Beyond the Black Box: Physics-Based and Conservation-Informed Evidence for ACVRL1 Variant Classification in Hereditary Hemorrhagic Telangiectasia and Pulmonary Arterial Hypertension

Missense variants in ACVRL1 (ALK1) cause hereditary hemorrhagic telangiectasia type 2 (HHT2), heritable pulmonary arterial hypertension (PAH), or both in the same patient, and many remain variants of uncertain significance. AlphaMissense classifies them accurately but does not show which evidence drives a call. I annotated 899 ACVRL1 missense variants with protein-stability estimates (FoldX, DynaMut2), a GREMLIN-style conservation score, and structural context. I then trained random-forest classifiers on 311 variants with labels traceable to ClinVar or a publication (296 pathogenic-like, 15 benign-like). A structural/conservation-only classifier (Model 1b) reached a cross-validated ROC-AUC of 0.954 (95% CI 0.916–0.984), and a model adding sequence-based predictors (Model 1) reached 0.979, versus 0.983 for AlphaMissense on the same variants. On variants first reported in 2025–2026, Model 1b scored 0.976, Model 1 0.992, and AlphaMissense 0.992; equivalence within ±0.05 AUC was supported for Model 1 only. Benign variants were the main weakness: Model 1b correctly classified 10/15, against 14/15 for AlphaMissense. Conservation drove most calls, with predicted destabilization contributing to a minority. An interpretable, gene-specific classifier can approach AlphaMissense for ACVRL1 while showing the structural or evolutionary evidence behind each call. Its calls are hypotheses, not functional evidence, and can prioritize VUS for functional testing.