About Standard Model Biomedicine

What are Standard Models?

A fundamentally different way to build a world model. We believe no one company can build the data and powerful AI needed to feed the coming Bio AI revolution in a silo. Standard Models are different, we build alongside the leading LLM AI, top foundation models, and patient atlases, then further power them with proprietary data and a bespoke model architecture that creates a unifying joint embedding of “all” — assembling a nimble, scalable AI architecture for encoding and decoding patient data.

Standard Models are world models whose output are digital representations of patients; predicting and projecting how their unique biology will change in response to treatments. Allowing our AI to effectively model and modulate many aspects of clinical trial design years before their clinical endpoints.

Standard Models are world models of human biology and clinical trials, which:

  • Model clinical trial design, allowing every potential trial input to be modeled against the others in a phase zero setting
  • Provide in silico digital twins that predict patients’ journeys across indication, disease and time
  • Project treatment outcomes, disease progression times on treatment & overall survival
  • Provide prognostic suggestions of the next best treatment

The Standard Model Biomedicine has built one of the world's first Joint Embedding Predictive Architectures, “JEPA” of the patient. First, merging the top performing “built for purpose” foundation models, into a single patient representation. Then, training the new multimodal Standard Model with our massive corpora of proprietary data, including 8 modalities; DNA, RNA, CT images, whole slide images, and other biomarkers to further refine the model. Finally, the Standard Model encodes and decodes each patient's digital twin benefitting from the predictive and prognostic data from the longitudinal patient electronic records from millions of patients already encoded in the model.

How Standard Model Works

Put simply: the Standard Model turns everything generated for one patient: assessment, imaging, pathology, genomics, and interventions into a single patient-level representation, then predicts what will happen under any intervention. Our users then adapt it to their relevant clinical endpoints: survival, response, toxicity, with a lightweight task head trained on their own labels, without the need to retrain the underlying Standard Model.

The Standard Models

Standard Model-V1 is a lightweight open-weight patient encoder and modality-specific decoder for EHR, CT, genomics, and proteomics on Hugging Face, alongside the methods, models, and clinical evaluations underwriting the platform. You can be running the open EHR model on your own machine in about a minute — see install & quickstart.

Standard Model-V2 (available through early access) is a JEPA model; matrixing multimodal encoders into a standard model, powered by partnerships with leading pharma and academic institutes. SM-V2 enables indication screening, dose optimization, tox prediction, regulatory preparation, outcome prediction and zero shot inference. Standard Model-V2 is accessed through a lightweight user agreement, and deployed directly in our user’s AWS or other secure environment. Unlike other world models available for license, SMB users have the flexibility to use the model off-the-shelf, or to fine-tune the Standard Model weights.

SMB-Trial Morph: AI Statistician that optimizes site selection, I/E criteria, enrollment, patient stratification, virtual control arms, response monitoring, AE prediction, imaging analysis, clinical decision support, dose response simulation, and other pipeline go/no-go decisions.

SMB-Observer: monitors phase IV studies, LoT sequencing, pharmacovigilance, indication expansion, post-approval changes, and enables users to quickly identify new market fits, analyze retrospective concordance, identify new indications & combinations, and for repurposing exercises.

Who we are

We are a team of researchers and engineers drawn from leading academics and pharma (BMS, Genentech, AbbVie, Sanofi, Stanford, Tempus, MIT, Yale, UCSD, MD Anderson, and Mount Sinai) with deep understanding of clinical workflows and published work spanning computational neuroscience, clinical foundation models, and federated learning for drug development.

Standard Model Biomedicine is based in San Francisco, California.

For partnerships, model access, or press contact us at info@standardmodel.bio.