Discovery of Novel Rare Disease Genes Through Scalable Prioritization and Geometric Scoring of AlphaFold3 Structures

Abstract

Rare diseases collectively represent a significant global health burden, with approximately 80% possessing a suspected genomic basis. Despite advances in sequencing, many patients remain without a molecular diagnosis, often due to the difficulty of functionalizing variants of uncertain significance (VUS) and thus implicating new undiscovered disease genes. This thesis presents an end-to-end framework designed to bridge the gap between clinical genomic variation and biophysical mechanisms by using protein-protein interactions as a way to implicate candidate disease genes. Utilizing data from the Children’s Rare Disease Collaborative at Boston Children’s Hospital, we used an existing interactome as well as agentic architectures as independent methods for the prioritization of candidate interactions for downstream structural prediction. Agentic platform ToolUniverse maintained 98.8% coverage across datasets of up to 25,000 queried interactions, successfully recovering known biological complexes within the top 200 ranked results, while effectively excluding all non-interacting pairs. To refine these prioritized lists, we developed the Interface Confidence Score (ICS), an XGBoost-based regressor trained on AlphaFold 3 (AF3) features. On a held-out test set, the ICS achieved a Spearman rank correlation of 0.74 and an accuracy of 95.9% when tasked with separating binding/non-binding proteins. Comparative analysis revealed that the ICS significantly outperformed ipTM in separation power, correctly rejecting 88.9% of "hallucinated" high confidence AF3 interfaces. When applied to 989 Predictomes structures, the ICS identified critical unique hits that were discarded by the previously established Structured Prediction Omics Classifier threshold. A high resolution case study of the MCM2-7 complex further validated the ICS's ability to provide granular, subcomplex stability rankings that aligned with known hexameric assembly mechanics. Structural and transcriptomic validation of our lead candidates provided a mechanistic rationale for their associated phenotypes. Collectively, this work demonstrates that integrating agentic reasoning with geometrically grounded structural scoring provides a transformative roadmap for resolving the molecular basis of rare pediatric disorders and closing the diagnostic gap.