DETAILED ACTION
Applicant’s response, filed 5/12/2026, has been fully considered. The following rejections and/or objections are either reiterated or newly applied. They constitute the complete set presently being applied to the instant application.
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 1-23 are currently pending and examined on the merits.
Priority
The instant application claims priority to U.S. Provisional Application 63/208,904 filed on 9 June 2021. At this point in examination, the effective filing date of claims 1-23 is 9 June 2021.
Information Disclosure Statement
The information disclosure statements (IDS) submitted on 30 November 2022 and 8 December 2022 are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statements have been considered by the examiner.
The listing of references in the specification is not a proper information disclosure statement. 37 CFR 1.98(b) requires a list of all patents, publications, or other information submitted for consideration by the Office, and MPEP § 609.04(a) states, "the list may not be incorporated into the specification but must be submitted in a separate paper." Therefore, unless the references have been cited by the examiner on form PTO-892, they have not been considered.
Drawings
The objections to Figures 1B and 1E are withdrawn, in view of the claim amendments.
New ground for objection was necessitated by amendment to the drawings, received 5/12/2026. The drawings are objected to as failing to comply with 37 CFR 1.84(p)(4) because of the following:
Reference character “190” in Figure 1E has been used to designate “Internal Coordinates L (
b
i
,
a
i
,
d
i
)” and “Generated Conformation
C
K
”
Corrected drawing sheets in compliance with 37 CFR 1.121(d) are required in reply to the Office action to avoid abandonment of the application. Any amended replacement drawing sheet should include all of the figures appearing on the immediate prior version of the sheet, even if only one figure is being amended. Each drawing sheet submitted after the filing date of an application must be labeled in the top margin as either “Replacement Sheet” or “New Sheet” pursuant to 37 CFR 1.121(d). If the changes are not accepted by the examiner, the applicant will be notified and informed of any required corrective action in the next Office action. The objection to the drawings will not be held in abeyance.
Specification
The objection to the specification is withdrawn, in view of the amendments to the specification submitted 5/12/2026.
Claim Objections
The objections to claims 10-11 are withdrawn, in view of the claim amendments filed 5/12/2026.
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 1-23 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. The claims recite: (a) mathematical concepts, (e.g., mathematical relationships, formulas or equations, mathematical calculations); and (b) mental processes, i.e., concepts performed in the human mind, (e.g., observation, evaluation, judgement, opinion). Any newly recited portions herein are necessitated by claim amendment.
Subject matter eligibility evaluation in accordance with MPEP 2106:
Eligibility Step 1: Claims 1-21 are directed to a method (process) that uses molecular graph data for a molecule as input and provides a report that includes at least one selected conformation for the molecule as an output. Claim 22 is directed to one or more non-transitory computer-readable storage media (machine). Claim 23 is directed to a system (machine). Therefore, these claims are encompassed by the categories of statutory subject matter, and thus satisfy the subject matter eligibility requirements under Step 1.
[Step 1: YES]
Eligibility Step 2A: First, it is determined in Prong One whether a claim recites a judicial exception, and if so, then it is determined in Prong Two whether the recited judicial exception is integrated into a practical application of that exception.
Eligibility Step 2A, Prong One: In determining whether a claim is directed to a judicial exception, examination is performed that analyzes whether the claim recites a judicial exception, i.e., whether a law of nature, natural phenomenon, or abstract idea is set forth described in the claim.
Claims 1-3, 5-14, and 16-23 recite the following steps which fall within the mental processes and/or mathematical concepts groups of abstract ideas, as noted below.
Independent claims 1, 22, and 23 further recite:
generating a plurality of conformations for the molecule with the machine learning platform (i.e., mental processes, mathematical concepts);
wherein generating the plurality of conformations comprises, for each generated conformation, operating a molecular graph generator to obtain the molecule graph data and latent code data to construct the generated conformation of the molecule with a set of internal coordinates (i.e., mental processes);
converting the internal coordinates into initial Cartesian coordinates (i.e., mental processes, mathematical concepts);
predicting target inter-atomic distance information and edge coefficients for graph edges including virtual edges between second, third, and fourth neighboring nodes (i.e., mental processes);
iteratively refining the initial Cartesian coordinates by performing a distance geometry optimization process using the predicted target inter-atomic distance information and the predicted edge coefficients to correct local distance geometry of at least one molecular substructure (i.e., mental processes, mathematical concepts);
produce refined Cartesian coordinates defining the generated conformation (i.e., mental processes);
selecting at least one conformation for the molecule based on at least one parameter related to molecular conformations (i.e., mental processes);
preparing a report that includes the selected at least one conformation for the molecule (i.e., mental processes).
Dependent claim 2 further recites:
further comprising the machine learning platform predicting lengths for each molecular graph bond of the molecule for each conformation (i.e., mental processes).
Dependent claim 3 further recites:
providing the at least one selected conformation of the molecule that has a lower energy compared to other generated conformations of the molecule (i.e., mental processes).
Dependent claim 5 further recites:
inputting molecule graph data of the molecule and a set of latent vectors into a generator (i.e., mental processes);
outputting a conformation of the molecule as a sequence of internal coordinates (i.e., mental processes);
distinguishing real conformations from generated conformations with predicted energy differences (i.e., mental processes);
mapping conformations into latent space (i.e., mental processes);
conforming the latent space to be similar to a prior distribution (i.e., mental processes).
Dependent claim 6 further recites:
further comprising a conformation generation (i.e., mental processes);
generating internal coordinates of a first conformation from the molecule graph data and noise (i.e., mental processes);
predicting bond lengths and a bond-wise loss function weight of the first conformation (i.e., mental processes);
converting the internal coordinates to Cartesian coordinates for the first conformation (i.e., mental processes, mathematical concepts);
computing the Cartesian coordinates for unit direction and unit normal vectors for the conformation (i.e., mental processes, mathematical concepts);
modulating bond length of the conformation to the predicted bond lengths (i.e., mental processes).
Dependent claim 7 further recites:
representing the molecular graph by nodes and edge feature sets (i.e., mental processes);
extending the molecular graph with auxiliary nodes and edges to make a proposed generative model (i.e., mental processes);
introducing virtual edges between second, third, and/or fourth neighboring nodes (i.e., mental processes);
setting each node to include a description of: atom type, charge, and chiral tag (i.e., mental processes);
setting each edge feature to include a first graph subset that has chemical bond type and bond stereochemistry (i.e., mental processes);
setting each edge feature to include a second graph subset that has a spanning tree traversal process and having defining edge features to be in the spanning tree and information regarding whether a source node appears earlier in the spanning tree traversal process than a destination node (i.e., mental processes).
Dependent claim 8 further recites:
estimating one or more of the following conformation properties for each generated molecule: asphericity, eccentricity, inertial shape factor, two normalized principal moments ratios, three principal moments of inertia, gyration radius or spherocity index (i.e., mental processes).
Dependent claim 9 further recites:
operating a molecular graph generator to obtain molecular graph data and latent code data to construct a conformation of a molecule with a set of internal coordinates, to convert the internal coordinates into Cartesian coordinates, and perform at least one optimization to correct local distance geometry of at least one molecular substructure (i.e., mental processes, mathematical concepts);
operating a conformation discriminator to distinguish between real conformations of a molecule from synthetic conformations of the molecule (i.e., mental processes);
operating a latent variables discriminator to map conformations into the latent space and to make the latent space similar to a normal prior distribution (i.e., mental processes).
Dependent claim 10 further recites:
determining a reconstruction loss between an original conformation of a molecule compared to a reconstructed conformation of the molecule by adversarial analysis between the molecular graph generator against the conformation discriminator and latent variables discriminator (i.e., mental processes, mathematical concepts).
Dependent claim 11 further recites:
constructing a first conformation having a rotation and translation invariant representation (i.e., mental processes);
predicting distances between neighboring atoms of the first conformation (i.e., mental processes)
Dependent claim 12 further recites:
considering a potential energy of a plurality of conformations (i.e., mental processes);
selecting physically plausible conformations based on the potential energy of each selected conformation (i.e., mental processes).
Dependent claim 13 further recites:
modeling at least one provided conformation of the molecule with a biological target (i.e., mental processes);
determine whether or not the at least one provided conformation modulates the biological target (i.e., mental processes).
Dependent claim 14 further recites:
operating a graph convolution block (i.e., mental processes);
update representations of nodes and edges of a molecule graph data (i.e., mental processes);
update node states (i.e., mental processes);
update hidden states of edges (i.e., mental processes).
Dependent claim 16 further recites:
encoding discrete features of nodes and edge features with embedding layers, each edge feature including a first graph subset that has a chemical bond type and bond stereochemistry (i.e., mental processes);
applying a sequence of graph convolution blocks to the discrete features to obtain an embedding of the molecular graph of the molecule (i.e., mental processes).
Dependent claim 17 further recites:
further comprising an encoder: obtaining a description of a conformation from molecular graph data of a molecule (i.e., mental processes);
further comprising an encoder: transforming the conformation with a sequence of graph convolution blocks to obtain node-wise latent codes (i.e., mental processes).
Dependent claim 18 further recites:
further comprising a latent variables discriminator: distinguishing generated latent codes of real conformations from noise (i.e., mental processes);
further comprising a latent variables discriminator: determining: node-wise latent codes being independent of each other; and node-wise latent codes following the normal distribution (i.e., mental processes).
Dependent claim 19 further recites:
further comprising a conformation discriminator: controlling quality of generated objects by: assessing a likelihood of one or more conformations (i.e., mental processes);
further comprising a conformation discriminator: controlling quality of generated objects by: determining a quality of the one or more conformations based on potential energy estimations (i.e., mental processes).
Dependent claim 20 further recites:
further comprising a conformation discriminator: obtaining one aggregated value for the whole molecular conformation (i.e., mental processes, mathematical concepts).
Dependent claim 21 further recites:
determining an ability to synthesize generated molecular conformation (i.e., mental processes).
The abstract ideas recited in the claims are evaluated under the broadest reasonable interpretation (BRI) of the claim limitations when read in light of and consistent with the specification. As noted in the foregoing section, the claims are determined to contain limitations that can practically be performed in the human mind with the aid of a pencil and paper, and therefore recite judicial exceptions from the mental process grouping of abstract ideas. Additionally, the recited limitations that are identified as judicial exceptions from the mathematical concepts grouping of abstract ideas are abstract ideas irrespective of whether or not the limitations are practical to perform in the human mind.
Therefore, claims 1-3, 5-14, and 16-23 recite an abstract idea.
[Step 2A, Prong One: YES]
Eligibility Step 2A, Prong Two: In determining whether a claim is directed to a judicial exception, further examination is performed that analyzes if the claim recites additional elements that, when examined as a whole, integrates the judicial exception(s) into a practical application (MPEP 2106.04(d)). A claim that integrates a judicial exception into a practical application will apply, rely on, or use the judicial exception in a manner that imposes a meaningful limit on the judicial exception. The claimed additional elements are analyzed to determine if the abstract idea is integrated into a practical application (MPEP 2106.04(d)(I); MPEP 2106.05(a-h)). If the claim contains no additional elements beyond the abstract idea, the claim fails to integrate the abstract idea into a practical application (MPEP 2106.04(d)(III)).
The judicial exceptions identified in Eligibility Step 2A, Prong One are not integrated into a practical application because of the reasons noted below.
Claims 2-3, 5-8, 10-14, 16-19, and 21 do not recite any elements in addition to the judicial exception, and thus are part of the judicial exception.
Claim 1 recites inputting the molecule graph data into a machine learning platform. The machine learning platform is obtaining data for further analysis, which is considered a well-understood, routine, and conventional activity. Data gathering steps are extra-solution activity as they collect the data needed to carry out the JE. It does not impose any meaningful limitation on the JE or how the JE is performed (MPEP 2106.04/.05, citing Intellectual Ventures LLC v. Symantee Corp, McRO, TLI communications, OIP Techs. Inc. v. Amason.com Inc., Electric Power Group LLC v. Alstrom S.A.). Therefore, the claimed additional element does not integrate the abstract ideas into a practical application.
Claim 1 recites obtaining molecule graph data for a molecule. Data gathering steps are not an abstract idea, they are extra-solution activity, as they collect the data needed to carry out the JE. The data gathering does not impose any meaningful limitation on the JE, or how the JE is performed. The additional limitation (data gathering) must have more than a nominal or insignificant relationship to the identified judicial exception. (MPEP 2106.04/.05, citing Intellectual Ventures LLC v. Symantee Corp, McRO, TLI communications, OIP Techs. Inc. v. Amason.com Inc., Electric Power Group LLC v. Alstrom S.A.).
Claim 9 recites operating a stochastic encoder to construct an irredundant latent space of latent data of input molecules and prevent mode collapse. The limitation recites using a stochastic encoder, which provide nothing more than mere instructions to implement an abstract idea on a generic computer. See MPEP 2106.05(f). Therefore, the claimed additional element does not integrate the abstract ideas into a practical application.
Claim 15 recites inputting condition data into the machine learning platform. The machine learning platform is obtaining data for further analysis, which is considered a well-understood, routine, and conventional activity. Data gathering steps are extra-solution activity as they collect the data needed to carry out the JE. It does not impose any meaningful limitation on the JE or how the JE is performed (MPEP 2106.04/.05, citing Intellectual Ventures LLC v. Symantee Corp, McRO, TLI communications, OIP Techs. Inc. v. Amason.com Inc., Electric Power Group LLC v. Alstrom S.A.). Therefore, the claimed additional element does not integrate the abstract ideas into a practical application.
Claim 20 recites passing molecular graph embeddings through a plurality of SchNet layers to obtain node representations. The limitation of passing molecular graph embeddings through SchNet layers provides nothing more than mere instructions to implement an abstract idea on a generic computer. See MPEP 2106.05(f). Therefore, the claimed additional element does not integrate the abstract ideas into a practical application.
Claims 22 and 23 recite the additional non-abstract element (EIA) of a general-purpose computer system or parts thereof:
One or more non-transitory computer readable media storing instructions (claim 22);
A computer system comprising: one or more processors; and one or more non-transitory computer readable media storing instructions (claim 23).
The EIA do not provide any details of how specific structures of the computer elements are used to implement the JE. The claims require nothing more than a general-purpose computer to perform the functions that constitute the judicial exceptions. The computer elements of the claims do not provide improvements to the functioning of the computer itself (as in DDR Holdings, LLC v. Hotels.com LP); they do not provide improvements to any other technology or technical field (as in Diamond v. Diehr); nor do they utilize a particular machine (as in Eibel Process Co. v. Minn. & Ont. Paper Co.). Hence, these are mere instructions to apply the JE using a computer, and therefore the claim does not recite integrate that JE into a practical application.
Thus, the additionally recited elements merely invoke a computer as a tool, and/or amount to insignificant extra-solution data gathering activity, and as such, when all limitations in claims 1-23 have been considered as a whole, the claims are deemed to not recite any additional elements that would integrate a judicial exception into a practical application. Claims 1, 9, 15, 20, and 22-23 contain additional elements that would not integrate a judicial exception into a practical application and are further probed for inventive concept in Step 2B.
[Step 2A, Prong Two: NO]
Eligibility Step 2B: Because the claims recite an abstract idea, and do not integrate that abstract idea into a practical application, the claims are probed for a specific inventive concept. The judicial exception alone cannot provide that inventive concept or practical application (MPEP 2106.05). Identifying whether the additional elements beyond the abstract idea amount to such an inventive concept requires considering the additional elements individually and in combination to determine if they amount to significantly more than the judicial exception (MPEP 2106.05A i-vi).
The claims do not include any additional elements that are sufficient to amount to significantly more than the judicial exception(s) because of the reasons noted below.
With respect to the recited obtaining molecule graph data for a molecule (claim 1): The limitations identified above as non-abstract elements (EIA) related to data gathering do not rise to the level of significantly more than the judicial exception. Activities such as data gathering do not improve the functioning of a computer, or comprise an improvement to any other technical field. The limitations do not require or set forth a particular machine, they do not affect a transformation of matter, nor do they provide an unconventional step (citing McRO and Trading Technologies Int’l v. IBG). Data gathering steps constitute a general link to a technological environment. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception are insufficient to provide significantly more (as discussed in Alice Corp.,).
The additional element of inputting the molecule graph data into a machine learning platform (claim 1) is conventional. Evidence for conventionality is shown by Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022. Ganea et al. reviews “We propose GeoMol – an end-to-end, non-autoregressive and SE(3)-invariant machine learning approach to generate distributions of low-energy molecular 3D conformers.” (Abstract, lines 6-8). Also, further reviews “We generate a representative set of low-energy 3D conformers from the input molecular graph.” (Figure 1). This shows that GeoMol is the machine learning platform that takes in molecule graph data, which makes it a conventional element in the art.
The additional element of operating a stochastic encoder to construct an irredundant latent space of latent data of input molecules and prevent mode collapse (claim 9) is conventional. Evidence for conventionality is shown by Winter et al. (CoRR, 2021, 1-6), as provided in the IDS filed 11/30/2022. Winter et al. reviews “The overall goal of the proposed model is to find functions
f
Θ
and
g
Θ
that map a conformation
Ξ
G
of a molecule
G
to and from a fixed-sized latent representation
z
Ξ
∈
R
F
z
, respectively.” (Page 2, Section “2.2 Conformation Autoencoder”, lines 1-2). Also, further reviews “by training a probabilistic model on a large dataset of molecular conformations, we demonstrate how our model can be used to generate diverse sets of energetically favorable conformations for a given molecule.” (Abstract, lines 5-8). This shows that an encoder model is used to construct an irredundant latent space of latent representation data and prevents mode collapse because the model generates diverse sets of conformations. Therefore, operating a stochastic encoder is a conventional element in the art.
The additional element of inputting condition data into the machine learning platform (claim 15) is conventional. Evidence for conventionality is shown by Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022. Ganea et al. reviews “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc. For each molecular graph G, we have a variable-size set of low-energy ground truth 3D conformers
{
C
l
*
}
l
that we predict with a model
{
C
k
}
k
≝
ζ
(
G
)
.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This shows that the machine learning platform takes in condition data, where the condition data includes at least one conformation of the molecule. Therefore, inputting condition data into a machine learning platform is a conventional element in the art.
The additional element of passing molecular graph embeddings through a plurality of SchNet layers to obtain node representations (claim 20) is conventional. Evidence for conventionality is shown by Schutt et al. (The Journal of Chemical Physics, 2018, 148(24), 1-11), as provided in the IDS filed 11/30/2022. Schutt et al. reviews “SchNet is a variant of the earlier proposed Deep Tensor Neural Network (DTNN) and therefore shares a number of their essential building blocks. Among these are atom embeddings interaction refinements and atom-wise energy contributions. At each layer, the atomistic system is represented atom-wise being refined using pairwise interactions with the surrounding atoms. In the DTNN framework, interactions are modeled by tensor layers, i.e., atom representations and interatomic distances are combined using a parameter tensor.” (Page 2, Section “II. Method”, lines 1-10). This shows that molecular graph embeddings are refined by passing through SchNet layers to obtain node or atom representations, making it a conventional element in the art.
With respect to claims 22 and 23: The limitations identified above as non-abstract elements (EIA) related to general-purpose computer systems do not rise to the level of significantly more than the judicial exception. These elements do not improve the functioning of the computer itself, or comprise an improvement to any other technical field (Trading Technologies Int’l v. IBG, TLI Communications). They do not require or set forth a particular machine (Ultramercial v. Hulu, LLC., Alice Corp. Pty. Ltd v. CLS Bank Int’l), they do not affect a transformation of matter, nor do they provide an unconventional step. Simply appending well-understood, routine, conventional activities previously known to the industry, specified at a high level of generality, to the judicial exception are insufficient to provide significantly more (as discussed in Alice Corp., CyberSource v. Retail Decisions, Parker v. Flook, Versata Development Group v. SAP America).
[Step 2B: NO]
Therefore, claims 1-23 are patent ineligible under 35 U.S.C. § 101.
Response to Arguments
Applicant’s arguments, see pages 12-17, filed 5/12/2026, with respect to claims 1-23, have been fully considered but they are not persuasive.
With respect to the Applicant’s argument that the amended claims 1 and 22-23 recite limitations that define a particular machine-learning-based molecular conformation generation technique and not an observation, judgment, or opinion that can practically be performed mentally (pg. 15, para. 3 of Applicant’s Remarks), this argument is not persuasive. The amended claims recite using molecular graph information and latent code information to construct a molecular conformation with internal coordinates, converting internal coordinates to Cartesian coordinates, predicting inter-atomic distance information and edge coefficients, refining Cartesian coordinates by performing a distance geometry optimization process, and producing the refined Cartesian coordinates, all of which are steps that can be performed mentally or by using pen and paper. Furthermore, the courts do not distinguish between mental processes that are performed entirely in the human mind and mental processes that require a human to use a physical aid (e.g., pen and paper or a slide rule) to perform the claim limitation. See, e.g., Benson, 409 U.S. at 67, 65, 175 USPQ at 674-75, 674 (noting that the claimed "conversion of binary- coded decimal numerals to pure binary numerals can be done mentally,” i.e., "as a person would do it by head and hand"). See MPEP 2106.04(a)(2)(III).
With respect to the Applicant’s argument that the amended claims 1 and 22-23 use the alleged mathematical operations in a specific technological process for generating physically meaningful molecular conformations from molecule graph data and therefore impose meaningful limits on any alleged abstract idea by applying the claimed operations to a concrete technological task: generating and refining three-dimensional molecular conformations (pg. 15-16, para. 4-5 of Applicant’s Remarks), this argument is not persuasive. In terms of Step 2A, Prong Two, the elements recited in the amended claims do not recite any additional elements in addition to the judicial exception and therefore fail to integrate the abstract ideas into a practical application. See MPEP 2106.04.II.A.2.
With respect to the Applicant’s argument that claims 22 and 23 do not rely on the recitation of computer-readable media, processors, or a computer system alone for eligibility, but instead use computer implementation as part of a concrete technological process for generating molecular conformations and not as a mere instruction to apply an abstract idea on a generic computer (pg. 16-17, para. 3 of Applicant’s Remarks), this argument is not persuasive. As stated above in Step 2A, Prong Two, the claims require nothing more than a general-purpose computer to perform the functions that constitute the judicial exceptions. The computer elements of the claims do not provide improvements to the functioning of the computer itself (as in DDR Holdings, LLC v. Hotels.com LP); they do not provide improvements to any other technology or technical field (as in Diamond v. Diehr); nor do they utilize a particular machine (as in Eibel Process Co. v. Minn. & Ont. Paper Co.). See MPEP 2106.04(a)(I).
With respect to the Applicant’s argument that under Step 2B, the amended claims recite significantly more than any alleged judicial exception (pg. 17, para. 2 of Applicant’s Remarks), this argument is not persuasive. Step 2B evaluates whether additional elements amount to an inventive concept. However, the elements recited in the amended claims do not recite any additional elements in addition to the judicial exception. Therefore, conventionality cannot be determined. See MPEP 2106.05(d).
Therefore, the rejection to claims 1-23 under 35 U.S.C. § 101 is maintained with modifications as necessitated by amendment of the claims, filed 5/12/2026.
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.
The rejection of claims 1-4, 11-13, 15-16, 19, and 22-23 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022, is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claims 5-6, 9, 14, and 18 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Winter et al. (CoRR, 2021, 1-6), as provided in the IDS filed 11/30/2022, is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 7 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Xu et al. (International Conference on Learning Representations, 2021, 1-17), as provided in the IDS filed 11/30/2022, referred to as Xu [A], is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 8 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Landrum G., rdkit.Chem.Descriptors3D module, RDKit: Open-Source Cheminformatics and Machine Learning, https://www.rdkit.org/docs/source/rdkit.Chem.Descriptors3D.html#, 2018, accessed on 6 January 2026, 1-3), is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 10 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Xu et al. (Proceedings of the 38th International Conference on Machine Learning, 2021, 139, 1-13), as provided in the IDS filed 11/30/2022, referred to as Xu [B], is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 17 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Mansimov et al. (Nature, Scientific Reports, 2019, 9(1), 1-15), as provided in the IDS filed 11/30/2022, is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 20 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Schutt et al. (The Journal of Chemical Physics, 2018, 148(24), 1-11), as provided in the IDS filed 11/30/2022, is withdrawn in view of the claim amendments filed 5/12/2026.
The rejection of claim 21 under 35 U.S.C. 103 as being unpatentable over Ganea et al. (NeurIPS, 2021, 1-13) as applied to claims 1-4, 11-13, 15-16, 19, and 22-23 above, in view of Shi et al. (International Conference on Machine Learning, PMLR, 2020, 1-10), is withdrawn in view of the claim amendments filed 5/12/2026.
Claims 1-2, 4, 14, 18, and 22-23 are rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6), as provided in the IDS filed 11/30/2022, in view of Xu et al. (International Conference on Learning Representations, 2021, 1-17), as provided in the IDS filed 11/30/2022; refer to as Xu [A]. This rejection is newly recited and necessitated by claim amendment.
With respect to claims 1, 22, and 23:
Claim 22 recites one or more non-transitory computer readable media storing instructions. Claim 23 recites a computer system comprising one or more processors and one or more non-transitory computer readable media storing instructions.
Broadly claiming an automated means to replace a manual function to accomplish the same result does not distinguish over the prior art. See Leapfrog Enters., Inc. v. Fisher-Price, Inc., 485 F .3d 1157, 1161, 82 USPQ2d 1687, 1691 (Fed. Cir. 2007) (“Accommodating a prior art mechanical device that accomplishes [a desired] goal to modern electronics would have been reasonably obvious to one of ordinary skill in designing children’s learning devices. Applying modern electronics to older mechanical devices has been commonplace in recent years.”); In re Venner, 262 F. 2d 91, 95, 120 USPQ 193, 194 (CCPA 1958); see also MPEP § 2144.04. Furthermore, implementing a known function on a computer has been deemed obvious to one of ordinary skill in the art if the automation of the known function on a general purpose computer is nothing more than the predictable use of prior art elements according to their established functions. KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 417, 82 USPQ2d 1385, 1396 (2007); see also MPEP § 2143, Exemplary Rationales D and F. Likewise, it has been found to be obvious to adapt an existing process to incorporate Internet and Web browser technologies for communicating and displaying information because these technologies had become commonplace for those functions. Muniauction, Inc. v. Thomson Corp., 532 F.3d 1318, 1326-27, 87 USPQ2d 1350, 1357 (Fed. Cir. 2008).
Regarding the recited obtaining molecule graph data for a molecule, Winter et al. discloses utilizing an internal coordinate representation, also known as the Z-matrix, to represent a three-dimensional arrangement of atoms of a molecule (pg. 2, Section "2.1 Representing Molecular Conformations"). In this notation, a molecules spatial arrangement (conformation) Ξ is defined by the set of distances
D
=
{
d
1
,
…
,
d
N
D
} between bonded atoms (bond length), the angles
Ф
=
{
ϕ
1
,
…
,
ϕ
N
Ф
}
of three connected atoms (bond angles), and the torsion angles (dihedral angles)
Ψ
=
{
ψ
1
,
…
,
ψ
N
Ψ
}
of three consecutive bonds. This teaches molecule graph data for a molecule.
Regarding the recited inputting the molecule graph data into a machine learning platform, Winter et al. discloses a conformation autoencoder that comprises a conformation-independent part comprising a Graph Neural Network that utilizes the molecular graph to extract node-level features for a given molecule, and a conformation-dependent part that utilizes these features either to encode the internal coordinates of a specific molecular conformation into a latent representation (conformation embedding) or to reconstruct a conformation (pg. 2-3, Sections "2.2 Conformation Autoencoder" and "2.2.1 Molecular Graph Encoder"). This teaches inputting molecular graph data into a machine learning platform.
Regarding the recited generating a plurality of conformations for the molecule with the machine learning platform, wherein the plurality conformations are specific to the molecule, each conformation having internal coordinates defining positions of atoms of the molecule, Winter et al. discloses a conformation encoder that learns to extract a latent representation
z
Ξ
of a set of internal coordinates Ξ for a given molecule using a permutation invariant function
f
Θ
parameterized by a neural network and conditioned on node embeddings
H
=
{
h
1
,
…
h
N
}
extracted from the molecular graph encoder (pg. 2, Figure 1; pg. 3-4, Section "2.2.2 Conformation Encoder"; pg. 4, Figure 2). This teaches encoding conformations for a molecule with machine learning where each conformation contains internal coordinates defining positions of atoms of for bond length, bond angle, and dihedral angle of the molecule.
Regarding the recited wherein generating the plurality of conformations comprises, for each generated conformation, operating a molecular graph generator to obtain the molecule graph data and latent code data to construct the generated conformation of the molecule with a set of internal coordinates, Winter et al. discloses a conformation encoder that learns to extract a latent representation
z
Ξ
of a set of internal coordinates Ξ for a given molecule using a permutation invariant function
f
Θ
parameterized by a neural network and conditioned on node embeddings
H
=
{
h
1
,
…
h
N
}
extracted from the molecular graph encoder (pg. 2, Figure 1; pg. 3-4, Section "2.2.2 Conformation Encoder"; pg. 4, Figure 2). Also, further discloses the molecular graph encoder extracting node-level representations of a molecular graph (pg. 3, Section "2.2.1 Molecular Graph Encoder"). This teaches operating a molecular graph generator to obtain molecule graph data and latent code data, which are used to construct molecule conformations with a set of internal coordinates.
Regarding the recited converting the internal coordinates into initial Cartesian coordinates, Winter et al. discloses a conformation decoder that reconstructs a molecular conformation back from its latent representation using three additional neural networks for each type of internal coordinate, respectively (pg. 2, Figure 1; pg. 4, Section "2.2.3 Conformation Decoder"). Because the conformation encoder initially takes in internal coordinates to produce latent representations, the conformation decoder takes in the latent representations to reproduce internal coordinates. Furthermore, internal coordinate representations are invariant to rotations and rigid translations and can always be transformed to and from Cartesian coordinates (pg. 2, Section "2.1 Representing Molecular Conformations"). This teaches that internal coordinates can be converted into initial Cartesian coordinates.
Regarding the recited predicting target inter-atomic distance information and edge coefficients for graph edges including virtual edges between second, third, and fourth neighboring nodes, Winter et al. discloses a molecular graph encoder that utilizes a Graph Neural Network (GNN) to extract node-level representations of a conformation-independent molecular graph, which is defined as an undirected Graph
G
=
(
V
,
E
)
with vertices or nodes
v
i
∈
V
and edges
e
i
j
=
(
v
i
,
v
j
)
∈
E
connecting
v
i
and
v
j
, where nodes
v
i
∈
R
F
v
represent atoms with features and edges
e
i
j
∈
R
F
e
represent bonds between atoms (pg. 2-3, Section "2.2 Conformation Autoencoder"; pg. 3, Section "2.2.1 Molecular Graph Encoder"). Also, further discloses representing a three-dimensional arrangement of atoms of a molecule using the internal coordinate representations, where a molecular spatial arrangement (conformation) Ξ is defined by the set of distances
D
=
{
d
1
,
…
,
d
N
D
} between bonded atoms (bond length), the angles
Ф
=
{
ϕ
1
,
…
,
ϕ
N
Ф
}
of three connected atoms (bond angles), and the torsion angles (dihedral angles)
Ψ
=
{
ψ
1
,
…
,
ψ
N
Ψ
}
of three consecutive bonds (pg. 2, Section "2.1 Representing Molecular Conformations"; pg. 2, Figure 1). This teaches target inter-atomic distance information as bond lengths, as well as a molecular graph comprising nodes and edges, teaching edge coefficients for graph edges.
Winter et al. does not disclose including virtual edges between second, third, and fourth neighboring nodes.
However, Xu [A] discloses conformation generation, which includes pre-processing molecular graphs by extending auxiliary edges or connecting atoms that are 2 or 3 hops away with virtual bonds, labeled differently from the real bonds of the original graph (pg. 3, Section "B Data Preprocess"). This teaches including virtual edges between neighboring nodes under certain conditions.
It would have been prima facie obvious to one of ordinary skill in the art to combine the inter-atomic distance information and edge coefficients disclosed by Winter et al. with adding virtual edges disclosed by Xu [A]. One would be motivated to combine inter-atomic distance information and edge coefficients with virtual edges because Xu [A] discloses that the extra edges from virtual bonds contribute to reducing the degrees of freedom in the 3D coordinates, with the edges between second neighbors helping to fix the angles between atoms, and those between third neighbors fixing dihedral angles (pg. 3, Section "B Data Preprocess"). This means including virtual edges between neighboring nodes serves as an improvement to the method of molecular conformation generation by reducing the number of parameters in the 3D reconstruction of the generated molecular graphs. There is a likelihood of success, since both methods generate molecular conformations and are well known in the field of computational chemistry.
Winter et al. does not disclose iteratively refining the initial Cartesian coordinates by performing a distance geometry optimization process using the predicted target inter-atomic distance information and the predicted edge coefficients to correct local distance geometry of at least one molecular substructure.
However, Xu [A] discloses parameterizing the conditional distribution of distances
p
θ
(
d
|
G
)
with the continuous normalizing flow, named Conditional Graph Continuous Flow (CGCF), which includes a dynamic
f
θ
implemented by Message Passing Neural Networks (MPNN) that takes node attributes, edge attributes, and bond lengths as input to compute node and edge embeddings fed into another neural network to calculate how each distance is changed (pg. 4-5, Section “3.2 Flow-based Generative Model”, para. 1). Also, further discloses converting the calculated pairwise distances into 3D structures by defining conformations as a conditional distribution, which uses parameters that control the variance of desired Cartesian coordinates over all edges (pg. 5, Section “3.2 Flow-based Generative Model”, para. 2-3). Xu [A] discloses employing an optimization procedure such as stochastic gradient descent to search for realistic conformations R with local maximum probability using the conditional distribution (pg. 5, Section "3.2 Flow-based Generative Model"; pg. 5-6, Section "3.4 Sampling"). This teaches refining Cartesian coordinates by performing a distance geometry optimization process using inter-atomic distance information and edge coefficients to correct distance geometry of molecular substructures.
Winter et al. does not disclose produce refined Cartesian coordinates defining the generated conformation.
However, Xu [A] discloses employing an optimization procedure such as stochastic gradient descent to search for realistic conformations R with local maximum probability, which comprises parameters that control the variance of desired Cartesian coordinates (pg. 5, Section "3.2 Flow-based Generative Model"; pg. 5-6, Section "3.4 Sampling"). This teaches producing conformations with refined Cartesian coordinates.
Winter et al. does not disclose selecting at least one conformation for the molecule based on at least one parameter related to molecular conformations.
However, Xu [A] discloses visualizations of generated conformations based on different molecular graphs and their atomic pairwise distances (pg. 4, Section "3.1 Overview", para. 1, lines 1-4; pg. 7, Figure 2; pg. 15, Section "G More Generated Samples"). This teaches a selection of conformations for molecules based on at least one parameter related to molecular conformations.
Winter et al. does not disclose preparing a report that includes the selected at least one conformation for the molecule.
However, Xu [A] discloses visualizations of generated conformations based on different molecular graphs (pg. 7, Figure 2; pg. 15, Section "G More Generated Samples"). This teaches an output of conformations for molecules.
With respect to claim 2:
Xu [A] does not disclose further comprising the machine learning platform predicting lengths for each molecular graph bond of the molecule for each conformation.
However, Winter et al. discloses a conformation decoder that reconstructs a molecular conformation back from its latent representation, using three additional neural networks to predict bond lengths, bond angles, and dihedral angles, respectively (pg. 2, Figure 1; pg. 4, Section "2.2.3 Conformation Decoder"). This teaches predicting bond lengths for each molecular graph bond of the molecule for each conformation.
With respect to claim 4:
Winter et al. does not disclose the report including a conformation space that is comprised of a plurality of overlaid selected conformations for the molecule.
However, Xu [A] discloses visualizations of generated conformations based on different molecular graphs (pg. 7, Figure 2; pg. 15, Section "G More Generated Samples"). This teaches an output of a plurality of conformations for molecules in a conformation space.
With respect to claim 14:
Xu [A] does not disclose operating a graph convolution block.
However, Winter et al. discloses “We utilize a Graph Neural Network to extract a node-level representation of a molecular graph.” (Page 3, Section “2.2.1 Molecular Graph Encoder”, paragraph 2, lines 1-2). This suggests that the Graph Neural Network is the graph convolution block.
Xu [A] does not disclose update representations of nodes and edges of a molecule graph data.
However, Winter et al. discloses “Given a molecular graph with initial node and edge features defined by the atoms and bonds of the molecule, a GNN iteratively updates node embeddings by aggregating localized information of connected nodes respectively.” (Page 3, Section “2.2.1 Molecular Graph Encoder”, paragraph 2, lines 2-4). Also, further discloses “To incorporate edge attributes
e
i
,
j
(bond-type information) in the model we also utilize the so-called edge-conditioned graph convolution (EConv) layer (Simonovsky and Komodakis, 2017), defined by the following update rule:
h
i
'
=
Θ
h
i
+
∑
j
∈
N
(
i
)
h
j
∙
f
Θ
(
e
i
,
j
)
” (Page 3, Section “2.2.1 Molecular Graph Encoder”, paragraph 4, lines 1-3). This suggests that node and edge representations are updated.
Xu [A] does not disclose update node states.
However, Winter et al. discloses “we use the Graph Attention Network (GAT) (Veličković et al., 2017) framework which updates the node embeddings
h
i
” (Page 3, Section “2.2.1 Molecular Graph Encoder”, paragraph 3, lines 1-3). This suggests that node states are updated.
Xu [A] does not disclose update hidden states of edges.
However, Winter et al. discloses “To incorporate edge attributes
e
i
,
j
(bond-type information) in the model we also utilize the so-called edge-conditioned graph convolution (EConv) layer (Simonovsky and Komodakis, 2017), defined by the following update rule:
h
i
'
=
Θ
h
i
+
∑
j
∈
N
(
i
)
h
j
∙
f
Θ
(
e
i
,
j
)
” (Page 3, Section “2.2.1 Molecular Graph Encoder”, paragraph 4, lines 1-3). This suggests that hidden states of edges are also updated.
With respect to claim 18:
Xu [A] does not disclose further comprising a latent variables discriminator.
However, Winter et al. discloses “Our proposed model converts the discrete spatial arrangements of atoms in a given molecular graph (conformation) into and from a continuous fixed-sized latent representation. We demonstrate that in this latent representation, similar conformations cluster together while distinct conformations split apart.” (Abstract, lines 1-5). This describes a latent variables discriminator that converts molecular graph data into latent variable representations before distinguishing conformations by similarity.
Xu [A] does not disclose distinguishing generated latent codes of real conformations from noise.
However, Winter et al. discloses “Our proposed model converts the discrete spatial arrangements of atoms in a given molecular graph (conformation) into and from a continuous fixed-sized latent representation. We demonstrate that in this latent representation, similar conformations cluster together while distinct conformations split apart.” (Abstract, lines 1-5). This suggests distinguishing latent representations of real conformations from noise, where real conformations cluster together while noise split apart.
Xu [A] does not disclose determining node-wise latent codes being independent of each other.
However, Winter et al. discloses “Our proposed model converts the discrete spatial arrangements of atoms in a given molecular graph (conformation) into and from a continuous fixed-sized latent representation. We demonstrate that in this latent representation, similar conformations cluster together while distinct conformations split apart.” (Abstract, lines 1-5). This suggests determining node-wise latent representations being independent of each other based on how similar their conformations are.
Xu [A] does not disclose determining node-wise latent codes following the normal distribution.
However, Winter et al. discloses “Our proposed model converts the discrete spatial arrangements of atoms in a given molecular graph (conformation) into and from a continuous fixed-sized latent representation. We demonstrate that in this latent representation, similar conformations cluster together while distinct conformations split apart.” (Abstract, lines 1-5). Also, further discloses “Our proposed model can easily be extended to a probabilistic generative model by employing the ideas from Kingma and Welling (2013), effectively defining the model as a variational auto encoder.” (Page 3, Section “2.2 Conformation Autoencoder”, lines 9-11). This suggests operating a latent variables discriminator that maps conformations into latent representations. The model can be extended to a probabilistic generative model that can conform the latent representations to be similar to a normal distribution, therefore determining latent codes following the normal distribution.
Claims 3, 5, 7, 9, 11-13, 15-16, and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022. This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 3:
Winter et al. and Xu [A] do not disclose wherein the at least one parameter related to molecular conformations includes an energy of each conformation, the method comprising providing the at least one selected conformation of the molecule that has a lower energy compared to other generated conformations of the molecule.
However, Ganea et al. discloses “In this work, we assume that the low-energy states are implicitly defined by the given dataset, i.e., our training data consist of molecular graphs and corresponding sets of energetically favorable 3D conformations. Low-energy structures are the most stable configurations and, thus, expected to be observed most often experimentally.” (Page 1, Section “Problem & importance”, paragraph 1, lines 7-12). This suggests that the conformations selected are based on low-energy states and only low energy structures are considered compared to other generated conformations.
It would have been prima facie obvious to one of ordinary skill in the art to modify the molecular conformation generation method disclosed by Winter et al. and Xu [A] to incorporate molecular conformation selection by energy disclosed by Ganea et al. One would be motivated to incorporate selecting conformations based on energy because the conformer generation model GeoMol disclosed by Ganea et al. is the fastest method from the considered baselines (pg. 9, Figure 7; pg. 10, para. 4). Therefore, incorporating molecular conformation selection by energy into the molecular conformation generation method will improve its speed. There is a likelihood of success, since all methods generate molecular conformations and are well known in the field of computational chemistry.
With respect to claim 5:
Winter et al. and Xu [A] do not disclose inputting molecule graph data of the molecule and a set of latent vectors into a generator.
However, Ganea et al. discloses “We generate a representative set of low-energy 3D conformers from the input molecular graph.” (Figure 1). Also, further discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests that the generator takes in molecular graph data and latent vectors encoded using node and edge features to generate low-energy 3D conformers.
Xu [A] and Ganea et al. do not disclose outputting a conformation of the molecule as a sequence of internal coordinates.
However, Winter et al. discloses “The conformation-independent part comprises a Graph Neural Network utilizing the molecular graph to extract node-level features for a given molecule. The conformation-dependent part utilizes these extracted node-level features either to encode the internal coordinates of a specific molecular conformation into a latent representation (conformation embedding) or to reconstruct a conformation by predicting the internal coordinates of sets of connected atoms, given their respective node features and a conformation embedding.” (Page 3, Section “2.2 Conformation Autoencoder”, lines 2-7). This suggests that node-level features from molecular graph data are used to encode internal coordinates of a molecular conformation.
Winter et al. and Xu [A] do not disclose distinguishing real conformations from generated conformations with predicted energy differences.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). Also, further discloses “The energy values from GeoMol are the lowest among the ML methods, indicating greater stability of generated conformers, especially for the druglike molecules.” (Supplementary Information, Page 20, Appendix K Energy calculations, lines 1-3). This suggests that real conformations are distinguished from generated conformations based on their low energy values.
Xu [A] and Ganea et al. do not disclose mapping conformations into latent space.
However, Winter et al. discloses “The overall goal of the proposed model is to find functions
f
Θ
and
g
Θ
that map a conformation
Ξ
G
of a molecule
G
to and from a fixed-sized latent representation
z
Ξ
∈
R
F
z
, respectively.” (Page 2, Section “2.2 Conformation Autoencoder”, lines 1-2). This suggests functions that can map a conformation to a fixed-sized latent representation or latent space.
Xu [A] and Ganea et al. do not disclose conforming the latent space to be similar to a prior distribution.
However, Winter et al. discloses “Our proposed model can easily be extended to a probabilistic generative model by employing the ideas from Kingma and Welling (2013), effectively defining the model as a variational auto encoder.” (Page 3, Section “2.2 Conformation Autoencoder”, lines 9-11). This suggests that the autoencoder model for molecular conformations can be extended to a probabilistic generative model that can conform latent space to be similar to a prior distribution.
With respect to claim 7:
Winter et al. and Xu [A] do not disclose representing the molecular graph by nodes and edge feature sets.
However, Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests the molecular graph is represented by nodes and edge feature sets.
Winter et al. and Ganea et al. do not disclose extending the molecular graph with auxiliary nodes and edges to make a proposed generative model.
However, Xu [A] discloses “We also follow the previous work (Simm & Hernández-Lobato, 2020) to expand the molecular graph with auxiliary bonds, which is elaborated in Appendix B. For the molecular 3D representation, each atom in
V
is assigned with a 3D position vector
r
∈
R
3
. We denote
d
u
v
=
r
u
-
r
v
2
as the Euclidean distance between the
u
t
h
and
v
t
h
atom. Therefore, we can represent all the positions
{
r
v
}
v
∈
V
as a matrix
R
∈
R
|
V
|
×
3
and all the distances between connected nodes
{
d
u
v
}
e
u
v
∈
ε
as a vector
d
∈
R
|
ε
|
.” (Page 3, Section “2.1 Problem Definition”, paragraph 1, lines 5-10). Also, further discloses “Since the bonds existing in the molecular graph are not sufficient to characterize a conformation, we pre-process the graphs by extending auxiliary edges” (Page 13, Appendix B Data Preprocess, lines 3-4). This suggests the molecular graph is extended with auxiliary edges as well as auxiliary nodes.
Winter et al. and Ganea et al. do not disclose introducing virtual edges between second, third, and/or fourth neighboring nodes.
However, Xu [A] discloses “Specifically, the atoms that are 2 or 3 hops away are connected with virtual bonds, labeled differently from the real bonds of the original graph.” (Page 13, Appendix B Data Preprocess, lines 4-6). This suggests introducing virtual edges between nodes that are 2 or 3 hops away, which are the second or third neighboring nodes.
Winter et al. and Xu [A] do not disclose setting each node to include a description of: atom type, charge, and chiral tag.
However, Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests that each node is set to include a description of atom type, charge, and chiral tag.
Winter et al. and Xu [A] do not disclose setting each edge feature to include a first graph subset that has chemical bond type and bond stereochemistry.
However, Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests that each edge feature includes a graph subset representing chemical bond type and bond stereochemistry.
Winter et al. and Xu [A] do not disclose setting each edge feature to include a second graph subset that has a spanning tree traversal process and having defining edge features to be in the spanning tree and information regarding whether a source node appears earlier in the spanning tree traversal process than a destination node.
However, Ganea et al. discloses “We first fix one BFS traversal of the graph and assemble the LS in the order given by this traversal. If the current node is not part of a cycle, then we can perform the assembling as described before. However, when we first encounter a node that is part of a cycle of nodes
X
1
,
X
2
,
…
,
X
n
, we will jointly compute all the 3D coordinates of this cycle and attach the entire ring structure to the current partial conformer.” (Supplementary Information, Page 16, Appendix E Details of the full conformer assembly procedure at test time, paragraph 4, lines 3-7). Also, further discloses “The key step is assembling two sets of 3D points: the set containing the LS of node X (and possibly other atom coordinates added in previous steps), denoted as
S
X
∶
=
{
p
1
,
…
,
p
n
}
⊂
R
3
, and the set containing the LS of node Y, denoted as
S
Y
∶
=
{
q
1
,
…
,
q
m
}
⊂
R
3
. Assume that X and Y are connected by a bond/edge.” (Supplementary Information, Page 16, Appendix E Details of the full conformer assembly procedure at test time, paragraph 2, lines 1-4). Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests that the Breadth First Search (BFS) spanning tree traversal process for reconstructing a molecular conformation comprises of defining edge features while connecting the nodes. This process also considers whether the first encounter of a node appears before a cycle of nodes, including the destination node.
With respect to claim 9:
Xu [A] and Ganea et al. do not disclose operating a molecular graph generator to obtain molecular graph data and latent code data to construct a conformation of a molecule with a set of internal coordinates, to convert the internal coordinates into Cartesian coordinates.
However, Winter et al. discloses “The conformation-independent part comprises a Graph Neural Network utilizing the molecular graph to extract node-level features for a given molecule. The conformation-dependent part utilizes these extracted node-level features either to encode the internal coordinates of a specific molecular conformation into a latent representation (conformation embedding) or to reconstruct a conformation by predicting the internal coordinates of sets of connected atoms, given their respective node features and a conformation embedding.” (Page 3, Section “2.2 Conformation Autoencoder”, lines 2-7). Also, further discloses “This representation is invariant to rotations and rigid translations and can always be transformed to and from Cartesian coordinates.” (Page 2, Section “2.1 Representing Molecular Conformations”, lines 8-9). This describes operating a molecular graph generator, or Graph Neural Network, that takes in molecular graph data and latent code data encoded using node-level features to generate a molecular conformation with internal coordinates. This also suggests that internal coordinates can be converted into Cartesian coordinates.
Winter et al. and Ganea et al. do not disclose perform at least one optimization to correct local distance geometry of at least one molecular substructure.
However, Xu [A] discloses parameterizing the conditional distribution of distances
p
θ
(
d
|
G
)
with the continuous normalizing flow, named Conditional Graph Continuous Flow (CGCF), which includes a dynamic
f
θ
implemented by Message Passing Neural Networks (MPNN) that takes node attributes, edge attributes, and bond lengths as input to compute node and edge embeddings fed into another neural network to calculate how each distance is changed (pg. 4-5, Section “3.2 Flow-based Generative Model”, para. 1). Also, further discloses converting the calculated pairwise distances into 3D structures by defining conformations as a conditional distribution, which uses parameters that control the variance of desired Cartesian coordinates over all edges (pg. 5, Section “3.2 Flow-based Generative Model”, para. 2-3). Xu [A] discloses employing an optimization procedure such as stochastic gradient descent to search for realistic conformations R with local maximum probability using the conditional distribution (pg. 5, Section "3.2 Flow-based Generative Model"; pg. 5-6, Section "3.4 Sampling"). This teaches performing a distance geometry optimization using inter-atomic distance information and edge coefficients to correct distance geometry of molecular substructures.
Winter et al. and Xu [A] do not disclose operating a conformation discriminator to distinguish between real conformations of a molecule from synthetic conformations of the molecule.
However, Ganea et al. discloses “In this work, we assume that the low-energy states are implicitly defined by the given dataset, i.e., our training data consist of molecular graphs and corresponding sets of energetically favorable 3D conformations. Low-energy structures are the most stable configurations and, thus, expected to be observed most often experimentally.” (Page 1, Section “Problem & importance”, paragraph 1, lines 7-12). This suggests that a conformation discriminator distinguishes between real conformations from synthetic conformations based on low-energy states.
Xu [A] and Ganea et al. do not disclose operating a stochastic encoder to construct an irredundant latent space of latent data of input molecules and prevent mode collapse.
However, Winter et al. discloses “The overall goal of the proposed model is to find functions
f
Θ
and
g
Θ
that map a conformation
Ξ
G
of a molecule
G
to and from a fixed-sized latent representation
z
Ξ
∈
R
F
z
, respectively.” (Page 2, Section “2.2 Conformation Autoencoder”, lines 1-2). Also, further discloses “by training a probabilistic model on a large dataset of molecular conformations, we demonstrate how our model can be used to generate diverse sets of energetically favorable conformations for a given molecule.” (Abstract, lines 5-8). This describes operating an encoder to construct an irredundant latent space of latent representation data. This also suggests preventing mode collapse because the model generates diverse sets of conformations.
Xu [A] and Ganea et al. do not disclose operating a latent variables discriminator to map conformations into the latent space and to make the latent space similar to a normal prior distribution.
However, Winter et al. discloses “Our proposed model converts the discrete spatial arrangements of atoms in a given molecular graph (conformation) into and from a continuous fixed-sized latent representation. We demonstrate that in this latent representation, similar conformations cluster together while distinct conformations split apart.” (Abstract, lines 1-5). Also, further discloses “Our proposed model can easily be extended to a probabilistic generative model by employing the ideas from Kingma and Welling (2013), effectively defining the model as a variational auto encoder.” (Page 3, Section “2.2 Conformation Autoencoder”, lines 9-11). This suggests operating a latent variables discriminator that maps conformations into latent space. The model can be extended to a probabilistic generative model that can conform latent space to be similar to a normal prior distribution.
With respect to claim 11:
Winter et al. and Xu [A] do not disclose constructing a first conformation having a rotation and translation invariant representation.
However, Ganea et al. discloses “It models conformers in an SE(3)-invariant (translation/rotation) manner by design.” (Page 3, Section “Our key contributions & model in a nutshell”, second bullet, line 1). This suggests modeling a conformation having a rotation and translation invariant representation.
Winter et al. and Xu [A] do not disclose predicting distances between neighboring atoms of the first conformation.
However, Ganea et al. discloses “It explicitly models and predicts essential molecular geometry elements: torsion angles and local 3D structures (bond distances and bond angles adjacent to each atom).” (Page 3, Section “Our key contributions & model in a nutshell”, third bullet, lines 1-2). This suggests predicting bond distances between atoms of a conformation, which is an essential molecular geometry element.
With respect to claim 12:
Winter et al. and Xu [A] do not disclose considering a potential energy of a plurality of conformations.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). This suggests that potential energy of a plurality of generated conformers is considered.
Winter et al. and Xu [A] do not disclose selecting physically plausible conformations based on the potential energy of each selected conformation.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). Also, further discloses “The energy values from GeoMol are the lowest among the ML methods, indicating greater stability of generated conformers, especially for the druglike molecules.” (Supplementary Information, Page 20, Appendix K Energy calculations, lines 1-3). This suggests that generated conformers are selected based on the potential energy values of each conformation.
With respect to claim 13:
Winter et al. and Xu [A] do not disclose modeling at least one provided conformation of the molecule with a biological target.
However, Ganea et al. discloses “We expect that such differentiable structure generators will significantly impact small molecule conformer generation along with many related applications (e.g., protein-ligand binding), thus speeding up areas such as drug discovery.” (Page 10, Section “Conclusion”, lines 2-5). This implies that GeoMol can assist in modeling a conformation with a biological target for applications such as protein-ligand binding.
Winter et al. and Xu [A] do not disclose determine whether or not the at least one provided conformation modulates the biological target.
However, Ganea et al. discloses “We expect that such differentiable structure generators will significantly impact small molecule conformer generation along with many related applications (e.g., protein-ligand binding), thus speeding up areas such as drug discovery.” (Page 10, Section “Conclusion”, lines 2-5). This suggests that the model of a conformation generated by GeoMol can be used to determine whether or not it modulates a biological target, which speeds up drug discovery.
With respect to claim 15:
Winter et al. and Xu [A] do not disclose inputting condition data into the machine learning platform, wherein the condition data is at least one conformation of the molecule.
However, Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc. For each molecular graph G, we have a variable-size set of low-energy ground truth 3D conformers
{
C
l
*
}
l
that we predict with a model
{
C
k
}
k
≝
ζ
(
G
)
.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). This suggests inputting condition data into the machine learning platform, where the condition data includes at least one conformation of the molecule.
With respect to claim 16:
Winter et al. and Xu [A] do not disclose encoding discrete features of nodes and edge features with embedding layers, each edge feature including a first graph subset that has a chemical bond type and bond stereochemistry.
However, Ganea et al. discloses “Our input is any molecular graph
G
=
(
V
,
E
)
with node and edge features,
x
v
∈
R
f
,
∀
v
∈
V
and
e
u
,
v
∈
R
f
,
∀
(
u
,
v
)
∈
E
representing atom types, formal charges, bond types, etc.” (Page 4, Section “Problem setup & notations”, paragraph 1, lines 1-3). Also, further discloses “It explicitly and deterministically distinguishes reflected structures (enantiomers) by solving tetrahedral stereocenters using oriented volumes and local chiral descriptors” (Page 3, Section “Our key contributions & model in a nutshell”, sixth bullet, lines 1-2). This suggests that GeoMol encodes node and edge features with embedding layers, where edge features include bond type and bond stereochemistry.
Winter et al. and Xu [A] do not disclose applying a sequence of graph convolution blocks to the discrete features to obtain an embedding of the molecular graph of the molecule.
However, Ganea et al. discloses “Given an input graph G, an MPNN [Gilmer et al., 2017, Battaglia et al., 2018, Yang et al., 2019] computes node embeddings
h
v
∈
R
d
,
∀
v
∈
V
using
T
layers of iterative message passing” (Page 4, Section “2.2 Message passing neural networks (MPNNs)”, paragraph 1, lines 1-2). Also, further discloses “we also compute a molecular embedding:
h
m
o
l
≝
M
L
P
(
∑
v
∈
V
h
v
)
.” (Pages 4-5, Section “2.2 Message passing neural networks (MPNNs)”, paragraph 1, lines 5-6). This suggests that message passing neural networks (MPNNs) are the graph convolution blocks used to obtain embeddings of the molecular graph of the molecule.
With respect to claim 19:
Winter et al. and Xu [A] do not disclose further comprising a conformation discriminator.
However, Ganea et al. discloses “In this work, we assume that the low-energy states are implicitly defined by the given dataset, i.e., our training data consist of molecular graphs and corresponding sets of energetically favorable 3D conformations. Low-energy structures are the most stable configurations and, thus, expected to be observed most often experimentally.” (Page 1, Section “Problem & importance”, paragraph 1, lines 7-12). This suggests that a conformation discriminator distinguishes between conformations based on low-energy states.
Winter et al. and Xu [A] do not disclose controlling quality of generated objects.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). This suggests controlling the quality of generated conformers based on computed energies.
Winter et al. and Xu [A] do not disclose assessing a likelihood of one or more conformations.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). Gauging the plausibility of generated conformers suggests assessing a likelihood of conformations.
Winter et al. and Xu [A] do not disclose determining a quality of the one or more conformations based on potential energy estimations.
However, Ganea et al. discloses “To gauge the plausibility of generated conformers, we compute the energies as defined by the MMFF force field within RDKit for conformers generated with ML-based methods before force field fine tuning.” (Supplementary Information, Page 19, Appendix K Energy calculations, lines 1-3). This suggests computing energy estimations in order to determine a quality of conformations.
Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022, and Simm et al. (Proceedings of the 37th International Conference on Machine Learning, 2020, 119, 1-19), as provided in the IDS filed 11/30/2022. This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 6:
Winter et al. and Xu [A] do not disclose further comprising a conformation generation.
However, Ganea et al. discloses “We propose GeoMol – an end-to-end, non-autoregressive and SE(3)-invariant machine learning approach to generate distributions of low-energy molecular 3D conformers.” (Abstract, lines 6-8). This describes a machine learning model that generates molecular conformations, which suggests a conformation generation.
Winter et al. and Xu [A] do not disclose predicting bond lengths and a bond-wise loss function weight of the first conformation.
However, Ganea et al. discloses modeling and predicting essential molecular geometry elements, which includes torsion angles and local 3D structures (bond distances and bond angles adjacent to each atom) (pg. 3, Section “Our key contributions & model in a nutshell”, third bullet, lines 1-2). Also, further discloses hyperparameter weights per bond distance in a loss function (pg. 15, Section “D Details of the loss function”). This teaches bond lengths and a bond-wise loss function weight of conformations.
Winter et al., Xu [A], and Ganea et al. do not disclose modulating bond length of the conformation to the predicted bond lengths.
However, Simm et al. discloses using a Euclidean Distance Geometry algorithm to translate a list of lower and upper bounds of distances between pairs of atoms to a set of Cartesian coordinates so that the bounds are fulfilled (pg. 3-4, Section “2.3 Conformation Generation through Euclidean Distance Geometry”). A decoder is used to model the bounds. Together with corresponding chemical elements, a conformation can be obtained. Also, further discloses predicted distances can be transformed into molecular conformations (pg. 2, Section “Method”). This teaches modulating bond lengths of a conformation to use predicted distances.
It would have been prima facie obvious to one of ordinary skill in the art to modify the molecular conformation generation method disclosed by Winter et al. and Xu [A] to incorporate bond length prediction disclosed by Ganea et al. and bond length modulation disclosed by Simm et al. One would be motivated to incorporate bond length prediction and modulation because the conformer generation model GeoMol disclosed by Ganea et al. is the fastest method from the considered baselines (pg. 9, Figure 7; pg. 10, para. 4). Therefore, incorporating bond length prediction into the molecular conformation generation method will improve its speed. The GraphDG model disclosed by Simm et al. generates samples significantly closer to the ground-truth distribution than other methods (pg. 6, Section “5.1. Distributions Over Distances”). This means incorporating bond length modulation from this model into the molecular conformation generation method will improve its accuracy. There is a likelihood of success, since all methods generate molecular conformations and are well known in the field of computational chemistry.
Xu [A], Ganea et al., and Simm et al. do not disclose generating internal coordinates of a first conformation from the molecule graph data and noise.
However, Winter et al. discloses “We utilize the internal coordinate representation, also known as Z-matrix. In this notation, a molecules spatial arrangement (conformation)
Ξ
is defined by the set of distances
D
=
{
d
1
,
…
,
d
N
D
}
between bonded atoms (bond length), the angles
Ф
=
{
Φ
1
,
…
,
Φ
N
Ф
}
of three connected atoms (bond angles) and the torsion angles (dihedral angles)
Ѱ
=
{
ѱ
1
,
…
,
ѱ
N
D
}
of three consecutive bonds (see Figure 1). This representation is invariant to rotations and rigid translations and can always be transformed to and from Cartesian coordinates.” (Page 2, Section “2.1 Representing Molecular Conformations”, lines 3-9). This suggests that internal coordinates are generated from molecule data, including noise.
Xu [A], Ganea et al., and Simm et al. do not disclose converting the internal coordinates to Cartesian coordinates for the first conformation.
However, Winter et al. discloses “We utilize the internal coordinate representation, also known as Z-matrix. In this notation, a molecules spatial arrangement (conformation)
Ξ
is defined by the set of distances
D
=
{
d
1
,
…
,
d
N
D
}
between bonded atoms (bond length), the angles
Ф
=
{
Φ
1
,
…
,
Φ
N
Ф
}
of three connected atoms (bond angles) and the torsion angles (dihedral angles)
Ѱ
=
{
ѱ
1
,
…
,
ѱ
N
D
}
of three consecutive bonds (see Figure 1). This representation is invariant to rotations and rigid translations and can always be transformed to and from Cartesian coordinates.” (Page 2, Section “2.1 Representing Molecular Conformations”, lines 3-9). This suggests that the internal coordinate representation can be converted to Cartesian coordinates.
Xu [A], Ganea et al., and Simm et al. do not disclose computing the Cartesian coordinates for unit direction and unit normal vectors for the conformation.
However, Winter et al. discloses “We utilize the internal coordinate representation, also known as Z-matrix. In this notation, a molecules spatial arrangement (conformation)
Ξ
is defined by the set of distances
D
=
{
d
1
,
…
,
d
N
D
}
between bonded atoms (bond length), the angles
Ф
=
{
Φ
1
,
…
,
Φ
N
Ф
}
of three connected atoms (bond angles) and the torsion angles (dihedral angles)
Ѱ
=
{
ѱ
1
,
…
,
ѱ
N
D
}
of three consecutive bonds (see Figure 1). This representation is invariant to rotations and rigid translations and can always be transformed to and from Cartesian coordinates.” (Page 2, Section “2.1 Representing Molecular Conformations”, lines 3-9). This describes a Z-matrix to Cartesian conversion, which involves computation of Cartesian coordinates for unit direction and unit normal vectors.
Claim 8 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Landrum G., rdkit.Chem.Descriptors3D module, RDKit: Open-Source Cheminformatics and Machine Learning, https://www.rdkit.org/docs/source/rdkit.Chem.Descriptors3D.html#, 2018, accessed on 6 January 2026, 1-3). This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 8:
Winter et al. and Xu [A] do not disclose estimating one or more of the following conformation properties for each generated molecule: asphericity, eccentricity, inertial shape factor, two normalized principal moments ratios, three principal moments of inertia, gyration radius or spherocity index.
However, Landrum discloses a software module that retrieves descriptors from a molecule’s 3D structure, including functions rdkit.Chem.Descriptors3D.Asphericity(*x, **y) for molecular asphericity (Page 1, lines 2-3), rdkit.Chem.Descriptors3D.Eccentricity(*x, **y) for molecular eccentricity (Page 1, lines 14-15), rdkit.Chem.Descriptors3D.InertialShapeFactor(*x, **y) for inertial shape factor (Page 1, lines 25-26), rdkit.Chem.Descriptors3D.NPR1(*x, **y) for normalized principal moments ratio 1 (Pages 1-2, lines 36-37), rdkit.Chem.Descriptors3D.NPR2(*x, **y) for normalized principal moments ratio 2 (Page 2, lines 9-10), rdkit.Chem.Descriptors3D.PMI1(*x, **y) for first (smallest) principal moment of inertia (Page 2, lines 18-19), rdkit.Chem.Descriptors3D.PMI2(*x, **y) for second principal moment of inertia (Page 2, lines 25-26), rdkit.Chem.Descriptors3D.PMI3(*x, **y) for third (largest) principal moment of inertia (Page 2, lines 32-33), rdkit.Chem.Descriptors3D.RadiusOfGyration(*x, **y) for radius of gyration (Page 2, lines 39-40), and rdkit.Chem.Descriptors3D.SpherocityIndex(*x, **y) for molecular spherocity index (Page 3, lines 11-12).
It would have been prima facie obvious to combine the molecular conformation generation method of Winter et al. and Xu [A] with the software module of Landrum. One would be motivated to make this combination because the module disclosed by Landrum is part of RDKit, which is a collection of cheminformatics and machine learning software written in C++ and Python. This RDKit module comprises of functions that compute for conformation properties of a molecule, which is a known technique that is applicable to a known method of molecular conformation generation disclosed by Winter et al. and Xu [A]. There is a likelihood of success, since both the methods and the module are commonly used together to generate molecular conformations and are well known in the field of computational chemistry before the effective filing date of the claimed invention.
Claim 10 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Xu et al. (Proceedings of the 38th International Conference on Machine Learning, 2021, 139, 1-13), as provided in the IDS filed 11/30/2022; refer to as Xu [B]. This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 10:
Winter et al. and Xu [A] do not disclose determining a reconstruction loss between an original conformation of a molecule compared to a reconstructed conformation of the molecule by adversarial analysis between the molecular graph generator against the conformation discriminator and latent variables discriminator.
However, Xu [B] discloses “we model the distribution of conformations
R
conditioning on molecular graph
G
(i.e.
p
(
R
|
G
)
) with a conditional variational autoencoder (CVAE) (Kingma & Welling, 2013), in which a latent variable
z
is introduced to model the uncertainty in molecule conformation generation. The CVAE model includes a prior distribution of latent variable
p
ѱ
(
z
|
G
)
and a decoder
p
Ѳ
(
R
|
z
,
G
)
to capture the conditional distribution of
R
given
z
. During training, we also involve an additional inference model (encoder)
q
Ф
(
z
|
R
,
G
)
. The encoder and decoder are jointly trained to maximize the evidence lower bound (ELBO) of the data log-likelihood” (Page 3, Section “3.1. Overview”, paragraph 1, lines 2-12). Also, further discloses “The ELBO can be interpreted as the sum of the negative reconstruction error
L
r
e
c
o
n
(the first term) and a latent space prior regularizer
L
p
r
i
o
r
(the second term).” (Page 3, Section “3.1. Overview”, paragraph 2, lines 1-3). This suggests conducting an adversarial analysis between the molecular graph generator against the conformation discriminator and latent variables discriminator. The decoder implies a molecular graph generator as it captures the conditional distribution, generating a reconstructed conformation
R
given latent variables discriminator
z
and original conformation
G
. The encoder implies a conformation discriminator as an inference model used to distinguish between conformations. Training both the encoder and decoder maximizes the evidence lower bound (ELBO), which improves adversarial robustness and is part of adversarial analysis. ELBO also comprises the determination of reconstruction loss
L
r
e
c
o
n
.
It would have been prima facie obvious to one of ordinary skill in the art to combine the teachings from Winter et al. and Xu [A] with the teachings of Xu [B]. One would be motivated to make this combination because experimental results demonstrate the superior performance of the ConfVAE framework disclosed by Xu [B] over all state-of-the-art baselines on several standard benchmarks (Page 9, Section “6. Conclusion”, lines 6-9). By incorporating an end-to-end training objective via bilevel optimization, we consistently achieved a better result on all four metrics in coverage and matching scores that measure diversity and quality, respectively (Page 7, Section “Results”, col. 2, lines 6-8). This means combining the molecular conformation generation method with determining reconstruction loss will provide higher quality results. There is a likelihood of success, since all methods generate molecular conformations and are well known in the field of computational chemistry.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Mansimov et al. (Nature, Scientific Reports, 2019, 9(1), 1-15), as provided in the IDS filed 11/30/2022. This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 17:
Xu [A] does not disclose further comprising an encoder.
However, Winter et al. discloses a conformation encoder that learns to extract a latent representation
z
Ξ
of a set of internal coordinates Ξ for a given molecule using a permutation invariant function
f
Θ
parameterized by a neural network and conditioned on node embeddings
H
=
{
h
1
,
…
h
N
}
extracted from the molecular graph encoder (pg. 2, Figure 1; pg. 3-4, Section "2.2.2 Conformation Encoder"; pg. 4, Figure 2). This teaches an encoder.
Xu [A] does not disclose obtaining a description of a conformation from molecular graph data of a molecule.
However, Winter et al. discloses a conformation encoder that learns to extract a latent representation
z
Ξ
of a set of internal coordinates Ξ for a given molecule using a permutation invariant function
f
Θ
parameterized by a neural network and conditioned on node embeddings
H
=
{
h
1
,
…
h
N
}
extracted from the molecular graph encoder (pg. 2, Figure 1; pg. 3-4, Section "2.2.2 Conformation Encoder"; pg. 4, Figure 2). Also, further discloses the molecular graph encoder extracting node-level representations of a molecular graph (pg. 3, Section "2.2.1 Molecular Graph Encoder"). This teaches obtaining node embeddings from molecular graphs of molecules, which are used to construct molecule conformations.
Xu [A] does not disclose transforming the conformation with a sequence of graph convolution blocks to obtain node-wise latent codes.
However, Winter et al. discloses a conformation encoder that learns to extract a latent representation
z
Ξ
of a set of internal coordinates Ξ for a given molecule using a permutation invariant function
f
Θ
parameterized by a neural network and conditioned on node embeddings
H
=
{
h
1
,
…
h
N
}
extracted from the molecular graph encoder (pg. 2, Figure 1; pg. 3-4, Section "2.2.2 Conformation Encoder"; pg. 4, Figure 2). Also, further discloses the molecular graph encoder that uses a Graph Neural Network (GNN) to extract node-level representations of a molecular graph (pg. 3, Section "2.2.1 Molecular Graph Encoder"). This teaches using a molecular graph encoder and a conformation encoder to transform a conformation into node-wise latent representations.
Winter et al. and Xu [A] do not disclose wherein the latent codes are stochastic and sampled with reparameterization from a normal distribution parameterized by outputs of the encoder.
However, Mansimov et al. discloses “With the choice of the Gaussian latent variables
z
i
, we can use the reparameterization trick to compute the gradient of the stochastic approximation to the lower bound in Eq. (7) with respect to all the parameters of the three distributions. This property allows us to train this model on a large dataset using stochastic gradient descent (SGD).” (Page 4, Section “Training the conditional variational graph autoencoder”, paragraph 1, lines 1-4). Also, further discloses “the MPNN outputs a Normal distribution for each latent variable
z
i
.” (Page 4, Section “Posterior parameterization”, paragraph 2, lines 1-2). This suggests that the latent variables are stochastic and modeled as normal distributions to be sampled with reparameterization.
It would have been prima facie obvious to one of ordinary skill in the art to modify the molecular conformation generation method of Winter et al. and Xu [A] to incorporate the teachings of Mansimov et al. One would be motivated to make this modification because compared to other methods, the CVGAE model disclosed by Mansimov et al. always succeeds at generating the specified number of conformations for any of the molecules in the test set (Page 6, Section “Results”, lines 2-3). This means that the model disclosed by Mansimov et al. is consistent in producing desired conformations and therefore incorporating reparameterization in the molecular conformation generation method would provide high chances of success for generating desired conformations. There is a likelihood of success, since all methods generate molecular conformations and are well known in the field of computational chemistry.
Claim 20 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Ganea et al. (NeurIPS, 2021, 1-13), as provided in the IDS filed 11/30/2022, and Schutt et al. (The Journal of Chemical Physics, 2018, 148(24), 1-11), as provided in the IDS filed 11/30/2022. This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 20:
Winter et al. and Xu [A] do not disclose further comprising a conformation discriminator.
However, Ganea et al. discloses “In this work, we assume that the low-energy states are implicitly defined by the given dataset, i.e., our training data consist of molecular graphs and corresponding sets of energetically favorable 3D conformations. Low-energy structures are the most stable configurations and, thus, expected to be observed most often experimentally.” (Page 1, Section “Problem & importance”, paragraph 1, lines 7-12). This suggests that a conformation discriminator distinguishes between conformations based on low-energy states.
Winter et al., Xu [A], and Ganea et al. do not disclose passing molecular graph embeddings through a plurality of SchNet layers to obtain node representations.
However, Schutt et al. discloses “SchNet is a variant of the earlier proposed Deep Tensor Neural Network (DTNN) and therefore shares a number of their essential building blocks. Among these are atom embeddings interaction refinements and atom-wise energy contributions. At each layer, the atomistic system is represented atom-wise being refined using pairwise interactions with the surrounding atoms. In the DTNN framework, interactions are modeled by tensor layers, i.e., atom representations and interatomic distances are combined using a parameter tensor.” (Page 2, Section “II. Method”, lines 1-10). This suggests that molecular graph embeddings are refined through SchNet layers to obtain node or atom representations.
Winter et al., Xu [A], and Ganea et al. do not disclose obtaining one aggregated value for the whole molecular conformation.
However, Schutt et al. discloses “Finally, a given property
P
of a molecule or material is predicted from the obtained atom-wise representations. We compute atom-wise contributions
P
^
i
from the fully connected prediction network (see blue layers in Fig. 1). Depending on whether the property is intensive or extensive, we calculate the final prediction
P
^
by summing or averaging over the atomic contributions, respectively.” (Page 3, Section “E. Property prediction”, paragraph 1, lines 1-7). This suggests that node representations obtained from the SchNet layers are used to compute one aggregated value for the whole molecular conformation.
It would have been prima facie obvious to one of ordinary skill in the art to combine the molecular conformation generation method of Winter et al. and Xu [A] with the conformation discriminator disclosed by Ganea et al. and the SchNet layers disclosed by Schutt et al. One would be motivated to make this combination because the conformer generation model GeoMol disclosed by Ganea et al. is the fastest method from the considered baselines (pg. 9, Figure 7; pg. 10, para. 4). Not only does SchNet disclosed by Schutt et al. yield fast and accurate predictions of molecular properties, but it also allows for examining the learned representation using local chemical potentials (Page 9, Section “Conclusion”, paragraph 2, lines 10-12). Therefore, combining and incorporating both the conformation discriminator and the SchNet layers would improve speed in the molecular conformation generation method. There is a likelihood of success, since molecular conformation generation methods and SchNet are commonly used in molecular discovery and are well known in the field of computational chemistry.
Claim 21 is rejected under 35 U.S.C. 103 as being unpatentable over Winter et al. (CoRR, 2021, 1-6) and Xu et al. (International Conference on Learning Representations, 2021, 1-17), referred to as Xu [A], as applied to claims 1-2, 4, 14, 18, and 22-23 above, in view of Shi et al. (International Conference on Machine Learning, PMLR, 2020, 1-10). This rejection is newly recited and necessitated by claim amendment.
Winter et al. and Xu [A] are applied to claims 1-2, 4, 14, 18, and 22-23 above.
With respect to claim 21:
Winter et al. and Xu [A] do not disclose determining an ability to synthesize generated molecular conformation, wherein the generated molecular conformation has at least one three dimensional restriction.
However, Shi et al. discloses “At the
i
t
h
transformation step, we first calculate the probabilities of all possible actions and sort them, and then select the top
k
ranked valid actions for each candidate graph in
S
i
-
1
,
j
in
S
. Once this is done for
k
graphs in
S
, the top
k
graphs among all the generated
k
2
graphs are then selected as the candidates for the next
(
i
+
1
)
t
h
transformation step. During this beam search, a translation branch will stop if
i
reaches the predefined maximum transformation step or
a
i
1
indicates a termination. In this scenario, the current graph will be added into a set
G
, and the whole beam search terminates once all translation branches stop. When the beam search finishes, the top
k
graphs in
G
, ranked by their likelihoods, will be collected as the final predicted graphs.” (Page 6, Section “3.3.3. Generation”, paragraph 2, lines 3-16). This suggests determining the ability of the top
k
graphs to synthesize through the beam search process. They are ranked by their likelihoods, which implies at least one three dimensional restriction for the generated molecular conformations.
It would have been prima facie obvious to one of ordinary skill in the art to combine the molecular conformation generation method of Winter et al. and Xu [A] with the teachings of Shi et al. One would be motivated to make this combination because experimental results disclosed by Shi et al. show that G2Gs significantly outperforms existing template-fee approaches by up to 63% in terms of the top-1 accuracy and achieves a performance close to that of the state-of-the-art template-based approaches, but does not require domain knowledge and is much more scalable (Abstract, lines 15-21). Therefore, combining determination of an ability to synthesize molecular conformations with the molecular conformation generation method would improve scalability and accuracy. There is a likelihood of success, since molecular conformation generation methods and frameworks for retrosynthesis prediction are commonly used in drug discovery and are well known in the field of computational chemistry.
Response to Arguments
Applicant’s arguments, see pg. 17-22, filed 5/12/2026, with respect to the rejection(s) of claim(s) 1-23 under 35 U.S.C. § 103 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Winter et al..
Conclusion
No claims are allowed.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Jammy Luo whose telephone number is (571)272-2358. The examiner can normally be reached Monday - Friday, 9:00 AM - 5:00 PM 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, Larry D Riggs can be reached at (571)270-3062. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of 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.N.L./Examiner, Art Unit 1686
/LARRY D RIGGS II/Supervisory Patent Examiner, Art Unit 1686
Prosecution Insights