A Bayesian framework for longitudinal EHR and genetic discovery.

Latent disease signatures, reconstructed from longitudinal diagnoses and polygenic risk, separate the phenotypic footprint of monogenic predisposition: LDLR, APOB and LPA on the cardiovascular side, TTN on heart failure and arrhythmia, BRCA2 on breast cancer.

In carriers of monogenic predisposition, penetrance is still framed as a cumulative risk at a given age, for want of tools able to track how a genotype is expressed across a lifetime. This work presents ALADYNOULLI, a Bayesian framework that reconstructs time-varying disease signatures from longitudinal diagnoses, age and 36 polygenic risk scores, applied to three independent biobanks — UK Biobank (427,239 participants), Mass General Brigham (48,069) and All of Us (208,263), totalling more than 683,000 individuals and up to 52 years of follow-up. Carriers of familial hypercholesterolaemia activate the cardiovascular signature before their first coronary event (OR 1.63; p = 0.017), and gene-based rare variant burden tests across 18,464 genes link LDLR, APOB and LPA to that same signature, TTN to the heart failure and arrhythmia signature (p = 1×10⁻²¹), BRCA2 to the breast cancer signature and PKD1 to the inflammation signature. Within a single diagnostic category the model separates clearly distinct trajectory subgroups (Cohen's d up to 4.25), and a signature-based GWAS identifies 151 significant loci including 23 cardiovascular loci absent from the corresponding single-trait GWAS. The authors note that the first recorded diagnostic code serves as a proxy for the true diagnosis date, which may make prevalent cases appear incident.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

What matters here is not predictive performance but a conceptual shift: penetrance becomes observable as a trajectory rather than a cumulative risk, and the fact that familial hypercholesterolaemia carriers separate out before their first event suggests that an at-risk genotype leaves a phenotypic footprint detectable ahead of overt disease. Two limits bound its relevance to our practice: resolution stays at the PheCode level, 348 common conditions, so the framework says nothing about rare disease diagnosis, and the associations rest on biobank cohorts without family segregation or functional validation. None of it transfers to the clinic today, but this is the kind of work that prefigures how we will reframe genetic counselling once longitudinal records become usable.

Analysis by Dr Thibaut Benquey