11/ Finally, we will present a RaCoon poster at the Computational structural biology EMBO workshop this week: www.embl.org/about/info/c...

10/ RaCoon is fully open-source with a public web server.
Try it → evaluate variant pathogenicity with interpretable, residue-aware probabilities.
bio3d.cs.huji.ac.il/webserver/ra...

9/ Why does this matter?
Mis-calibrated probabilities = inconsistent evidence for clinical classification (and you may also improve AUROC).

8/ 🎯 Performance
RaCoon improves ESM1b substantially:
• ClinVar AUROC: 0.930 → 0.941
• ProteinGym AUROC: 0.912 → 0.924
• Per-protein AUROC improves too
• Calibration error (ECE/MCE) drops across all subgroups

7/ RaCoon 4-step pipeline:

1. Split variants by key residue properties
2. Fit GMMs to benign vs. pathogenic scores in each subgroup
3. Convert raw ESM1b LLRs into calibrated probabilities
4. No direct label exposure → labels only used for prior estimation

6/ RaCoon performs multicalibration of ESM1b, producing reliable probabilities across all relevant subgroups using minimal supervision.

5/ This motivated RaCoon (Residue-aware calibration via conditional distributions).
(It calibrates like a raccoon sorts trash: by categories. 😉)

4/ Surprisingly calibrating per variant subgroup (i.e. interface) not only improves miscalibration but also increase global AUROC across most models.

3/ We calibrate VEPs at the residue level since variant effects depend strongly on local residue properties. We find that model entropy distribution can guide calibration.

2/ In a calibrated model a model score of 0.8 means ~80% of similar variants are pathogenic. But VEPs often fail, especially for variants in disordered regions or protein-protein interfaces.