# US 2023/0054552 A1 · Optimizing Proteins Using Model Based Optimizations > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 2023/0054552 A1 - Document ID: US20230054552A1 - Status: Published application - Kind: A1 - Application number: 17/975,466 - Filed: 2022-10-27 - Published: 2023-02-23 - Earliest claimed: 2020-04-27 - Assignees: Flagship Pioneering Innovations VI, LLC - Applicants: Flagship Pioneering Innovations VI, LLC - Inventors: Kevin Kaichuang Yang; Jacob D. Feala; Maxim Baranov; Brinda Monian - Web page: https://patentgrep.com/patent/US20230054552A1 ## CPC - G16B 15/30: Drug targeting using structural data; Docking or binding prediction - G06N 3/08: Learning methods - G16B 15/20: Protein or domain folding - G16B 40/20: Supervised data analysis - G16B: BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract Humanizing proteins can be a laborious process, often involving trial and error or other non-systematic methods. To improve humanization, neural networks can be employed to generate new protein sequences having higher probabilities of being humanized. In an embodiment, a method includes evaluating the immunogenicity of a sampling of protein sequences. The method can include weighting the sampling of protein sequences from the generative model according to an estimated probability of a particular generated protein sequence having a deviation in immunogenicity than a particular percentile of immunogenicity of the sampling of protein sequences. The method can further include generating a protein sequence weighted sampling of protein sequences. The generated protein sequence representing a protein has an altered immunogenicity. Such a generated protein has a higher likelihood of being humanized. ## Claims ### Claim 1 A method of generating a sequence of a protein predicted to have altered immunogenicity compared to a sampling of protein sequences, the protein having a desired function, comprising: evaluating the immunogenicity of a sampling of protein sequences; selecting a set of protein sequences from the sampling by weighting the sampling of protein sequences according to an estimated probability of a particular generated protein sequence having a deviation in immunogenicity of a particular percentile of immunogenicity of the sampling of protein sequences, wherein the sampling of protein sequences is from a generative model and the generative model is configured to preserve the desired function; and generating a protein sequence based on the selected set of protein sequences, the protein sequence representing a protein predicted to have an altered immunogenicity. ### Claim 2 (depends on 1) The method of claim 1 further comprising: retraining the generative model based on the selected set of protein sequences; wherein generating the protein sequence is based on the retrained generative model that is retrained based on the selected set of protein sequences. ### Claim 3 (depends on 2) The method of claim 2, further comprising: retraining the generative model two or more iterations, wherein each iteration includes: a) evaluating the immunogenicity of a sampling of protein sequences generated by the retrained generative model, and b) further retraining the generative model by weighting the sampling of protein sequences generated by the retrained generative model according to their evaluated immunogenicity in the retraining. ### Claim 4 (depends on 2) The method of claim 2, wherein the sampling of protein sequences is unique from any protein sequences present in a training set of protein sequences used to train the generative model. ### Claim 5 A method of generating a sequence of a protein predicted to have altered immunogenicity compared to a sampling of protein sequences, the protein having a desired function, comprising: evaluating the immunogenicity of a sampling of protein sequences; weighting the sampling of protein sequences according to an estimated probability of a particular generated protein sequence having a deviation in immunogenicity of a particular percentile of immunogenicity of the sampling of protein sequences, resulting in a first weighted sampling of protein sequences; evaluating the immunogenicity of a sampling of protein sequences using first sampling of protein sequences; reweighting the sampling of protein sequences from the first weighted sampling of protein sequences evaluated for immunogenicity according to their evaluated immunogenicity, resulting in a second weighted sampling of protein sequences; generating a protein sequence from the second weighted sampling of protein sequences, the protein sequence representing a protein predicted to have an altered immunogenicity; and providing a generative model that preserves the desired function, the generative model comprising an encoder neural network and a decoder neural network. ### Claim 6 (depends on 5) The method of claim 5, wherein evaluating the immunogenicity of the sampling of protein sequences further includes evaluating the immunogenicity from a generative model, wherein the generative model preserves the desired function, the generative model further comprising an encoder neural network and a decoder neural network. ### Claim 7 (depends on 5) The method of claim 5, wherein the sampling of protein sequences does not comprise any protein sequences present in a training set of protein sequences used to weight the first weighted sampling of protein sequences, and the first weighted sampling of protein sequences does not comprise any sequences present in a training set of protein sequences used to weight the first weighted sampling of protein sequences. ### Claim 8 (depends on 5) The method of claim 5, further comprising iterating the steps of evaluating the immunogenicity of a sampling of protein sequences and providing a second weighted sampling of protein sequences. ### Claim 9 (depends on 5) The method of claim 5, wherein immunogenicity is evaluated using the predicted affinity of k-mers of a subject sequence for a MHC molecule. ### Claim 10 (depends on 9) The method of claim 9, wherein k is greater than or equal to 5 and less than or equal to 13. ### Claim 11 (depends on 9) The method of claim 9, wherein the predicted affinity for a k-mer of a subject sequence for the MHC molecule is evaluated by one or more of MHCFlurry, NetMHCpan, or NetMHCIIpan. ### Claim 12 (depends on 9) The method of claim 9, further comprising combining the predicted affinity of k-mers by at least one of (a) the reciprocal of the sum of the reciprocal of each predicted affinity of k-mers, and (b) the lowest predicted affinity of the k-mers. ### Claim 13 (depends on 5) The method of claim 5, wherein the immunogenicity of the sampling of protein sequences from the first weighted sampling of protein sequences or second weighted sampling of protein sequences is evaluated experimentally. ### Claim 14 (depends on 5) The method of claim 5, wherein the altered immunogenicity is one or more of increased affinity and reduced affinity for a human MHC, including at least one of human MHC I, such as human MHC I encoded by allele HLA-A0201, and MHC II. ### Claim 15 (depends on 5) The method of claim 5, wherein the desired function is selected from enzymatic activity, asparaginase activity, target binding affinity (other than an MHC), thermal stability, solubility, size of protein, intracellular location of protein, activation of a cellular signaling pathway or gene expression program, and expression level. ### Claim 16 (depends on 5) The method of claim 5, wherein immunogenicity is evaluated using residue-wise predictions of a relative likelihood for MHC II to bind to the protein predicted to have the altered immunogenicity. ### Claim 17 (depends on 5) The method of claim 5, wherein a candidate model accepts an amino acid sequence as input and outputs one or more residue-wise predictions of relative likelihood for MHCII to bind to a peptide containing each residue. ### Claim 18 (depends on 5) The method of claim 5, wherein the estimated probability is a likelihood of MHC II presentation, and the likelihood is based on both a probability of peptide presentation of MHC-associated peptide proteomics (MAPPS). ### Claim 19 (depends on 5) The method of claim 5, wherein the first weighted sampling of protein sequences is used to provide a first updated generative model by updating a generative model, and second weighted sampling of protein sequences is used to provide a second updated generative model that retrains the first updated generative model. ### Claim 20 A method of generating a sequence of a protein predicted to have altered immunogenicity compared to a sampling of protein sequences, the protein having a desired function, comprising: providing a generative model that preserves the desired function, the generative model comprising an encoder neural network and a decoder neural network; evaluating the immunogenicity of a sampling of protein sequences from a generative model, wherein the generative model preserves the desired function, the generative model further comprising an encoder neural network and a decoder neural network; providing a first updated generative model by retraining the generative model by weighting the sampling of protein sequences from the generative model according to an estimated probability of a particular generated protein sequence having a deviation in immunogenicity of a particular percentile of immunogenicity of the sampling of protein sequences; evaluating the immunogenicity of a sampling of protein sequences using the first updated generative model; providing a second updated generative model by retraining the first updated generative model by weighting the sampling of protein sequences from the first updated generative model evaluated for immunogenicity according to their evaluated immunogenicity in the retraining; and generating a protein sequence from the further updated generative model, the protein sequence representing a protein predicted to have an altered immunogenicity. ### Claim 21 (depends on 20) The method of claim 20, wherein the generative model was trained using a set of naturally occurring sequences predicted to have the desired function. ### Claim 22 (depends on 20) The method of claim 20, wherein the encoder neural network and decoder neural network were trained to minimize the negative log-likelihood of a reconstruction of an input the input to the encoder by the output of the decoder. ### Claim 23 (depends on 20) The method of claim 20, wherein the encoder neural network and decoder neural network were trained to minimize the Kullback-Leibler-divergence (KLD) between the encoder and Normal (0,1). ### Claim 24 (depends on 20) The method of claim 20, wherein the encoder neural network and decoder neural network employ at least two hidden layers. ### Claim 25 (depends on 20) The method of claim 20, wherein a neural network of the generative model evaluates the immunogenicity of the sampling of protein sequences by an oracle outputting at least one of a value and an uncertainty. ### Claim 26 An isolated protein sequence comprising an amino acid sequence that differs from a sample of amino acid sequences by at least one amino acid addition, deletion, or substitution relative to the sample generated by: evaluating the immunogenicity of a sampling of protein sequences; selecting a set of protein sequences from the sampling by weighting the sampling of protein sequences according to an estimated probability of a particular generated protein sequence having a deviation in immunogenicity of a particular percentile of immunogenicity of the sampling of protein sequences, wherein the sampling of protein sequences is from a generative model and the generative model is configured to preserve the desired function; and generating a protein sequence based on the selected set of protein sequences, the protein sequence representing a protein predicted to have an altered immunogenicity.