University of Cambridge

Conditional flow matching model that reconstructs 4D left-ventricular geometry and motion from a 12-lead ECG and basic demographic variables.

Tissue reconstruction model placing dissociated single cells into spatial coordinates by predicting pairwise distances in a learned embedding space.

Spatial transcriptomics foundation model giving gene-, cell- and neighborhood-scale embeddings zero-shot, plus in-silico gene knockout in tissue.

ECG foundation model that reads any subset of the 12 standard leads natively, encoding recordings as variable-size spatiotemporal graphs.

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

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.

Single-cell perturbation prediction model trained only on synthetic priors, inferring drug targets, intervention strengths, and regulatory graphs.

De novo small molecule design for protein pockets, steering a discrete flow-matching generator with gradients from a frozen Boltz-2 affinity model.

Zero-shot generative framework that turns 3D pharmacophores into synthesis-ready DNA-encoded libraries of purchasable building blocks.

Transformer that predicts protein-RNA binding affinity from Boltz-2 pre-structural embeddings via cross-modal attention, with no 3D structure step.

Polarizable machine-learning interatomic potential extending MACE with long-range electrostatics, trained on 100M OMol25 DFT calculations.

Vision-omics finetuning that aligns pathology foundation models with spatial transcriptomics so morphology features predict local gene expression.

Generative foundation model that imputes genes and denoises spatial transcriptomics, conditioned on H&E histology, scRNA-seq, and spatial priors.

RNA inverse-folding model that generates sequences predicted to fold into a target 3D backbone, capturing non-canonical pairs and tertiary motifs.

Vision-language encoders for chest CT that align 3D volumes with radiology reports using contrastive, report-generation, and masked-image objectives.

Single-molecule localisation microscopy encoder that embeds nanoscale point clouds into a 128-dimension latent space to compare protein architecture.

Histopathology model predicting homologous recombination deficiency from H&E slides in ovarian cancer, reaching 0.846 AUC and 0.938 specificity.

Centrosome segmentation framework chaining YOLOv11 detection, U-Net refinement, and StarDist cell boundaries across immunofluorescence and IHC tissue.

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

Multi-state protein inverse folding model that designs one sequence for two conformations, improving sequence recovery 12% over ProteinMPNN.

Bacterial genomics foundation model reading whole genomes as ordered protein sequences. Predicts operons, gene essentiality, and phenotypic traits.

Synthesizable 3D molecule generation that jointly samples building blocks, reactions, and atomic coordinates, returning a synthesis route per design.

Sensor-language foundation models aligning wearable biosignals with text for zero-shot activity recognition, retrieval, and sensor captioning.

Brain foundation model unifying EEG and MEG in a single encoder via a shared discrete tokenizer that transfers across sensor layouts and montages.

Brain MRI segmentation foundation model trained on 66,000+ image-label pairs across 14 MRI sub-modalities, with a hypergraph dynamic adapter.

Antibody sequence generation model that samples paired human VH/VL chains, covering inpainting, inverse folding, and CDR grafting in one network.

GPCR peptide agonist screening with a graph neural network over AlphaFold-Multimer active-state complexes and interatomic contact graphs.

Latent diffusion model generating 3D drug-like molecules and inorganic crystals from one shared all-atom autoencoder and Transformer denoiser.

Refines AlphaFold2 predictions against cryo-EM, cryo-ET, and X-ray data by optimizing coevolutionary embeddings rather than atomic coordinates.

Crossmodal diffusion model synthesizing bulk tumor gene expression from H&E whole-slide images, so grading and survival prediction need no RNA assay.

Surrogate machine learning force field reusing a reference model's node features from earlier steps, running molecular dynamics eight times faster.

DNA methylation foundation model reconstructing genome-wide profiles from sparse input. Outperforms GrimAge2 aging clocks on mortality prediction.

Spatial transcriptomics prediction from H&E slides, inferring spot-level expression and tumor microenvironment composition in breast cancer.

Pharmacophore-conditioned generator that emits molecules as synthetic trees of Enamine building blocks, so every design carries a synthesis route.

Nanobody thermostability predictor stacking regressors over frozen language model embeddings to estimate apparent melting temperature from sequence.

Respiratory acoustic foundation models pretrained on roughly 136K cough and breathing recordings for disease detection and lung function estimation.

Protein function prediction model that conditions a T5 encoder-decoder on retrieved homologs to assign EC numbers, GO terms and Pfam families.

Neuroscience literature language model that predicts experimental outcomes from study methods, LoRA-tuned on 1.3 billion tokens of papers.

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.

Protein structure and complex prediction from sequence, in a three-track network that reasons over alignments, distances, and 3D coordinates at once.

Antibody paratope prediction model that identifies antigen-contacting residues from heavy and light CDR sequences alone, using CNN and RNN layers.