MAPS

Mechanistic Atlas of Protein Sequences
Map Report Learn more

Which probes move under mutation — and do the movers predict pathogenicity?

Each of the 38 ESMC-6B L78 functional probes reads one property at a residue. For every … variant of uncertain significance we read the probe on the wild-type residue and on the mutant, take the shift (WT − mutant), and split the VUS by their predicted pathogenicity (ESMC L78 covariance probe): … predicted-pathogenic (≥ 0.5) vs … predicted-benign. The question this page answers: which reads change the most, and do they change more for the variants the model calls pathogenic?

Signed shift in probe units (binary: Δ probability; regression: Δ value; secondary structure: Jensen–Shannon divergence, ≥ 0). Histograms clipped to the 0.5–99.5 percentile.

1 · Probes that move most for likely-pathogenic VUS

Ranked by the fraction of predicted-pathogenic VUS whose read shifts more than 10% of the probe's own dynamic range. These are the channels carrying the most mutational signal where the model already suspects pathogenicity.

2 · Pathogenic vs benign — which reads shift more when the model calls pathogenic

Bars show mean |shift| among predicted-pathogenic minus predicted-benign VUS. Bars to the right = the probe moves more for pathogenic-predicted variants (the discriminating channels); to the left = moves more for benign.

3 · Per-probe distributions & the variants that move them most

Overlaid shift distributions — pathogenic-predicted vs benign-predicted VUS (each normalised to its own peak). Below each: the predicted-pathogenic VUS whose read this probe shifts the most. Click a variant to open it on the map. Sorted by pathogenic-side sensitivity.

Probes are leakage-controlled linear heads on the ESMC-6B layer-78 residue embedding, trained on an MMseqs cluster split; the covariance probe is the layer-78 EVEE-style pathogenicity probe (2-seed ensemble, homology-disjoint AUROC ≈ 0.94). A probe that shifts more for pathogenic-predicted VUS is reading a property the pathogenicity call depends on — but this is descriptive of the embedding geometry, a mechanism lens, not an independent pathogenicity prediction.