Mechanism-First, Context-Aware Pathogenicity Prediction: A Novel Human-AI Collaborative Framework for Genetic Variant Interpretation

Abstract

The current landscape of in silico pathogenicity prediction is dominated by powerful but fundamentally limited tools. While foundational methods like SIFT and PolyPhen-2 remain in wide use, their performance often struggles with the nuances of biological context, and even advanced "meta-predictors" like REVEL or ClinPred face a trade-off between sensitivity and specificity. It is this challenge—the need for a model that is both highly sensitive and highly specific, grounded in biological mechanism rather than pure statistics—that the AdaptiveInterpreter system was designed to address. Here, we present the AdaptiveInterpreter framework, a novel, mechanism-first prediction model developed through a unique collaborative process between a human strategist and a cohort of AI collaborators. Our system models four primary mechanisms of protein failure and integrates deep biological context to generate predictions grounded in plausible mechanistic narratives, validated using novel Directional Agreement Logic (DAL). The present study focuses on missense variants; splice, intronic, and UTR variants are planned for future development. We validated our framework on a comprehensive dataset of **109,939 variants across 93 genes** (44 ACMG Secondary Findings v3.2 + 49 Discovery genes, n=15,007 with definitive ClinVar labels). The model achieves a **Positive Predictive Value (PPV) of 87.2%**, **Negative Predictive Value (NPV) of 85.8%**, **sensitivity of 99.8%**, and **specificity of 53.5%**. Agreement with ClinVar is **89.6%**. Critically, post-hoc analysis revealed that all 23 initial dangerous misclassifications (ClinVar P/LP → AI B/LB) were flagged by our conservation safety mechanism (MISSING_CONSERVATION), indicating insufficient data for confident benign classification. When properly accounting for this safety clamp, **no dangerous misclassifications were observed** (0/15,007 variants). Among ClinVar VUS, **62.8%** (59,587/94,932) were resolved to definitive classifications, demonstrating exceptional ability to resolve clinical uncertainty. The AdaptiveInterpreter framework represents both a significant advance in genomic variant interpretation and a powerful new paradigm for human-AI collaborative science. Full repo available here: https://github.com/menelly/adaptive_interpreter

Keywords