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

The AdaptiveInterpreter framework is a mechanism-first pathogenicity prediction system developed through a collaborative research process between human and AI investigators. Unlike traditional in-silico predictors—often limited by context-blind statistical inference—AdaptiveInterpreter explicitly models four mechanistic modes of protein dysfunction and integrates deep biological context, producing interpretable, mechanistically grounded variant classifications.

In this study, we validate AdaptiveInterpreter across 109,939 variants in 93 genes, including 15,007 variants with definitive ClinVar labels. The framework achieves 99.8% sensitivity, 87.2% PPV, 85.8% NPV, and 89.6% overall agreement with ClinVar. A multi-layered safety architecture (including the conservation clamp) ensures that no observed dangerous misclassifications occur, even in challenging contexts of missing data.
AdaptiveInterpreter resolves 62.8% of ClinVar VUS (59,587 variants), demonstrating substantial potential to reduce clinical uncertainty.

This paper presents the computational and methodological core of the AdaptiveInterpreter system—its architecture, mechanistic logic, validation metrics, and safety controls. A companion biological discovery paper introduces two novel mechanistic insights derived from the system: the Semi-Dominant Hypothesis and the CASCADE phenomenon, both published separately at
https://doi.org/10.5281/zenodo.18109872.

Together, these works illustrate a new model of scientific collaboration in which human investigators and AI research agents jointly generate hypotheses, build tools, and uncover novel biological mechanisms.