Tempus Unveils Early Results from Multimodal Foundation Models, Demonstrating Potential to Accelerate Precision Oncology and Drug Development
Tempus a technology company focused on advancing precision medicine through artificial intelligence, has announced new findings from its multimodal foundation model research at the 2026 American Society of Clinical Oncology (ASCO) Annual Meeting. The results highlight the company’s progress in developing large-scale AI systems capable of generating clinically meaningful insights across oncology without requiring disease-specific retraining.
The announcement marks another step in Tempus’ broader strategy to leverage its vast repository of multimodal healthcare data to improve patient care, support drug development, and accelerate the discovery of new biomarkers and therapeutic opportunities. By combining genomic, clinical, imaging, and molecular data within a unified AI framework, Tempus aims to transform how healthcare organizations analyze complex patient populations and predict treatment outcomes.
Building a Comprehensive Multimodal Foundation Model
At the core of Tempus’ initiative is one of the industry’s largest healthcare data ecosystems. The company has accumulated more than 500 petabytes of molecularly grounded healthcare data, representing over 45 million de-identified patient records. Among these are approximately 1.5 million patients with sequenced genomic data and more than 400,000 cancer patient records containing integrated genomic, transcriptomic, imaging, and clinical information.
Using this extensive data infrastructure, Tempus has developed a multimodal transformer-based foundation model designed to understand patient trajectories across time and multiple data types. The latest version of the model was trained using approximately 2.5 million longitudinal patient records, incorporating more than 250 million pages of clinical documentation, 450,000 digitized medical images, and over 500,000 genomic and transcriptomic sequences.
Unlike traditional machine learning systems that are often developed for a single prediction task, Tempus’ multimodal architecture is designed to address thousands of clinical prediction objectives simultaneously. The model integrates information from several billion-parameter foundation models, creating unified patient representations that can be applied across numerous clinical and research scenarios.
According to the company, this approach dramatically reduces both the time and data requirements needed to generate clinically relevant insights. The system can be used for applications ranging from patient risk assessment and clinical trial design to biomarker discovery and the development of next-generation diagnostic tools.
Zero-Shot Learning Enables Rapid Insight Generation
One of the most significant aspects of Tempus’ research is the model’s ability to perform in a “zero-shot” setting. In artificial intelligence, zero-shot learning refers to a model’s ability to make predictions or generate insights without being specifically trained on the exact task being evaluated.
Tempus applied its patient trajectory model, along with a series of agentic AI workflows, to analyze a cohort of more than 1.2 million de-identified patient records containing rich multimodal data. Notably, the model was able to generate clinically valuable predictions without additional fine-tuning or task-specific training.
This capability could have significant implications for healthcare and life sciences organizations, as it may enable faster hypothesis generation and validation while reducing the need for extensive model development efforts for each individual research question.
Evaluating Outcomes in EGFR-Mutant Lung Cancer Patients
As a proof-of-concept study, Tempus evaluated patients with epidermal growth factor receptor (EGFR)-mutated non-small cell lung cancer (NSCLC) who were treated with osimertinib, a third-generation EGFR inhibitor that serves as a frontline standard-of-care therapy for many patients with advanced disease.
The objective was to determine whether the multimodal foundation model could accurately predict treatment outcomes and identify patients who may experience poorer responses despite receiving the current standard treatment. Importantly, researchers sought to assess whether the model could uncover predictive signals beyond known clinical and molecular biomarkers.
Without undergoing any additional task-specific training, the model demonstrated strong prognostic performance. It achieved a concordance index (C-index) of 0.802 for overall survival, indicating a high level of predictive accuracy. Statistical analysis showed a highly significant result, with a p-value below 0.001.
The model also successfully stratified patients into high-risk and low-risk groups. Researchers reported a hazard ratio of 4.536 between the two populations, suggesting a substantial difference in survival outcomes based on the model’s risk predictions.
Beyond overall survival predictions, the model continued to provide clinically relevant insights even after accounting for a wide range of established molecular biomarkers. The evaluation included more than 30 EGFR-specific genomic features, including common mutations such as L858R and exon 19 deletions.
The AI system maintained significant prognostic value across various clinical and molecular subgroups. For example, among patients carrying TP53 mutations, the model identified survival differences associated with a hazard ratio of 5.96. It also successfully stratified progression-free survival among patients without central nervous system metastases, generating a hazard ratio of 1.94.
These findings suggest that the model is capturing complex biological and clinical relationships that extend beyond currently recognized biomarkers, potentially enabling the discovery of previously unrecognized predictors of treatment response.
Potential Applications in Clinical Trial Development
In addition to patient risk prediction, Tempus highlighted the potential value of its multimodal foundation models for pharmaceutical research and clinical development programs.
Clinical trial design remains one of the most challenging and costly aspects of drug development. Selecting appropriate patient populations, identifying meaningful endpoints, and accurately predicting treatment outcomes can significantly influence a trial’s success.
Tempus reported that its patient trajectory models have demonstrated an ability to predict outcomes in cohorts designed to mirror landmark clinical studies in non-small cell lung cancer. Specifically, the company evaluated patient populations resembling those enrolled in major practice-changing trials, including KEYNOTE-189, FLAURA-2, and DESTINY.
According to the company, the multimodal foundation model consistently outperformed traditional statistical approaches such as Cox proportional hazards modeling in these retrospective evaluations. Because the model can generate predictions without extensive task-specific training, it may provide biopharmaceutical companies with a powerful tool for assessing clinical trial feasibility and identifying risk factors before patient enrollment begins.
The ability to anticipate trial outcomes and evaluate potential patient stratification strategies could help reduce development costs, improve study design, and increase the likelihood of successful therapeutic development.
Advancing the Future of Precision Medicine
Tempus believes that foundation models and agentic AI workflows represent a transformative advancement for precision medicine. By integrating diverse healthcare data into a single analytical framework, these technologies have the potential to accelerate the transition from research discovery to clinical implementation.
The company’s leadership views these systems as tools capable of supporting rapid hypothesis generation, identifying novel biomarkers, and uncovering treatment-response patterns that may otherwise remain hidden within complex healthcare datasets.
Eric Lefkofsky, Founder and Chief Executive Officer of Tempus, emphasized the broader implications of the technology, noting that foundation models combined with agentic workflows could significantly shorten the timeline between scientific discovery and practical clinical application. He also highlighted the encouraging performance of the company’s general-purpose multimodal models, which are already demonstrating results that exceed those achieved by many highly specialized predictive systems.
As Tempus continues to expand its multimodal data resources and refine its AI infrastructure, the company expects these foundation models to play an increasingly important role in supporting physicians, researchers, and life sciences organizations. Potential applications include improved clinical trial design, enhanced biomarker development, more accurate patient stratification, and the creation of next-generation diagnostics.
Transforming Healthcare Through Data-Driven Intelligence
Tempus’ latest ASCO presentation underscores a growing trend within healthcare: the convergence of large-scale multimodal data and advanced artificial intelligence. By transforming millions of patient records into actionable insights, foundation models may help address some of the most pressing challenges in oncology, including treatment selection, outcome prediction, and therapeutic discovery.
With one of the world’s largest multimodal healthcare datasets and an expanding portfolio of AI-enabled solutions, Tempus is positioning itself at the forefront of precision medicine innovation. The company’s early findings suggest that large-scale multimodal foundation models could become valuable tools for improving clinical decision-making and accelerating the development of more effective cancer therapies in the years ahead.
About Tempus
Tempus is a technology company advancing precision medicine through the practical application of artificial intelligence in healthcare. With one of the world’s largest libraries of multimodal data, and an operating system to make that data accessible and useful, Tempus provides AI-enabled precision medicine solutions to physicians to deliver personalized patient care and in parallel facilitates discovery, development and delivery of optimal therapeutics. The goal is for each patient to benefit from the treatment of others who came before by providing physicians with tools that learn as the company gathers more data
