Search

Abstract Computational approaches for predicting the pathogenicity of genetic variants have advanced in recent years. These methods enable researchers to determine the possible clinical impact of rare and novel variants. Historically these prediction methods used hand-crafted features based on structural, evolutionary, or physiochemical properties of the variant. In this study we propose a novel framework that leverages the power of pre-trained protein language models to predict variant pathogenicity. We show that our approach VariPred (Variant impact Predictor) outperforms current state-of-the-art methods by using an end-to-end model that only requires the protein sequence as input. Using one of the best-performing protein language models (ESM-1b), we establish a robust classifier that requires no calculation of structural features or multiple sequence alignments. We compare the performance of VariPred with other representative models including 3Cnet, Polyphen-2, REVEL, MetaLR, FATHMM and ESM variant. VariPred performs as well as, or in most cases better than these other predictors using six variant impact prediction benchmarks despite requiring only sequence data and no pre-processing of the data.

Background: Zinc binding proteins make up a significant proportion of the proteomes of most organisms and, within those proteins, zinc performs rôles in catalysis and structure stabilisation. Identifying the ability to bind zinc in a novel protein can offer insights into its functions and the mechanism by which it carries out those functions. Computational means of doing so are faster than spectroscopic means, allowing for searching at much greater speeds and scales, and thereby guiding complimentary experimental approaches. Typically, computational models of zinc binding predict zinc binding for individual residues rather than as a single binding site, and typically do not distinguish between different classes of binding site—missing crucial properties indicative of zinc binding. Methods: Previously, we created ZincBindDB, a continuously updated database of known zinc binding sites, categorised by family (the set of liganding residues). Here, we use this dataset to create ZincBindPredict, a set of machine learning methods to predict the most common zinc binding site families for both structure and sequence. Results: The models all achieve an MCC ≥ 0.88, recall ≥ 0.93 and precision ≥ 0.91 for the structural models (mean MCC = 0.97), while the sequence models have MCC ≥ 0.64, recall ≥ 0.80 and precision ≥ 0.83 (mean MCC = 0.87), with the models for binding sites containing four liganding residues performing much better than this. Conclusions: The predictors outperform competing zinc binding site predictors and are available online via a web interface and a GraphQL API.