OmicFormer: a statistical priors-informed transformer for accurate and generalizable omics prediction of diseases and complex traits

An interpretable omnigenic neural network architecture for the human genome

Genetic prediction of complex phenotypes typically relies on additive linear models, which scale well but cannot capture non-additive effects or deeply integrate molecular and clinical data. Domain-specific neural networks have driven advances in images, text, and other modalities, but genome-scale neural networks remain challenging because genotypes are sparse and high-dimensional, effective sample sizes are limited, and generic architectures lack interpretability. Here, we introduce the omnigenic neural network, a biologically structured architecture inspired by the omnigenic model of complex traits. The model learns hierarchical representations of biological processes, accommodates multimodal inputs, supports transfer learning, and enables multitask prediction. Models trained in the UK Biobank and evaluated in the All of Us cohort for ischemic heart disease, type 2 diabetes, and schizophrenia outperformed published PGS Catalog and PRS-CSx scores. A multitask model trained across 36 cardiovascular endpoints further outperformed corresponding single-phenotype models and baselines. The architecture provides systems-level interpretability by quantifying the contributions of biological processes, which were consistent with established disease mechanisms. It also captures non-linear interactions between variants. Analysis of these interactions using Integrated Hessians revealed patterns concordant with previously reported epistatic associations. Together, these findings establish the omnigenic neural network as a flexible framework for interpretable, multimodal, and multitask genomic prediction.

Biochemically Constrained Multi‐Omics Integration Reveals Protein–Metabolite Dependencies Across Diseases

ABSTRACT Integrating proteomic and metabolomic data is essential for understanding complex diseases, yet current approaches that rely primarily on statistical associations often overlook the structured biochemical relationships between molecular entities and suffer from discriminative instability in small clinical cohorts. Here, we present ProMetNet, a biochemically constrained framework that incorporates pathway‐derived connectivity from the Reactome database into neural network architecture. By encoding protein–metabolite relationships based on reaction topology, ProMetNet models structured cross‐omics dependencies rather than relying solely on statistical correlations, reducing spurious associations while preserving global molecular context and improving robustness in data‐limited settings. Across four heterogeneous disease cohorts, including Alzheimer's disease, type 2 diabetes, COVID‐19, and glioblastoma, ProMetNet consistently outperforms evaluated multi‐omics integration methods, including MOGONET, P‐NET, PEARL, and MOINER, maintaining high discriminative performance under substantial data downsampling. In addition to classification accuracy, the framework prioritizes biologically plausible protein–metabolite dependencies that are not captured by conventional differential or correlation‐based analyses. Importantly, pathway‐level signals identified by ProMetNet demonstrate consistent discriminative performance in independent large‐scale population data from the UK Biobank (N = 47,507), supporting their robustness and generalizability. Together, these results establish ProMetNet as a biologically grounded and interpretable framework for multi‐omics integration, enabling robust identification of structured molecular dependencies across diseases.

Genetic decoding reveals druggable biology implicitly learned by a medical-history foundation model

A generalized supervised contrastive learning framework for integrative multi-omics prediction models

MB-SupCon-cont improves prediction accuracy by incorporating a generalized contrastive loss function that defines similarity and dissimilarity for continuous responses using three distance-based weighting methods, and provides superior representation learning and improves data visualization in lower-dimensional spaces.

Imputation-free transformer learning enables robust Alzheimer’s disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts

Accurate diagnostic classification and disease-severity prediction for Alzheimer’s disease are hampered by the pervasive incompleteness and heterogeneity of real-world clinical data. Left unaddressed, these barriers prevent reliable disease modelling and hinder effective clinical evaluation. Conventional imputation strategies introduce systematic bias, distort inter-feature relationships, and yield overconfident predictions, limitations that are especially consequential in diagnostic settings. Here, we propose NITROGEN‖, an imputation-free transformer that jointly models within-patient feature dependencies and between-patient relational structure through masked and intersample attention, enabling robust multimodal learning directly from partially observed records. We trained NITROGEN on the Alzheimer’s Disease Neuroimaging Initiative (ADNI; N=7858 scans), and evaluated it, without further retraining, on two independent cohorts: the Open Access Series of Imaging Studies (OASIS-3; N=2675 scans) and the Australian Imaging, Biomarkers and Lifestyle (AIBL; N=1286 scans) study. Across all cohorts and various tasks including binary and multi-class diagnostic classification as well as continuous cognitive score prediction, NITROGEN showed robust probability calibration and uncertainty quantification advantages over tree-based ensemble methods, while maintaining competitive discriminative and continuous cognitive score prediction performance. Moreover, cross-cohort and cross-method analyses identified cortical thickness in the temporal pole, age, and APOE genotype as important, though not individually sufficient, features for robust AD status classification. To address prediction reliability under incomplete data, we further introduced a modality-aware uncertainty adjustment that augments predictive uncertainty in proportion to the estimated importance of absent modalities, enabling models to express calibrated confidence when critical diagnostic information is unavailable. Taken together, our results show that imputation-free attention learning preserved meaningful discrimination under external cohort shift, while revealing the expected performance degradation on more distributionally different cohorts. These findings further demonstrate that evaluating machine learning models for neurological disease along dimensions of calibration, interpretability, and cross-cohort reliability—not predictive accuracy alone—is essential for better clinical deployment. Code to reproduce our experiments is publicly available at https://github.com/cschneuw/nitrogen.

Mass univariate aggregation methods for machine learning in neuroscience

Machine learning is a ubiquitous part of the modern neuroimaging toolkit, particularly for research aimed towards precision medicine goals of improving individual-level diagnosis and treatment. However, the high dimensionality of neuroimaging data poses significant challenges for constructing interpretable predictive models. Several established methods, such as Connectome-based Predictive Modeling (CPM), Polyneuro Risk Scores (PNRS, inspired by Polygenic Risk Scores), and Polyconnectomic Scoring (PCS), offer an interpretable approach, which we term “Mass Univariate Aggregation” (MUA). MUA approaches evaluate each feature independently and then use a linear combination of weighted features to derive predictions, providing directly interpretable weights for individual features. Despite the existence and widespread usage of various MUA approaches in neuroimaging, tools for their application are still fragmented, and there does not yet exist an open-access unified tool for implementing, evaluating, or comparing these models within a standardized machine learning workflow. Here, we present a unified, flexible, and accessible configurable pipeline that can be used for implementing CPM, PNRS, PCS, and many new MUA configurations facilitated by user-specified parameters. Built in Python as an add-on for scikit-learn, our configurable pipeline enables researchers to leverage the functionality and standards provided by a widely used open-access machine learning tool. We validated the configurable pipeline by replicating the outcomes achieved by existing CPM and PNRS implementations, utilizing resting-state functional connectivity data from the Human Connectome Project to predict fluid intelligence (n = 1067). We further validated the pipeline’s PCS implementation, confirming PCS computation with external connectome summary statistics (CSS) matrices using the same data, and CSS derivation using simulated data. Although designed to fill a gap in neuroimaging, our open-source, configurable pipeline provides a standardized platform for applying the MUA methods to any machine learning setting that features high-dimensional data.