Verily Health and NVIDIA Release Forecast 1.0, Integrating Genomics and EHR
Key Achievements of Forecast 1.0
Verily Health has publicly released Forecast 1.0, a multimodal model integrating electronic health records (EHR) and genomics, developed in collaboration with NVIDIA Corporation (NVDA) and the NIH All of Us Research Program. Adding polygenic risk scores (PRS) to EHR improved type 2 diabetes prediction performance by 15.5% and enabled evaluation of 5–10-year long-term risk. This differentiates it from existing EHR-only models by learning both static genetic information and time-varying clinical histories within a single model. However, the results are from risk stratification research and do not represent clinical trials validating therapeutic efficacy.
GPU-Driven Development Efficiency
Pre-training using NeMo AutoModel on NVIDIA H100 GPU was three times more efficient than Hugging Face Accelerate on the same hardware. This reduced computational costs and iterative development time when combining large-scale genomics and longitudinal EHR. The training and evaluation code for Forecast 1.0 is available on GitHub, and All of Us data usage is designed to be performed by registered researchers in a secure research environment. Platform competitiveness is increasingly determined not only by single-prediction accuracy but also by data access rights, computational efficiency, and reproducible deployment systems.
Connection to the Diabetes Treatment Market
The global diabetes drug market is projected to reach USD 101.48 billion in 2025, with type 2 diabetes expected to account for 95.62% of the market in 2026. Standard treatments include Ozempic (semaglutide, GLP-1 receptor agonist), Mounjaro (tirzepatide, GIP·GLP-1 dual receptor agonist), and Jardiance (empagliflozin, SGLT2 inhibitor). The U.S. FDA approved Ozempic for type 2 diabetes on December 5, 2017, Mounjaro on May 13, 2022, and Jardiance on August 1, 2014. Forecast 1.0 is not a competing therapeutic but a research platform supporting high-risk group identification, clinical trial patient recruitment, and biomarker discovery.
Commercial Value and Validation Challenges
Verily Health is set to become an independent company in 2026, no longer under Alphabet's control, though Alphabet will retain a meaningful minority stake. Therefore, Forecast 1.0 is interpreted as a differentiating asset for Verily's precision medicine platform, Pre, targeting hospitals, pharmaceutical companies, and insurers. Competitors include EHR-based models like BEHRT and Med-BERT, as well as vision and text-focused medical AI models, but Forecast directly combines genomic risk scores with long-term clinical histories. The current stage is a public research model and does not correspond to Phase 1·2·3 or FDA·EMA·PMDA drug approval stages. For real-world clinical adoption, external cohort validation and assessments of bias, privacy, and clinical utility are required.
Forecast 1.0 improves long-term type 2 diabetes prediction performance by 15.5% and pre-training efficiency by three times, offering a platform foundation that could reduce pharmaceutical companies' biomarker exploration and clinical trial patient selection costs. It could serve as an auxiliary infrastructure for early identification of appropriate patients for marketed therapies like Ozempic, Mounjaro, and Jardiance in the USD 101.48 billion global diabetes drug market in 2025. For researchers, the All of Us-based code and 5–10-year prediction framework provide a reproducible benchmark, while NVIDIA gains a pathway to expand GPU and NeMo demand in medical AI. Short-term value will emerge from public model adoption and partnerships with pharmaceutical and hospital entities, while long-term enterprise value will depend on external cohort validation and clinical utility proof at the regulatory level.
Source: BioPharma Dive (rss)