Machine learning can be used to distinguish protein families and generate new proteins belonging to those families.

Abstract

Proteins are classified into families based on evolutionary relationships and common structure-function characteristics. Availability of large data sets of gene-derived protein sequences drives this classification. Sequence space is exponentially large, making it difficult to characterize family differences. In this work, we show that Machine Learning (ML) methods can be trained to distinguish between protein families. A number of supervised ML algorithms are explored to this end. The most accurate is a Long Short Term Memory (LSTM) classification method that accounts for the sequence context of the amino acids. Sequences for a number of protein families where there are sufficient data to be used in ML are studied. By splitting the data into training and testing sets, we find that this LSTM classifier can be trained to successfully classify the test sequences for all pairs of the families. Also investigated is whether the addition of structural information increases the accuracy of the binary comparisons. It does, but because there is much less available structural than sequence information, the quality of the training degrades. Another variety of LSTM, LSTM_wordGen, a context-dependent word generation algorithm, is used to generate new protein sequences based on seed sequences for the families considered here. Using the original sequences as training data and the generated sequences as test data, the LSTM classification method classifies the generated sequences almost as accurately as the true family members do. Thus, in principle, we have generated new members of these protein families.