SIMLINK Enables Accurate Variant Pathogenicity Prediction through Modeling the Gene-Variant-Feature Association Structure.

Explicit modelling of the gene-variant-feature association structure through a knowledge graph of more than 8 million triplets, with joint learning of a linear and a non-linear component using graph neural networks.

Variant pathogenicity prediction faces two limitations: the lack of explicit modelling of the gene-variant-feature association structure, and the failure to disentangle the linear component of the variant-pathogenicity relationship. SIMLINK builds a variant-centred knowledge graph comprising more than 8 million triplets, then jointly learns the linear and non-linear components using a linear model and graph neural networks. Trained on ClinVar variants, it outperforms state-of-the-art methods including CADD and AlphaMissense on independent test sets, for both missense and synonymous variants, with the effect of allele frequency on performance also assessed. Applied to variants implicated in autism spectrum disorder, it separated high- from low-confidence variants, the genes harbouring top-ranked variants being highly pathogenic. The source code is freely available.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The contribution worth noting is the claimed performance on synonymous variants, a genuine blind spot of missense predictors such as AlphaMissense, although its translation into diagnostic gain remains entirely to be demonstrated. The usual caveat is not addressed in the abstract: a model trained on ClinVar and evaluated on ClinVar-derived sets remains exposed to circularity and gene-level bias, which better overall performance does not rule out. Before any use in classification, calibration into ACMG evidence weights would be needed, and this work does not provide it.

Analysis by Dr Thibaut Benquey