Protein database search using learned representations

Abstract

A method for efficient search of protein sequence databases for proteins that have sequence, structural, and/or functional homology with respect to information derived from a search query. The method involves transforming the protein sequences into vector representations and searching in a vector space. Given a database of protein sequences and a learned embedding model, the embedding model is applied to each amino acid sequence to transform it into a sequence of vector representations. A query sequence is also transformed into a sequence of vector representations, preferably using the same learned embedding model. Once the query has been embedded in this manner, proteins are retrieved from the database based on distance between the query embedding and the protein embeddings contained within the database. Rapid and accurate search of the vector space is carried out using exact search using metric data structures, or approximate search using locality sensitive hashing.

What is claimed here follows below: 
     
         1 . A method, comprising:
 receiving a protein mutagenesis dataset containing sequence and phenotype pairs for a set of protein sequence variants;   training a sequence-to-phenotype prediction model by:
 passing the protein sequence variants through a language model to obtain a vector representation of an amino acid at each position of each variant sequence; 
 concatenating or pooling embeddings for each position and applying a dimensional reduction to generate a set of sequence representations; and 
 training a statistical model to take the set of sequence representations as input to predict one or more phenotype values; and 
   following training, receiving a set of amino acid sequences and applying the language model and the statistical model to predict one or more phenotype values for the set of amino acid sequences.   
     
     
         2 . The method as described in  claim 1  wherein the statistical model is a Gaussian process regression model.