Abstract
Missense variants in ACVRL1 (ALK1) cause either hereditary hemorrhagic telangiectasia type 2 (HHT2) or heritable pulmonary arterial hypertension (PAH), amid a large, growing burden of variants of uncertain significance. Deep-learning predictors such as AlphaMissense now approach expert-level accuracy but return an unexplained probability, with no actionable mechanistic rationale. I assembled 899 ACVRL1 missense variants annotated with physics-based (FoldX and DynaMut2 ΔΔG, short-timescale molecular dynamics), evolutionary (a GREMLIN-style conservation score across vertebrate orthologs), and electrostatic features, and trained random-forest classifiers using only this interpretable evidence. The structural/conservation-only classifier achieved 5-fold cross-validated ROC-AUC 0.946 and, on a genuine temporally external holdout, 0.965 — statistically indistinguishable from AlphaMissense across every paired DeLong comparison performed (p = 0.29–0.83). Worked examples show that, unlike a single opaque score, the pipeline decomposes each call into a specific, falsifiable mechanism — destabilization, conservation-constraint violation, or charge disruption — concordant with independent BMP9–SMAD functional-assay data used as proof of concept, and in one case generating a testable hypothesis where AlphaMissense disagrees with every other line of evidence. Mechanistically interpretable evidence can match deep-learning pathogenicity prediction for this clinically important receptor while additionally providing a per-variant biological rationale that black-box scores cannot.