MAGI preprint proposes mechanistic variant annotation pipeline built on genomic transformer models

A preprint deposited on bioRxiv on 3 June 2026 introduces MAGI (Mechanistic Annotation of Genomic Impacts), a computational pipeline designed to address a widely acknowledged limitation of current variant interpretation tools: that most existing predictors output a pathogenicity probability without explaining the underlying molecular mechanism.

MAGI leverages a genomic transformer model to generate sequence-based predictions of how a given variant alters regulatory grammar — for instance, by disrupting splice sites, transcription factor binding motifs, or chromatin accessibility signals. The method is positioned as a research and educational tool for generating mechanistic hypotheses about variants of uncertain significance (VUS), rather than as a clinical decision support application.

The authors describe a pipeline that takes a variant as input, queries the transformer model for predicted functional consequences across multiple molecular layers, and returns an annotation summarising the most probable mechanistic impacts. The preprint includes benchmarking against curated variant datasets and case studies illustrating how the framework could be applied to interpret non-coding variants, which remain particularly challenging for conventional prediction tools.

The authors acknowledge that genomic foundation models have limitations: they are trained on patterns in reference genome sequences and may not capture context-specific regulation. **This is a preprint that has not yet been peer-reviewed; its methods, benchmarking claims, and conclusions should be treated as preliminary.**