DETAILED ACTION
The amendment filed on 01/22/2026 has been entered and fully considered. Claims 1-3 and 5-31 are pending. Claims 18-31 have been withdrawn from consideration. Claims 1-3 and 5-17 are considered on merits, of which claim 1 is amended.
Response to Amendment
In response to amendment, the examiner modifies rejection over the prior art established in the previous Office action.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Rejections - 35 USC § 103
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
Claim(s) 1-3, 6-7 and 9-17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Mason et al. (WO 2020/208555, IDS) (Mason) in view of Ofer et al. (Computational and Structural Biotechnology Journal, 2021) (Ofer) and Tam et al. (US 2022/0101113) (Tam).
Regarding claim 1, Mason discloses a computing system for identifying protein sequence variants of interest (abstract), the computing system comprising:
one or more processors (Fig. 1, par [51]); and
one or more non-transitory computer-readable media having stored thereon (par [65]):
a machine-learned model trained using training data (Fig. 1, par [4][51]),
wherein the training data includes one or more training protein sequence variants, each having a respective measured binding characteristic representing an ability of each to bind to a corresponding respective binding partner (par [4][70]), and
wherein the machine-learned model is configured to output a predicted protein binding characteristic of an input protein sequence variant (par [49]),
wherein the machine-learned language model is configured to receive, as the input protein sequence variant, a amino-acid sequence that includes (ii) one or more complementarity-determining-region (CDR) delimiting that delimit concatenated CDR segments within the amino-acid sequence (par [6] [83]); and
wherein the training data includes sequences derived from developability screening that includes at least stability, expression level and aggregation propensity (par [51] [63]), and
instructions that, when executed by the one or more processors, cause the computing system to:
process one or more protein sequence variants with the machine-learned model to generate one or more predicted binding characteristics, each corresponding to a respective one of the one or more protein sequence variants (par [49][50]);
analyze the one or more predicted binding characteristics to identify one or more protein sequence variants of interest from among the one or more protein sequence variants, each of the one or more protein sequence variants of interest having a respective one or more desired properties (The present disclosure describes systems and methods to make predictions of protein sequence-phenotype relationships and can be employed for the identification of therapeutic antibodies with one or more desired parameters, such as antigen specificity or affinity)(par [49][50]); and
provide the one or more protein sequence variants of interest as an output (par [49]).
Masom teaches that “The present disclosure describes systems and methods to make predictions of protein sequence-phenotype relationships and can be employed for the identification of therapeutic antibodies with one or more desired parameters, such as antigen specificity or affinity.” (par [0050]). Masom further teaches that “The training data 118 can be a set of variants that is selected by physical screening of a rationally designed library of variants based on a selected parameter (e.g., antigen binding). For example, in some embodiments, the training data includes numerical values. In some embodiments, the numerical values correspond to binding kinetic values for a set of variants. In some embodiments, the numerical values correspond to numerical value results for biophysical assays (e.g., melting temperature for thermal stability, or AC-SINS for solubility). Exemplary methods for generation of the training data is described in further detail (see, e.g., FIG. 4A)” (par [70]). Thus, Masom teaches that wherein the training data includes one or more training protein sequence variants, each having a respective measured binding characteristic representing an ability of each to bind to a corresponding respective binding partner.
Mason broadly discloses the use of machine learning models (one or more machine leaning model) to classify protein sequences and generate predictions regarding binding properties (par [50] [70]). Mason specifically teaches that such models may include CNNs, RNNs, SVMs, random forests, or logistic regression, and describes the system in functional terms, i.e., generating predictions of binding and selecting candidate variants based on those predictions (par [93] [105]).
Mason does not disclaim or exclude the use of other types of machine learning models. Rather, Mason provides a flexible computational framework capable of incorporating improved model architectures as they become available (Fig. 2-3, par [81]). Thus, the principle of operation—i.e., using machine learning to predict and classify properties of antibody/protein variants—is preserved regardless of the specific model architecture selected.
Mason does not specifically disclose that the machine-learned model is a machine-learned language model.
However, in the analogous art of machine learning and protein sequence, Ofer discloses that the most advanced machine learning model is machine-learned language model (abstract). Ofer teaches that “Proteins, which can be represented as strings of amino-acid letters, are a natural fit to many NLP methods. … We present methods for encoding the information of proteins as text and analyzing it with NLP methods, reviewing classic concepts such as bag-of-words, k-mers/n-grams and text search, as well as modern techniques such as word embedding, contextualized embedding, deep learning and neural language models. In particular, we focus on recent innovations such as masked language modeling, self-supervised learning and attention-based models.” (abstract). Ofer further teaches that “Language models have been used to understand and predict viral mutations that evade neutralizing antibodies [35]. Language generation models can also be applied to synthetic protein design, as in ProGen and other works [56,110,4]. For example we might generate peptides with the gene-ontology attribute ‘‘defense response to virus” and an initial primer sequence amenable to binding a sequence of interest, such as the ACE2 receptor targeted by SARS-CoV-2” (page 1755, par 3).
A finding of obviousness does not require certainty. See, e.g., In re O'Farrell, 853 F.2d 894, 903-904 (Fed. Cir. 1988) ("Obviousness does not require absolute predictability of success .... all that is required is a reasonable expectation of success."). Further, "obviousness cannot be avoided simply by a showing of some degree of unpredictability in the art so long as there was a reasonable probability of success." Pfizer, Inc. v. Apotex, Inc., 480 F.3d 1348, 1364 (Fed. Cir. 2007). In this case, Ofer teaches that “Proteins, which can be represented as strings of amino-acid letters, are a natural fit to many NLP methods” (abstract). “A crucial advantage of language models for pretraining is that they are self-supervised (as in the masked language task): the model predicts an explicit ground truth, but it doesn’t require labelled data, making it usable on any corpus, at potentially massive scale [77,21].” (page 1754, par 1). Here, Ofer encourages one of ordinary skill in the art to take the advantage of the language model. Thus, it would have been obvious to one of ordinary skill in the art to modify Mason and include the machine-learned language model, in order to utilize massive scale data.
The court has held that the test for obviousness is not whether the features of a secondary reference may be bodily incorporated into the structure of the primary reference; nor is it that the claimed invention must be expressly suggested in any one or all of the references. Rather, the test is what the combined teachings of the references would have suggested to those of ordinary skill in the art. See In re Keller, 642 F.2d 413, 208 USPQ 871 (CCPA 1981).
In this case, Ofer teaches that “A crucial advantage of language models for pretraining is that they are self-supervised (as in the masked language task): the model predicts an explicit ground truth, but it doesn’t require labelled data, making it usable on any corpus, at potentially massive scale [77,21].” (page 1754, par 1). Since language model has the advantages of self-supervised, predicts an explicit ground truth, and does not require labelled data, and the language model is built using Deep learning technique, it would motivate one of ordinary skill in the art to upgrade the Deep learning in to language model for high throughput secerning. That would give Mason capability to handle larger scale data with more precision.
Ofer expressly teaches that protein sequences can be represented as strings amenable to natural language processing and specifically highlights that language models (e.g., transformer-based models) provide predictive advantages in analyzing protein stability, binding, and function (abstract).
Ofer further notes that language models address limitations of prior approaches by enabling contextualized embeddings and self-supervised learning to capture sequence-function relationships (abstract).
A person of ordinary skill in the art would have recognized that applying the improved sequence-modeling capabilities of Ofer’s language models to Mason’s antibody optimization system would predictably improve accuracy and throughput of variant classification. This provides a clear rationale consistent with KSR Int’l v. Teleflex, 550 U.S. 398 (2007).
Mason does not specifically teach an amino-acid sequence that includes (i) a species associated with the input protein sequence variant. However, Ofer teaches that protein language models can detect taxonomic origin and utilize organism information associated with protein sequences. Specifically, Ofer teaches that protein language models may be used for “detection of the taxonomic origin of proteins,” (page 1754, par 3), and that sequence-based predictions may include organism information associated with sequences (page 1754, par 3).
Because species identity is a known and biologically relevant characteristic of protein sequences affecting protein properties, a person of ordinary skill in the art would have found it obvious to include species information in the input representation provided to a machine learning model to improve prediction accuracy.
Ofer does not specifically teach tokenizing the amino acid sequences. However, Tam expressly teaches tokenization performed by transformer-based language neural networks.
Specifically, Tam teaches that during training, transformer-based language neural networks “receive input words … and encode input words … into vectors in a vector space,” wherein “a vector includes one or more tokens corresponding to one or more words,” and further teaches that “an input layer performs tokenization of input words … into vectors in a vector space,” and that “an input layer uses byte-pair encoding (BPE) to tokenize input words.” (par [0075]).
Tam further teaches that such neural networks “encode query phrase into a first vector of tokens using byte-pair encoding (BPE)” and encode a target phrase into a second vector of tokens using BPE (par [0113]).
Additionally, Tam teaches that transformer-based language neural networks use “byte-pair encoding (BPE), as a tokenizer.” (par [0072]).
Thus, Tam explicitly teaches that transformer-based language models receive tokenized input sequences.
Ofer teaches applying language modeling techniques, including transformer-based architectures, to protein sequences and teaches that proteins are represented as symbolic sequences suitable for language modeling techniques.
Accordingly, it would have been obvious to apply the tokenization techniques taught by Tam when implementing the protein language modeling techniques taught by Ofer in the protein sequence prediction system taught by Mason.
Regarding claim 2, Mason discloses that wherein (i) at least one of the one or more training protein sequence variants is an antibody sequence variant, and the corresponding respective binding partner is an antigen; and the one or more protein sequence variants are antibody sequence variants (par [50]).
Regarding claim 3, Mason discloses that wherein the training data includes multi-species sequence data comprising human sequence data (par [62]).
antibody-antigen-specific weights corresponding to the different antibody-antigen pair.
Regarding claim 6, Mason does not specifically teach the one or more non-transitory computer-readable media having stored thereon further instructions that, when executed by the one or more processors, cause the computing system to perform at least one of the listed functions in the instant claim.
However, Ofer discloses
(i) pre-train the machine-learned model using a self-supervised pre-training objective to analyze the one or more training protein sequence variants, wherein the pre-training includes generating a set of universal model weights (Fig. 2, page 1754, par 1).
(ii) pre-train the machine-learned model using a self-supervised pre-training objective to analyze the one or more training protein sequence variants, wherein the pre-training includes generating a set of universal model weights, wherein the self-supervised pre-training objective is a masked language model objective (Fig. 2, page 1754, par 1).
Ofer teaches that “Language models have been used to understand and predict viral mutations that evade neutralizing antibodies [35]. Language generation models can also be applied to synthetic protein design, as in ProGen and other works [56,110,4]. For example we might generate peptides with the gene-ontology attribute ‘‘defense response to virus” and an initial primer sequence amenable to binding a sequence of interest, such as the ACE2 receptor targeted by SARS-CoV-2” (page 1755, par 3). Ofer further teaches that “A crucial advantage of language models for pretraining is that they are self-supervised (as in the masked language task): the model predicts an explicit ground truth, but it doesn’t require labelled data, making it usable on any corpus, at potentially massive scale [77,21].” (page 1754, par 1). Thus, it would have been obvious to one of ordinary skill in the art to modify Mason and include the machine-learned model being pre-trained using a masked language model objective to analyze the one or more training protein sequence variants, the pre-training generating a set of universal model weights, in order to utilize massive scale data.
Mason teaches deep learning as machine learning (par [47][108][111]). Many Language models are built using deep learning techniques. Thus, an upgrade machine learning from deep learning to language model would have been obvious to one of ordinary skill in the art.
Regarding claim 7,
(i) high-throughput screening,
(ii) low-throughput screening,
(iii) high accuracy targeted screening,
(iv) a surface plasmon resonance (SPR) technique,
(v) an isothermal titration calorimetry (ITC) technique,
(vi) a biolayer interferometry (BLI) technique, or
(vii) a microscale thermophoresis (MST) technique.
are known assay methods widely used in the art. Mason teaches high-throughput screening (par [7]).
Regarding claim 9, Mason discloses the one or more nontransitory computer-readable media having stored thereon further instructions that, when executed by the one or more processors, cause the computing system to:
determine at least one respective measured binding characteristic based on an environmental condition (pH-dependant antigen binding) (par [116]).
Regarding claim 10, Mason discloses that wherein the one or more protein sequence variants include at least one of:
a reference antibody (par [120]),
a commercial antibody,
a non-commercial antibody,
a clinical antibody (par [110]),
a non-clinical antibody,
a research-grade antibody,
a diagnostic-grade antibody,
a publicly-available antibody,
an antibody derived from patient samples (par [62]),
a de novo antibody discovered in vivo,
a de novo antibody discovered in vitro,
a de novo antibody discovered in silica (par [50]).
Regarding claim 11, Mason discloses that wherein the one or more protein sequence variants include at least one sequence variant selected from the group consisting of a monoclonal antibody (par [137]), a human antibody (par [62]), a humanized antibody, a camelised antibody, a chimeric antibody, single-chain Fvs (scFv), disulfide-linked Fvs(sdFv), Fab fragments, F (ab') fragments, anti-idiotypic (anti-Id) antibody and epitopebinding fragments of any of the above.
Regarding claim 12, Mason discloses that wherein the one or more non-transitory computer-readable media having stored thereon further instructions that, when executed by the one or more processors, cause the computing system to:
generate the one or more protein sequence variants by programmatically mutating (deep mutation scanning) (par [6]):
(i) one or more amino acids of at least one protein in the one or more protein sequence variants (par [6]);
(ii) one or more regions of the at least one of the one or more protein sequence variants, selected from the group consisting of complementarity determining regions (CDR), heavy chain variable region (VH), light chain variable region (VL), framework (FR), or constant domain of an antibody (par [6]);
(iii) one or more CDR selected from the group consisting of CDR1, CDR2 and CDR3 of the VH (par [6]); or
(iv) one or more CDR selected from the group consisting of CDR1, CDR2 and CDR3 of the VL (par [6]).
Regarding claim 13, Mason discloses that wherein an isotype of at least one of the one or more protein sequence variants is selected from the group consisting of lgG, lgE, lgM, lgD, lgA and lgY (par [46]).
Regarding claim 14, Mason discloses that wherein at least one of the one or more predicted binding characteristics is expressed as an equilibrium dissociation constant (KD) and is improved by 2-, 3-, 4-, 5-, 6-, 7-, 8-, 9-, or 10-fold or more relative to at least one of the one or more protein sequence variants (par [108]).
Regarding claim 15, Mason discloses that wherein respective desired properties of at least one variant of interest in the one or more variants of interest include at least one of:
(i) an increase in at least one predicted binding equilibrium of the variant of interest (par [108]);
(ii) a decrease in at least one predicted binding equilibrium of the variant of interest;
(iii) an upper bound of at least one predicted binding equilibrium of the variant of interest (par [108]);
(iv) a lower bound of at least one predicted binding equilibrium of the variant of interest;
(v) an increase in equilibrium toward a first antigen of a first predicted binding equilibrium of the variant of interest and a decrease in equilibrium toward a second antigen of a second predicted binding equilibrium of the variant of interest (par [99]);
(vi) ability of a cytokine sequence of a variant of interest to increase or decrease binding equilibrium towards receptors;
(vii) suitability of a variant of interest for use as a next-generation antibody scaffold and/or antibody mimetic scaffold;
(viii) ability of a variant of interest in an Fe region of an antibody to bind to an Fe receptor;
(ix) a developability of the variant of interest as indicated by tolerability upon administration; or
(x) an ability of a protein to interact with another protein (par [108]).
Regarding claim 16, Mason discloses that wherein the predicted binding characteristics include at least one of:
(i) a numerical dissociation constant (Kd) (the numerical values correspond to binding kinetic values for a set of variants) (par [70]);
(ii) a surrogate/correlate to Kd (the numerical values correspond to binding kinetic values for a set of variants) (par [70]);
(iii) a numerical association constant (Ka) (the numerical values correspond to binding kinetic values for a set of variants) (par [70]); or
(iv) a surrogate/correlate to Ka (the numerical values correspond to binding kinetic values for a set of variants) (par [70]).
Regarding claim 17, Mason discloses that wherein the non-transitory computer-readable media stores at least one of:
(i) an artificial neural network (par [55]);
(ii) a transformer neural network;
(iii) a convolutional neural network (par [55]);
(iv) a recurrent neural network (par [55]);
(v) a deep learning network;
(vi) an autoencoder (par [112]);
(vii) a regression model (par [55]);
(viii) a plug-and-play language model;
(iv) a generative model (par [112]); or
(x) a genetic algorithm.
Claim(s) 5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Masonin view of Ofer as applied to claims 1-3, 6-7 and 9-17 above, and further in view of Kovaltsuk et al. (The Journal of Immunology, 2018) (Kovaltsuk).
Regarding claim 5, Examiner is not sure whether Mason discloses that the one or more nontransitory computer-readable media having stored thereon further instructions that, when executed by the one or more processors, cause the computing system to:
obtain at least one of the one or more training protein sequence variants from at least one of:
(i) Observed Antibody Space (OAS) database;
(ii) Uniref90 protein database;
(iii) any Uniref-derived dataset;
(iv) a BFD dataset;
(v) a Mgnify dataset;
(vi) any metagenomic dataset derived from a JGI or EBI compendiums;
(vii) any corpus of assembled protein sequences; or
(viii) any dataset of natural antibody sequences, which might be obtained by BCR-sequencing or other means.
However, Kovaltsuk discloses Observed Antibody Space (OAS) database for data mining next generation sequencing of antibody repertoires (title). It would have been obvious to one of ordinary skill in the art to select the Observed Antibody Space (OAS) database for obtaining at least one of the one or more training protein sequence variants, because the selection is based on its suitability for the intended use.
Claim(s) 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Mason view of Ofer as applied to claims 1-3, 6-7 and 9-17 above, and further in view of Browniee (Machine Learning Mastery, 2021).
Regarding claim 8, Mason does not specifically teach re-train the machine-learned model using data output by a different at least one binding assay corresponding to a different antibody-antigen pair, wherein the re-training includes generating a different set of antibody-antigen-specific weights corresponding to the different antibody-antigen pair
However, Browiee teaches that model re-training enables the model in production to make the most accurate predictions with the most up-to-date data. Model re-training use the existing model as a starting point and retraining it (page 1). It adapts the model to the current data so that the re-trained model will improve the performance (page 1). Thus, it would have been obvious to one of ordinary skill in the art to re-train the machine-learned model of Mason using data output by a different at least one binding assay corresponding to a different antibody-antigen pair, wherein the re-training includes generating a different set of antibody-antigen-specific weights corresponding to the different antibody-antigen pair, in order to enable the model in production to make the most accurate predictions with the most up-to-date data.
Response to Arguments
Applicant's arguments filed 01/22/2026 have been fully considered but they are not persuasive.
Applicant argues that neither Mason nor Ofer teaches or suggests that the machine-learned language model is configured to receive a tokenized amino-acid sequence that includes a species-conditioning token and one or more complementarity-determining-region (CDR) delimiting tokens, and therefore asserts that claim 1 is patentable.
These arguments are not persuasive.
Response Regarding Tokenized Input Sequence
Applicant argues that the cited references fail to teach a tokenized input sequence. However, Tam expressly teaches tokenization performed by transformer-based language neural networks.
Specifically, Tam teaches that during training, transformer-based language neural networks “receive input words … and encode input words … into vectors in a vector space,” wherein “a vector includes one or more tokens corresponding to one or more words,” and further teaches that “an input layer performs tokenization of input words … into vectors in a vector space,” and that “an input layer uses byte-pair encoding (BPE) to tokenize input words.” (par [0075]).
Tam further teaches that such neural networks “encode query phrase into a first vector of tokens using byte-pair encoding (BPE)” and encode a target phrase into a second vector of tokens using BPE (par [0113]).
Additionally, Tam teaches that transformer-based language neural networks use “byte-pair encoding (BPE), as a tokenizer.” (par [0072]).
Thus, Tam explicitly teaches that transformer-based language models receive tokenized input sequences.
Ofer teaches applying language modeling techniques, including transformer-based architectures, to protein sequences and teaches that proteins are represented as symbolic sequences suitable for language modeling techniques.
Accordingly, it would have been obvious to apply the tokenization techniques taught by Tam when implementing the protein language modeling techniques taught by Ofer in the protein sequence prediction system taught by Mason.
Response Regarding Species-Conditioning Token
Applicant further argues that the cited references fail to teach conditioning on species via species-conditioning tokens. However, Ofer teaches that protein language models can detect taxonomic origin and utilize organism information associated with protein sequences. Specifically, Ofer teaches that protein language models may be used for “detection of the taxonomic origin of proteins,” (page 1754, par 3), and that sequence-based predictions may include organism information associated with sequences (page 1754, par 3).
Because species identity is a known and biologically relevant characteristic of protein sequences affecting protein properties, a person of ordinary skill in the art would have found it obvious to include species information in the input representation provided to a machine learning model to improve prediction accuracy.
Further, Tam teaches that language models encode input sequences into tokens representing elements of input sequences (par [0072][0075][0113]).
Accordingly, representing species information using tokens in the tokenized sequence input represents an obvious implementation choice consistent with the token-based input processing taught by Tam.
Response Regarding CDR Delimiting Tokens
Applicant also argues that neither Mason nor Ofer teaches delimiting complementarity-determining-region (CDR) segments using tokens.
However, Mason expressly teaches that antibody binding properties are determined by sequence regions including complementarity-determining regions (CDRs), and teaches providing antibody sequence variants including CDR sequences as input to machine learning models for prediction of binding characteristics (Mason par [0006], [0049]). Thus, Mason teaches providing CDR sequence regions as model input.
Tam teaches that transformer-based language neural networks perform tokenization of input sequences and encode input sequences into vectors of tokens using tokenization techniques such as byte-pair encoding (Tam par [0072], [0075], [0113]).
Because Mason teaches providing CDR sequence regions as input and Tam teaches tokenizing input sequences into tokens, it would have been obvious to tokenize the CDR sequence regions taught by Mason when implementing the language modeling techniques taught by Ofer using the tokenization techniques taught by Tam.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to XIAOYUN R XU, Ph. D. whose telephone number is (571)270-5560. The examiner can normally be reached M-F 8am-5pm.
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, Lyle Alexander can be reached on 571-272-1254. 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.
/XIAOYUN R XU, Ph.D./ Primary Examiner, Art Unit 1797