DETAILED ACTION
This office action is in response to submission of application on 08/16/2023.
Claims 1-20 are presented for examination.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 11/03/2023 is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 102
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.
Claims 1, 5, 7-12 and 16-18 are rejected under 35 U.S.C. 102(a)(2) as being anticipated by
Gómez-Bombarelli et al. (hereinafter Gómez), “Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules” (published 12/05/2017).
Regarding Claim 1,
Gómez teaches A processor comprising: one or more circuits to: (Examiner notes that this limitation is interpreted as implementation of the disclosed steps in a generic computing environment. Pg. 21, section ‘Supplementary Materials’: “The code and full training data sets will be made available…” The use of computer code necessitates implementation on a computer.)
receive a data structure representing a first chemical species; (Pg. 4, figure 1: “A diagram of the proposed autoencoder for molecular design, including the joint property prediction model. Starting from a discrete molecular representation, such as a SMILES string…” The model receives a SMILES string molecular representation (i.e. a data structure representing a first chemical species).)
encode, using at least one machine learning model, the data structure into a latent space of a fixed size to determine an encoded representation of the data structure; (Pg. 5, section ‘Training an autoencoder’: “Starting from a large library of string-based representations of molecules, we trained a pair of deep neural networks: an encoder network to convert each string into a fixed-dimensional vector, and a decoder network to convert vectors back into strings (Figure 1a). Such encoder decoder pairs are known as autoencoders.” An encoder of a deep neural network autoencoder (i.e. a machine learning model) is trained to convert a string (i.e. data structure) into a fixed-dimensional vector (i.e. a latent space of a fixed size) to determine its encoded representation.)
apply noise to the encoded representation to determine a modified encoded representation; and (Pg. 7, section ‘Using variational autoencoders to produce a latent representation’: “The intuition is that adding noise to the encoded molecules forces the decoder to learn how to decode a wider variety of latent points and find more robust representations.” Noise is added to the encoded molecule representation to determine a modified encoded molecule representation.)
decode the modified representation using the at least one machine learning model to determine a modified data structure representing a second chemical species different from the first chemical species. (Pg. 5, section ‘Training an autoencoder’: “Starting from a large library of string-based representations of molecules, we trained a pair of deep neural networks: an encoder network to convert each string into a fixed-dimensional vector, and a decoder network to convert vectors back into strings (Figure 1a). Such encoder decoder pairs are known as autoencoders.” A decoder of a deep neural network autoencoder (i.e. machine learning model) is trained to convert the modified fixed-dimensional vector (i.e. modified representation) to a string (i.e. a modified data structure representing a second chemical species different from the first chemical species).)
Regarding Claim 5, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein the receiving the data structure representing a first chemical species comprises receiving a plurality of simplified molecular-input line-entry system (SMILES) forms representing the first chemical species. (Pg. 5, section ‘Representation and Autoencoder Framework’: “Before building an encoder that produces a continuous latent representation, we must choose which discrete molecular representation to use for the input and output. To leverage the power of recent advances in sequence-to-sequence autoencoders for modeling text, we used the SMILES representation, a commonly-used text encoding for organic molecules… Starting from a large library of string-based representations of molecules…” A library of SMILES string molecular representations are used as the data structures representing the chemical species.)
Regarding Claim 7, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein:
the data structure has dimensions N
×
D, the latent space has dimensions K
×
D, and the modified data structure has dimensions M
×
D, wherein N is a variable tokens number for the data structure, D is an embeddings dimension, K is the fixed size of the latent space, and M is a variable tokens number for the modified data structure. (Pg. 13, section ‘Autoencoder architecture’: “Our SMILES-based text encoding used a subset of 35 different characters for ZINC and 22 different characters for QM9. For ease of computation, we encoded strings up to a maximum length of 120 characters for ZINC and 34 characters for QM9, although in principle there is no hard limit to string length. Shorter strings were padded with spaces to this same length.” Pg. 7, section ‘Representation of molecules in latent space’: “The latent space representations for the QM9 and ZINC datasets had 156 dimensions and 196 dimensions respectively.” For the ZINC dataset, the SMILES strings (i.e. the data structure and modified data structure) have length 120 (i.e. dimensions N
×
D and M
×
D, where N=120 and M=120 are the number of tokens in the data structure and D=1 is the embedding dimension). The latent space has dimension 196 (i.e. K
×
D, where K=196 is the fixed size of the latent space and D=1 is the embedding dimension).)
Regarding Claim 8, Gómez teaches The processor of claim 7, as shown above.
Gómez also teaches wherein M equals N. (Pg. 13, section ‘Autoencoder architecture’: “For ease of computation, we encoded strings up to a maximum length of 120 characters for ZINC and 34 characters for QM9, although in principle there is no hard limit to string length. Shorter strings were padded with spaces to this same length… The last layer of the RNN decoder defines a probability distribution over all possible characters at each position in the SMILES string.” For the ZINC dataset, the encoded strings have length 120 (i.e. N=120), and the decoder produces strings of the same length (i.e. M=120).)
Regarding Claim 9, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein the at least one machine learning model comprises an encoder to encode the data structure into the latent space and a decoder to determine the modified data structure from the modified encoded representation. (Pg. 5, section ‘Training an autoencoder’: “Starting from a large library of string-based representations of molecules, we trained a pair of deep neural networks: an encoder network to convert each string into a fixed-dimensional vector, and a decoder network to convert vectors back into strings (Figure 1a). Such encoder decoder pairs are known as autoencoders.” The autoencoder (i.e. machine learning model) comprises an encoder to convert each string into a fixed-dimensional vector (i.e. encode the data structure into the latent space) and a decoder to convert vectors back into strings (i.e. determine the modified data structure from the modified encoded representation).)
Regarding Claim 10, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein the different chemical species satisfy one or more criteria comprising at least one of matching a data structure representing a chemical species in a database, existing in a chemically stable form; or being capable of synthesis. (Pg. 6-7, section ‘Using variational autoencoders to produce a latent representation’: “To ensure that points in the latent space correspond to valid realistic molecules, we modified our autoencoder and its objective into a variational autoencoder (VAE)… Two autoencoder system were trained; one with 108,000 molecules from the QM9 dataset of molecules with fewer than 9 heavy atoms and another with 250,000 drug-like commercially available molecules extracted at random from the ZINC database.” Molecules (i.e. chemical species) used as model input are drawn from the QM9 dataset and ZINC database (i.e. match a data structure representing a chemical species in a database), and molecules decoded from the latent space correspond to valid realistic molecules (i.e. exist in a chemically stable form or are capable of synthesis).)
Regarding Claim 11, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein the processor is comprised in at least one of: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for performing digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing deep learning operations; a system implemented using an edge device; a system implemented using a robot; a system implemented using a language model; a system implemented using a large language model (LLM); a system for performing generative Al operations; a system for performing conversational Al operations; a system for generating synthetic data; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources. (Pg. 3-4, Introduction: “We apply such generative models to chemical design, using a pair of deep networks trained as an autoencoder to convert molecules represented as SMILES strings into a continuous vector representation.” Pg. 21, section ‘Supplementary Materials’: “The code and full training data sets will be made available…” Deep neural networks are implemented using computer code, which necessitates a computing system for performing deep learning operations.)
Claims 12 and 16-17 are system claims containing substantially the same elements as system claims 1, 5, and 11, respectively. Gómez teaches the elements of claims 1, 5, and 11, as shown above.
Gómez also teaches A system comprising: one more processing units to execute operations (Examiner notes that this limitation is interpreted as implementation of the disclosed steps in a generic computing environment. Pg. 21, section ‘Supplementary Materials’: “The code and full training data sets will be made available…” The use of computer code necessitates implementation on a computer.)
Claim 18 is a method claim containing substantially the same elements as system claim 1. Gómez teaches the elements of claim 1, as shown above.
Claim Rejections - 35 USC § 103
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.
Claims 2, 13, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Gómez in view of
Kang et al. (hereinafter Kang), “Analysis of training and seed bias in small molecules generated with a conditional graph-based variational autoencoder” (published 09/03/2021).
Regarding Claim 2, Gómez teaches The processor of claim 1, as shown above.
Gómez does not appear to explicitly disclose the remaining features of claim 2.
However, Kang teaches wherein the one or more circuits are to apply noise to the encoded representation by applying noise sampled from a Gaussian distribution with a defined standard deviation according to a target amount of modification of the second chemical species relative to the first chemical species. (Pg. 4, section ‘Protocols for Molecule Generation’: “Using our trained conditional VAEs, we generated molecules using two sets of protocols: 1) direct (random) sampling from the latent space, and 2) seed-based sampling based on designated input molecules… In the latter case, a seed molecule and corresponding activity condition were fed through the encoder to produce an embedding, which was then augmented with Gaussian noise… we prepared four different sets of molecules at various distances from the variational mean with noise standard deviations 0.5, 1.0, 2.0, and 4.0 (indicated with s0.5, s1.0, s2.0, and s4.0, respectively).” Pg. 4, section ‘Generative Model Can Learn Unconditioned Molecular Features’: “The distributions do vary somewhat as a function of standard deviation: at points, sampling at small standard deviations (e.g., s0.5) generates molecules closer in size to the underlying seed molecules than does sampling at larger standard deviations.” Gaussian noise is applied to the embedding of the seed molecule (i.e. the encoded representation). The standard deviation of the gaussian noise can be defined by different values, which will result in different degrees of similarity between the seed molecule and the generated molecule (i.e. amount of modification of the second chemical species relative to the first chemical species).)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Gómez and Kang. Gómez teaches an autoencoder for molecule generation and optimization in continuous latent space. Kang teaches an autoencoder for molecule generation by applying gaussian noise to the latent representation of a seed molecule. One of ordinary skill would have motivation to combine Gómez and Kang because Gómez’s continuous latent space allows for new molecules to be generated by perturbing known molecules (“Continuous representations allow us to automatically generate novel chemical structures by performing simple operations in the latent space, such as decoding random vectors, perturbing known chemical structures, or interpolating between molecules”, Gómez, pg. 2, Abstract), and Kang provides a mechanism for performing this perturbation by applying gaussian noise with an adjustable standard deviation, with the stated benefit that “molecules of relatively large, drug-like size are efficiently generated through the seed-based approach” (Kang, pg. 4, section ‘Generative Model Can Learn Unconditioned Molecular Features’).
Claim 13 is a system claim containing substantially the same elements as system claim 2. Gómez and Kang teach the elements of claim 2, as shown above.
Claim 19 is a method claim containing substantially the same elements as system claim 2. Gómez and Kang teach the elements of claim 2, as shown above.
Claims 3-4, 14-15, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Gómez in view of Kang and
Livne et al. (hereinafter Livne), “SentenceMIM: A Latent Variable Language Model” (published 09/03/2021).
Regarding Claim 3, Gómez teaches The processor of claim 1, as shown above.
Gómez does not appear to explicitly disclose the remaining features of claim 3.
However, Livne teaches wherein: the latent space comprises one or more clusters of encoded representations of chemical species; and (Pg. 2, section 1: “Here we propose MIM learning to enable high dimensional representations, while encouraging low latent entropy (under certain conditions) to promote clustering of semantically similar observations.” Latent (i.e. encoded) representations of similar observations (i.e. chemical species) are clustered in the latent space.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Gómez and Livne. Gómez teaches a variational autoencoder for molecule generation and optimization in continuous latent space which operates on string data representing molecules. Livne teaches applying a mutual information machine (MIM)—an improved variational autoencoder—to text data. One of ordinary skill would have motivation to combine Gómez and Livne because “MIM is a recently introduced LVM framework that shares the same underlying architecture as VAEs, but uses a different learning objective that is robust against posterior collapse… sMIM provides better reconstruction than VAE models, matching the reconstruction accuracy of AEs, but with semantically meaningful representations, comparable to VAEs” (Livne, pg. 1-2, section 1).
Kang teaches the one or more circuits are to determine the modified representation using the one or more clusters. (Pg. 12, section ‘Molecular Similarity is Correlated with Latent-Space Similarity’: “Notably, latent vectors sampled during seed-based generation are tightly clustered around their reference seeds;” A seed-based latent vector (i.e. modified representation) is generated (i.e. determined) within the reference seed’s cluster.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Gómez, Livne, and Kang. Gómez teaches a variational autoencoder for molecule generation and optimization in continuous latent space which operates on string data representing molecules. Livne teaches applying a mutual information machine (MIM)—an improved variational autoencoder—to text data. Kang teaches an autoencoder for molecule generation by applying gaussian noise to the latent representation of a seed molecule. One of ordinary skill would have motivation to combine Gómez, Livne, and Kang because “molecules of relatively large, drug-like size are efficiently generated through the seed-based approach” (Kang, pg. 4, section ‘Generative Model Can Learn Unconditioned Molecular Features’).
Regarding Claim 4, Gómez, Livne, and Kang teach The processor of claim 3, as shown above.
Livne also teaches wherein the at least one machine learning model is updated, at least in part, by a mutual information machine (MIM) training process comprising: (Pg. 1, Abstract: “SentenceMIM is a probabilistic auto-encoder for language data, trained with Mutual Information Machine (MIM) learning…” The autoencoder (i.e. machine learning model) is trained (i.e. updated) by mutual information machine (MIM) learning.)
receiving first and second training sample distributions, the first and second training sample distributions comprising data structures representing a plurality of [chemical species]; (Pg. 3, section 2.2: “The MIM framework begins with two anchor distributions,
P
(
x
)
and
P
(
z
)
, for observations and the latent space, from which one can draw samples. They are fixed and not learned. MIM also has a parameterized encoder-decoder pair,
q
θ
(
z
|
x
)
and
p
θ
(
x
|
z
)
, and parametric marginal distributions
q
θ
(
x
)
and
p
θ
(
z
)
. These parametric elements define joint encoding and decoding model distributions:
q
θ
x
,
z
=
q
θ
(
z
|
x
)
q
θ
(
x
)
,
p
θ
x
,
z
=
p
θ
(
x
|
z
)
p
θ
(
z
)
…” Anchor/sample distribution
P
(
x
)
for observations is a first training sample distribution, and encoding model distribution
q
θ
x
,
z
is a second training sample distribution.)
encoding the first and second training sample distributions into the latent space using the at least one machine learning model to determine updated encoded representations; and (Pg. 3, section 2.2: “For language modeling we therefore use A-MIM learning, a MIM variant that minimizes a loss defined on the encoding and decoding distributions, with samples drawn from an encoding sample distribution, denoted
M
S
q
(
x
,
z
)
; i.e.,
M
S
q
x
,
z
=
q
θ
(
z
|
x
)
P
(
x
)
.” As can be seen in equation 7, the loss function used by MIM training includes the term
M
S
q
(
x
,
z
)
, which represents the observation sample distribution and associated latent encodings (i.e. the first training sample distribution encoded into latent space), and the term
q
θ
x
,
z
, which represents the encoding model distribution (i.e. the second training sample distribution encoded into latent space).)
clustering the updated encoded representations by similarity of chemical species using the at least one machine learning model (Pg. 2, section 1: “Here we propose MIM learning to enable high dimensional representations, while encouraging low latent entropy (under certain conditions) to promote clustering of semantically similar observations.” Latent (i.e. encoded) representations of similar observations (i.e. similar chemical species) are clustered by the model.)
with a variational upper bound on differences between the first and second training sample distributions. (Pg. 3, section 2.2: “MIM learning entails the minimization of the cross-entropy between a sample distribution and the model encoding and decoding distributions (Livne et al., 2019). This simple loss constitutes a variational upper bound on a regularized Jensen-Shannon divergence…” MIM learning includes a variational upper bound on the Jensen-Shannon divergence (i.e. differences) between the sample distribution (i.e. first training sample distribution) and the model encoding distribution (i.e. second training sample distribution).)
Gómez teaches data structures representing a plurality of chemical species (Pg. 5, section ‘Representation and Autoencoder Framework’: “To leverage the power of recent advances in sequence-to-sequence autoencoders for modeling text, we used the SMILES representation, a commonly-used text encoding for organic molecules… Starting from a large library of string-based representations of molecules…” Data structures used by the model represent organic molecules (i.e. chemical species).)
Claims 14-15 are system claims containing substantially the same elements as system claims 3-4, respectively. Gómez, Livne, and Kang teach the elements of claim 3-4, as shown above.
Regarding Claim 20, Gómez teaches The method of claim 18, as shown above.
Gómez does not appear to explicitly disclose the remaining features of claim 20.
However, Livne teaches further comprising clustering encoded representations of [chemical species] according to [chemical] similarity. (Pg. 2, section 1: “Here we propose MIM learning to enable high dimensional representations, while encouraging low latent entropy (under certain conditions) to promote clustering of semantically similar observations.” Latent (i.e. encoded) representations of observations are clustered according to similarity.)
Gómez teaches data representing chemical species (Pg. 5, section ‘Representation and Autoencoder Framework’: “To leverage the power of recent advances in sequence-to-sequence autoencoders for modeling text, we used the SMILES representation, a commonly-used text encoding for organic molecules… Starting from a large library of string-based representations of molecules…” Data used by the model represent organic molecules (i.e. chemical species).)
Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Gómez in view of
Hoffman et al. (hereinafter Hoffman), “Optimizing Molecules using Efficient Queries from Property Evaluations” (published 10/18/2021).
Regarding Claim 6, Gómez teaches The processor of claim 1, as shown above.
Gómez also teaches wherein the one or more circuits are to:
evaluate a physicochemical property of the second chemical species represented by the modified data structure by inputting the modified [data structure] into a function trained with physicochemical property data and outputting a physicochemical property score for the second chemical species; and (Pg. 11, section ‘Optimization of molecules via properties’: “We next optimized molecules in the latent space from the autoencoder which was jointly trained for property prediction. We used a Gaussian process model to predict target properties for molecules given the latent space representation of the molecules as an input. The 2,000 molecules used for training the Gaussian process were selected to be maximally diverse.” Target molecular properties (i.e. physicochemical properties) are predicted (i.e. property scores are output) by inputting the modified latent representation to a Gaussian process model (i.e. a function) trained with molecular property data.)
further modify the encoded representation responsive to the physiochemical property score not satisfying a target criterion. (Pg. 11, section ‘Optimization of molecules via properties’: “Using this model, we optimized in the latent space to find a molecule that maximized our objective.” The molecule’s latent space representation (i.e. encoded representation) is optimized (i.e. further modified) if the objective is not maximized (i.e. the physicochemical property score does not satisfy a target criterion).)
As shown above, Gómez teaches predicting a physicochemical property by inputting the latent representation into a function. Gómez does not appear to explicitly disclose predicting a physicochemical property by inputting the modified data structure (i.e. the decoded representation) into a function
However, Hoffman teaches predicting a physicochemical property by inputting the modified data structure into a function (Pg. 4, section ‘Molecule Optimization Formulation via Guided Search’: “[O]ur QMO [query-based molecule optimization] framework incorporates molecular property prediction models and similarity metrics at the sequence level as external guidance. Specifically, for any given sequence
x
∈
X
m
, we use a set of
I
separate prediction models
{
f
i
(
x
)
}
i
=
1
I
to evaluate the properties of interest for MO [molecule optimization]… Our QMO framework covers two practical cases in MO: (i) optimizing molecular similarity while satisfying desired chemical properties and (ii) optimizing chemical properties with similarity constraints.” Pg. 5, section ‘Query-based Molecule Optimization (QMO) Procedure’: “When solving (1), an iterate
z
(
t
)
is considered as a valid solution if its decoded sequence
D
e
c
(
z
(
t
)
)
satisfies the property conditions
f
i
(
D
e
c
(
z
(
t
)
)
)
≥
τ
i
for all
i
∈
[
I
]
.” The decoded sequence
D
e
c
(
z
(
t
)
)
of latent representation
z
(
t
)
is input to the property prediction model
f
i
for molecule optimization.)
It would have been obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Gómez and Hoffman. Gómez teaches an autoencoder for molecule generation and optimization in continuous latent space. Hoffman teaches molecule optimization in continuous latent space by inputting decoded representations to property prediction models. One of ordinary skill would have motivation to combine Gómez and Hoffman because Hoffman’s framework for molecule optimization “reduces the problem complexity by decoupling representation learning and guided search. It applies to any plug-in (pre-trained) encoder-decoder with continuous latent representations. It is also a unified and principled approach that incorporates multiple predictions and evaluations made directly at the molecule sequence level into guided search without further model fitting” (Hoffman, pg. 2, Introduction).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BENJAMIN M ROHD whose telephone number is (571)272-6445. The examiner can normally be reached Mon-Thurs 8:00-6:00 EST.
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, Viker Lamardo can be reached at (571) 270-5871. 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.
/B.M.R./Examiner, Art Unit 2147
/ERIC NILSSON/Primary Examiner, Art Unit 2151
Prosecution Insights