Integrating Genomics into Multimodal EHR Foundation Models

Amar, J.; Liu, E.; Breschi, A.; Zhang, L.; Kheradpour, P.; Li, S.; Soleymani Lehmann, L.; Giulianelli, A.; Edwards, M.; Nola, D.; Mani, R.; Vats, P.; Tetreault, J.; Chen, T. J.; McLean, C. Y.

doi:10.1101/2025.10.26.684668

bioRxiv · 10.1101/2025.10.26.684668

Integrating Genomics into Multimodal EHR Foundation Models

Amar, J.·Liu, E.·Breschi, A.·Zhang, L.·Kheradpour, P.·Li, S.·Soleymani Lehmann, L.·Giulianelli, A.·Edwards, M.·Nola, D.·Mani, R.·Vats, P.·Tetreault, J.·Chen, T. J.·McLean, C. Y.

View original source ↗Export metadata ↓

Abstract

AO_SCPLOWBSTRACTC_SCPLOWThis paper introduces an innovative Electronic Health Record (EHR) foundation model that integrates Polygenic Risk Scores (PRS) as a foundational data modality, moving beyond traditional EHR-only approaches to build more holistic health profiles. Leveraging the extensive and diverse data from the All of Us (AoU) Research Program, this multimodal framework aims to learn complex relationships between clinical data and genetic predispositions. The methodology extends advancements in generative AI to the EHR foundation model space, enhancing predictive capabilities and interpretability. Evaluation on AoU data demonstrates the models predictive value for the onset of various conditions, particularly Type 2 Diabetes (T2D), and illustrates the interplay between PRS and EHR data. The work also explores transfer learning for custom classification tasks, showcasing the architectures versatility and efficiency. This approach is pivotal for unlocking new insights into disease prediction, proactive health management, risk stratification, and personalized treatment strategies, laying the groundwork for more personalized, equitable, and actionable real-world evidence generation in healthcare.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Amar, J., Liu, E., Breschi, A., Zhang, L., Kheradpour, P., Li, S., Soleymani Lehmann, L., Giulianelli, A., Edwards, M., Nola, D., Mani, R., Vats, P., Tetreault, J., Chen, T. J., McLean, C. Y.. 2025-10-27. Integrating Genomics into Multimodal EHR Foundation Models. https://doi.org/10.1101/2025.10.26.684668

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Inferring cascade drivers of VEXAS syndrome by a causal machine learning tool CauNagi

VEXAS syndrome is an adult-onset severe autoinflammatory disease caused by somatic mutations in UBA1, yet the cascade mechanisms linking primitive hematopoietic abnormalities to mature myeloid dysfunctions remain largely unknown. Identifying master regulators of a progressive disease, a black-box process, from complex transcriptomic data also remains challenging. To address this challenge, we developed CauNagi, a computational framework for prioritizing cascade candidate regulators (CCRs). CauNagi integrates a causal representation learning module derived from CausCell with an iterative deep learning backbone adapted from UNAGI; in addition, CauNagi extends these two components with a unique downstream module for CCRs analysis designed to characterize regulatory propagation across hierarchical cellular states. Mechanistically, CauNagi iteratively integrates causal disentangled representation learning with (1) disease-stage cell-state trajectory reconstruction and (2) dynamic regulatory analysis. Benchmarking on single-cell transcriptomic datasets showed that CauNagi preserved cell-type structure in idiopathic pulmonary fibrosis (IPF) and enriched known acute myeloid leukemia(AML)-associated genes among its top-ranked global regulators. When applied to VEXAS syndrome, CauNagi readily revealed inflammatory responses, endoplasmic reticulum stress, and myeloid bias, consistent with the disease features. Furthermore, the CCRs analysis module of CauNagi assisted us in identifying 36 causal drivers, with SPI1, NFKB1, STAT3, and FOS prioritized as high-confidence regulatory hubs linking aberrant myeloid differentiation and inflammatory programs. These findings were further supported by an independent single-cell transcriptomic dataset from a murine VEXAS model. Overall, CauNagi provides a computationally efficient and systematic framework for identifying candidate causal regulators. Beyond hematopoietic diseases, CauNagi may also be applicable to other progressive disorders for which multistage single-cell transcriptomic datasets are available. CauNagi is available at https://github.com/steamed-stuffed-bun/CauNagi.

bioinformatics↗

bioRxiv2026

Inferential boundaries of age prediction: why prediction does not establish biological age measurement

Chronological-age clocks reconstruct age from biological measurements, yet their outputs are interpreted as biological age, gaps as ageing acceleration and intervention-associated decreases as rejuvenation. We show that age supervision identifies an age-task statistic, not a biological-age construct, and establish how this distinction changes biomarker construction and validation. Even at the population optimum, the same observable distribution and age-prediction performance admit incompatible biological-age interpretations. Resolving this ambiguity requires assumptions or evidence beyond the age task. Squared-error age loss penalizes within-age output dispersion without defining its biological direction. Given age and background, a gap re-expresses the compressed score; exact age recovery eliminates it even when heterogeneity remains in the measurements. Shared biological covariance permits genuine prognostic value without establishing construct identity. After allogeneic haematopoietic stem-cell transplantation, recipient-blood scores showed excess donor-lineage affiliation under a score-pairing null. In NHANES, age-trained scores improved held-out five-year mortality prediction beyond age and background, yet direct modelling of source measurements and mortality supervision at matched scalar capacity yielded further gains. The intended biological object must therefore guide study design, measurement selection and representation; validation must establish the claimed measurement relation rather than rely on age-prediction success alone.

bioinformatics↗

bioRxiv2026

DisenTE: Sparse Pattern-Context Modeling for Interpretable Translation-Efficiency Matrix Completion

Partially observed object-by-context matrices arise across data-rich science, where dominant object effects can obscure smaller but informative context-dependent variation. We study this problem in a translation-efficiency atlas of 9,494 5' UTRs across 78 cellular and tissue contexts. We present DisenTE, a sequence-conditioned neural model that combines separate sequence and context branches with a sparse low-rank pattern-context channel. Each module pairs a sequence-derived activation with context-specific deployment weights, forming a dictionary whose sequence and context components can be examined separately. Under five-fold within-panel entry masking, DisenTE achieves a UTR-centered residual Spearman correlation of 0.641 +/- 0.005, compared with 0.304 +/- 0.003 for the strongest reference model. The learned dictionary retains 11 of 20 candidate modules. CTM 6 has the largest overlap with an external TOP set and a cap-proximal pyrimidine pattern; CTMs 5 and 7 also overlap the set but have purine-containing consensuses. The evidence supports CTM 6 as a TOP sequence anchor and CTMs 5 and 7 as TOP-set-associated factors. On this dataset, DisenTE improves completion over the evaluated references and provides module-level summaries of its fitted context-dependent variation.

bioinformatics↗