Tempus AI's multimodal foundation model, trained on 2.5 million patient records and 450,000 medical images, predicted overall survival with a C-index of 0.802 in a zero-shot analysis of EGFR-mutant lung cancer patients — outperforming standard approaches without any fine-tuning.
Tempus AI Inc. (NASDAQ: TEM) presented initial results from its multimodal foundation model efforts at the 2026 American Society of Clinical Oncology Annual Meeting in Chicago, demonstrating the ability to generate clinically actionable insights from more than 500 petabytes of molecularly grounded data. The transformer-based model, trained on 2.5 million longitudinal records encompassing over 250 million pages of clinical notes, 450,000 digitized medical images, and 500,000 genomic and transcriptomic sequences, is designed to address thousands of prediction objectives anchored in overall survival and progression-free survival without requiring additional data or model fine-tuning.
"Our general purpose models are already outperforming highly tuned smaller models, which bodes well for the ability of our novel biological multimodal foundation models to improve clinical trial design and biomarker development," said Eric Lefkofsky, Founder and Chief Executive Officer of Tempus.
In a primary proof of concept, the model analyzed EGFR-mutant non-small cell lung cancer patients treated with osimertinib, the frontline standard of care third-generation EGFR inhibitor. Without any pre-specified training, the model achieved a C-index of 0.802 for overall survival (p value < 0.001) and produced a hazard ratio of 4.536 (95% CI: 3.289 to 6.255) between high- and low-risk subgroups. The results held independent of molecular and clinical subgroups, significantly stratifying survival in TP53-positive patients (HR of 5.96) and progression-free survival in patients without central nervous system metastasis (HR of 1.94).
The company also demonstrated the model's utility for drug development by successfully predicting outcomes of patient cohorts mirroring three practice-changing clinical trials — KEYNOTE-189, FLAURA-2, and DESTINY — outperforming standard Cox proportional hazards modeling in a zero-shot setting. Tempus said the architecture significantly reduces the time and data required to produce hundreds of clinically relevant insights for clinical trial design, patient risk prediction, and novel multimodal diagnostics.
Data Scale and Competitive Positioning
Tempus has built one of the largest multimodal datasets in precision medicine, with more than 45 million total de-identified patient journeys, 1.5 million with sequenced data, and over 400,000 cancer records with full genomic, transcriptomic, imaging, and clinical data. The company's data library exceeds 500 petabytes, giving it a scale advantage over competitors such as BostonGene, which presented nine abstracts at ASCO 2026 covering its own AI-powered tumor and immune biology models.
BostonGene's Kassandra cell deconvolution platform and Tumor Portrait test, developed in collaboration with MD Anderson Cancer Center and the Parker Institute for Cancer Immunotherapy, focus on tumor microenvironment classification and immunotherapy response prediction. Tempus's approach differs by using a single transformer-based foundation model trained across multiple data modalities to generate insights without task-specific fine-tuning.
Regulatory and Financial Milestones
The foundation model announcement follows Tempus's receipt of FDA approval for a tumor-only indication of its xT CDx next-generation sequencing platform on May 29, making it the first laboratory to hold FDA companion diagnostic approval for both tumor-only and tumor-normal comprehensive genomic profiling. The 648-gene tissue-based assay previously required a matched normal sample; the expanded label allows testing when blood or saliva specimens are not viable.
Chief Financial Officer Jim Rogers said the approval enables Tempus to migrate its entire DNA solid tumor portfolio to FDA-approved assays under unified ADLT pricing, with an estimated $200 average selling price benefit beginning in 2027.
Investor Implications
Tempus shares face a dual narrative: the foundation model results validate the company's AI capabilities and data moat, while the FDA approval provides a clear near-term revenue catalyst through ASP expansion. The company's ability to generate predictive insights from its multimodal data without task-specific training could reduce drug development costs for biopharmaceutical partners, though the technology remains early-stage with no disclosed commercialization timeline for the foundation model itself. Tempus did not disclose the training cost for the model or provide independent validation of its benchmark results.
This article is for informational purposes only and does not constitute investment advice.