FINGERS-7B: A Multi-Omic Foundation Model for Precision Biomarker Discovery

Arvid E. Gollwitzer, Adrián Noriega de la Colina, Deepak A. Subramanian, Giovanni Traverso

Keywords: foundation models, gut-brain axis, Alzheimer's disease, dementia prevention, precision-prevention, biomarker discovery, microbiome, multi-omics, causal inference, hierarchical attention, metagenomics

TL;DR: FINGERS-7B is a 7B-parameter gut microbiome foundation model that discovers diagnostic, prognostic, and therapeutic biomarkers for Alzheimer's and Parkinson's, achieving AUC=0.92 for preclinical AD via four druggable gut-brain axes.

Abstract: Precision-prevention offers a transformative opportunity to reduce the growing global burden of dementia by aligning interventions with individual biological risk profiles. Building on the foundational successes of multidomain randomized trials and advances in biomarker discovery, the next phase of dementia prevention requires computational frameworks that are biologically informed, harmonized across populations, and capable of resolving who benefits from which intervention, under what conditions, and at what intensity. Here we introduce **FINGERS-7B**, a 7-billion-parameter multi-omic foundation model pretrained on 8 trillion quality-aware, hierarchically structured, and semantically meaningful tokens and 300K metabolite profiles curated from public gut-brain-relevant metagenomic archives, and fine-tuned on clinical data from the World-Wide FINGERS (WW-FINGERS) dementia prevention network. By integrating genomic, metabolomic, and clinical data within a hierarchical set attention architecture, FINGERS-7B shifts multi-omic research beyond associative biomarker discovery toward causally-motivated modeling, enabling mediation-consistent analyses that move from describing correlations toward predicting intervention effects. Applied to preclinical Alzheimer's disease (AD) detection, FINGERS-7B achieves AUC=0.92, substantially outperforming established microbiome analysis pipelines (AUC $\approx$ 0.68–0.76) and matching leading proteomic biomarkers. Our model identifies gut microbiome signatures predictive of cognitive decline 3–5 years before symptom onset (AUC=0.89) and stratifies intervention responders, with high-benefit subjects showing 2.3$\times$ greater cognitive improvement under multidomain lifestyle intervention. Mechanistic analysis reveals four druggable gut-brain axes with quantified mediation proportions (23–41%), from metabolite-mediated neuroinflammation to bacterial amyloid molecular mimicry. Cross-cohort validation across three WW-FINGERS trials ($n$=4,950) confirms robustness across populations. This work implements the precision-prevention framework, establishing the gut microbiome as a modifiable mechanistic layer connecting lifestyle interventions to brain health outcomes, and providing research-grade computational tools for the next generation of dementia prevention trials.

Submission Number: 111