From {triangleup}{triangleup}G Prediction to Mutation Prioritization with Conformal Prediction

ProbJanusDDG uses conformal prediction with a protein-stability predictor to assign directional calls for uncharacterized mutations on the S669L low-homology benchmark: 34.2% called at 50% with 94.8% sign accuracy.

• Why it matters: Point estimates can’t show whether uncertainty still allows both stabilizing and destabilizing effects, limiting which predicted stability changes deserve experimental follow-up.
• What they did: The method calibrates conformal intervals by assigning equal total weight to each protein and uses standard and adaptive intervals to decide whether intervals lie entirely above or below zero.
• The result: On S669L, adaptive intervals give directional calls to 34.2% (50% level) and 15.5% (70% level) with 94.8% and 99.0% sign accuracy, while mutation-level coverage is 91.6% (standard) and 92.5% (adaptive) at 90%.