University of Oxford

Protein perturbation model conditioning substitution-effect prediction on a learned protein-level regime coordinate over 202 million sequences.

Protein-ligand co-folding model on the OpenFold3 architecture, trained on PDB structures through June 2025 with inference-time chemical steering.

Single-cell foundation model adding a gated cell-level contrastive objective to masked expression pretraining for transferable frozen cell embeddings.

Breast cancer histopathology foundation model distilled from three general-purpose PFMs, over 30x smaller with comparable balanced accuracy and AUC.

Cross-species single-cell ageing-state classifier that transfers mouse age labels to human HSC and CD8+ T cells, reaching 0.953 held-out AUROC.

Fine-tuned Enformer derivative that annotates cis-regulatory elements from DNA sequence, emitting enhancer, promoter, and insulator class labels.

Pretrained antibody structure predictor that outputs full paired heavy/light 3D structures faster than protein language models generate embeddings.

Renal pathology foundation model self-supervised on a million kidney biopsy tiles spanning glomeruli, interstitium, and surrounding structures.

Renal pathology foundation model whose learning unit is a whole detected glomerulus, pretrained on over a million of them across four biopsy stains.

Prime editing efficiency prediction that quantifies per-pegRNA uncertainty, pairing a Dirichlet outcome model with conformal coverage guarantees.

Synthetic 3D MRI generation spanning brain, prostate, breast and abdomen, with the T1, T2 or FLAIR contrast selected at sampling time.

All-atom structure prediction for arbitrary biomolecular complexes of proteins, nucleic acids, and ligands, with code and weights under a BSD license.

Splicing variant effect prediction across 49 human tissues and 15 developmental stages, from four weeks post conception to adulthood.

Antibody inverse folding by discrete diffusion, sampling sequences that fold onto a given backbone. Generates varied CDRs at high sequence recovery.

Multimodal foundation model for cardiac biosignals, pretrained by masked modeling on ECG, PPG, and clinical text from ~1.7 million individuals.

Contrastive language-image model for fMRI functional decoding, predicting cognitive tasks, concepts, and domains from brain activation maps.

Protein fitness prediction with end-to-end differentiable homology search, replacing MSA construction with vector search over 62M UniRef50 sequences.

3D MRI organ segmentation foundation model built on Swin-UNETR and trained on the UKBOB whole-body dataset covering 72 organs and skeletal structures.

3D CT image synthesis with paired anatomical masks, driven by a rectified-flow latent diffusion model that samples a whole volume in 30 steps.

2B-parameter medical vision-language model that uses reinforcement learning to show interpretable reasoning for radiology visual question answering.

Predicts haplotype-specific 3D genome organization and Hi-C contact maps from a single long-read Fiber-seq assay, using no DNA sequence as input.

Structure-based drug design generator pairing continuous flow matching with discrete Markov bridges to sample 3D ligands inside a protein pocket.

Multimodal medical imaging foundation model for zero-shot clinical diagnosis and report generation from chest X-ray and CT in English and Chinese.

Somatic copy-number alteration calling from tumor whole-genome sequencing, using Mamba and Hyena blocks over genome-scale inputs of about 1M loci.

Genetic language model predicting disease risk and cell-type-specific expression changes from up to 88 megabases of an individual's genome sequence.

Text-guided medical image synthesis across OCT, fundus, X-ray, CT and MRI. Synthetic data lifts downstream clinical tasks by 12-17%.

Antibody language model that generates paired heavy and light variable domains, with a developability-conditioned variant for manufacturable designs.

SAM2-based foundation model that segments 2D and 3D medical images by treating volumes and image sets as video object tracking.

Self-supervised CNN pretrained on 700,000 person-days of UK Biobank accelerometer data for human activity recognition across devices and cohorts.

Antibody language model trained on paired and unpaired OAS sequences to suggest non-germline mutations instead of reverting them to germline.

Codon-level BERT model that captures genomic signals invisible to amino acid models, outperforming billion-parameter PLMs with just 86M parameters.

Vision-language pretraining for 3D CT volumes, aligning scans with their radiology reports for zero-shot classification, retrieval, and segmentation.

De novo protein backbone generation by SE(3) flow matching, with motif-scaffolding built in. Samples a designable backbone in seconds on one GPU.

De novo protein backbone generation by SE(3) diffusion over rigid-body frames. Samples designable monomers up to 500 residues without pretraining.

Autoregressive protein language model scoring variant effects zero-shot, blending sequence likelihood with homolog statistics retrieved at inference.

Antibody-specific language model trained on the OAS database for restoring missing residues and generating high-quality sequence representations.