AI in Oncology · AI Models

Predicts the 3D structure of a protein from its amino-acid sequence at near-experimental accuracy; the AlphaFold Protein Structure Database (with EMBL-EBI) provides predicted structures for over 200 million proteins, including cancer-relevant targets and mutants.

Diffusion-based successor that predicts joint structures of proteins with DNA, RNA, ligands, ions and modified residues — the interaction types that matter for drug design and for understanding oncogenic complexes.

BERT pretrained from scratch on PubMed abstracts and PMC full text with a biomedical vocabulary; the workhorse encoder for named-entity recognition, relation extraction and classification over oncology literature and reports.

CLIP-style vision-language model pretrained on PMC-15M — 15 million figure–caption pairs from biomedical papers — with a PubMedBERT text tower and a ViT-B image tower; supports zero-shot classification and retrieval across pathology, radiology and more.

Clinical Histopathology Imaging Evaluation Foundation model: trained on 60,530 whole-slide images across 19 anatomical sites and validated on 19,491 slides from 24 hospitals for cancer detection, tumour-origin prediction, genomic profiling and survival.

Vision-language foundation model for pathology: an image encoder and a text encoder trained together on 1.17 million histopathology image–caption pairs, enabling zero-shot classification and image–text retrieval without labelled slides.

Deep-learning variant caller that turns aligned sequencing reads into pileup images and classifies genotypes with a CNN; widely used for germline calling, with the companion DeepSomatic extending the approach to tumour–normal somatic variants.

Protein language models from 8 million to 15 billion parameters trained on UniRef sequences; embeddings power variant-effect and function prediction, and ESMFold predicts structure directly from a single sequence without MSAs.

Transformer pretrained on about 30 million single-cell transcriptomes (Genecorpus-30M) that learns gene-network context from ranked expression, enabling few-shot prediction of gene dosage effects, cell states and candidate therapeutic targets.

1.1-billion-parameter ViT-g/14 pathology encoder trained on more than 500,000 H&E slides (hundreds of millions of tiles), released under Apache-2.0 — one of the few large pathology foundation models with a permissive licence.

Presentation predictor trained on the mono-allelic peptidome of 95 cell lines — 186,464 peptides across 95 HLA alleles, fifteen of which had no described motif before — which is the dataset that changed this field more than any architectural idea.

Open-weight medical vision-language models built on Gemma 3: the 4B variant reads chest X-rays, dermatology, ophthalmology and histopathology images alongside text; the 27B variant targets medical text reasoning. Meant as a starting point for developers to fine-tune, not as a finished clinical product.

Segment Anything adapted to medical images: a promptable segmentation model fine-tuned on over 1.5 million image–mask pairs across 10 imaging modalities and more than 30 cancer types.

Open-source pan-allele presentation predictor whose distinctive idea is a separate antigen-processing model: it reads the peptide together with fifteen amino acids of flanking sequence on each side, because the proteasome cut depends on what lies around the cut site.

Allele-specific LSTM networks — one per allele, 148 for class I — that read peptides letter by letter, so no alignment or padding is needed, and that are trained by transfer learning from the data-richest allele.

Position-weight-matrix predictor built on the observation that class I ligands show almost no dependence between positions; version 3.0 closes the circle by having a neural network predict the matrix from the same 34 groove residues.

The class II counterpart of NetMHCpan, released in the same paper: predicts presentation by HLA-DR, -DQ and -DP, whose groove is open at both ends, so peptides are longer and the binding core has to be found inside a longer sequence.

The reference pan-allele predictor of MHC class I antigen presentation: one small neural network covers more than 11,000 MHC molecules because the groove itself is part of the input — a 34-residue pseudosequence next to the peptide.

Self-configuring segmentation framework: given a labelled dataset it chooses preprocessing, network topology and training schedule automatically, and it remains the baseline to beat on most medical segmentation challenges, including tumour tasks.

ViT-L pathology encoder trained with DINOv2 on PANCAN-XL — 456 million tiles from 58,359 whole-slide images that mix public cohorts (TCGA, CPTAC, GTEx and others) with private data — positioned for biomarker prediction.

Whole-slide foundation model with 1.3 billion parameters, pretrained on 1.3 billion tiles from 171,189 slides of real-world clinical data; pairs a DINOv2 tile encoder with a LongNet slide encoder that reasons over an entire slide.

Generative pretrained transformer for single-cell multi-omics, trained on over 33 million cells, supporting cell-type annotation, batch integration, perturbation response prediction and gene-network inference.

Command-line tool that segments 117 anatomical structures (v2) in any CT scan with nnU-Net models trained on more than 1,200 clinical CTs — the standard way to get organ masks for radiomics, dose planning and tumour-context features.

Graph neural network for zero-shot drug repurposing that scores drug–disease indications and contraindications over a medical knowledge graph of 17,080 diseases, including ones with no approved treatment.

Successor to Virchow: ViT-H/14 pretrained on 3.1 million whole-slide images from about 225,000 patients across 45 countries, at mixed magnifications (5×–40×), with pathology-specific augmentations.

General-purpose self-supervised vision encoder for H&E histopathology tiles, pretrained on more than 100 million tiles from over 100,000 whole-slide images; the reference foundation model for pathology feature extraction.

632-million-parameter vision transformer pretrained on 1.5 million whole-slide images from about 100,000 patients — the largest pathology pretraining set at its release — and used to build a pan-cancer detection model covering 17 cancer types, including rare ones.

Classifies the pathogenicity of every possible single amino-acid substitution in the human proteome — 71 million missense variants — by fine-tuning AlphaFold on population variant frequencies; a resource for interpreting variants of uncertain significance.

Three-stage LLM framework (retrieval, criterion-level matching, ranking) that matches a patient summary to clinical trials from ClinicalTrials.gov; in the paper it cut clinician screening time by more than 40% in a pilot user study.

Google's medical large language model, the first to reach expert-level scores on USMLE-style questions (86.5% on MedQA); available only through Google Cloud to selected partners, and largely succeeded by Gemini-based medical models.