Foundation Models in Precision Oncology: Where the Evidence Actually Is

Foundation models are now being pulled into precision oncology from four directions at once: single-cell biology, digital pathology, radiology, and narrative clinical records. The word "foundation" makes those efforts sound more unified than they are.

They are not one market. They are not one evidence base. They are not even one kind of clinical claim.

The useful question is not whether foundation models will matter in oncology. They will. The useful question is where the evidence is already strong enough to shape product design, where it is still preclinical, and where the term is being used as a valuation wrapper.

For the broader AI literacy layer, this is exactly why the /course surface starts with model capability and clinical boundaries before use-case enthusiasm. Precision oncology is a tempting place to overclaim because the need is real and the data are complex.

The first divide: biology, tissue, image, and text

In genomics and single-cell biology, the strongest foundation-model argument is representation learning. Gene expression, perturbation response, cell-state transitions, and pathway structure are too high-dimensional for simple feature engineering to carry the field indefinitely.

Geneformer, published in Nature, pretrained on roughly 30 million single-cell transcriptomes and showed that transfer learning can support network-biology predictions in data-limited settings (Nature). scGPT pushed the same direction for single-cell multi-omics, framing generative pretraining as a way to learn reusable cell and gene representations (Nature Methods).

That is real signal. It does not mean a model can choose therapy. It means oncology research workflows can gain better priors for mechanism, perturbation, and cell-state structure.

Digital pathology is the second strong lane. CONCH used more than one million pathology image-caption pairs to train a visual-language foundation model for computational pathology (Nature Medicine). Virchow scaled pathology pretraining further and reported strong pan-cancer and rare-cancer detection results across multiple cancer types (Nature Medicine).

This is closer to clinical workflow than single-cell modeling, but the same boundary applies: performance on benchmark tasks is not the same as a deployed, monitored, liability-bearing diagnostic service.

Radiology foundation models are promising but less settled. RadFM and related work show the shape of generalist radiology modeling across 2D and 3D medical data (arXiv). The practical problem is not just model capability. It is local scanner variation, acquisition protocol, report-label quality, and the hard requirement that the model behave under real workflow pressure.

Narrative models are the broadest and most dangerous lane. LLMs can summarize, retrieve, explain, and draft. In oncology, that creates immediate value around trial matching, prior authorization packets, survivorship instructions, and molecular-tumor-board preparation. But narrative fluency is a weak proxy for clinical correctness. The model can sound like an oncologist while missing the decision logic that actually matters.

Where the evidence is real

The strongest current evidence supports foundation models as infrastructure, not autonomous decision-makers.

They can compress high-dimensional biological data into reusable representations. They can reduce the amount of labeled data needed for a downstream pathology or imaging task. They can make retrieval across complex oncology records less brittle. They can help create evaluation harnesses that expose where smaller task-specific systems fail.

That is enough to matter.

It is not enough to skip the normal ladder of clinical validation. A precision-oncology system still needs locked use cases, dataset provenance, test-set separation, local validation, drift monitoring, and human review. The more the output touches treatment selection, staging, diagnosis, prognosis, or toxicity management, the more the evidence burden rises.

NCCN-style evidence discipline is still the right mental model: level of evidence and degree of consensus matter. A model paper is not a guideline. A benchmark is not a treatment recommendation. A retrospective slide dataset is not prospective clinical utility.

The false shortcut

The trap is to treat foundation models as a way around data quality.

They are not. They are data quality multipliers.

If the tumor registry is incomplete, the sequencing report is trapped in a PDF, the line-of-therapy field is inferred inconsistently, and the pathology report uses local language that does not map cleanly to structured concepts, a foundation model can make the surface look more intelligent while preserving the underlying error.

That matters because precision oncology is not a generic prediction problem. It is a chain of evidence:

• What is the diagnosis?
• What is the stage and treatment context?
• What molecular alteration is present?
• Was the assay adequate?
• What prior therapies have been given?
• What guideline, label, trial, or biological rationale supports the next step?
• What patient-specific constraint changes the option set?

A model that helps with one link in that chain is valuable. A model that blurs the links is unsafe.

The next 12 months

The most credible near-term deployments will not look like "AI oncologist" products. They will look like narrow infrastructure inside expert workflows.

The first lane is pathology pre-screening and triage: not replacing pathologists, but prioritizing cases, retrieving similar patterns, and improving biomarker workflow efficiency.

The second lane is molecular-tumor-board preparation: extracting structured context, linking alterations to evidence sources, and making the discussion more complete before the meeting begins.

The third lane is trial matching: using narrative and structured data to reduce missed eligibility signals, while keeping final eligibility review human.

The fourth lane is administrative precision oncology: prior authorization, documentation completeness, and payer evidence packets. This connects directly to the argument in Agentic Prior Authorization in Oncology: if the model can assemble evidence faster without inventing evidence, it can reduce delay without practicing medicine.

What builders should do

The builder's posture should be conservative but not timid.

Use foundation models where they improve representation, retrieval, summarization, or structured packet assembly. Do not use them as black-box treatment arbiters. Keep every claim attached to a source object. Keep PHI boundaries explicit. Separate educational surfaces from clinical systems. Validate against local data before using outputs in a real clinical workflow.

The center of gravity is shifting. Precision oncology will become model-mediated because the data burden has outgrown unaided human workflow. But the systems that win will not be the ones with the largest "foundation model" label. They will be the ones that preserve the chain of evidence from tissue to treatment decision.

That is where the evidence actually is.

---

Editorial boundary: This article is educational analysis for clinicians and health-IT leaders. It is not medical advice, does not recommend care for any individual patient, and uses no PHI.