Models and datasets

Aidoc CARE is one radiology foundation model, pretrained on CT scans without labels, whose task-specific heads have each been FDA-cleared to flag urgent findings in emergency scans so radiologists read those first. Its oncology relevance is indirect, catching incidental masses; the regulatory evidence covers triage, not diagnostic accuracy for tumours.

Reads a million letters of DNA at once and predicts how a mutation changes gene regulation, splicing and chromatin.

Arc's growing library of cell data, the fuel for virtual cell models.

Atlas is a pathology foundation model trained on 1.2 million slides from two of the world's largest hospitals.

BioEmu is a Microsoft generative diffusion model that predicts the range of shapes a protein moves between, not one static structure, thousands of times faster than molecular dynamics simulation. For cancer drug discovery that can reveal transient pockets, as in KRAS, that static predictors miss, but its outputs are approximate and validated mainly on small proteins.

Turns a cell's gene expression into a sentence so a normal language model can reason about it; a 27-billion-parameter version proposed a cancer immunotherapy idea that was confirmed in the lab.

A model pretrained on 148,000 CT scans to segment organs and triage findings.

ESM3 is a generative protein model that designed a working fluorescent protein far from any natural sequence.

A DNA language model trained on 9.3 trillion bases that can flag cancer-causing BRCA1 variants without being told about them.

Midnight is a pathology model that matched the leaders while training on far fewer slides.

A model that reads slides and clinical text together to predict who will respond to immunotherapy.

Predicts how cells will respond to a drug or gene knockout, trained on over 100 million perturbed cells.

Tahoe-100M is the biggest single-cell dataset ever released, built to teach AI how cancer cells respond to drugs.

CZI's open cross-species cell models and a reasoning model trained on them.

Predicts the 3D shape of proteins together with DNA, RNA, small molecules and antibodies, the starting point for much modern drug design.

Open-source structure models that match AlphaFold 3, with Boltz-2 also predicting how strongly a drug binds.

CellFM is an 800-million-parameter single-cell model trained on 100 million human cells.

Structure and antibody-design models from Chai Discovery, with Chai-2 reporting high zero-shot antibody hit rates.

A pathology model trained across 19 cancer types that predicts survival and mutations from slides.

A model trained on millions of hospital records that forecasts a patient's next diagnoses.

An open 1.1-billion-parameter pathology model from a French startup, among the strongest on public benchmarks.

Hibou is a family of open pathology foundation models under a permissive licence.

Google's medical versions of its Gemini models, able to reason over text, images, and long records.

Adaptations of Meta's Segment Anything model that outline tumours and organs on any scan with a click.

Merlin is a model trained on 15,000 CT scans with their reports that can find and describe hundreds of findings.

Nicheformer is a model trained on both dissociated and spatial data so it learns how a cell's neighbourhood shapes it.

DNA language models trained on thousands of genomes for variant and regulatory prediction.

A model trained on billions of cell microscopy images to read what a drug or gene knockout does to a cell.

Owkin's open pathology models trained on TCGA and its federated hospital network.

PLUTO is PathAI's compact pathology foundation model, a vision transformer pretrained at several magnifications on 195 million tiles from 158,000 slides, so one network serves slide-level and biomarker quantification tasks at whatever resolution each needs. It runs inside PathAI's AISight product, but its weights are proprietary, so outside groups cannot benchmark or adapt it.

An open pathology model trained on 1.3 billion image tiles from a US health system, modelling whole slides at gigapixel scale.

scFoundation is a 100-million-parameter model trained on 50 million cells, from China's BioMap.

A GPT-style model for single-cell data that predicts cell types, perturbation responses, and gene networks.

Models trained on Tempus's paired genomic, pathology, imaging and outcome data to predict response and prognosis.

TITAN is a model that summarises a whole slide, not just tiles, and can write a draft pathology report.

Two open academic pathology models: UNI reads tissue images, CONCH links images with pathology text.

A pathology foundation model trained on millions of slides that can detect cancer and predict biomarkers from an ordinary H&E slide.

Scored all 71 million possible single-letter protein changes in humans as likely harmful or benign.

Geneformer is a transformer trained on about 30 million single cells that encodes each cell as a ranked list of its genes, so deleting a gene in silico shows which genes matter in a disease. It was the first single-cell foundation model in general use, though benchmarks find only modest gains over linear baselines on some tasks.

Uses text embeddings of gene descriptions from a general LLM to represent cells, and performs surprisingly well.

An open generalist model that answers questions about 2D and 3D scans.

The tools that design entirely new proteins to bind a chosen target, now used for cancer binders and antibodies.

Predicts a person's six-year lung cancer risk from one low-dose CT, even when no nodule is visible.

Universal Cell Embedding maps any cell from any species into one shared space without retraining.

Models that predict how DNA sequence controls gene activity, used to interpret non-coding cancer mutations.

Reads a mammogram to estimate five-year breast cancer risk, consistently across races and devices.

A digital twin is a computer model of one patient's tumour and body, updated with each scan and blood test, used to forecast how the disease will respond to each option before it is tried.

America's answer to UK Biobank, built for diversity.

AACR Project GENIE is real-world tumour sequencing data shared by leading cancer centres.

Which genes each cancer cell line cannot live without. The map of synthetic-lethal targets.

Mathematics from population genetics shows that resistant cells almost always exist before treatment in large tumours, and that combining drugs with different resistance mutations from the start can succeed where the same drugs in sequence fail.

Growing tumours compress themselves and their surroundings; models of this solid stress explain collapsed vessels, poor drug delivery and stiff stroma, and point to drugs that soften the tumour so treatment can get in.

Cancer cells are treated as players whose success depends on what neighbouring cells do, which lets researchers predict how a tumour's mix of cell types shifts under treatment and design schedules that steer it.

Mechanistic computer models that join a drug's pharmacology to the biology of the tumour and the body, used by developers and regulators to pick doses, predict combinations and explain why a trial failed.

The public archive of cancer scans that most radiology AI is trained and tested on.

Instead of hitting a tumour as hard as possible, adaptive therapy gives just enough drug to keep it in check and stops when it shrinks, so drug-sensitive cells survive to compete with resistant ones. A pilot trial in prostate cancer roughly doubled the time to progression on abiraterone.

The reference atlas of cancer genomes that most cancer biology since 2008 is built on.

UK Biobank is the richest population cohort for linking genes, blood and imaging to who later develops cancer.

Instead of equations for average behaviour, agent-based models simulate every cell as an individual with rules for dividing, moving, dying and signalling, producing virtual tumours in which immune attack, drug delivery and evolution can be watched and tested.

The speed at which a molecular marker falls during treatment predicts outcome better than a single level: BCR-ABL halving time in chronic myeloid leukaemia and circulating tumour DNA slopes in solid tumours are now used to judge response within weeks.

Gliomas grow by both dividing and migrating through the brain, and a two-parameter equation fitted to a patient's MRI scans estimates how far invisible cells have spread, which can guide how much brain to irradiate and how fast the tumour will grow.

Models of how tumours recruit blood vessels, and of Rakesh Jain's idea that anti-angiogenic drugs at the right dose normalise rather than destroy those vessels, improving drug and oxygen delivery for a window of days.

Predator-prey style equations describe how immune cells hunt tumour cells, and they reproduce dormancy, escape and the delayed, sometimes explosive, responses seen with immunotherapy; they now help design combination and scheduling trials.

Curves that turn a radiation dose into a probability: how likely the tumour is to be eradicated and how likely a nearby organ is to be damaged. They underlie dose constraints, dose escalation trials and comparisons between treatment plans.

Resistant cells arise by chance mutation as a tumour grows, so the chance of a cancer already containing resistant cells rises with its size. The model argued for treating early and for alternating non-cross-resistant drugs.

Because tumours regrow fastest when small, the best way to finish them is to give the same chemotherapy doses closer together. The idea, from Larry Norton and Richard Simon, was proved in breast cancer by the CALGB 9741 trial and made two-weekly chemotherapy a standard.

Peter Nowell's 1976 idea that a tumour is an evolving population of competing clones is now measured directly by sequencing several regions or repeated blood samples, and models of branching evolution predict which clones will drive relapse.

Radiotherapy works through repair, redistribution, reoxygenation and repopulation between fractions; the discovery that tumours speed up their regrowth during a course explained why gaps in treatment cost cures and led to accelerated schedules.

Equations that describe how a drug's concentration rises and falls in the body and how that concentration translates into effect and toxicity; the reason doses are given per square metre, why some drugs are infused over days, and how children's doses are set.

Tumours do not grow exponentially forever: growth slows as they enlarge, following a curve Benjamin Gompertz devised for human mortality in 1825 and Anna Kane Laird fitted to tumours in 1964. It explains why small tumours are the most chemosensitive and why doubling times lengthen.

A dose of chemotherapy kills a constant fraction of cancer cells, not a constant number, so each cycle removes the same proportion, which is why treatment continues after the tumour has disappeared from scans.

Mathematical oncology writes down how tumours grow, evolve, respond to treatment and interact with the immune system as equations or simulations, then uses them to design doses, schedules and trials. The models themselves are records, each with what it was fitted to and what it was used to decide.

Chemotherapy doses are usually written per square metre of body surface, a convention from 1958 that scales drug clearance between species and people; it is imprecise, and for many newer drugs flat or weight-based doses have replaced it.

How long a tumour takes to double in volume, measured from two scans; it separates cancers from benign nodules in lung screening, sorts aggressive from indolent disease and estimates how long a tumour has been present.

From Paget's seed-and-soil idea to models that estimate when metastases were seeded from a primary and how long they lay dormant, these frameworks explain late relapse and argue for treating micrometastases early.

Find a mouse or dish model that matches a tumour type or mutation.

cBioPortal lets you browse the mutations, copy number, and expression of tens of thousands of tumours without writing code.

CIViC is an open, Wikipedia-style database of what cancer mutations mean for treatment.

COSMIC, the Catalogue of Somatic Mutations in Cancer, is the largest curated catalogue of mutations found in tumours, run by the Wellcome Sanger Institute. It also holds the Cancer Gene Census of around 750 genes, the reference mutational signatures and a cell line project; academic use is free, industry needs a licence.

CPTAC measures the proteins, not just the genes, of thousands of tumours.

CELLxGENE and the Human Cell Atlas hold single-cell data from healthy and diseased tissue, browsable and downloadable.

Real-world evidence at scale: what happened to patients with a given genomic profile on a given treatment.

The Human Protein Atlas shows where in the body each protein is found, which tells you whether an ADC target is safe.

NCI CDAS holds the lung screening trial images that trained Sybil and most lung-nodule AI.

The Imaging Data Commons is TCIA in the cloud, ready for large-scale model training.

Tells you, for a given mutation, whether an approved or investigational drug exists and how strong the evidence is.

Open Targets scores how strongly each gene is linked to each disease, with the evidence behind it.

Pathology AI benchmarks are the open challenge datasets on which every pathology model is scored: CAMELYON16 and 17 for lymph node metastasis detection, PANDA for prostate grading with 11,000 biopsies, and TCGA slide-level tasks used to compare foundation models. Licences are mostly CC BY-NC-SA or set per challenge.