BiobankFM is an open-source multimodal foundation AI model that learns from population-scale human biobanks—including UK Biobank, NIH All of Us, FinnGen, BioBank Japan, and other large cohorts—to enable disease prediction, therapeutic target discovery, biomarker identification, patient stratification, and precision medicine.
UK Biobank (UKB) and NIH All of Us frontier AI models are next-generation biomedical foundation models trained on large-scale, multimodal population datasets—including genomics, electronic health records, medical imaging, laboratory tests, wearables, proteomics, metabolomics, and longitudinal clinical outcomes—to learn a general representation of human biology rather than perform a single prediction task. Analogous to how GPT serves as a foundation model for language, these models aim to become foundation models for human health that can be adapted for disease risk prediction, therapeutic target discovery, biomarker identification, patient stratification, clinical trial optimization, and precision medicine.
UK Biobank contributes exceptionally deep longitudinal phenotyping and imaging data from approximately 500,000 participants, while All of Us provides greater ancestral diversity and comprehensive U.S. clinical data from more than one million participants, making them complementary resources for building AI that generalizes across populations.
Current biomedical AI is largely fragmented into task-specific models that fail to capture the interconnected nature of human biology across diseases, populations, and data modalities. A single foundation model can learn universal biological representations from millions of patient-years of data and transfer that knowledge across many downstream applications, eliminating the need to build separate models for each disease or research question. This enables more accurate target identification, earlier disease prediction, biomarker discovery, patient stratification, treatment response prediction, and safer, more efficient clinical trials, ultimately improving decision-making throughout drug discovery and development.
Building such a model requires integrating heterogeneous data into a unified patient-centric representation:
- Data harmonization — After obtaining access to UKB and/or All of Us, data from genomics, EHR, laboratory measurements, imaging, multi-omics, wearables, and clinical outcomes are harmonized into longitudinal patient timelines.
- Modality encoders — Specialized encoders, such as transformers for genomics and EHR, vision transformers for imaging, and modality-specific neural networks for molecular data, are used to generate representations that are fused through multimodal transformer architectures.
- Self-supervised pretraining — The model is pretrained using self-supervised learning objectives, including masked prediction, contrastive learning, and temporal forecasting, allowing it to learn general biological patterns without relying on disease-specific labels.
- Fine-tuning and validation — The model is fine-tuned for applications such as disease prediction, therapeutic target prioritization, biomarker discovery, and clinical decision support, with rigorous validation across independent cohorts (e.g., training on UK Biobank and testing on All of Us) to ensure robustness, fairness, and generalizability.
These frontier AI models convert massive, fragmented biomedical datasets into a reusable decision-making engine for human biology. Instead of repeatedly developing isolated AI models, researchers and biopharma organizations can leverage a single pretrained foundation model across virtually every stage of drug discovery and development—from target discovery and validation to clinical trial design and portfolio prioritization. This improves predictive performance, reduces dependence on labeled datasets, enhances generalization across diseases and populations, accelerates scientific discovery, shortens development timelines, reduces R&D risk, and ultimately increases the probability of developing safe and effective medicines.
For pharmaceutical companies, these models represent a strategic shift from AI solving individual problems to AI serving as a foundational platform for data-driven drug discovery and precision medicine.