Goodfire's MAPS Tool Reveals Why 2.1 Million Genetic Variants Cause Disease
Goodfire's MAPS uses ESM-C 6B to explain 2.1 million protein variants, predicting not just pathogenicity but the specific biological mechanism disrupted
- MAPS is a new atlas from Goodfire covering 2.1 million missense variants, predicting both pathogenicity and the specific protein property disrupted.
- It extends Goodfire's prior EVEE work from DNA (Evo 2) to proteins, using ESM-C 6B, a protein language model trained on billions of sequences.
- The core technique is covariance probing: lightweight classifiers trained on frozen ESM-C embeddings that match or exceed AlphaMissense on missense benchmarks.
- Unlike AlphaMissense, MAPS produces a disruption profile per variant, identifying whether a mutation breaks a binding site, fold, DNA-contact region, or other property.
- The interactive atlas is free and live at maps.goodfire.com; custom variant analysis is available via Goodfire's Silico platform.
- MAPS is complementary to EVEE, which covers all ClinVar variant types from the DNA level, including regulatory and splicing mutations.
Every human carries millions of genetic variants. The vast majority are harmless background noise, but a small fraction cause serious disease. The hard part is telling them apart, and harder still is understanding why a particular mutation is dangerous. Goodfire just released MAPS (Mechanistic Atlas of Protein Sequences), a new tool that covers 2.1 million missense variants and, for each one, predicts not just whether it is harmful but which specific protein property it disrupts.
The problem that makes genomic medicine so hard
A missense variant is a single-letter change in DNA that swaps one amino acid for another in a protein. Genome sequencing efforts have led to the discovery of tens of millions of protein missense variants found in the human population, with the majority having no annotated role. When a clinician sequences a patient's genome, most of the flagged variants land in a category called VUS, or variants of uncertain significance, meaning their effect on health is simply unknown. That classification gives clinicians almost nothing to act on.
Sequence-based AI approaches have become highly accurate at predicting variants that are detrimental to protein function, but they do not inform on mechanisms of disruption. Tools like AlphaMissense from Google DeepMind are strong at scoring pathogenicity, but AlphaMissense lacks interpretability, does not assess the functional impact of variants, and provides pathogenicity scores that are not disease specific. MAPS is designed to close exactly that gap.
From DNA to protein: extending the probing approach
MAPS builds directly on Goodfire's prior work, EVEE (Evo Variant Effect Explorer). EVEE showed that probing the internal representations of Evo 2, a 7-billion-parameter genomic foundation model trained on DNA, could achieve state-of-the-art pathogenicity prediction. EVEE provides pathogenicity predictions for all 4.2 million ClinVar variants using embedding-based probes trained on Evo 2, achieving 0.997 overall AUROC on 839k ClinVar variants.