Disease variant prediction with deep generative models of evolutionary data

Abstract

Quantifying the pathogenicity of protein variants in human disease-related genes would have a profound impact on clinical decisions, yet the overwhelming majority (over 98%) of these variants still have unknown consequences [1][2][3] . In principle, computational methods could support the largescale interpretation of genetic variants. However, prior methods 4-7 have relied on training machine learning models on available clinical labels. Since these labels are sparse, biased, and of variable quality, the resulting models have been considered insufficiently reliable 8 . By contrast, our approach leverages deep generative models to predict the clinical significance of protein variants without relying on labels. The natural distribution of protein sequences we observe across organisms is the result of billions of evolutionary experiments 9,10 . By modeling that distribution, we implicitly capture constraints on the protein sequences that maintain fitness. Our model EVE (Evolutionary model of Variant Effect) not only outperforms computational approaches that rely on labelled data, but also performs on par, if not better than, high-throughput assays which are increasingly used as strong evidence for variant classification [11][12][13][14][15][16][17][18][19][20][21][22][23] . After thorough validation on clinical labels, we predict the pathogenicity of 11 million variants across 1,081 disease genes, and assign high-confidence reclassification for 72k Variants of Unknown Significance 8 . Our work suggests that models of evolutionary information can provide a strong source of independent evidence for variant interpretation and that the approach will be widely useful in research and clinical settings.

TP53_PROF: a machine learning model to predict impact of missense mutations in TP53

“…TP53_PROF’s classification was compared with six scores that provided formal classification cutoff in the TP53_UMD database (Polyphen2 HumVar, Polyphen2 HumDiv, Sift, Condel, Provean and Mutassessor). EVE, a recently released deep generative model of evolutionary data [ 45 ], CHASM, a cancer-specific algorithm that showed peak performances in recent analysis [ 14 , 15 ] and Revel, another in-silico tool that presented with best balanced accuracy in a recently published algorithms comparison [ 13 , 59 ], were also used for this analysis with different defined cutoffs for LOF provided by the scores documentation (see Methods). Annotations based on the three experimental validations done on the set of 41 variants (presented in Supplementary Table S7A available online at http://bib.oxfordjournals.org/ ) were used as the truth set.…”

“…For EVE, annotations were used for EVE90pct and EVE75pct, which account for the percentage of certainty. The EVE90pct sets 10% of the variants as ‘uncertain,’ and EVE75pct sets 25% of the variants as ‘uncertain’ [ 45 ].…”

Vasor: Accurate prediction of variant effects for amino acid substitutions in MDR3

“…) is a recently developed unsupervised computational method, which trained Bayesian variational autoencoders on multiple sequence alignments to classify variant effects based on a variant-specific, computed evolutionary index followed by a fitted global-local mixture of Gaussian Mixture Models (17).…”

“…Here, we created an MDR3-specific variant dataset and trained a machine learning algorithm using several established general prediction tools, namely EVE, EVmutation, PolyPhen-2, I-Mutant2.0, MUpro, MAESTRO, and PON-P2 (17)(18)(19)(20)(21)(22)(23), as well as half-sphere exposure, posttranslational modification (PTM) site influence, and secondary structure disruption as features to obtain an MDR3-specific prediction tool for help in classifying variants as benign or pathogenic (see Fig. 1 for a graphical overview).…”

Extracting phylogenetic dimensions of coevolution reveals hidden functional signals

“…Just as evolutionarily conserved individual residues are generally crucial to a protein’s proper function, the statistical covariation (arising from correlated evolution, i.e. coevolution) between pairs of residues 1 , 2 carries information that is useful for predicting structural contacts 3 – 7 and protein–protein interactions 8 – 11 and their interfaces 12 , intuiting novel protein conformations 5 , understanding protein allostery 13 , interpreting variants 14 , 15 , identifying functional domains 16 – 19 , and reprogramming protein specificity 20 . However, despite the increasing prevalence of sequencing data, sampling of the phylogenetic tree is fundamentally limited and biased.…”