Research · Hackel Lab · University of Minnesota
CALIVA
A calibrated missense-pathogenicity predictor built only from protein structure, evolution, and language-model features, with no other predictor's score among its inputs. It matches the field's strongest tools while being the best-calibrated model tested.
Calibrated pathogenicity from primary features alone
Interpreting missense variants of uncertain significance is a major bottleneck in clinical genetics. The most accurate computational tools are metapredictors that combine the outputs of many upstream predictors. Although accurate, they aren't transparent about which biological signal drives a call, at risk of circularity with the ClinVar labels used to evaluate them, and rarely calibrated, even though predicted probabilities strongly inform clinical triage.
CALIVA is built exclusively from primary features: AlphaFold structure with FoldX thermodynamics, ESM-2 protein-language embeddings, multiple-sequence-alignment conservation, and macromolecular interface annotation, while ingesting no other predictor's score. On a protein-grouped ClinVar test set it reaches 0.961 AUROC, statistically equivalent to both REVEL and AlphaMissense, while achieving the lowest calibration error of any model tested (ECE 0.014, roughly three times better than AlphaMissense). A CatBoost + graph-attention ensemble, it stays interpretable at the level of its inputs via SHAP, so every prediction can be traced back to conservation, substitution chemistry, thermodynamic stability, and language-model context.