Machine learning modeling of family wide enzyme-substrate specificity screens

Abstract

Biocatalysis is a promising approach to sustainably synthesize pharmaceuticals, complex natural products, and commodity chemicals at scale. However, the adoption of biocatalysis is limited by our ability to select enzymes that will catalyze their natural chemical transformation on non-natural substrates. While machine learning and in silico directed evolution are well-posed for this predictive modeling challenge, efforts to date have primarily aimed to increase activity against a single known substrate, rather than to identify enzymes capable of acting on new substrates of interest. To address this need, we curate 6 different high-quality enzyme family screens from the literature that each measure multiple enzymes against multiple substrates. We compare machine learning-based compound-protein interaction (CPI) modeling approaches from the literature used for predicting drug-target interactions. Surprisingly, comparing these interaction-based models against collections of independent (single task) enzyme-only or substrate-only models reveals that current CPI approaches are incapable of learning interactions between compounds and proteins in the current family level data regime. We further validate this observation by demonstrating that our no-interaction baseline can outperform CPI-based models from the literature used to guide the discovery of kinase inhibitors. Given the high performance of non-interaction based models, we introduce a new structure-based strategy for pooling residue representations across a protein sequence. Altogether, this work motivates a principled path forward in order to build and evaluate meaningful predictive models for biocatalysis and other drug discovery applications.

Turnover number predictions for kinetically uncharacterized enzymes using machine and deep learning

“…The ESM-1 b model was trained in a self-supervised fashion, i.e., 10-15 % of the amino acids in a sequence were masked at random, and the model was trained to predict the identity of the masked amino acids. It has been shown that the resulting representations contain rich information about the structure and the function of the proteins 17,30,31 . Using the pre-trained ESM-1 b model 17 , we calculated these 1280-dimensional representations for all enzymes in our dataset, in the following referred to as ESM-1b vectors.…”

Contrasting drugs from decoys

“…High performing DTI prediction methods should be able to generalize broadly to unseen types of drugs and targets, while also discriminating between highly similar molecules with different binding properties. Previous work demonstrated the utility of PLMs to improve the generalizability of DTI prediction methods[6, 20]; we now add a contrastive learning approach which improves specificity. The contrastive approach taken by CON-Plex is directly enabled by the architecture of the base PLM-enabled lexicographic model— to compute the triplet distance loss, the protein and drugs must be co-embedded, and the distance between them must be meaningful and simply computed.…”

“…Trained over hundreds of millions of protein sequences, PLMs apply the distributional hypothesis [2, 3, 18, 4] and learn a rich implicit featurization of proteins that has proven useful in a variety of tasks [12, 19, 21]. Sledzieski et al [20], and later Goldman et al [6] independently, have demonstrated the power of PLMs for DTI prediction. The use of pre-trained PLMs unlocks the full richness and diversity of data across the protein universe, whereas models trained solely on DTI data can leverage only the very limited percentage of protein space that has been tested experimentally for interactions.…”

EP-Pred: A Machine Learning Tool for Bioprospecting Promiscuous Ester Hydrolases

“…Therefore, methods to directly identify substrate promiscuity from sequence alone would greatly increase the efficiency of bioprospecting for new esterase candidates, a task ideal for machine learning algorithms (see for example, other recently developed methods for activity predictions [ 12 ]). Several studies have already predicted enzyme substrate promiscuity using molecular descriptors [ 13 ] or machine learning [ 14 , 15 , 16 ] approaches, although for other enzyme families. In addition, there are several differences in our approach.…”