update the initial pair embeddings while generating the structure parameters for the iteration generating the features comprises generating, from updated pair embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of pair embeddings generating the combined input comprises combining the transformed set of pair embeddings and the initial pair embeddings for the subsequent iteration).
With respect to claims 9 and 34, a template sequence is defined in the specification as an MSA sequence for an amino acid chain in the protein where the folded structure of the template sequence is known (in one embodiment, para. 0121). A pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; p. 2, col. 2, para. 1; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
With respect to claims 9 and 34, Jorgensen et al. teaches a graph neural network method for predicting properties of molecules that has an edge update network that iteratively refines an initial bond embedding alongside node representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom. Therefore Jorgensen et al. teaches a method of updating edges (i.e. relationships between nodes) (p. 2, para. 3; abstract; Fig. 1; generating the combined input comprises modifying the initial pair embeddings for the subsequent iteration by adding the protein and the structure prediction defined by the embeddings at the preceding iteration to the set of one or more template sequences and the corresponding known structures).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Ju et al discloses that learning residue co-evolution directly from MSA will contribute to more accurate predictions of protein tertiary structures (p. 7, col. 1, para. 5). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Senior et al. to include the learned pairwise representations taught by Ju et al. and to update those representations at successive neural network layers according to Jorgensen et al. because doing so would permit information associated with relationships between residues to be progressively refined during iterative prediction, thereby improving the representations available for predicting structure and thus improving accuracy of predictions. Furthermore, one of ordinary skill in the art would predict that the learned pairwise representations taught by Ju et al. and the update method taught by Jorgensen et al. could be readily added to the method and system of Senior et al. with a reasonable expectation of success because the methods and systems of Senior et al. and Ju et al. both pertain to predicting protein structures by neural network analysis and Senior et al. and Jorgensen et al. both pertain to predicting molecular structures by neural network analysis. The invention is therefore prima facie obvious.
Regarding claims 9 and 34, Senior et al., Ju et al. and Jorgensen et al. are silent to: the initial pair embedding is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences. However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Adhikari.
With respect to claims 9 and 34, Adhikari et al. discloses DNCON2, a deep learning-based tool to predict protein residue-residue contacts and inter-residue distance distributions using protein amino acid sequence, multiple sequence alignments (MSAs). Adhirkari et al. discloses doing this in two levels: 1) 5 separate CNNs evaluate an array of features including protein length and predicted secondary structure and coevolutionary information to predict contact maps and 2) a final CNN takes these five predictions as inputs and compiles them in to a single, highly refined contact map. that the prediction system can generate structure fragments trained on a database of known protein structures (p. 1467, col. 1, para. 4 – p. 1468, col. 2, para. 1; Fig. 1; the initial pair embedding is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Adhikari et al. discloses that their method of predicting the contact map of proteins from existing sequence and structure data can predict all the contacts in a protein at once from the entire input information of a protein, which is more effective and easier to train and use than local fixed window-based approaches such as deep belief networks (p. 1471, col. 2, para. 2). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Senior et al. Ju et al. and Jorgensen et al. to include the method of predicting the contact map of proteins taught by Adhikari et al. to generate the initial input data more easily and effectively. Furthermore, one of ordinary skill in the art would predict that the method of predicting contact maps taught by Adhikari et al. could be readily added to the method and system of Senior et al., Ju et al. and Jorgensen et al. with a reasonable expectation of success because both methods pertain to computational prediction of structures of proteins using existing data and furthermore, the contact map disclosed by Adhikari et al. functions as a 2D pair embedding (or a 2D pairwise representation) that is in the same format that is input in to the method of Senior et al., Ju et al. and Jorgensen et al. The invention is therefore prima facie obvious.
10. Claims 7-8 and 32-33 are rejected under 35 U.S.C. 103 as being unpatentable over Senior et al. (WO2020058174A1; 11/08/2023 IDS document), as applied to claims 1-3, 5-6, 26-28, 30-31 and 35-36 above, in view of Ju et al. (BioRxiv 7 October 2020, p. 1-9) and Jorgensen et al. (arXiv:1806.03146; 2018, p. 1-10) and further in view of Hayder et al. (2017, 2017 IEEE conference on computer vision and pattern recognition, p. 587 – 595). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-3, 5-6, 26-28, 30-31 and 35-36 have been taught by Senior et al. above.
Pertaining to claims 7 and 32, Senior et al. discloses that in some implementations, the structure parameters are a sequence of three-dimensional (3D) numerical coordinates (e.g., represented as 3D vectors) where each coordinate represents the position (in some given frame of reference) of a corresponding atom in an amino acid from the amino acid sequence (para. 0157; at each iteration, the structure parameters specify, for each amino acid, a predicted 3-D spatial location of a specified atom in the amino acid in the structure of the protein).
Pertaining to claims 7 and 32, Senior et al. discloses that the system can include a distance engine that is configured to process the predicted structure to generate a structure distance map. The structure distance map characterizes a respective distance (e.g., measured in angstroms) between each pair of amino acids in the amino acid sequence when it is folded in accordance with the predicted structure (para. 0205; generating the features comprises: generating, from the predicted 3-D spatial locations for the amino acids specified by the structure parameters at the preceding iteration, a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein).
Regarding claims 8 and 33, Senior et al. teaches that in some implementations the one or more scoring neural networks comprise a distance prediction neural network configured to process the representation of the sequence of amino acids to generate a distance map for the given protein. In implementations the distance map defines, for each of a plurality of pairs of amino acids in the sequence, a respective probability distribution over possible distance ranges between the pair of amino acids (para. 0012; at each iteration, the structure parameters specify a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein).
Pertaining to claims 7, 8, 32 and 33, Senior is silent to: the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein and generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration (claims 7 and 32) and the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein and generating, a transformed distance map from the distance map specified by the structure parameters at the preceding iteration and generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration (claims 8 and 33). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Ju et al. and Jorgensen et al.
Pertaining to claims 7 and 32, pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
Pertaining to claims 7 and 32, combining a transformed distance map and an initial pair embedding involves concatenating or stacking or math functions such as fusion or aggregation. Jorgensen et al. teaches using tensor concatenation for their edge update networks. During the interaction steps the edge update function takes the concatenation of the current edge representation and the hidden states of the sending and receiving vertices as the direct input to a two-layered feed-forward neural network (Equations 6 and 7 (semicolon represents concatenation function); p. 3, para. 2-5 (section 2.1); generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration).
Regarding claims 8 and 33, a pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
Regarding claims 8 and 33, Jorgensen et al. teaches a machine learning method for predicting properties of molecules. It is a neural message passing model with an edge update network which allows the information exchanged between atoms to depend on the hidden state of the receiving atom; therefore, generating a network that repeatedly updates learned edge representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom (p. 2, para. 3; abstract; Fig. 1; generating, a transformed distance map from the distance map specified by the structure parameters at the preceding iteration).
Regarding claims 8 and 33, combining a transformed distance map and an initial pair embedding involves concatenating or stacking or math functions such as fusion or aggregation. Jorgensen et al. teaches using tensor concatenation for their edge update networks. During the interaction steps the edge update function takes the concatenation of the current edge representation and the hidden states of the sending and receiving vertices as the direct input to a two-layered feed-forward neural network (Equations 6 and 7 (semicolon represents concatenation function); p. 3, para. 2-5 (section 2.1); generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration.
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Ju et al discloses that learning residue co-evolution directly from MSA will contribute to more accurate predictions of protein tertiary structures (p. 7, col. 1, para. 5). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Senior et al. to include the learned pairwise representations taught by Ju et al. and to update those representations at successive neural network layers according to Jorgensen et al. because doing so would permit information associated with relationships between residues to be progressively refined during iterative prediction, thereby improving the representations available for predicting structure and thus improving accuracy of predictions. Furthermore, one of ordinary skill in the art would predict that the learned pairwise representations taught by Ju et al. and the update method taught by Jorgensen et al. could be readily added to the method and system of Senior et al. with a reasonable expectation of success because the methods and systems of Senior et al. and Ju et al. both pertain to predicting protein structures by neural network analysis and Senior et al. and Jorgensen et al. both pertain to predicting molecular structures by neural network analysis. The invention is therefore prima facie obvious.
Pertaining to claims 7-8 and 32-33, Senior et al., Ju et al. and Jorgensen et al. are silent to: generating, from the distance map, a transformed distance map that has a same dimensionality as the initial pair embeddings (claims 7 and 32); generating the features comprises: generating a transformed distance map that has a same dimensionality as the initial pair embeddings (claims 8 and 33). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Hayden et al.
Pertaining to claims 7 and 32, to combine a distance map (shape: N x N) with a pair embedding (shape: N x N x D, where D is the embedding dimension), one must either expand/transform the distance map to match the embedding’s depth, or flatten the embedding to match the map’s 2D structure. Hayden et al. teaches Boundary-Aware Instance Segmentation (BAIS) framework, which includes Boundary-Aware Mask Representation, a method of converting traditional binary mask output of a shape N x N into a multi-valued map of shape N x N x D, which involves converting each scalar distance value into a distinct “bin” index, followed by one-hot encoding to transform the bin index into a vector of length B (where B is the total number of bins), followed by embedding in N x N x D by multiplying the B-dimensional one-hot vector by B x D embedding matrix to project each distance onto a D-dimensional continuous vector space. (p. 589, col. 1, para. 3 – p. 590, col. 1, para. 2; generating, from the distance map, a transformed distance map that has a same dimensionality as the initial pair embeddings).
Regarding claims 8 and 33, to combine a distance map (shape: N x N) with a pair embedding (shape: N x N x D, where D is the embedding dimension), one must either expand/transform the distance map to match the embedding’s depth, or flatten the embedding to match the map’s 2D structure. Hayden et al. teaches Boundary-Aware Instance Segmentation (BAIS) framework, which includes Boundary-Aware Mask Representation, a method of converting traditional binary mask output of a shape N x N into a multi-valued map of shape N x N x D, which involves converting each scalar distance value into a distinct “bin” index, followed by one-hot encoding to transform the bin index into a vector of length B (where B is the total number of bins), followed by embedding in N x N x D by multiplying the B-dimensional one-hot vector by B x D embedding matrix to project each distance onto a D-dimensional continuous vector space. (p. 589, col. 1, para. 3 – p. 590, col. 1, para. 2; generating the features comprises: generating a transformed distance map that has a same dimensionality as the initial pair embeddings).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Hayder et al. discloses that their instance segmentation approach to jointly detect, segment and classify every individual object in an image, outperforms other segmentation approaches (as measured by the mean average precision (mAP) of the models) and is robust to noisy object proposals (p. 588, col. 1, para. 2-3; Tables 1-2). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Senior et al., Ju et al. and Jorgensen et al. to include the distance dimensionality transformation method taught by Hayder et al. because it yields robust models that outperform models generated using other segmentation approaches. Furthermore, one of ordinary skill in the art would predict that the distance dimensionality transformation method taught by Hayder et al. could be readily added to the method of Senior et al., Ju et al. and Jorgensen et al. with a reasonable expectation of success because the methods and systems of Senior et al. and Ju et al. both pertain to predicting protein structures by neural network analysis and Senior et al. and Jorgensen et al. both pertain to distance transformations for neural network-based predictions in image analysis. The invention is therefore prima facie obvious.
11. Claims 1-3, 26-28 and 35-36 are rejected under 35 U.S.C. 103 as being unpatentable over Greener et al. (Nature Communications, 2019, Vo. 10, p. 1-13), in view of Ju et al. (BioRxiv 7 October 2020, p. 1-9).The italicized text corresponds to the instant claim limitations.
Regarding claims 1, 26 and 35, Greener et al. discloses DMPfold, which uses deep learning to predict protein structure and runs on a standard desktop computer. Greener et al. further discloses DMPfold predicts proteins comprised of chains of amino acids (p. 2, col. 1, para. 5 – col. 2, para, 2; p. 10, col. 1, para. 3 – col. 2, para. 2; a method performed by one or more computers for predicting a structure of a protein comprising one or more chains, wherein each chain comprises a sequence of amino acids).
Regarding claims 1, 26 and 35, Greener et al. discloses that the DMPfold pipeline is an iterative method with 3 or more iterations. Greener et al. further discloses that initially, structural predictions are made using DMP inputs, which are covariance data contact predictions and other data (Fig. 8; p. 10, col. 1, para. 3-5; p. 11, col. 1, para. 2 – col. 2, para. 2; the method comprising: at a first iteration of a sequence of iterations that comprises the first iteration followed by one or more subsequent iterations: obtaining a network input for the first iteration that characterizes the protein).
Pertaining to claims 1, 26 and 35, Greener et al. discloses that the DMPfold pipeline involves: 1) Initially predicting inter-residue Cβ distances, H-bonds and torsion angles from DMP inputs, 2) using these predictions to generate models with CNS software, and 3) selecting and using a single model as additional input to refine the distances and H-bonds. After 3 iterations a final set of models is returned. Greener et al. further discloses that neural networks are used in the steps to predict inter-residue distance probability distributions (between Cβ atoms or Cα atoms for glycine), main chain hydrogen bond (H-bond) donor/acceptor pairs and torsion angles (Fig. 8; p. 10, col. 1, para. 3; processing the network input for the first iteration to generate structure parameters for the first iteration that define an initial predicted structure for the protein).
Regarding claims 1, 26 and 35, Greener et al. discloses that at the first iteration, structures are predicted using DMP inputs, but for subsequent iterations, the same distance and H-bond procedures are used as with the first iteration but also a ‘best structure’ prediction is used as an additional input to refine the distances and H-bonds; in this way, the combined contact prediction and structure generation
procedure can evolve a better prediction at each iteration. (Fig. 8; p. 11, col. 1, para. 2 – col. 2, para. 3; at each subsequent iteration in the sequence of iterations: obtaining a network input for the subsequent iteration that characterizes the protein).
With respect to claims 1, 26 and 35, Greener et al. discloses that additional input features added to modeling at each iteration include a Cβ-Cβ distance matrix calculated from the seed structure; this allows new distances and H-bonds to be predicted using prior information of likely Cβ-Cβ distances from the previous iteration of 3-D modelling. Greener et al. further discloses that, in this way, the combined contact prediction and structure generation procedure can evolve a better prediction at each iteration. Greener et al. further discloses, deep neural networks are used in the first and subsequent iteration steps to predict inter-residue distance probability distributions (between Cβ atoms or Cα atoms for glycine), main chain hydrogen bond (H-bond) donor/acceptor pairs, torsion angles and distance (p. 11, col. 1, para. 2 – col. 2, para. 3; Fig. 8; p. 10, col. 1, para. 3 – col. 2, para. 2; generating, from (i) the structure parameters generated at a preceding iteration that precedes the subsequent iteration in the sequence, (ii) one or intermediate outputs while generating the structure parameters at the last iteration, or (iii) both, features for the subsequent iteration; processing the features and the network input for the subsequent iteration using the protein structure prediction neural network to generate structure parameters for the subsequent iteration that define another predicted structure for the protein).
Regarding claims 1, 26 and 35, Greener et al. discloses that a final structure is generated after model convergence, which usually takes 3-5 iterations (Fig. 8, p. 11, col. 1, para. 2 – col. 2, para. 3; determining a final predicted structure for the protein from the structure parameters for the last iteration in the sequence).
Pertaining to claim 26, Greener et al. discloses that their tool runs on a standard desktop computer and that the source code, documentation, trained neural network models and Pfam 3-D models are available from Github repository (p. 2, col. 2, para. 2; a system comprising: one or more computers; and one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform [the claimed method]).
Regarding claim 35, Greener et al. discloses that their tool runs on a standard desktop computer and that the source code, documentation, trained neural network models and Pfam 3-D models are available from Github repository (p. 2, col. 2, para. 2; one or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform [the claimed method]).
Regarding claims 1, 26 and 35, Greener et al. is silent to using a protein structure prediction neural network; and generated by the protein structure prediction neural network. However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Ju et al.
Regarding claims 1, 26 and 35, Ju et al. teaches an approach for predicting protein structures called proFOLD. At the core of this approach is a deep neural network framework called CopulaNet specially designed to learn inter-residue distances directly from MSA using three key modules, namely, MSA encoder, coevolution aggregator, and distance estimator (p. 2, col. 1, para. 4 – p. 3, col. 2, para. 1; Fig. 2; using a protein structure prediction neural network; and generated by the protein structure prediction neural network).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Ju et al. discloses that learning residue co-evolution directly from MSA will contribute to more accurate predictions of protein tertiary structures (p. 7, col. 1, para. 5). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Greener et al. to include the learned pairwise representations taught by Ju et al. and to update those representations at successive neural network layers according to Jorgensen et al. because doing so would permit information associated with relationships between residues to be progressively refined during iterative prediction, thereby improving the representations available for predicting structure and thus improving accuracy of predictions. Furthermore, one of ordinary skill in the art would predict that the learned pairwise representations taught by Ju et al. could be readily added to the method and system of Greener et al. with a reasonable expectation of success because the methods and systems of Greener et al. and Ju et al. both pertain to predicting protein structures using neural network analysis. The invention is therefore prima facie obvious.
Pertaining to claims 2, 27 and 36, Greener et al. discloses that DMP inputs comprising covariance data and contact predictions etc. are the same at each iteration of the method. The inputs are combined from different sources of covariation data from alignments using a deep residual neural network to predict contacts, along with the raw residue-residue covariance matrix as employed in the DeepCov contact prediction method. (Fig. 8; p. 10, col. 1, para. 3 the method of claim 1/system of claim 26, wherein the network input for each iteration in the sequence of iterations is the same network input).
Regarding claims 3 and 28, Greener et al. discloses that at each iteration after the first, the same input data is used as well as additional input features from the predicted structure (including a Cβ-Cβ distance matrix) and that these data are recombined to generate new distances and H-bonds to be predicted using prior information of likely Cβ-Cβ distances from the previous iteration of 3-D modelling. (p. 11, col. 1, para. 2 – col. 2, para. 3; Fig. 8; the method of claim 1/system of claim 26, wherein processing the features and the network input for the subsequent iteration using the protein structure prediction neural network to generate structure parameters for the subsequent iteration that define a predicted structure for the protein comprises: generating a combined input from the features and the network input for the subsequent iteration, processing the combined input using the protein structure prediction neural network to generate the structure parameters for the subsequent iteration).
12. Claims 4, 9, 29 and 34 are rejected under 35 U.S.C. 103 as being unpatentable over Greener et al. (Nature Communications, 2019, Vo. 10, p. 1-13), in view of Ju et al. (BioRxiv 7 October 2020, p. 1-9), as applied to claims 1-3, 26-28 and 35-36 above, and further in view of Jorgensen et al. (arXiv:1806.03146; 2018, p. 1-10) and Adhikari et al. (Bioinformatics, Vol. 34, 2018, p. 1466-1472). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-3, 26-28 and 35-36 have been taught by Greener et al. and Ju et al. above.
Pertaining to claims 4 and 29, pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1 (or system of claim 26) wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
With respect to claims 9 and 34, a template sequence is defined in the specification as an MSA sequence for an amino acid chain in the protein where the folded structure of the template sequence is known (in one embodiment, para. 0121). A pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; p. 2, col. 2, para. 1; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
Pertaining to claims 4, 9, 29 and 34, Greener et al. and Ju et al. are silent to: at each iteration, the structure prediction neural network is configured to repeatedly update the initial pair embeddings while generating the structure parameters for the iteration, generating the features comprises generating, from updated pair embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of pair embeddings; and generating the combined input comprises combining the transformed set of pair embeddings and the initial pair embeddings for the subsequent iteration (claims 4 and 29); that is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences; and generating the combined input comprises modifying the initial pair embeddings for the subsequent iteration by adding the protein and the structure prediction defined by the embeddings at the preceding iteration to the set of one or more template sequences and the corresponding known structures (claims 9 and 34). However, these limitations were known in the art at the time of the effective filing date of the invention as taught Jorgensen et al. and Adhikari et al.
Pertaining to claims 4 and 29, Jorgensen et al. teaches a graph neural network method for predicting properties of molecules that has an edge update network that iteratively refines an initial bond embedding alongside node representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom. Therefore Jorgensen et al. teaches a method of updating edges (relationships between nodes) (p. 2, para. 3; abstract; Fig. 1; the structure prediction neural network is configured to repeatedly update the initial pair embeddings while generating the structure parameters for the iteration generating the features comprises generating, from updated pair embeddings generated while generating the structure parameters at the preceding iteration, a transformed set of pair embeddings generating the combined input comprises combining the transformed set of pair embeddings and the initial pair embeddings for the subsequent iteration).
With respect to claims 9 and 34, Jorgensen et al. teaches a graph neural network method for predicting properties of molecules that has an edge update network that iteratively refines an initial bond embedding alongside node representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom. Therefore Jorgensen et al. teaches a method of updating edges (relationships between nodes) (p. 2, para. 3; abstract; Fig. 1; generating the combined input comprises modifying the initial pair embeddings for the subsequent iteration by adding the protein and the structure prediction defined by the embeddings at the preceding iteration to the set of one or more template sequences and the corresponding known structures).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Jorgensen et al. teaches their model with an edge update network has improved accuracy on formation energy prediction of molecules and materials across three benchmark datasets (p. 8, para. 2 – p. 9, para. 1). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Greener et al. and Ju et al. to update the learned pairwise representations taught by Ju et al. at successive neural network layers according to Jorgensen et al. because doing so would improve accuracy of predictions. Furthermore, one of ordinary skill in the art would predict the update method taught by Jorgensen et al. could be readily added to the method and system of Greener et al. and Ju et al. with a reasonable expectation of success because the methods and systems of Greener et al. and Ju et al. and Jorgensen et al. pertain to predicting molecular structures by neural network analysis and the method of Jorgensen et al. is applicable to applications where pairwise distances are used as edge features (Jorgensen et al. p. 2, para. 2). The invention is therefore prima facie obvious.
Regarding claims 9 and 34, Greener et al., Ju et al. and Jorgensen et al. are silent to: the initial pair embedding is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences. However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Adhikari et al.
With respect to claims 9 and 34, Adhikari et al. discloses DNCON2, a deep learning-based tool to predict protein residue-residue contacts and inter-residue distance distributions using protein amino acid sequence, multiple sequence alignments (MSAs). Adhikari et al. discloses doing this in two levels: 1) 5 separate CNNs evaluate an array of features including protein length and predicted secondary structure and coevolutionary information to predict contact maps and 2) a final CNN takes these five predictions as inputs and compiles them in to a single, highly refined contact map. that the prediction system can generate structure fragments trained on a database of known protein structures (p. 1467, col. 1, para. 4 – p. 1468, col. 2, para. 1; Fig. 1; the initial pair embedding is generated using a set of one or more template sequences and corresponding known structures for each of the template sequences).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Adhikari et al. discloses that their method of predicting the contact map of proteins from existing sequence and structure data can predict all the contacts in a protein at once from the entire input information of a protein, which is more effective and easier to train and use than local fixed window-based approaches such as deep belief networks (p. 1471, col. 2, para. 2). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Greener et al. Ju et al. and Jorgensen et al. to include the method of predicting the contact map of proteins taught by Adhikari et al. to generate the initial input data more easily and effectively. Furthermore, one of ordinary skill in the art would predict that the method of predicting contact maps taught by Adhikari et al. could be readily added to the method and system of Greener et al., Ju et al. and Jorgensen et al. with a reasonable expectation of success because both methods pertain to computational prediction of structures of proteins using existing data and furthermore, the contact map disclosed by Adhikari et al. functions as a 2D pair embedding (or a 2D pairwise representation) that is in the same format that is input in to the method of Senior et al., Ju et al. and Jorgensen et al. The invention is therefore prima facie obvious.
13. Claims 7-8 and 32-33 are rejected under 35 U.S.C. 103 as being unpatentable over Greener et al. (Nature Communications, 2019, Vo. 10, p. 1-13), in view of Ju et al. (BioRxiv 7 October 2020, p. 1-9) as applied to claims 1-3, 26-28 and 35-36 above, and further in view of Jorgensen et al. (arXiv:1806.03146; 2018, p. 1-10) and Hayder et al. (2017, 2017 IEEE conference on computer vision and pattern recognition, p. 587 – 595). The italicized text corresponds to the instant claim limitations.
The limitations of claims 1-3, 26-28 and 35-36 have been taught by Greener et al. and Ju et al. above.
Pertaining to claims 7 and 32, Greener et al. discloses that the DMPfold model predicts inter-atomic distance bounds, hydrogen bonds and torsion angles and errors that are used as constraints to predict 3-D protein structures. Greener et al. further discloses that this is performed at each iteration of the method (Fig. 1; Fig. 9; p. 2, col. 1, para. 5 – col. 2, para. 2; Fig. 8; at each iteration, the structure parameters specify, for each amino acid, a predicted 3-D spatial location of a specified atom in the amino acid in the structure of the protein).
Pertaining to claims 7 and 32, Greener et al. discloses generating an initial distance map and distance maps at each iteration of the method (Fig. 6; generating the features comprises: generating, from the predicted 3-D spatial locations for the amino acids specified by the structure parameters at the preceding iteration, a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein).
Pertaining to claims 7 and 32, pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
Regarding claims 8 and 33, Greener et al. discloses that the DMPfold model predicts inter-atomic distance bounds, hydrogen bonds and torsion angles and errors that are used as constraints to predict 3-D protein structures. Greener et al. further discloses that this is performed at each iteration of the method. Greener et al. further discloses generating an initial distance map and distance maps at each iteration of the method (Fig. 1; Fig. 9; p. 2, col. 1, para. 5 – col. 2, para. 2; Fig. 8; Fig. 6; at each iteration, the structure parameters specify a distance map that characterizes, for each pair of amino acids in the protein, a respective estimated distance between the pair of amino acids in the structure of the protein).
Regarding claims 8 and 33, a pair embedding is interpreted to mean a learned vector representation for each residue pair. Ju et al. teaches a generating a learned vector representation for each residue pair called a ‘co-evolution feature”. Ju et al. teaches doing this using an approach called ProFOLD for inter-residue distance estimation through learning the conditional joint-residue distributions directly from multiple sequence alignments (MSAs). Ju et al. further discloses that their approach uses a deep neural network framework CopulaNet, which consists of three key elements: 1) MSA encoder: a 1D convolution neural network processes the MSA to generate residue representations. 2) Co-evolution aggregator: To model co-evolution between residues in positions I and j, it computes the outer product of these learned residue representations, which transforms the 1D sequence features into the 2D pairwise representation. This resulting pairwise interaction map operates on a dimension of N x N (where N is the length of the protein sequence) across D feature channels (the embedding dimension) 3) Distance estimator: N x N x D pair feature tensor is fed into a 2D resNet to predict the final inter-residue contacts and distances. Under broadest reasonable interpretation, this learned vector representation corresponds to the claimed initial pair embedding (p. 2, col. 1, para. 2-3; p. 7, col. 1, para. 6 – col. 2, para. 7; Fig. 2; the method of claim 1/system of claim 26, wherein: the respective network input for each iteration comprises a respective initial pair embedding for each pair of amino acids in the protein).
Pertaining to claims 7, 8, 32 and 33, Greener et al. and Ju et al. are silent to: generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration). (claims 7 and 32); from the distance map specified by the structure parameters at the preceding iteration); and generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration (claims 8 and 33). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Jorgensen et al.
Pertaining to claims 7 and 32, combining a transformed distance map and an initial pair embedding involves concatenating or stacking or math functions such as fusion or aggregation. Jorgensen et al. teaches using tensor concatenation for their edge update networks. During the interaction steps the edge update function takes the concatenation of the current edge representation and the hidden states of the sending and receiving vertices as the direct input to a two-layered feed-forward neural network (Equations 6 and 7 (semicolon represents concatenation function); p. 3, para. 2-5 (section 2.1); generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration).
Regarding claims 8 and 33, Jorgensen et al. teaches a machine learning method for predicting properties of molecules. It is a neural message passing model with an edge update network which allows the information exchanged between atoms to depend on the hidden state of the receiving atom. a network that repeatedly updates learned edge representations. In this method, the edge feature depends on the representation of the atoms that the edge connects, thus information exchanged between atoms also depends on the receiving atom (p. 2, para. 3; abstract; Fig. 1; generating, a transformed distance map from the distance map specified by the structure parameters at the preceding iteration).
Regarding claims 8 and 33, combining a transformed distance map and an initial pair embedding involves concatenating or stacking or math functions such as fusion or aggregation. Jorgensen et al. teaches using tensor concatenation for their edge update networks. During the interaction steps the edge update function takes the concatenation of the current edge representation and the hidden states of the sending and receiving vertices as the direct input to a two-layered feed-forward neural network (Equations 6 and 7 (semicolon represents concatenation function); p. 3, para. 2-5 (section 2.1); generating the combined input comprises combining the transformed distance map and the initial pair embeddings for the subsequent iteration.
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Jorgensen et al. teaches their model with an edge update network has improved accuracy on formation energy prediction of molecules and materials across three benchmark datasets (p. 8, para. 2 – p. 9, para. 1). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Greener et al. and Ju et al. to update the learned pairwise representations taught by Ju et al. at successive neural network layers according to Jorgensen et al. because doing so would improve accuracy of predictions. Furthermore, one of ordinary skill in the art would predict the update method taught by Jorgensen et al. could be readily added to the method and system of Greener et al. and Ju et al. with a reasonable expectation of success because the methods and systems of Greener et al. and Ju et al. and Jorgensen et al. pertain to predicting molecular structures by neural network analysis and the method of Jorgensen et al. is applicable to applications where pairwise distances are used as edge features (Jorgensen et al. p. 2, para. 2). The invention is therefore prima facie obvious.
Pertaining to claims 7-8 and 32-33, Greener et al., Ju et al. and Jorgensen et al. are silent to: generating, from the distance map, a transformed distance map that has a same dimensionality as the initial pair embeddings (claims 7 and 32); generating the features comprises: generating a transformed distance map that has a same dimensionality as the initial pair embeddings (claims 8 and 33). However, this limitation was known in the art at the time of the effective filing date of the invention as taught by Hayden et al.
Pertaining to claims 7 and 32, to combine a distance map (shape: N x N) with a pair embedding (shape: N x N x D, where D is the embedding dimension), one must either expand/transform the distance map to match the embedding’s depth, or flatten the embedding to match the map’s 2D structure. Hayden et al. teaches Boundary-Aware Instance Segmentation (BAIS) framework, which includes Boundary-Aware Mask Representation, a method of converting traditional binary mask output of a shape N x N into a multi-valued map of shape N x N x D, which involves converting each scalar distance value into a distinct “bin” index, followed by one-hot encoding to transform the bin index into a vector of length B (where B is the total number of bins), followed by embedding in N x N x D by multiplying the B-dimensional one-hot vector by B x D embedding matrix to project each distance onto a D-dimensional continuous vector space. (p. 589, col. 1, para. 3 – p. 590, col. 1, para. 2; generating, from the distance map, a transformed distance map that has a same dimensionality as the initial pair embeddings).
Regarding claims 8 and 33, to combine a distance map (shape: N x N) with a pair embedding (shape: N x N x D, where D is the embedding dimension), one must either expand/transform the distance map to match the embedding’s depth, or flatten the embedding to match the map’s 2D structure. Hayden et al. teaches Boundary-Aware Instance Segmentation (BAIS) framework, which includes Boundary-Aware Mask Representation, a method of converting traditional binary mask output of a shape N x N into a multi-valued map of shape N x N x D, which involves converting each scalar distance value into a distinct “bin” index, followed by one-hot encoding to transform the bin index into a vector of length B (where B is the total number of bins), followed by embedding in N x N x D by multiplying the B-dimensional one-hot vector by B x D embedding matrix to project each distance onto a D-dimensional continuous vector space. (p. 589, col. 1, para. 3 – p. 590, col. 1, para. 2; generating the features comprises: generating a transformed distance map that has a same dimensionality as the initial pair embeddings).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Hayder et al. discloses that their instance segmentation approach to jointly detect, segment and classify every individual object in an image, outperforms other segmentation approaches (as measured by the mean average precision (mAP) of the models) and is robust to noisy object proposals (p. 588, col. 1, para. 2-3; Tables 1-2). Therefore, one of ordinary skill in the art would have been motivated to modify the iterative protein-structure prediction network taught by Greener et al., Ju et al. and Jorgensen et al. to include the distance dimensionality transformation method taught by Hayder et al. because it yields robust models that outperform models generated using other segmentation approaches. Furthermore, one of ordinary skill in the art would predict that the distance dimensionality transformation method taught by Hayder et al. could be readily added to the method of Greener et al., Ju et al. and Jorgensen et al. with a reasonable expectation of success because the methods and systems of Greener et al. and Ju et al. both pertain to predicting protein structures by neural network analysis and Greener et al. and Jorgensen et al. both pertain to distance transformations for neural network-based predictions in image analysis. The invention is therefore prima facie obvious.
Double Patenting
The nonstatutory double patenting rejection is based on a judicially created doctrine grounded in public policy (a policy reflected in the statute) so as to prevent the unjustified or improper timewise extension of the “right to exclude” granted by a patent and to prevent possible harassment by multiple assignees. A nonstatutory double patenting rejection is appropriate where the conflicting claims are not identical, but at least one examined application claim is not patentably distinct from the reference claim(s) because the examined application claim is either anticipated by, or would have been obvious over, the reference claim(s). See, e.g., In re Berg, 140 F.3d 1428, 46 USPQ2d 1226 (Fed. Cir. 1998); In re Goodman, 11 F.3d 1046, 29 USPQ2d 2010 (Fed. Cir. 1993); In re Longi, 759 F.2d 887, 225 USPQ 645 (Fed. Cir. 1985); In re Van Ornum, 686 F.2d 937, 214 USPQ 761 (CCPA 1982); In re Vogel, 422 F.2d 438, 164 USPQ 619 (CCPA 1970); In re Thorington, 418 F.2d 528, 163 USPQ 644 (CCPA 1969).
A timely filed terminal disclaimer in compliance with 37 CFR 1.321(c) or 1.321(d) may be used to overcome an actual or provisional rejection based on nonstatutory double patenting provided the reference application or patent either is shown to be commonly owned with the examined application, or claims an invention made as a result of activities undertaken within the scope of a joint research agreement. See MPEP § 717.02 for applications subject to examination under the first inventor to file provisions of the AIA as explained in MPEP § 2159. See MPEP § 2146 et seq. for applications not subject to examination under the first inventor to file provisions of the AIA . A terminal disclaimer must be signed in compliance with 37 CFR 1.321(b).
The filing of a terminal disclaimer by itself is not a complete reply to a nonstatutory double patenting (NSDP) rejection. A complete reply requires that the terminal disclaimer be accompanied by a reply requesting reconsideration of the prior Office action. Even where the NSDP rejection is provisional the reply must be complete. See MPEP § 804, subsection I.B.1. For a reply to a non-final Office action, see 37 CFR 1.111(a). For a reply to final Office action, see 37 CFR 1.113(c). A request for reconsideration while not provided for in 37 CFR 1.113(c) may be filed after final for consideration. See MPEP §§ 706.07(e) and 714.13.
The USPTO Internet website contains terminal disclaimer forms which may be used. Please visit www.uspto.gov/patent/patents-forms. The actual filing date of the application in which the form is filed determines what form (e.g., PTO/SB/25, PTO/SB/26, PTO/AIA /25, or PTO/AIA /26) should be used. A web-based eTerminal Disclaimer may be filled out completely online using web-screens. An eTerminal Disclaimer that meets all requirements is auto-processed and approved immediately upon submission. For more information about eTerminal Disclaimers, refer to www.uspto.gov/patents/apply/applying-online/eterminal-disclaimer.
14. Claims 1, 26 and 35 are rejected on the ground of nonstatutory double patenting as being unpatentable over claims 1, 10 and 16 of U.S. Patent No. 12,100,477 (reference patent; Senior et al.), in view of Greener et al. (Nature Communications, 2019, Vo. 10, p. 1-13).
Claim 1 lines 1-3 of the instant application are taught by claim 1 col. 20 lines 15-18 of the reference patent.
Claim 1 line 6 of the instant application (obtaining input data) are taught by claim 1 col. 20, lines 19-23 of the reference patent.
Claim 1 lines 7-9 of the instant application (combining with model parameters to generate parameters/features) are taught by claim 1 col. 20, lines 28-32 of the reference patent.
Claim 1 lines 11-12 of the instant application (obtain input data) are taught by claim 1 col. 20, lines 19-23 of the reference patent.
Claim 1 lines 17-19 of the instant application (combining input data with features) are taught by claim 1, col. 20, lines 28-30 of the reference patent.
Claim 1 lines 20-21 of the instant application (determining a final structure) are taught by claim 1, col. 20, lines 33-35 of the reference patent.
Claim 26 lines 1-7 of the instant application are taught by claim 10 of the reference patent. Note that the limitations of claim 10 of the reference application include all limitations of claim 1 of the reference application, and thus claim 10 teaches all other limitations of claim 26 except those indicated below.
Claim 35 lines 1-4 of the instant application are taught by claim 16 of the reference patent. Note that the limitations of claim 16 of the reference application include all limitations of claim 1 of the reference application, and thus claim 16 teaches all other limitations of claim 35 except those indicated below.
Regarding claims 1, 26 and 35, the reference patent does not teach iteratively repeating the predictions of protein structures. That is, the reference patent does not teach the limitations directed to: at a first iteration of a sequence of iterations that comprises the first iteration followed by one or more subsequent iterations (claims 1, 26 and 35); at each subsequent iteration in the sequence of iterations (claims 1, 26 and 35); and generating from (i) the structure parameters generated at a preceding iteration that precedes the subsequent iteration in the sequence, (ii) one or intermediate outputs generated by the protein structure prediction neural network while generating the structure parameters at the last iteration, or (iii) both, features for the subsequent iteration (claims 1, 26 and 35). However, these limitations were known in the art at the time of the effective filing date of the invention as taught by Greener et al.
Regarding claims 1, 26, and 35, Greener et al. discloses an iterative pipeline for predicting protein structures including obtaining and integrating a network input with structural parameters generated by a neural network model. Greener et al. discloses that in the iterative DMPfold pipeline, initially inter-residue Cβ distances, H-bonds and torsion angles are predicted from DeepMetaPSICOV (DMP) inputs using a neural network; these are used to generate models with CNS, and a single model is used as additional input to refine the distances and H-bonds. After 3 iterations a final set of models is returned (Fig. 8; p. 10, col. 1, para. 3 – p. 11, col. 2, para. 3; at a first iteration of a sequence of iterations that comprises the first iteration followed by one or more subsequent iterations (claims 1, 26 and 35); at each subsequent iteration in the sequence of iterations (claims 1, 26 and 35); and generating from (i) the structure parameters generated at a preceding iteration that precedes the subsequent iteration in the sequence, (ii) one or intermediate outputs while generating the structure parameters at the last iteration, or (iii) both, features for the subsequent iteration (claims 1, 26 and 35)).
An invention would have been prima facie obvious to one of ordinary skill in the art at the effective filing date of the invention if some motivation in the prior art would have led that person to combine the prior art teachings to arrive at the claimed invention. Greener et al. discloses that the benefit of iterations to DMPfold (the protein structure prediction method) is shown by the fact that in an example structure determination, 19 of 22 domains showed a higher template modeling score (TM-score) at the last iteration than at the first iteration (Fig. 2d; p. 3, col. 2, para. 1 – p. 5, col. 1, para. 1). Therefore, one of ordinary skill in the art would have been motivated to modify the protein-structure prediction network taught by the reference patent (Jumper et al.) to include iterative prediction refinement taught by Greener et al. in order to improve the accuracy of structure predictions. Furthermore, one of ordinary skill in the art would predict that the iterative model refinement structure taught by Greener et al. could be readily added to the method of the reference patent (Jumper et al.) with a reasonable expectation of success because the methods and systems of Greener et al. and Jumper et al. both pertain to predicting protein structures by neural network analysis using the same data types. The invention is therefore prima facie obvious.
15. Claims 1, 26 and 35 are provisionally rejected on the ground of nonstatutory double patenting as being unpatentable over claims 1, 4, 9, 12, 17 and 20 of copending Application No. 18/940,983.
Claim 1 lines 1-3 of the instant application are taught by claim 1 lines 1-2 of the reference application.
Claim 1 lines 4-5 of the instant application are taught by claim 1 lines 5-6 of the reference application.
Claim 1 line 6 of the instant application is taught by claim 1 lines 3-4 of the reference application.
Claim 1 lines 7-19 of the instant application (processing and generating features to iteratively refine a structure prediction) are taught by claim 1 lines 7-15 and claim 4 of the reference application.
Claim 1 lines 20-21 of the instant application (determining a final model) are taught by claim 1, lines 16-19 of the reference application.
Claim 26 lines 1-7 of the instant application are taught by claim 9 lines 1-6 of the reference application.
Claim 26 lines 8-25 of the instant application are taught by claims 9 and 12 of the reference application. Note that the limitations of claims 9 and 12 of the reference application include all limitations of claims 1 and 4 of the reference application, and thus are also addressed above.
Claim 35 lines 1-4 of the instant application are taught by claim 17 lines 1-3 of the reference patent.
Claim 35 lines 5-22 of the instant application are taught by claims 17 and 20 of the reference application. Note that limitations of claims 17 and 20 of the reference application include all limitation of claims 1 and 4 of the reference application and thus are also addressed above.
This is a provisional nonstatutory double patenting rejection.
Conclusion
16. No claims are allowed.
E-mail Communications Authorization
17. Per updated USPTO Internet usage policies, Applicant and/or applicant's representative is encouraged to authorize the USPTO examiner to discuss any subject matter concerning the above application via Internet e-mail communications. See MPEP 502.03. To approve such communications, Applicant must provide written authorization for e-mail communication by submitting the following statement via EFS-Web (using PTO/SB/439) or Central Fax (571-273-8300): "Recognizing that Internet communications are not secure, / hereby authorize the USPTO to communicate with the undersigned and practitioners in accordance with 37 CFR 1.33 and 37 CFR 1.34 concerning any subject matter of this application by video conferencing, instant messaging, or electronic mail. / understand that a copy of these communications will be made of record in the application file."
Written authorizations submitted to the Examiner via e-mail are NOT proper. Written authorizations must be submitted via EFS-Web (using PTO/SB/439) or Central Fax (571-273- 8300). A paper copy of e-mail correspondence will be placed in the patent application when appropriate. E-mails from the USPTO are for the sole use of the intended recipient, and may contain information subject to the confidentiality requirement set forth in 35 USC § 122. See also MPEP 502.03.
Inquiries
18. Any inquiry concerning this communication or earlier communications from the examiner should be directed to JENNIFER J SMITH whose telephone number is (571)272-7801. The examiner can normally be reached Monday-Friday 7:30 AM - 4:00 PM.
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, Olivia Wise can be reached at (571) 272-2249. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/J.J.S./Examiner, Art Unit 1685
/OLIVIA M. WISE/Supervisory Patent Examiner, Art Unit 1685
Prosecution Insights