What 2 Million Downloads Taught Us About The Usage of Foundation Models for Biology

At MIT, researchers compared pathology foundation models to predict progression-free survival across more than 2,600 patients with breast, lung and colon cancer. H‑Optimus‑1 was the top-performing model, with time-dependent AUCs approaching 0.8. The most powerful thing is not only the result: choosing the right foundation model allowed the team to stop spending time on the backbone and move forward with the actual clinical research. Read the MIT case study.

At Oxford, researchers used H‑Optimus to predict spatial gene-expression signatures from routine H&E slides in inflammatory bowel disease. The goal was very practical: identify the most informative slides and regions before spending time and money on expensive spatial transcriptomics assays. Read the Oxford case study.

At a European medical centre, researchers used H‑Optimus‑1 as the foundation for a new diagnostic test in solid-tumour oncology. In external cohorts, the test achieved an AUC of 0.95 and showed the potential to reduce molecular testing by more than 75%. It is now being prepared for commercial deployment as a diagnostic. More to come on this very soon!

These examples show the breadth of high impact patient use-cases but also the potential for commercial opportunities to arise out of pioneering research based on open models.

Benchmarking was the most common declared use case, and for good reason. Independent comparison creates trust. H‑Optimus‑1 is currently ranked first and H‑Optimus‑0 third on PathBench, one of the most comprehensive evaluations in the field, covering 229 clinically relevant tasks.

But even 229 tasks capture only part of what these models can do.

Most established benchmarks were designed for familiar pathology problems: classification, detection, molecular prediction and survival. They are less able to evaluate the functionality emerging from newer models – connecting morphology to omics, combining different biological modalities, selecting the next experiment, adapting efficiently to a small proprietary dataset or becoming part of an agentic research workflow.

This is going to become a bigger issue. If a new model does not fit neatly into an existing benchmark, that does not necessarily mean it is less useful. It may mean the benchmark was designed for the previous generation of models.

We still need rigorous comparison. But we also need new ways to measure whether a model helps scientists ask better questions, generate new biological insight or reach an answer faster.

The range of applications has reinforced our belief in general-purpose foundation models.

The same model can provide embeddings for a segmentation workflow, support biomarker discovery, help predict patient outcomes or become the backbone of a more specialised model. With fine-tuning and additional data, it can be adapted to a specific disease or institution. Through inference, it can extract useful representations from slide collections that were previously very difficult to analyse at scale.

We see this reuse across gastrointestinal, breast, thoracic, genitourinary, gynaecological and skin cancers, as well as CNS tumours, haematological malignancies and sarcomas. Applications also extend beyond oncology to inflammatory and immune-mediated diseases, fibrosis, kidney disease, infectious disease and cardiovascular conditions and even outside of human biology.

That creates the potential for groundbreaking results. But we have also learned that the most valuable applications are not always the most glamorous.

Predicting a new biomarker is exciting. But helping a researcher choose the right 20 regions for an expensive experiment can save months. Automatically excluding unusable slides can unblock an entire data pipeline. A strong feature extractor can prevent every lab from having to rebuild the same foundational capability from scratch.

Often, the “boring” workflow problem is the one that determines whether ambitious science can happen at all.

We learnt where researchers repeatedly struggle: limited labelled data, expensive molecular assays, scarce linked clinical data, scanner and staining variation, compute constraints, and models that are technically available but still difficult to integrate.

We also saw the gaps in our own models. Performance is not equally strong across every tissue or application. Rare biological signals remain difficult. Researchers often need better tooling, documentation and deployment options as much as they need a better checkpoint.

Releasing the models did more than distribute our work. It gave us a much clearer map of what the field still needs.

Perhaps the clearest lesson is that a useful model disappears into the workflow.

It becomes a line in a pipeline, an encoder inside another model, a recurring step in an experiment or a shared tool passed between research teams. By the time an institution starts thinking formally about deployment, governance or infrastructure, the model may already be used across several departments.

That is how H‑Optimus has spread: not through one top-down decision, but through thousands of individual decisions by researchers who found it useful.

The 2 million downloads are one measure of that journey. The more meaningful measure is how much work now has H‑Optimus somewhere inside it, and what researchers have been able to build from there.

None of this would exist without the patients. Every H‑Optimus model is built from tissue samples that people gave in the course of their own care, often at some of the hardest moments of their lives. Millions of these anonymised slides, carefully prepared and digitised by pathologists and hospital teams, are what we learn from: the variety in the patterns of human disease is what turns a neural network into a general representation of tissue. That representation now helps researchers in more than a thousand institutions find biomarkers, choose better experiments and design new diagnostics. We want to thank every patient whose data made this possible, along with the clinicians, pathologists and partner institutions who look after that data with care and consent. What started with them now returns to patients everywhere, as faster research, better diagnostics and medicines that reach the people who need them.