Skip to content

ALADYNOULLI

Summary

ALADYNOULLI is a Bayesian generative model designed to analyze longitudinal electronic health records (EHR) linked to germline genetics. Formulated as a mixture of probabilities rather than a probability of a mixture, it accommodates simultaneous and chronic conditions. It models individual patient health histories over time, recovering latent time-varying disease signatures and individual-specific signature loadings to improve dynamic risk prediction, enhance genome-wide and rare variant association studies, and reveal patient subgroups within diagnostic categories.

Methodology and Model Architecture

ALADYNOULLI integrates genetic data and longitudinal EHR codes by formulating patient trajectories through latent signatures. For patient $i$, disease $d$, and time $t$, the hazard of disease occurrence $\pi_{idt}$ is modeled as:

$$\pi_{idt} = \sum_k \theta_{ikt} \cdot \text{sigmoid}(\phi_{kdt} + \kappa)$$

where: - $\theta_{ikt}$ is individual $i$'s normalized, time-varying association (loading) with signature $k$ at time $t$ ($\sum_k \theta_{ikt} = 1$). - $\phi_{kdt}$ is the time-varying signature-specific probability of disease $d$ at time $t$. - $\kappa$ is a global calibration parameter.

The latent variables $\lambda_{ikt}$ (which map to $\theta_{ikt}$ via a softmax function) follow a Gaussian process prior incorporating genetic covariates (36 polygenic risk scores, sex, and 10 principal components; 47 features total):

$$\lambda_{ik} \sim \mathcal{GP}(\mathbf{r}k + \mathbf{\Gamma}_k \mathbf{g}_i, \Omega)$$

where: - $\mathbf{r}k$ is the signature-specific population reference level. - $\mathbf{\Gamma}_k$ represents the genetic effects of polygenic risk scores ($\mathbf{g}_i$) on signature $k$. - $\Omega$ is the temporal covariance kernel modeling smooth patient trajectories over time.

Performance and Validation

The framework was trained and validated across three biobanks (UK Biobank, Mass General Brigham, and All of Us; total $n > 683,000$): - Signature Stability: Recovered 21 biological signatures (20 disease-specific and 1 low-incidence signature) showing a median composition preservation index of 80% across independent biobanks. - Genetic Discovery: GWAS using ALADYNOULLI's continuous "lifetime signature exposure" metric identified 151 genome-wide significant loci, identifying 23 loci for the cardiovascular signature (including IL6R, SCARB1, SMAD3, and PDGFD) that were missed by single-trait clinical endpoint analyses. - Rare Variant Associations: Gene-based rare variant analyses linked the heart failure signature to TTN, the cardiovascular signature to LDLR, APOB, and LPA, and the cancer signature to BRCA2. - Risk Prediction: Outperformed the Pooled Cohort Equation (PCE), PREVENT, and Gail risk models at both 1-year and 10-year horizons.

See Also

  • Delphi-2M — A complementary foundation model for code-level EHR prediction.
  • REGENIE — Utilized for genome-wide and rare variant association mapping of signature phenotypes.

  • Natarajan Lab — Co-developer and co-supervising laboratory.

  • Gusev Lab — co-developer with expertise in statistical genetics and longitudinal EHR modeling.

  • Parmigiani Lab — co-developer with expertise in Bayesian biostatistics and prediction.

Citations

  • Urbut, S. M., Ding, Y., Nakao, T., Koyama, S., Misra, A., Jiang, X., Harish, A., Gaffney, L., Hornsby, W. E., Smoller, J. W., Gusev, A., Natarajan, P., & Parmigiani, G. (2026). A Bayesian framework for longitudinal EHR and genetic discovery. Nature. DOI: 10.1038/s41586-026-10780-5. Source paper: s41586-026-10780-5.pdf