Cross-protein transfer learning substantially improves disease variant prediction

Abstract

Genetic variation in the human genome is a major determinant of individual disease risk, but the vast majority of missense variants have unknown etiological effects. Various computational strategies have been proposed to predict the effects of missense variants across the human proteome, using many different predictive signals. Here, we present a robust learning framework for leveraging functional assay data to construct computational predictors of disease variant effects. We train cross-protein transfer (CPT) models using deep mutational scanning data from only five proteins and achieve state-of-the-art performance on unseen proteins across the human proteome. On human disease variants annotated in ClinVar, our model CPT-1 improves specificity at 95% sensitivity to 64%, from 31% for ESM-1v and 50% for EVE. Our framework combines general protein sequence models with vertebrate sequence alignments and AlphaFold2 structures, and it is adaptable to the future inclusion of other sources of information. We release predictions for all missense variants in 90% of human genes. Our results establish the utility of functional assay data for learning general properties of variants that can transfer to unseen proteins.

Pathogenic Morphological Signatures of Perturbations in Mitochondrial-Related Genes Revealed by Pooled Imaging Assay

“…Raft-Seq enables the selection of live single cells with various indel mutations from the primary phenotyping, and several isogenic clones are available for SLC25A19 and ATAD3A upon request. Additionally, it is worth mentioning the advent of tools in the realm of machine learning-enabled variant prediction, including AlphaMissense and CPT 46,47 . While such tools excel in predicting substitution variant effects, they face limitations in their ability to predict the functional consequences of indel missense variants.…”

Alignment-based protein mutational landscape prediction: doing more with less

“…In the last couple of years, a lot of attention has been drawn to optimizing, ensembling, clustering, subsampling, and pairing alignments toward improving protein 3D models (Petti et al, 2023), generating multiple functional conformations (Wayment-Steele et al, 2022), and resolving interactomes (Bret et al, 2023;Bryant et al, 2022). In the context of disease variants calling, Jagota and co-authors recently showed that vertebrate alignments exhibit a strong signal that can be used to boost specificity (Jagota et al, 2022).…”

Functional genomics in inborn errors of immunity

“…Indeed, we and others have demonstrated that large language models can be effectively tailored to predict the effects of variants on different alternative spliced isoforms. 67,68,70,140 Furthermore, these models can also be employed to functionalize short insertions and deletions (indels) with promising performance. 63 These recent results underscore the immense potential of large language models in the field of genomics, ultimately driving our understanding of genetic diseases and accelerating the discovery of potential therapies.…”