EPFL

Chemistry embedding model placing SMILES strings and scientific prose in one vector space, top-tier on 43 of 48 molecular and text benchmarks.

Enzyme-substrate specificity prediction by end-to-end co-folding, with no predefined binding pocket. AUROC 0.766 on unseen enzymes and substrates.

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

Bidirectional generative framework linking nuclear morphology and gene expression, built on a morphology model trained on 21 million segmented nuclei.

Diffusion model over coarse-grained protein topology that samples diverse folds, handing them to RFdiffusion to build de novo protein backbones.

Patient-level representation learning from scRNA-seq: a transformer set encoder with a diffusion decoder, fine-tuned for clinical prediction.

De novo transmembrane protein design by joint all-heavy-atom sequence and structure diffusion, validated by a 1.7 A crystal structure.

Model quality assessment network predicting per-residue lDDT, CAD-score, and RMSD for protein loops in predicted and designed structures.

Proteome-scale protein language model whose representations enable zero-shot protein-protein interaction and gene essentiality prediction.

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

CLIP-based vision-language foundation model for eye imaging, enabling zero-shot disease detection and cross-modal retrieval across 11 modalities.

Protein binder design that inverts the frozen Boltz-1 all-atom predictor, targeting small molecules, nucleic acids, metals, and modified residues.

Retrieval-augmented protein language model that conditions ESM-2 on homologous sequences via cross-attention for conditional sequence generation.

Histopathology foundation model pretrained with DINOv2 on tiles chosen by unsupervised hierarchical clustering over 350 million whole-slide tiles.

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

Generative diffusion model that reassembles dissociated single cells into tissues, predicting spatial coordinates from gene expression alone.

Spatial proteomics foundation model that embeds every marker through its protein sequence, so heterogeneous antibody panels share one representation.

Enhancer-promoter interaction prediction from DNA sequence and ATAC-seq alone. Spearman above 0.90 on cell types unseen during training.

Per-residue classifiers fine-tuned from ESM-2 and ProtT5 that label 20 UniProt protein features and read out what a missense variant disrupts.

Discrete flow-matching generator for molecular and cell graphs, with sampling decoupled from training so a checkpoint is retuned, not retrained.

De novo protein binder design pipeline that hallucinates binders through frozen AlphaFold2 weights, averaging a 46% experimental success rate.

Homology-aware but alignment-free protein language model using the Mamba state space architecture to condition on hundreds of unaligned homologs.

Ophthalmic imaging foundation model pretrained on 2.78M images across 11 modalities for diagnosis, prognosis, and visual question answering.

Self-supervised 3D cell segmentation for fluorescence microscopy, pairing WNet3D with Swin-UNetR to segment volumes without annotated training data.

Threshold-agnostic spot detection for 2D and 3D fluorescence microscopy, locating transcripts at subpixel accuracy via stereographic flow regression.

Nucleus instance segmentation for fluorescence microscopy and H&E histology, predicting a star-convex polygon per pixel to separate crowded nuclei.