bioRxiv2021
Classification of proteins into their respective functional categories remains a long-standing key challenge in computational biology. Machine Learning (ML) based discriminative algorithms have been used extensively to address this challenge; however, the presence of small-sized, noisy, unbalanced protein classification datasets where high sequence similarity does not always imply identical functional properties have prevented robust prediction performance. Herein we present a ML method, Ensemble method for enZyme Classification (EnZymClass), that is specifically designed to address these issues. EnZymClass makes use of 47 alignment-free feature extraction techniques as numerically encoded descriptors of protein sequences to construct a stacked ensemble classification scheme capable of categorizing proteins based on their functional attributes. We used EnZymClass to classify plant acyl-ACP thioesterases (TEs) into short, long and mixed free fatty acid substrate specificity categories. While general guidelines for inferring substrate specificity have been proposed before, prediction of chain-length preference from primary sequence has remained elusive. EnZymClass achieved high classification metric scores on the TE substrate specificity prediction task (average accuracy score of 0.8, average precision and recall scores of 0.87 and 0.89 respectively on medium-chain TE prediction) producing accuracy scores that are about twice as effective at avoiding misclassifications than existing similarity-based methods of substrate specificity prediction. By applying EnZymClass to a subset of TEs in the ThYme database, we identified two acyl-ACP TE, ClFatB3 and CwFatB2, with previously uncharacterized activity in E. coli fatty acid production hosts. We incorporated modifications into ClFatB3 established in prior TE engineering studies, resulting in a 4.2-fold overall improvement in observed C10 titers over the wildtype enzyme. EnZymClass can be readily applied to other protein classification challenges and is available at: https://github.com/deeprob/ThioesteraseEnzymeSpecificity Author SummaryThe natural diversity of proteins has been harnessed to serve specialized applications in various fields, including medicine, renewable chemical production, and food and agriculture. Acquiring and characterizing new proteins to meet a given application, however, can be an expensive process, requiring selection from thousands to hundreds of thousands of candidates in a database and subsequent experimental screening. Using amino acid sequence to predict a proteins function has been demonstrated to accelerate this process, however standard approaches require information on previously characterized proteins and their respective sequences. Obtaining the necessary amount of data to accurately infer sequence-function relationships can be prohibitive, especially with a low-throughput testing cycle. Here, we present EnZymClass, a model that is specifically designed to work with small to medium-sized protein sequence datasets and retain high prediction performance of function. We applied EnZymClass to predict the presence or absence of a desired function among acyl-ACP thioesterases, a key enzyme class used in the production of renewable oleochemicals in microbial hosts. By training EnZymClass on only 115 functionally characterized enzyme sequences, we were able to successfully detect two plant acyl-ACP thioesterases with the desired specialized function among 617 sequences in the ThYme database.