Goodfire's MAPS Tool Reveals Why 2.1 Million Genetic Variants Cause Disease

Every human carries millions of genetic variants. The vast majority are harmless, but a small fraction cause serious disease. Telling them apart is hard enough; understanding why a particular mutation is dangerous is harder still. Goodfire just released MAPS (Mechanistic Atlas of Protein Sequences), a tool covering 2.1 million missense variants that predicts not just whether a variant is harmful, but which specific protein property it disrupts.

Why variant interpretation is so difficult

A missense variant is a single-letter change in DNA that swaps one amino acid for another in a protein. Genome sequencing has uncovered tens of millions of such variants in the human population, most with no annotated role. When a clinician sequences a patient's genome, the majority of 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 tools have become accurate at predicting which variants harm protein function, but they rarely explain how. Google DeepMind's AlphaMissense scores pathogenicity well, but it lacks interpretability, does not assess functional impact, and produces scores that are not disease-specific. MAPS is built to close that gap.

Building on EVEE, moving to the protein level

MAPS extends Goodfire's earlier tool, EVEE (Evo Variant Effect Explorer). EVEE probes the internal representations of Evo 2, a 7-billion-parameter genomic foundation model trained on DNA, to predict pathogenicity across all 4.2 million ClinVar variants, achieving 0.997 AUROC on 839,000 of them.

DNA models have a structural blind spot: they do not natively represent how a mutation reshapes a protein's three-dimensional shape. MAPS applies the same probing framework to ESM-C 6B, a protein language model trained on billions of protein sequences. Operating at the protein level lets MAPS detect things Evo 2 cannot, such as whether a mutation destabilizes a binding pocket or disrupts a DNA-contact domain.

How the probing pipeline works

The core technique is covariance probing, introduced with EVEE. Rather than fine-tuning ESM-C or training a new model end-to-end on labeled variant data (which is scarce), the team trains lightweight classifiers called probes on top of ESM-C's frozen internal representations. The model already encodes rich biological knowledge; the probes learn to read it out.

• Pathogenicity probe: Takes the difference in ESM-C embeddings between the reference and mutant protein sequences and predicts whether the variant is harmful. MAPS matches or exceeds both EVEE and AlphaMissense on missense variant benchmarks.
• Annotation probes: A panel of probes trained to predict specific protein properties, including binding sites, fold stability, DNA-contact regions, and post-translational modification sites. By measuring how much each variant shifts these probe predictions, MAPS produces a structured disruption profile for every variant.
• Covariance pooling: A covariance-based probe captures second-order structure from sequence embeddings. Standard mean pooling blurs positional relationships; covariance pooling preserves them, capturing whether a motif appears within a specific structural context.

The output is a hypothesis, not just a score: this variant is likely harmful because it disrupts the protein's binding site, destabilizes its fold, or interferes with a DNA-contact interface. A complementary panel of probes trained on existing annotations identifies which genomic properties a variant disrupts, offering a general framework for interrogating foundation models across scientific disciplines.

What MAPS covers and who it's for

The atlas covers 2.1 million missense variants, all visualized in an interactive 2D map at maps.goodfire.com. Each variant is colored by predicted pathogenicity and linked to a full disruption profile. The site is free to explore with no login required.

MAPS focuses specifically on missense variants, the class where protein-level modeling adds the most signal. EVEE, which operates on DNA, covers a broader range including regulatory and splicing mutations. The two tools are designed to be read together: EVEE calls variants from the DNA perspective, MAPS from the protein perspective.

• Clinicians and genetic counselors can use MAPS to generate mechanistic hypotheses for VUS variants, giving a structured starting point for experimental follow-up.
• Rare disease researchers can query specific proteins and see which variants are predicted to disrupt binding sites versus fold stability, helping prioritize which mutations to study first.
• Computational biologists can replicate MAPS or run it on private variant sets using Goodfire's Silico research platform.

Interpretability as the translation layer

Most prior tools treat variant scoring as a black-box classification problem. Goodfire's approach treats it as an interpretability problem: the foundation model already encodes the biology, and the task is extracting that knowledge in human-readable form. A pathogenicity score tells you whether to worry. A disruption profile tells you what to do about it.

One caveat worth noting: systematic evaluation of embedding-based classifiers and frameworks for generating mechanistic interpretation from genomic foundation models has been limited. MAPS and EVEE together represent the most thorough attempt yet to fill that gap, but the annotation probes are bounded by what biological concepts are currently annotated. Truly novel mechanisms fall outside their scope.

Access and availability

The interactive atlas is live and free at maps.goodfire.com. To run MAPS on your own variant sets or replicate the pipeline, Goodfire provides access through their Silico platform. The EVEE preprint is available on bioRxiv and was developed in collaboration with Mayo Clinic. Both tools are free for research use, with Silico offering a path to custom experiments on private data.