Ongoing Research

Reliability ML Framework & Platform for GWAS-Based Disease Risk Modeling

Build any GWAS ML pipeline - Get every metric that matters

NHDSL is developing a reproducibility-first machine learning framework and software platform for genomic disease risk modeling. The project focuses on transforming GWAS-based machine learning from isolated performance reporting into structured reliability assessment using two complementary scores: the Biological Reliability Score (BRS) and the Computational Reliability Score (CRS).

Instead of relying on a single accuracy value or a single experimental run, the framework evaluates model discrimination, probability calibration, multi-run robustness, external dataset generalization, and feature stability together. The platform implements this methodology through modular pipelines, reproducible run artifacts, custom model integration, parallel repeated runs, and publication-ready scientific reports.

Core reliability dimensions:
Biological signal survival · External support · Pathway interpretation · Ranking quality · Calibration · Cross-run stability · Leakage-safe validation

BRS evaluates whether candidate SNPs, genes, and pathways remain biologically credible across runs, models, validation settings, external datasets, and annotation layers.

CRS evaluates whether machine-learning predictions are computationally reliable through ranking quality, calibration, cross-run stability, and generalization behavior.

GWAS data are evaluated through base models, ensemble integration, repeated seeded runs, integrity checks, and validation layers to produce two complementary reliability scores: the Biological Reliability Score (BRS) and the Computational Reliability Score (CRS).

Framework Methodology 

The framework treats variance as evidence. Instead of selecting a single best run, it evaluates how predictions, metrics, calibration behavior, and selected genomic features behave across models, seeds, and datasets. 

Reliable ML Platform

The platform implements the framework as a modular research system. Researchers can register datasets, build GWAS ML pipelines, swap pipeline steps, plug in custom estimators, execute repeated runs in parallel, evaluate multiple metric families, and generate publication-ready reports. 

The framework is implemented as a modular research platform for reproducible pipelines, custom model integration, parallel repeated runs, external validation, feature stability, and paper-ready reports.