Optimizing proteins using model based optimizations

Inventors

Yang, Kevin Kaichuang • Feala, Jacob D. • Baranov, Maxim • Monian, Brinda

Assignees

Flagship Pioneering Inc • Flagship Pioneering Innovations VI Inc • Generate Biomedicines Inc

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.

Core Innovation

The invention relates to a design by adaptive sampling (DbAS) framework for generating a protein sequence with altered immunogenicity relative to a sampling of protein sequences, while preserving a desired function. The approach evaluates the immunogenicity of protein sequences sampled from a generative model configured to preserve the desired function, and uses this evaluation to guide selection of sequences for generation. The core sampling is weighted according to an estimated probability that a generated protein sequence deviates in immunogenicity of a particular percentile of the sampling distribution.

Immunogenicity evaluation is performed using residue-wise predictions of a relative likelihood for MHC II to bind to the protein predicted to have the altered immunogenicity. The altered immunogenicity is described as reduced immunogenicity that improves a likelihood of humanization relative to the immunogenicity of the sampling of protein sequences. In the framework, the generative model is retrained using the selected or reweighted samples, enabling iterative improvement across sampling and evaluation rounds.

In disclosed embodiments, the generative model includes an encoder neural network and a decoder neural network, and is trained to preserve the desired function. The framework can include first and second weighted samplings and provides a second updated generative model by retraining the first updated generative model based on immunogenicity evaluated for retraining. An immunogenicity oracle is provided that outputs residue-wise predictions and is tied to MHC II binding, including likelihood of MHC II presentation based on MHC-associated peptide proteomics (MAPPs), such that the estimated probability corresponds to a likelihood of MHC II presentation.

Claims Coverage

The independent claims cover methods that evaluate immunogenicity of sampled protein sequences, select and/or reweight sampling according to an estimated probability of deviating in immunogenicity percentile from the sample distribution, and generate a protein sequence predicted to have altered immunogenicity with preserved desired function, using residue-wise predictions of relative likelihood for MHC II binding. Across the independent claims, the main inventive features amount to iterative probability-weighted adaptive sampling and generative-model-based function preservation, with MHC II residue-wise oracle evaluation driving the immunogenicity-directed selection and retraining.

Overall, the independent claims are directed to adaptive, probability-weighted selection and generation driven by immunogenicity evaluation using residue-wise MHC II binding likelihood, with a generative model configured to preserve a desired function. The claims further emphasize iterative or multi-stage updated generative models obtained via immunogenicity-guided retraining so the generated protein sequence is predicted to have reduced immunogenicity that improves likelihood of humanization relative to the sampling distribution.

Stated Advantages

Documented Applications

No documented applications found