DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Status
Claims 1-20 are currently pending and under exam herein.
Claims 1-20 are rejected.
Claim 8 is objected to.
Priority
The is no claim to domestic or foreign priority in the instant application, hence the effective filing date of claims 1-20 is June 17th, 2022.
Information Disclosure Statement
The information disclosure statements (IDS) were submitted on Jun 17th 2022, May 7th 2024, and Sep 18th 2024. The submissions are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statements are being considered by the examiner. A signed copy of the list of references cited from the IDS is included with this Office Action.
Drawings
The Drawings filed on June 17th 2022 are accepted.
Specification
The Specification filed on June 17th 2022 is accepted.
Claim Objections
Claim 8 is objected to because of the following informalities: contains an extra “and” in the phrase “constructing a graph with said neural network using and a node feature set”
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
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, 8, and 16 recites a system, a method, and a program respectively, for predicting structural features (abstract idea; mental process), producing predicted structures based on the predicted structural feature sets (abstract idea; mental process), and converting the predicted structures into predicted graphs with predicted edges (abstract idea; mental process), before comparing the predicted graphs and predicted edges to training graphs and training edges to obtain a comparison (abstract idea; mental process). After obtainment of the comparison, the claims then constructs a graph using a node feature set (abstract idea; mental process and/or mathematical concepts) before attempting to reduce missing edges in the graph with the comparisons (abstract idea; mental process). The claims are merely predicting features (ex: angles around a carbon atom) relevant to a molecule (protein in this case) to construct a prediction on the structure of the molecule (whether or not they have alpha helices, beta sheets, and how the amino acids are connected), which is a mental process that can be done in the human mind through the use of mathematical relations or with paper and pen. The claims then map the predicted structure to a graph, by modeling each amino acid as a node and connections to other amino acids as an edge, which utilizes mathematical concepts and in the broadest sense is also feasible by the human mind. Lastly, the claims compare the predicted graphs to known/trained graphs to get differentials between the two (see if edges are missing in the predicted versus the trained), again utilizing mathematical relations and can also can be done in the human mind. With the differentials, the claims then corrects constructed graphs to reduce missing edges, a process that can be performed mentally and with mathematical relations. Although, the claims do recite performing these steps in a generic computer environment/system, this does not bar the limitations from be practically performed in the human mind as well. Therefore, these claim limitations constitute both mental processes and mathematical concepts, falling broadly under the category of abstract ideas. Please see MPEP § 2106.04(a)(2) for more details.
Claims 6 and 14 recite labeling differences between the predicted graph and training graphs (abstract idea; mental process) and labeling differences between predicted edges and training edges (abstract idea; mental process). The claims are merely comparing between two data structures to find their differences and point them out through labeling, which is a simple process that can be done in the human mind either on a generic computer or with pen and paper.
Claims 7, 15, and 20 recite predicting a plurality of features selected from the group of dihedral angles, B-factor, solvent-accessible surface are, long-range angles and short-range angles (abstract idea; mental process and/or mathematical concept). The process of predicting these features, for example a dihedral angle (angle around the alpha carbon of an amino acid) based on information of substitutes around the molecule, in the broadest sense, is something that can be done in the human mind. In addition, the process of predicting these features also utilizes mathematical relations, hence they may also constitute mathematical concepts.
These recitations are similar to the concepts of collecting information, analyzing it and displaying certain results of the collection and analysis in Electric Power Group, LLC, v. Alstom (830 F.3d 1350, 119 USPQ2d 1739 (Fed. Cir. 2016)), organizing and manipulating information through mathematical correlations in Digitech Image Techs., LLC v Electronics for Imaging, Inc. (758 F.3d 1344, 111 U.S.P.Q.2d 1717 (Fed. Cir. 2014)) and comparing information regarding a sample or test to a control or target data in Univ. of Utah Research Found. v. Ambry Genetics Corp. (774 F.3d 755, 113 U.S.P.Q.2d 1241 (Fed. Cir. 2014)) and Association for Molecular Pathology v. USPTO (689 F.3d 1303, 103 U.S.P.Q.2d 1681 (Fed. Cir. 2012)) that the courts have identified as concepts that can be practically performed in the human mind or mathematical relationships. Therefore, these limitations fall under the “Mental process” and “Mathematical concepts” groupings of abstract ideas. While claims 1-20 recite performing some aspects of the analysis with an “neural network”, there are no additional limitations that indicate that this neural network 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 1-20 recite an abstract idea (Step 2A, Prong 1: YES).
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). The above judicial exceptions are not integrated into a practical application because the claims do not recite additional elements that reflects an improvement to technology or applies or uses the recited judicial exception to affect a particular treatment for a condition. Rather, the instant claims recite additional elements that amount to mere instructions to implement the abstract idea in a generic computing environment. Specifically, the claims recite the following additional elements:
Claims 1, 8, and 16 recite training a neural network
Claim 1 recites a system with a memory and a processor in communication with the memory
Claim 2, 9, and 17 recites that the neural network is a multi-scale neighborhood-based neural network
Claim 3, 11, and 18 recites that the model is a variational graph auto-encoder
Claim 4 and 12 recites inputting an amino acid code into the neural network used to produce the structures
Claims 5, 13, and 19 recite a molecular stimulation program
Claim 8 recites a computer-implemented method
Claim 10 further defines the multi-scale neighborhood-based neural network in claim 9 to be a multi-goal multi-scale neighborhood-based neural network
Claim 16 recites a computer program product with a computer readable storage medium and processor
While the additional elements further limit the claim limitations/judicial exceptions by specifying that the structures predicted relate to amino acids, and naming the analysis engines used, they do not indicate that the claimed neural network, autoencoder, and molecular stimulation program, requires anything other implementing the abstract ideas on generic computing components. 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. In general, linking the use of an abstract idea to a particular technological environment, such as a computer, does not integrate the abstract idea into a practical application based on MPEP 2106.06(h). As such, claims 1-20 are directed to an abstract idea as the additional elements do not integrate the judicial exceptions into a practical application (Step 2A, Prong 2: NO).
Claims found to be directed to a judicial exception under Step 2A, Prong 2, are then further evaluated to determine if the claims recite an inventive concept that provides significantly more than the judicial exception itself (Step 2B). In the instant application, claims 1-20 do not recite further limitations or specification to the additional elements that would indicate anything other than carrying out the steps in a generic computing environment, leading to a lack of inventive step. According to MEPE 2106.05(d), courts have held computer-implemented processes not to be significantly more than an abstract idea (and thus ineligible) where the claim as a whole amount to nothing more than generic computer functions merely used to implement an abstract idea, such as an idea that could be done by a human analog (i.e. by hand or by merely thinking). MPEP 2106.05(f) also discloses that mere instructions to apply the judicial exception cannot provide an inventive concept to the claims. 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 1-20 are not patent eligible.
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.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Heffernan et al. (Bioinformatics 33(10) Pgs. 2842-2849, Apr 18 2017) in view of Guo et al. (Bioinformatics Advances 1(1) vbab036, Nov 29 2021). The limitations of the instant claims are italicized below.
With respect to claim 1, Heffernan et al. teaches the use of long short-term memory (LTSM) bidirectional recurrent neural networks (BRNNs) to predict protein secondary structures, backbone angles, contact numbers and solvent accessibility (pg. 2842 Abstract: Results). Heffernan et al. teaches that the neural network model was trained on a system utilizing Google’s open-source TensorFlow library, a software library that has a memory, with a Nvidia GeForce Titan X GPU, a processor in communication with the memory (pg. 2844 right col para 1, A system, said system comprising: a memory; and a processor in communication with said memory). Heffernan et al. also discloses that the neural network was trained using Adam optimization, a method for efficient stochastic optimization that is able to compute adaptive learning rates and requires little hyperparameter tuning (pg. 2844 left col para 2, training a neural network). In addition, Heffernan et al. adds that the neural network is able to predict structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting structural feature sets with said neural network). Furthermore, Heffernan et al. highlights that the neural network utilizes an iterative process, that can take these structural features to further predict secondary structures (SS) of the protein (pg. 2844 right col para 3, producing predicted structures with said neural network using said structural feature sets).
Concerning claim 2, Heffernan et al. teaches that the neural network architecture is able to process various inputs (protein sequences, position specific substitution matrices, hidden Markov Model sequence profiles) of various sizes (pg. 2844 Fig 1 and right col para 3, said neural network is a multi-scale). In addition, Heffernan et al. specifically utilizes long short-term memory (LTSM) cells due to their ability to learn both distant and close intra-sequence dependencies, which helps to analyze the nodes in conjunction with their neighbors (pg. 2844 left col para 1, neighborhood-based neural network).
Regarding 4, Heffernan et al. teaches that 5789 proteins were used as input into the neural network for prediction of secondary structures (pg. 2842 right col para 3, inputting amino acid code into said neural network…wherein said amino acid code is used to produce said predicted structures). In addition, Heffernan et al. discloses that structural features predicted from the first iteration with these proteins are then also added as another initial input for a second round of hopefully more accurate predictions (pg. 2844 right col para 3, with said structural feature set).
With respect to claim 5, Heffernan et al. discloses that the secondary features are produced by predicting/stimulating molecular interactions between the amino acids of the protein sequence based on physio-chemical properties (PP) of amino acids (pg. 2844 right col para 3, using a molecular stimulation program to produce said predicted structures). On another note, applicant also provided an example of a well-known molecular stimulation program in the Specification of the instant application, called the Chemistry at Harvard Macromolecular Mechanics program (CHARMM) which has been well established for the use of protein structure stimulation (Specification [0066]).
Concerning claim 7, Heffernan et al. teaches the prediction of several structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting a plurality of features of said predicted structure selected from the group consisting of dihedral angles, B-factor, solvent-accessible surface area, long-range angles, and short-range angles).
Regarding claim 8, Heffernan et al. discloses a computer-implemented method of using long short-term memory (LTSM) bidirectional recurrent neural networks (BRNNs) to predict protein secondary structures, backbone angles, contact numbers and solvent accessibility (pg. 2842 Abstract: Results, a computer-implemented method). Heffernan et al. teaches that the neural network was trained using Adam optimization, a method for efficient stochastic optimization that is able to compute adaptive learning rates and requires little hyperparameter tuning (pg. 2844 left col para 2, training a neural network). In addition, Heffernan et al. adds that the neural network is able to predict structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting structural feature sets with said neural network). Furthermore, Heffernan et al. highlights that the neural network utilizes an iterative process, that can take these structural features to further predict secondary structures (SS) of the protein (pg. 2844 right col para 3, producing predicted structures with said neural network using said structural feature sets).
With respect to claim 9, Heffernan et al. teaches that the neural network architecture is able to process various inputs (protein sequences, position specific substitution matrices, hidden Markov Model sequence profiles) of various sizes (pg. 2844 Fig 1 and right col para 3, said neural network is a multi-scale). In addition, Heffernan et al. specifically utilizes long short-term memory (LTSM) cells due to their ability to learn both distant and close intra-sequence dependencies, which helps to analyze the nodes in conjunction with their neighbors (pg. 2844 left col para 1, neighborhood-based neural network).
Concerning claim 10, Heffernan et al. teaches that the neural network architecture is able to process various inputs (protein sequences, position specific substitution matrices, hidden Markov Model sequence profiles) of various sizes (pg. 2844 Fig 1 and right col para 3, multi-scale). In addition, Heffernan et al. specifically utilizes long short-term memory (LTSM) cells due to their ability to learn both distant and close intra-sequence dependencies, which helps to analyze the nodes in conjunction with their neighbors (pg. 2844 left col para 1, neighborhood-based). Lastly, Heffernan et al. discloses that the neural network utilizes multiple loss function (cross-entropy loss, and square loss) to solve the classification problem of secondary structures and regression problems of ASA, backbone angles, HSE, and CN (pg. 2844 right col para 2, multi-goal).
Regarding claim 12, Heffernan et al. teaches that 5789 proteins were used as input into the neural network for prediction of secondary structures (pg. 2842 right col para 3, inputting amino acid code into said neural network…wherein said amino acid code is used to produce said predicted structures). In addition, Heffernan et al. discloses that structural features predicted from the first iteration with these proteins are then also added as another initial input for a second round of hopefully more accurate predictions (pg. 2844 right col para 3, with said structural feature set).
With respect to claim 13, Heffernan et al. discloses that the secondary features are produced by predicting/stimulating molecular interactions between the amino acids of the protein sequence based on physio-chemical properties (PP) of amino acids (pg. 2844 right col para 3, using a molecular stimulation program to produce said predicted structures). On another note, applicant also provided an example of a well-known molecular stimulation program in the Specification of the instant application, called the Chemistry at Harvard Macromolecular Mechanics program (CHARMM) which has been well established for the use of protein structure stimulation (Specification [0066]).
Concerning claim 15, Heffernan et al. teaches the prediction of several structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting a plurality of features of said predicted structure selected from the group consisting of dihedral angles, B-factor, solvent-accessible surface area, long-range angles, and short-range angles).
Regarding claim 16, Heffernan et al. teaches the use of long short-term memory (LTSM) bidirectional recurrent neural networks (BRNNs) to predict protein secondary structures, backbone angles, contact numbers and solvent accessibility (pg. 2842 Abstract: Results). Heffernan et al. teaches that the neural network model was trained on a system utilizing Google’s open-source TensorFlow library, a software library that has a memory, with a Nvidia GeForce Titan X GPU, a processor in used to carry out the functions from the memory (pg. 2844 right col para 1, a computer program product, said computer product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions executable by a processor to cause said processor to perform a function). Heffernan et al. also discloses that the neural network was trained using Adam optimization, a method for efficient stochastic optimization that is able to compute adaptive learning rates and requires little hyperparameter tuning (pg. 2844 left col para 2, training a neural network). In addition, Heffernan et al. adds that the neural network is able to predict structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting structural feature sets with said neural network). Furthermore, Heffernan et al. highlights that the neural network utilizes an iterative process, that can take these structural features to further predict secondary structures (SS) of the protein (pg. 2844 right col para 3, producing predicted structures with said neural network using said structural feature sets).
With respect to claim 17, Heffernan et al. teaches that the neural network architecture is able to process various inputs (protein sequences, position specific substitution matrices, hidden Markov Model sequence profiles) of various sizes (pg. 2844 Fig 1 and right col para 3, said neural network is a multi-scale). In addition, Heffernan et al. specifically utilizes long short-term memory (LTSM) cells due to their ability to learn both distant and close intra-sequence dependencies, which helps to analyze the nodes in conjunction with their neighbors (pg. 2844 left col para 1, neighborhood-based neural network).
Concerning claim 19, Heffernan et al. discloses that the secondary features are produced by predicting/stimulating molecular interactions between the amino acids of the protein sequence based on physio-chemical properties (PP) of amino acids (pg. 2844 right col para 3, using a molecular stimulation program to produce said predicted structures). On another note, applicant also provided an example of a well-known molecular stimulation program in the Specification of the instant application, called the Chemistry at Harvard Macromolecular Mechanics program (CHARMM) which has been well established for the use of protein structure stimulation (Specification [0066]).
Regarding claim 20, Heffernan et al. teaches the prediction of several structural features about a protein such as accessible surface area (ASA), backbone angles, half sphere exposure (HSE), and contact number (CN) (pg. 2844 right col para 3, predicting a plurality of features of said predicted structure selected from the group consisting of dihedral angles, B-factor, solvent-accessible surface area, long-range angles, and short-range angles).
However, Heffernan et al. fails to teach the converting the predicted structures into predicted graphs with predicted edges (claims 1, 8, 16). Due to this, Heffernan et al. also does not teach comparing predicted graphs to training graphs and predicted edges to training edges to obtain a comparison (claims 1, 8, 16) along with labeling differences between the graphs and edges (claims 6, 14). Hence, Heffernan et al. does not disclose training a model, like a variational graph autoencoder (claims 3, 11, 18), with said comparison and constructing a graph with the neural network using a node feature set before reducing missing edges in the graph with the model (claims 1, 8, 16). Yet, these limitations were known in the art at the time of the effective filing date of the invention as taught by Guo et al. below.
With regards to claim 1, Guo et al. discloses the use of a neural network-based graph variational autoencoders to represent protein structures as “contact graphs” so that rich, local, and distal constraints are able to be captured and interpreted (pg. 1 Abstract). Specifically, Guo et al. teaches how contact maps (Cartesian coordinates of central alpha carbon atoms) can be extracted from protein structures, and converted to contact graphs (pg. 3 left col para 2 and pg. 2 Fig. 1, converting predicted structures into predicted graphs). Guo et al. further discloses how the process maps each amino acid as a node in the contact graph, and each connection between amino acids as an edge (pg. 3 left col para 3, predicted graphs with predicted edges). In addition, Guo et al. teaches how each node has a node attribute matrix (F), that stores not just the identity of the amino acids, but also their position-specific scoring matrix profile, their solvent accessibility, and secondary structures (pg. 3 left col para 4, a node feature set). This node attribute matrix (F) is then inputted into the convolution layer, based on a classical graph convolution network, to later generate contact graphs (pg. 3 right col para 3, constructing a graph with said neural network using a node feature set). Guo et al. further teaches that a training dataset is used to train the graph variational autoencoder, which based on the training contact graph, produces new generated contact maps with new node and edge attribute data (pg. 2 right col para 3, pg. 2 Fig. 1). Guo et al. also discloses that the model then compares the generated contact maps to the training graphs through a reconstruction objective, as seen in equation (1), where the first item on the right-hand-side is the expected reconstruction loss of the generated contact graphs, log p (E, F|Z) (pg. 3 left col para 6). In this case, E and F are the edge and node attribute tensors respectively, and Z is the stochastic latent variable vectors, so log p (E, F|Z) represents how well the generated edges and nodes from the latent representation match the true ones (pg. 3 left col para 6, comparing predicted graphs to training graphs and predicted edges to training edges to obtain a comparison). Guo et al. specifically discloses that the node attribution matrix F is fixed, hence the learning is primarily focuses on reconstruction of the edge matrix E (pg. 3 right col para 4, reducing missing edges in said graph with said model). Guo et al. even goes one step further by evaluating the quality of the generated dataset, by using the training dataset as a reference and calculating dissimilarity (Bhattacharyya distance, BD) on variables such as density, number of edges, average degree coefficient, and transitivity (pg. 4 right col para 4-6). Guo et al. demonstrated that the lowest BDs are obtained by their contact graph variation autoencoder (CO-VAE) and disentanglement enhanced contact graph variation autoencoder (DECO-VAE), i.e. the generated distribution is most similar to the distribution in the training dataset (pg. 5 left col para 1).
Concerning claim 3, Guo et al. teaches the use of graph variational autoencoders to generate protein structures (pg. 1 Abstract, said model is a variational graph auto-encoder).
Regarding claim 6, Guo et al. teaches the evaluation of the generated contact map distributions to the training dataset by looking at the differences between generated contact maps and the training contact maps (pg. 7 right col para 2). Guo et al. specifically looks at the difference in labeled variables such as NAT-C which represents the percentage of native contacts/edges in a contact graph, and NONNAT-C which represents the non-native contacts/edges found on a contact graph (pg. 6 right col para 1, labeling differences between predicted edges and training edges). In addition to these metrics, Guo et al. also looks at other labeled differences between the generated contact maps and the training contact maps such as density, number of edges, average degree coefficient, and transitivity (pg. 4 right col para 4-6, labelling differences between predicted graphs and training graphs).
With respect to claim 8, Guo et al. discloses the use of a neural network-based graph variational autoencoders to represent protein structures as “contact graphs” so that rich, local, and distal constraints are able to be captured and interpreted (pg. 1 Abstract). Specifically, Guo et al. teaches how contact maps (Cartesian coordinates of central alpha carbon atoms) can be extracted from protein structures, and converted to contact graphs (pg. 3 left col para 2 and pg. 2 Fig. 1, converting predicted structures into predicted graphs). Guo et al. further discloses how the process maps each amino acid as a node in the contact graph, and each connection between amino acids as an edge (pg. 3 left col para 3, predicted graphs with predicted edges). In addition, Guo et al. teaches how each node has a node attribute matrix (F), that stores not just the identity of the amino acids, but also their position-specific scoring matrix profile, their solvent accessibility, and secondary structures (pg. 3 left col para 4, a node feature set). This node attribute matrix (F) is then inputted into the convolution layer, based on a classical graph convolution network, to later generate contact graphs (pg. 3 right col para 3, constructing a graph with said neural network using a node feature set). Guo et al. further teaches that a training dataset is used to train the graph variational autoencoder, which based on the training contact graph, produces new generated contact maps with new node and edge attribute data (pg. 2 right col para 3, pg. 2 Fig. 1). Guo et al. also discloses that the model then compares the generated contact maps to the training graphs through a reconstruction objective, as seen in equation (1), where the first item on the right-hand-side is the expected reconstruction loss of the generated contact graphs, log p (E, F|Z) (pg. 3 left col para 6). In this case, E and F are the edge and node attribute tensors respectively, and Z is the stochastic latent variable vectors, so log p (E, F|Z) represents how well the generated edges and nodes from the latent representation match the true ones (pg. 3 left col para 6, comparing predicted graphs to training graphs and predicted edges to training edges to obtain a comparison). Guo et al. specifically discloses that the node attribution matrix F is fixed, hence the learning is primarily focuses on reconstruction of the edge matrix E (pg. 3 right col para 4, reducing missing edges in said graph with said model). Guo et al. even goes one step further by evaluating the quality of the generated dataset, by using the training dataset as a reference and calculating dissimilarity (Bhattacharyya distance, BD) on variables such as density, number of edges, average degree coefficient, and transitivity (pg. 4 right col para 4-6). Guo et al. demonstrated that the lowest BDs are obtained by their contact graph variation autoencoder (CO-VAE) and disentanglement enhanced contact graph variation autoencoder (DECO-VAE), i.e. the generated distribution is most similar to the distribution in the training dataset (pg. 5 left col para 1).
Concerning claim 11, Guo et al. teaches the use of graph variational autoencoders to generate protein structures (pg. 1 Abstract, said model is a variational graph auto-encoder).
Regarding claim 14, Guo et al. teaches the evaluation of the generated contact map distributions to the training dataset by looking at the differences between generated contact maps and the training contact maps (pg. 7 right col para 2). Guo et al. specifically looks at the difference in labeled variables such as NAT-C which represents the percentage of native contacts/edges in a contact graph, and NONNAT-C which represents the non-native contacts/edges found on a contact graph (pg. 6 right col para 1, labeling differences between predicted edges and training edges). In addition to these metrics, Guo et al. also looks at other labeled differences between the generated contact maps and the training contact maps such as density, number of edges, average degree coefficient, and transitivity (pg. 4 right col para 4-6, labelling differences between predicted graphs and training graphs).
With respect to claim 16, Guo et al. discloses the use of a neural network-based graph variational autoencoders to represent protein structures as “contact graphs” so that rich, local, and distal constraints are able to be captured and interpreted (pg. 1 Abstract). Specifically, Guo et al. teaches how contact maps (Cartesian coordinates of central alpha carbon atoms) can be extracted from protein structures, and converted to contact graphs (pg. 3 left col para 2 and pg. 2 Fig. 1, converting predicted structures into predicted graphs). Guo et al. further discloses how the process maps each amino acid as a node in the contact graph, and each connection between amino acids as an edge (pg. 3 left col para 3, predicted graphs with predicted edges). In addition, Guo et al. teaches how each node has a node attribute matrix (F), that stores not just the identity of the amino acids, but also their position-specific scoring matrix profile, their solvent accessibility, and secondary structures (pg. 3 left col para 4, a node feature set). This node attribute matrix (F) is then inputted into the convolution layer, based on a classical graph convolution network, to later generate contact graphs (pg. 3 right col para 3, constructing a graph with said neural network using a node feature set). Guo et al. further teaches that a training dataset is used to train the graph variational autoencoder, which based on the training contact graph, produces new generated contact maps with new node and edge attribute data (pg. 2 right col para 3, pg. 2 Fig. 1). Guo et al. also discloses that the model then compares the generated contact maps to the training graphs through a reconstruction objective, as seen in equation (1), where the first item on the right-hand-side is the expected reconstruction loss of the generated contact graphs, log p (E, F|Z) (pg. 3 left col para 6). In this case, E and F are the edge and node attribute tensors respectively, and Z is the stochastic latent variable vectors, so log p (E, F|Z) represents how well the generated edges and nodes from the latent representation match the true ones (pg. 3 left col para 6, comparing predicted graphs to training graphs and predicted edges to training edges to obtain a comparison). Guo et al. specifically discloses that the node attribution matrix F is fixed, hence the learning is primarily focuses on reconstruction of the edge matrix E (pg. 3 right col para 4, reducing missing edges in said graph with said model). Guo et al. even goes one step further by evaluating the quality of the generated dataset, by using the training dataset as a reference and calculating dissimilarity (Bhattacharyya distance, BD) on variables such as density, number of edges, average degree coefficient, and transitivity (pg. 4 right col para 4-6). Guo et al. demonstrated that the lowest BDs are obtained by their contact graph variation autoencoder (CO-VAE) and disentanglement enhanced contact graph variation autoencoder (DECO-VAE), i.e. the generated distribution is most similar to the distribution in the training dataset (pg. 5 left col para 1).
Concerning claim 18, Guo et al. teaches the use of graph variational autoencoders to generate protein structures (pg. 1 Abstract, said model is a variational graph auto-encoder).
It would have been prima facie obvious to one of ordinary skill in the art before the effect filing date of the invention to combine the LSTM BR neural network of Heffernan et al. with the variational graph autoencoder of Guo et al. to successfully produce a multi-level neural network that would be able to produce a distribution of possible protein structures for protein sequences with unknown structures. One of ordinary skill in the art would have been motivated to add the variational graph autoencoder of Guo et al. to complement the structural prediction neural network of Heffernan et al. to generate a variety of possible protein structures based on known data as proteins are complex molecules that can adopt various shapes based on conditions, and having a distribution of possible structures is a better representation. In addition, one of skill in the art before the effective filing date of the claimed invention would have a reasonable expectation of success at combining the two neural network based models as neural networks were well established both in the field of protein prediction and graphical representation. Furthermore, the combination of neural network layers through stacking or merging is a well-known and established technique in neural network model building.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WENYU YANG whose telephone number is (571)272-0035. The examiner can normally be reached 8:00am - 5:00 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.
/W.Y./Examiner, Art Unit 1685
/OLIVIA M. WISE/Supervisory Patent Examiner, Art Unit 1685
Prosecution Insights