Most disease risk models treat conditions as isolated events — a myocardial infarction here, a diabetes diagnosis there — missing the complex, overlapping patterns that define how illness actually unfolds across a lifetime. A new computational framework challenges that assumption by treating a patient's entire medical history as a coherent biological narrative, with implications for how polygenic risk scores are used in clinical practice and population health.
The ALADYNOULLI model, validated across three large independent biobanks — UK Biobank, Mass General Brigham, and the All of Us Research Program — integrates longitudinal EHR diagnoses, patient age, and germline genetic data simultaneously across 683,000 participants and 348 disease categories tracked over up to 52 years. Rather than assigning each diagnosis to a single disease bucket, the framework recovers 21 replicable latent disease signatures that capture co-occurring and time-varying illness patterns. These signatures demonstrate strong cross-cohort stability, with a median 80% compositional preservation across independent datasets — a notable benchmark for generalizability. The model's mathematical formulation as a mixture of probabilities, rather than a probability of a mixture, correctly handles the clinical reality of simultaneous and chronic conditions. Genetically, known risk carriers behaved as expected: familial hypercholesterolaemia variants enriched the cardiovascular signature, clonal haematopoiesis variants aligned with inflammation, and rare variants in LDLR, TTN, and BRCA2 tracked disease-specific signatures. A signature-based GWAS identified 151 genome-wide significant loci, including cardiovascular associations missed by conventional single-trait analyses.
This work sits at the intersection of two maturing fields — polygenic risk scoring and phenome-wide association studies — and meaningfully advances both. By modeling disease as a longitudinal, multi-dimensional trajectory rather than a binary outcome, the framework may ultimately enable earlier stratification of patients into biologically coherent subtypes before clinical thresholds are crossed. The large sample sizes, multi-cohort replication, and explicit genetic integration elevate this beyond incremental methodology. Key limitations include reliance on ICD-coded diagnoses (subject to documentation bias) and the challenge of translating latent statistical signatures into clinically actionable categories. This is a potentially paradigm-shifting contribution to computational medicine.