Genome-wide prediction of pathogenic gain- and loss-of-function variants from ensemble learning of a diverse feature set

Abstract

Gain-of-function (GOF) variants yield increased or novel protein function while loss-of-function (LOF) variants yield diminished protein function. GOF and LOF variants can result in markedly varying phenotypes even when occurring in the same gene. Experimental approaches for identifying GOF and LOF are slow and costly, and computational tools cannot accurately discriminate between GOF and LOF variants. We developed LoGoFunc, an ensemble machine learning method for predicting pathogenic GOF, pathogenic LOF and neutral variants, trained on 672 protein, gene, variant, and network annotations describing diverse biological characteristics. We analyzed features in terms of protein structure and function, splice disruption, and phenotypic associations, revealing previously unreported relationships between various biological phenomena and variant functional outcomes. For example, GOF and LOF variants demonstrate contrasting enrichments in protein structural and functional regions, and LOF variants are more likely to disrupt canonical splicing as indicated by splicing-related features employed by the model. Furthermore, by performing phenome-wide association studies (PheWAS), we identified strong associations between relevant phenotypes and high-confidence predicted GOF and LOF variants. LoGoFunc performs well on an independent test set of GOF and LOF variants, and we provide precomputed genome-wide GOF and LOF predictions for 71,322,505 missense variants.

PreMode predicts mode-of-action of missense variants by deep graph representation learning of protein sequence and structural context

“…Low-capacity models such as random forest and single layer perceptron fine-tuned ESM model is worse than pretrained PreMode except for SCN2A . We also compared it to LoGoFunc 42 , a method trained on G/LoF variants across genes. PreMode is slightly worse than LoGoFunc in SCN5A but better in all other genes ( Supplementary Figure 8a ).…”

A power-based sliding window approach to evaluate the clinical impact of rare genetic variants

“…Future studies with even larger sample sizes may reclassify some portions of the genes tested here, identifying some regions we erroneously excluded or identifying new regions to include. Other improvements will involve focusing on certain classes of variants in the gene, such as those computationally predicted to be gain or loss of function, those with functional data from screenings, and those in transcripts expressed in tissues of interest 19,26,27 .…”

Leveraging large-scale multi-omics to identify therapeutic targets from genome-wide association studies

“…Coding variants were defined as those impacting protein coding transcript annotated as missense variant or predicted to have "high" impact. We also retrieved predicted gain or loss of function (GoLoF) variants from LoGoFunc [47], and linked non-coding variants to genes using activity-by-contact (ABC) maps [41]. ABC scores represent the contribution of an enhancer to the regulation of gene, measured by multiplying the estimates of enhancer activity and threedimensional contact frequencies between enhancers and promoters.…”

“…We prioritized genes as putatively causal using a combination of evidence including MR, colocalization H4 posterior probabilities (PP) with molQTL, presence of an associated GoLoF variant [47] or other coding variants, distance to lead variant, and enhancer-promoter ABC scores [41]. Specifically, we ranked genes as follow: For a given locus, we then prioritized the best gene(s) as the one with the highest rank.…”

“…). However, the strongest enrichment was generally observed for genes with associated lead coding variants [47] ). The strong enrichment for nearest genes is expected given that the gene closest to the lead variant is often the causal gene.…”