Genetic variant interpretation contains an uncomfortable asymmetry: enormous amounts of public evidence exist, yet turning that evidence into a reviewable judgment remains slow, specialized work. FH Variant Triage explores how software can reduce that burden without pretending to replace the expert.
The project takes the form of a disease-specific research prototype that estimates whether variants in LDLR, APOB, and PCSK9 resemble ClinVar records labeled benign or pathogenic. It is explicitly a triage aid, not a diagnostic tool. Its purpose is to help experts decide which genetic variants deserve closer review without pretending to replace clinical judgment.
The question behind FH Variant Triage
Familial hypercholesterolemia provides a deliberately narrow test case. Restricting the first implementation to LDLR, APOB, and PCSK9 makes the biological and validation boundary easier to describe, while still addressing a condition where variant review can carry meaningful consequences. Researchers, variant scientists, clinicians, and patients may benefit indirectly through faster evidence review. Incorrect interpretation could cause harm, so outputs must stay inside a research and expert-review boundary.
How FH Variant Triage took shape
The workflow assembles public evidence, creates transparent feature representations, compares simple model baselines, and produces scores intended to prioritize review. Documentation, tests, model limitations, and reviewer guidance are treated as part of the system rather than material to add after the model is complete. Josiah selected the disease boundary, public-benefit posture, validation ladder, review boundary, and publication strategy, working with AI agents on implementation, testing, documentation, and methodology.
Method diagram · research boundary
The model prioritizes review; it does not finish interpretation.
Research prioritization only · never diagnosis or clinical decision-making
What the evidence supports
The public repository includes tests, a model card, methods, source documentation, limitations, and an external-validation request. Initial performance is prototype signal, not clinical proof. The project argues for a restrained form of medical AI: not an oracle, but an evidence-sorting instrument. Its credibility depends less on a single performance number than on calibration, external validation, transparent failure modes, and a handoff that keeps consequential interpretation with qualified people.
Independent validation, harder temporal and gene-held-out tests, calibration analysis, and blinded expert review remain necessary. The tool must not be used for diagnosis or clinical decision-making. Future work centers on stronger independent validation and a compact visual demonstration while keeping every claim inside the research-triage boundary.