{"id":"ai-oncology-roadmap","name":"AI in oncology roadmap: pattern readers → foundation models → agents in the workflow","route":"/roadmaps/ai-oncology-roadmap/","eras":[{"era":"1998-2016","title":"Computer-aided detection","description":"The first cleared cancer AI was computer-aided detection for mammography in 1998, which marked suspicious regions for the radiologist and, in large observational studies, did not improve accuracy. Rule-based decision support for treatment recommendations was tried and mostly abandoned. The lesson that survived: an algorithm has to be evaluated on the decision it changes, not on the pattern it finds.","status":"historic","refs":[{"id":"radiology-ai-screening","kind":"technology","name":"AI in radiology","route":"/technologies/radiology-ai-screening/","status":"established","tldr":"Software that reads scans alongside radiologists, catching cancers earlier and predicting who is at risk."},{"id":"mammography","kind":"technology","name":"Mammography & tomosynthesis","route":"/technologies/mammography/","status":"standard-of-care","tldr":"Low-dose breast X-ray used for screening. Newer 3D versions find more cancers with fewer false alarms."}],"trials":[],"papers":[]},{"era":"2017-2022","title":"Deep learning reaches cleared devices","description":"Convolutional networks trained on labelled images matched specialists on narrow tasks. Paige Prostate (2021) became the first FDA-authorised AI for reading pathology slides; radiology triage tools for haemorrhage and embolism were cleared by the dozen; Sybil and Mirai predicted future lung and breast cancer from today's scan. Whole-slide scanning became routine in large centres, which made slide-level AI possible at all.","status":"historic","refs":[{"id":"paige-prostate","kind":"drug","name":"Paige Prostate Detect","route":"/drugs/paige-prostate/","status":"approved","tldr":"The first AI for reading biopsy slides authorised by the FDA, which points pathologists to prostate cancer they might otherwise miss."},{"id":"paige","kind":"company","name":"Paige AI","route":"/companies/paige/","tldr":"MSK spin-out with the first FDA-cleared AI pathology product and the Virchow foundation model."},{"id":"digital-pathology-ai","kind":"technology","name":"Digital pathology & AI","route":"/technologies/digital-pathology-ai/","status":"established","tldr":"Scanning microscope slides and letting software measure things a pathologist cannot see, including predictions of who will benefit from a treatment."},{"id":"whole-slide-scanners","kind":"technology","name":"Whole-slide scanners and image management","route":"/technologies/whole-slide-scanners/","status":"established","tldr":"The scanners that turn glass slides into gigapixel images, and the software that stores and serves them, without which pathology AI cannot run."},{"id":"sybil","kind":"technology","name":"Sybil (MIT/MGH lung cancer risk from CT)","route":"/technologies/sybil/","status":"emerging","tldr":"Predicts a person's six-year lung cancer risk from one low-dose CT, even when no nodule is visible."},{"id":"mirai","kind":"technology","name":"Mirai (MIT breast cancer risk from mammograms)","route":"/technologies/mirai/","status":"emerging","tldr":"Reads a mammogram to estimate five-year breast cancer risk, consistently across races and devices."},{"id":"aidoc","kind":"company","name":"Aidoc","route":"/companies/aidoc/","tldr":"Radiology AI company with the first FDA-cleared foundation-model triage platform (CARE, January 2026); oncology-relevant for incidental findings and workflow."}],"trials":[],"papers":[]},{"era":"2020-2026","title":"Structure prediction and generative design","description":"AlphaFold made protein structure a lookup rather than a two-year experiment; AlphaFold 3 (2024), Boltz and Chai extended it to drug-protein and antibody complexes, and RFdiffusion and ESM3 design proteins that never existed. Insilico's generative chemistry produced the first AI-discovered drug to reach phase 2, Isomorphic's first oncology candidate was cleared for trials, and Recursion and Xaira are betting that image and perturbation data can find targets no hypothesis would. None has yet produced an approved cancer drug, which is the honest benchmark.","status":"current","refs":[{"id":"alphafold3","kind":"technology","name":"AlphaFold 3","route":"/technologies/alphafold3/","status":"established","tldr":"Predicts the 3D shape of proteins together with DNA, RNA, small molecules and antibodies, the starting point for much modern drug design."},{"id":"boltz","kind":"technology","name":"Boltz-1 / Boltz-2 (MIT, open)","route":"/technologies/boltz/","status":"emerging","tldr":"Open-source structure models that match AlphaFold 3, with Boltz-2 also predicting how strongly a drug binds."},{"id":"chai-1","kind":"technology","name":"Chai-1 / Chai-2","route":"/technologies/chai-1/","status":"emerging","tldr":"Structure and antibody-design models from Chai Discovery, with Chai-2 reporting high zero-shot antibody hit rates."},{"id":"rfdiffusion","kind":"technology","name":"RFdiffusion / RFdiffusion2 and ProteinMPNN (Baker Lab)","route":"/technologies/rfdiffusion/","status":"emerging","tldr":"The tools that design entirely new proteins to bind a chosen target, now used for cancer binders and antibodies."},{"id":"esm3","kind":"technology","name":"ESM3 (EvolutionaryScale)","route":"/technologies/esm3/","status":"emerging","tldr":"ESM3 is a generative protein model that designed a working fluorescent protein far from any natural sequence."},{"id":"ai-drug-design","kind":"technology","name":"AI-driven drug & target discovery","route":"/technologies/ai-drug-design/","status":"phase-2","tldr":"Using machine learning to pick targets, design molecules and antibodies, and predict which ADC will work."},{"id":"chemistry42","kind":"technology","name":"Chemistry42 and Pharma.AI (Insilico)","route":"/technologies/chemistry42/","status":"emerging","tldr":"Generative chemistry platform behind the first AI-discovered drug to reach phase 2, plus oncology candidates."},{"id":"insilico-medicine","kind":"company","name":"Insilico Medicine","route":"/companies/insilico-medicine/","tldr":"Generative-AI drug discovery company, listed in Hong Kong in December 2025, with a pan-KRAS candidate and a pan-TEAD inhibitor in the clinic."},{"id":"isomorphic-labs","kind":"company","name":"Isomorphic Labs","route":"/companies/isomorphic-labs/","tldr":"Alphabet's AlphaFold-derived drug design company; its first AI-designed oncology candidate was cleared for human trials in January 2026 after a $2.1B raise."},{"id":"recursion","kind":"company","name":"Recursion Pharmaceuticals","route":"/companies/recursion/","tldr":"Recursion is an AI-first biotech (merged with Exscientia in 2024) with a clinical oncology pipeline that includes an RBM39 degrader and a MEK inhibitor for familial adenomatous polyposis."},{"id":"phenom-2","kind":"technology","name":"Phenom-2 and Recursion OS","route":"/technologies/phenom-2/","status":"emerging","tldr":"A model trained on billions of cell microscopy images to read what a drug or gene knockout does to a cell."},{"id":"xaira-therapeutics","kind":"company","name":"Xaira Therapeutics","route":"/companies/xaira-therapeutics/","tldr":"Launched in 2024 with over $1 billion to build AI-native drug discovery from Baker-lab protein design."},{"id":"google-deepmind","kind":"company","name":"Google DeepMind (and Google Research)","route":"/companies/google-deepmind/","tldr":"Google DeepMind built AlphaFold, AlphaMissense, AlphaGenome and Med-Gemini, the reference models for structure, variants, and medical multimodal reasoning."}],"trials":[],"papers":[]},{"era":"2023-2026","title":"Foundation models and the first predictive tests","description":"Pathology models pretrained on millions of slides (Virchow, Prov-GigaPath, UNI and CONCH, H-optimus, TITAN) predict mutations, biomarkers and outcomes from a routine stain; MUSK adds clinical text. ArteraAI Prostate (2025) was the first AI test cleared to predict benefit from a treatment, and ArteraAI Breast followed in 2026. MASAI gave the first randomised evidence that AI-supported screening finds more cancers with less workload; Aidoc CARE (January 2026) was the first foundation-model triage platform cleared. Single-cell models (Geneformer, scGPT, State) and the Tahoe-100M dataset began the same arc for biology.","status":"current","refs":[{"id":"pathology-foundation-model","kind":"technology","name":"Pathology & radiology foundation models","route":"/technologies/pathology-foundation-model/","status":"emerging","tldr":"Pathology and radiology foundation models are AI networks pretrained without labels on over a million slides or scans (Virchow used 1.5 million), then adapted with small task heads to predict mutations, prognosis or treatment response from routine images. They power the FDA-cleared ArteraAI tools, but validation across hospitals and how regulators treat general-purpose models remain unsettled."},{"id":"virchow","kind":"technology","name":"Virchow / Virchow2 (Paige, MSK)","route":"/technologies/virchow/","status":"emerging","tldr":"A pathology foundation model trained on millions of slides that can detect cancer and predict biomarkers from an ordinary H&E slide."},{"id":"prov-gigapath","kind":"technology","name":"Prov-GigaPath (Microsoft, Providence)","route":"/technologies/prov-gigapath/","status":"emerging","tldr":"An open pathology model trained on 1.3 billion image tiles from a US health system, modelling whole slides at gigapixel scale."},{"id":"uni-conch","kind":"technology","name":"UNI and CONCH (Harvard, Mahmood Lab)","route":"/technologies/uni-conch/","status":"emerging","tldr":"Two open academic pathology models: UNI reads tissue images, CONCH links images with pathology text."},{"id":"h-optimus","kind":"technology","name":"H-optimus (Bioptimus)","route":"/technologies/h-optimus/","status":"emerging","tldr":"An open 1.1-billion-parameter pathology model from a French startup, among the strongest on public benchmarks."},{"id":"titan","kind":"technology","name":"TITAN (whole-slide multimodal model)","route":"/technologies/titan/","status":"emerging","tldr":"TITAN is a model that summarises a whole slide, not just tiles, and can write a draft pathology report."},{"id":"musk","kind":"technology","name":"MUSK (Stanford, vision-language pathology)","route":"/technologies/musk/","status":"emerging","tldr":"A model that reads slides and clinical text together to predict who will respond to immunotherapy."},{"id":"chief","kind":"technology","name":"CHIEF (Harvard, Yu Lab)","route":"/technologies/chief/","status":"emerging","tldr":"A pathology model trained across 19 cancer types that predicts survival and mutations from slides."},{"id":"artera-ai-prostate","kind":"drug","name":"ArteraAI Prostate","route":"/drugs/artera-ai-prostate/","status":"approved","tldr":"The first AI tool cleared by the FDA to predict both prognosis and treatment benefit from a routine biopsy slide, in prostate cancer."},{"id":"artera-ai-breast","kind":"drug","name":"ArteraAI Breast","route":"/drugs/artera-ai-breast/","status":"approved","tldr":"An FDA-cleared AI test (May 2026) that reads breast cancer slides to estimate recurrence risk in early hormone-positive disease."},{"id":"artera","kind":"company","name":"Artera","route":"/companies/artera/","tldr":"First company with FDA-cleared AI pathology tests that predict treatment benefit (prostate 2025, breast 2026)."},{"id":"masai","kind":"trial","name":"MASAI (Mammography Screening with Artificial Intelligence)","route":"/trials/masai/","status":"positive","tldr":"The first randomised trial of AI in breast screening found more cancers and cut radiologists' reading work almost in half without more false alarms."},{"id":"aidoc-care","kind":"technology","name":"Aidoc CARE (clinical radiology foundation model)","route":"/technologies/aidoc-care/","status":"emerging","tldr":"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."},{"id":"ct-fm","kind":"technology","name":"CT-FM (whole-body CT foundation model)","route":"/technologies/ct-fm/","status":"emerging","tldr":"A model pretrained on 148,000 CT scans to segment organs and triage findings."},{"id":"merlin-ct","kind":"technology","name":"Merlin (Stanford abdominal CT vision-language model)","route":"/technologies/merlin-ct/","status":"emerging","tldr":"Merlin is a model trained on 15,000 CT scans with their reports that can find and describe hundreds of findings."},{"id":"geneformer","kind":"technology","name":"Geneformer","route":"/technologies/geneformer/","status":"emerging","tldr":"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."},{"id":"scgpt","kind":"technology","name":"scGPT","route":"/technologies/scgpt/","status":"emerging","tldr":"A GPT-style model for single-cell data that predicts cell types, perturbation responses, and gene networks."},{"id":"state-arc","kind":"technology","name":"State (Arc Institute perturbation model)","route":"/technologies/state-arc/","status":"emerging","tldr":"Predicts how cells will respond to a drug or gene knockout, trained on over 100 million perturbed cells."},{"id":"tahoe-100m","kind":"collection","name":"Tahoe-100M","route":"/collections/tahoe-100m/","tldr":"Tahoe-100M is the biggest single-cell dataset ever released, built to teach AI how cancer cells respond to drugs."},{"id":"pathology-benchmarks","kind":"collection","name":"Pathology AI benchmarks (CAMELYON, PANDA, TCGA slide tasks)","route":"/collections/pathology-benchmarks/","tldr":"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."}],"trials":[{"id":"masai","name":"MASAI (Mammography Screening with Artificial Intelligence)","route":"/trials/masai/","outcomes":[{"endpoint":"Cancer detection rate (interim analysis)","unit":"per 1,000 screened","arms":[{"name":"AI-supported screening","n":39996,"value":6.1},{"name":"Standard double reading","n":40024,"value":5.1}],"source":"https://doi.org/10.1016/S1470-2045(23)00298-X"}],"setting":"Population screening in Sweden: AI-supported reading (Transpara, single or double reading by risk score) versus standard double reading","enrolled":105934,"enrolledNote":"ClinicalTrials.gov lists 100,000 participants (actual); the Lancet 2026 interval-cancer analysis reports 105,934 women randomly assigned between April 2021 and December 2022.","enrolledBasis":"randomised"}],"papers":[]},{"era":"2025-2028","title":"Language models enter the workflow","description":"The first widely deployed AI in cancer care is not a diagnosis but a time-saver: auto-contouring of organs and tumours for radiotherapy planning now runs in hundreds of centres. Language models are being tested to match patients to trials from the record at the moment a treatment is chosen, to draft tumour-board summaries and pathology reports, and to answer patient questions under supervision. Federated learning lets models train across hospitals without moving data. The evidence standard for each is still being written.","status":"emerging","refs":[{"id":"auto-contouring-ai","kind":"technology","name":"AI auto-contouring and adaptive planning","route":"/technologies/auto-contouring-ai/","status":"established","tldr":"Software that draws organs and tumours on scans automatically, saving hours per patient and making daily plan adaptation practical."},{"id":"limbus-ai","kind":"company","name":"Limbus AI","route":"/companies/limbus-ai/","tldr":"Limbus AI provides AI auto-contouring for radiotherapy, FDA-cleared and used across hundreds of centres."},{"id":"therapanacea","kind":"company","name":"TheraPanacea","route":"/companies/therapanacea/","tldr":"TheraPanacea is a Paris AI company whose ART-Plan does auto-contouring and synthetic CT in radiotherapy."},{"id":"ai-trial-matching","kind":"technology","name":"AI trial matching & clinical decision support","route":"/technologies/ai-trial-matching/","status":"established","tldr":"Software, increasingly LLM-based, that reads a patient's record and finds trials or guideline options they qualify for."},{"id":"trial-library","kind":"company","name":"Trial Library","route":"/companies/trial-library/","tldr":"Trial Library is software that helps everyday cancer clinics spot which patients might qualify for a clinical trial, refer them, and remove practical barriers like transport so more people, especially in under served communities, can join trials."},{"id":"massive-bio","kind":"company","name":"Massive Bio","route":"/companies/massive-bio/","tldr":"Massive Bio uses artificial intelligence to read a cancer patient's medical records and match them to clinical trials they may be eligible for, anywhere in the world, with doctors checking the results."},{"id":"med-gemini","kind":"technology","name":"Med-Gemini and MedLM (Google)","route":"/technologies/med-gemini/","status":"emerging","tldr":"Google's medical versions of its Gemini models, able to reason over text, images, and long records."},{"id":"foresight-ehr","kind":"technology","name":"Foresight (generative EHR model)","route":"/technologies/foresight-ehr/","status":"emerging","tldr":"A model trained on millions of hospital records that forecasts a patient's next diagnoses."},{"id":"federated-learning-medical-ai","kind":"technology","name":"Federated learning and privacy-preserving AI","route":"/technologies/federated-learning-medical-ai/","status":"emerging","tldr":"Federated learning trains one AI model across hospitals by exchanging model updates, not patient data, so a pathology or radiology model learns from every site while records stay behind each firewall. Owkin, NVIDIA FLARE and the MELLODDY pharma consortium use it; governance overhead and differing data across sites are the practical obstacles."},{"id":"owkin","kind":"company","name":"Owkin","route":"/companies/owkin/","tldr":"French AI biotech using federated learning across hospitals; first CE-marked AI for MSI prediction from H&E."},{"id":"tempus","kind":"company","name":"Tempus AI","route":"/companies/tempus/","tldr":"Genomic testing plus one of the largest multimodal clinical datasets, used for AI models and trial matching."},{"id":"multidisciplinary-tumour-board","kind":"technology","name":"Multidisciplinary tumour boards","route":"/technologies/multidisciplinary-tumour-board/","status":"standard-of-care","tldr":"Multidisciplinary tumour boards are regular meetings where surgeons, oncologists, radiologists and pathologists review each patient's case with the full dataset and agree a plan before treatment starts. They are mandatory in the UK and for accreditation in the US, Europe and Germany, and change the diagnosis or plan in 10 to 30% of cases, though randomised evidence is lacking."},{"id":"idea-tr1-ehr-point-of-care-trial-alert","kind":"idea","name":"Trial matching inside the electronic record at the moment a treatment is chosen","route":"/ideas/idea-tr1-ehr-point-of-care-trial-alert/","tldr":"When an oncologist opens the order screen to prescribe a new line of treatment, the record would show the trials this patient may fit, with the nearest open site and a one-click referral."}],"trials":[],"papers":[]},{"era":"2027-2032","title":"From prediction to prospective proof","description":"The field has thousands of retrospective models and a handful of prospective trials. The infrastructure being proposed: a registry of external validation datasets with mandatory reporting, AI-first reading for high-volume common diagnoses with pathologists handling exceptions, every routine CT checked opportunistically for early cancer signs with a tracked pathway, AI central reads to cut trial endpoint cost, and digital twins as virtual control arms where a randomised control is unethical. Regulators are building predetermined change control plans so that models can update without re-clearance.","status":"emerging","refs":[{"id":"idea-tr2-ai-external-validation-registry","kind":"idea","name":"A registry of external validation datasets for cancer AI models, with mandatory reporting","route":"/ideas/idea-tr2-ai-external-validation-registry/","tldr":"Cancer AI models are usually tested on data from the same hospital they were built on. A registry of independent test datasets, and a rule that every model reports performance on at least one, would show which models really work."},{"id":"idea-acc-ai-first-pathology-common-cases","kind":"idea","name":"AI-first reading for high-volume common cancer diagnoses, pathologist for the exceptions","route":"/ideas/idea-acc-ai-first-pathology-common-cases/","tldr":"Let validated AI make the first read on routine, high-volume samples like cervical smears and standard breast biopsy stains, so scarce pathologists spend their time on the difficult cases."},{"id":"idea-prev-opportunistic-ct-ai-registry","kind":"idea","name":"Every routine CT scan checked by AI for early cancer signs, with a tracked follow-up pathway","route":"/ideas/idea-prev-opportunistic-ct-ai-registry/","tldr":"Hundreds of millions of CT scans are done each year for other reasons. Software could check each one for early lung, kidney, liver and pancreas changes, but only if a follow-up system exists."},{"id":"idea-tr1-ai-central-imaging-reads","kind":"idea","name":"AI-assisted central imaging reads to cut endpoint cost and variability","route":"/ideas/idea-tr1-ai-central-imaging-reads/","tldr":"Measuring tumours on scans for trials is slow, expensive and inconsistent between readers. Software that measures lesions and flags changes, checked by a radiologist, could make trial endpoints cheaper and more reliable."},{"id":"digital-twins-trials","kind":"technology","name":"Digital twins and virtual control arms","route":"/technologies/digital-twins-trials/","status":"emerging","tldr":"Using a model of what would have happened to a patient on standard treatment, so fewer people have to be randomised to it."},{"id":"imaging-data-commons","kind":"collection","name":"NCI Imaging Data Commons (IDC)","route":"/collections/imaging-data-commons/","tldr":"The Imaging Data Commons is TCIA in the cloud, ready for large-scale model training."},{"id":"flatiron-foundation-cgdb","kind":"collection","name":"Flatiron Health and Foundation Medicine Clinico-Genomic Database","route":"/collections/flatiron-foundation-cgdb/","tldr":"Real-world evidence at scale: what happened to patients with a given genomic profile on a given treatment."},{"id":"b-ai-validation","kind":"bottleneck","name":"AI that is built but not validated or deployed","route":"/bottlenecks/b-ai-validation/","tldr":"Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients."}],"trials":[],"papers":[]},{"era":"2030+","title":"Patient-level models and the virtual cell","description":"The two long-range bets are a multimodal model that reads slides, scans, genomics and the record to recommend and monitor treatment, and a virtual cell accurate enough to run a drug experiment in silico before it is run in a dish. Both depend on data at a scale no single institution holds, on validation standards that do not yet exist, and on liability and consent questions that are open today. The companion roadmaps on the AI clinic and the virtual cell follow each in detail.","status":"speculative","refs":[{"id":"idea-multimodal-foundation-model","kind":"idea","name":"Patient-level multimodal foundation models for treatment selection","route":"/ideas/idea-multimodal-foundation-model/","tldr":"Train one AI on scans, pathology slides, genomics and treatment outcomes pooled across patients, including completed phase 3 trials, so it can predict which treatment will work for a new patient. Pathology and radiology models already exist separately; combining them with genomic and trial outcome data is the untested step."},{"id":"tempus-multimodal","kind":"technology","name":"Tempus multimodal models","route":"/technologies/tempus-multimodal/","status":"emerging","tldr":"Models trained on Tempus's paired genomic, pathology, imaging and outcome data to predict response and prognosis."},{"id":"pathos-ai","kind":"company","name":"Pathos AI","route":"/companies/pathos-ai/","tldr":"Pathos AI trains multimodal artificial intelligence models on millions of cancer patients' genomic, imaging, text and outcome records to choose which experimental drugs to develop and which patients to test them in. It has licensed a TROP2/HER3 bispecific antibody-drug conjugate and an oestrogen receptor degrader, and raised a 365 million dollar Series D in 2025."},{"id":"noetik","kind":"company","name":"Noetik","route":"/companies/noetik/","tldr":"Noetik builds artificial intelligence models of tumours from huge sets of tissue images and molecular data, to predict which patients will respond to a cancer drug and to find new targets."},{"id":"arc-virtual-cell-atlas","kind":"collection","name":"Arc Virtual Cell Atlas","route":"/collections/arc-virtual-cell-atlas/","tldr":"Arc's growing library of cell data, the fuel for virtual cell models."},{"id":"idea-bio1-in-silico-trials-dose","kind":"idea","name":"In silico trials to choose the dose before the first patient","route":"/ideas/idea-bio1-in-silico-trials-dose/","tldr":"Simulating thousands of virtual patients on a computer can suggest which dose and schedule to test, so fewer real patients receive doses that are too high or too low."},{"id":"ai-oncology-clinic","kind":"roadmap","name":"AI in the oncology clinic: from narrow cleared tools to multimodal decision support","route":"/roadmaps/ai-oncology-clinic/","tldr":"How AI is moving from single-task readers of scans and slides towards systems that weigh everything about a patient, and what regulators and evidence still require."},{"id":"virtual-cell","kind":"roadmap","name":"Virtual cell roadmap: from bulk omics to a predictive model of a cancer cell","route":"/roadmaps/virtual-cell/","tldr":"The attempt to build a computer model of a cell good enough to predict what a drug or mutation will do before anyone runs the experiment."}],"trials":[],"papers":[]},{"era":"What sets the pace","title":"Validation, data and compute","description":"The bottleneck is not model quality but the path from a published model to a deployed one: prospective evidence, external validation, regulatory status for updating models, payment, and data that can be shared. Records, scans and genomes sit in silos; real-world outcomes are weakly recorded, so there is little to learn from; and the workforce that would supervise AI is already short. Compute and model platforms are the one input that is not scarce.","status":"current","refs":[{"id":"b-ai-validation","kind":"bottleneck","name":"AI that is built but not validated or deployed","route":"/bottlenecks/b-ai-validation/","tldr":"Thousands of cancer AI models are published; a handful are in clinical use, and fewer have shown they help patients."},{"id":"b-data-silos","kind":"bottleneck","name":"Data silos","route":"/bottlenecks/b-data-silos/","tldr":"Records, scans, genomes and outcomes sit in separate systems that cannot talk. Every patient's experience is lost to the next."},{"id":"b-real-world-evidence","kind":"bottleneck","name":"Weak real-world evidence and registries","route":"/bottlenecks/b-real-world-evidence/","tldr":"We do not reliably know what happens to patients after approval, so we cannot tell which drugs deliver in practice."},{"id":"b-workforce","kind":"bottleneck","name":"Not enough oncologists, nurses, pathologists, physicists","route":"/bottlenecks/b-workforce/","tldr":"The number of people with cancer is rising faster than the workforce trained to treat them."},{"id":"b-reproducibility","kind":"bottleneck","name":"Preclinical results do not reproduce","route":"/bottlenecks/b-reproducibility/","tldr":"Fewer than half of landmark cancer biology findings reproduce when someone else tries."},{"id":"ai-compute-platforms","kind":"technology","name":"AI compute and model platforms for oncology","route":"/technologies/ai-compute-platforms/","status":"emerging","tldr":"AI compute platforms are the GPUs, model libraries, and cloud services that pathology, radiology, and drug-design AI run on."},{"id":"nvidia","kind":"company","name":"NVIDIA","route":"/companies/nvidia/","tldr":"Supplies the GPUs and the BioNeMo framework most biological foundation models are trained on, and co-developed Evo 2 with Arc."},{"id":"bionemo","kind":"technology","name":"NVIDIA BioNeMo","route":"/technologies/bionemo/","status":"established","tldr":"NVIDIA BioNeMo is a software stack, not a model: GPU-optimised training recipes and inference services on which protein, DNA and single-cell foundation models such as Evo 2, ESM and Geneformer are trained and served. The Arc Institute, Recursion and pharmaceutical companies use it, at the price of being tied to NVIDIA hardware and tooling."}],"trials":[],"papers":[]}],"watch":[]}