ETH Zurich

Histopathology tile encoders reach 22M parameters by distilling billion-parameter teachers through their frozen class and patch tokens alone.

Contrastive encoder aligning NMR metabolomics to the plasma proteome, adding proteome-level disease risk signal to cohorts with no proteomics.

EEG foundation model for continuous monitoring, using windowed alternating attention to hold KV-cache memory constant on recordings up to 14 hours.

Foundation model that predicts microRNA-mRNA target specificity from sequence, using a dual-encoder trained with a symmetric contrastive objective.

Transcription factor-DNA binding specificity prediction from sequence, with a physics-derived dual-encoder trained by symmetric contrastive learning.

Enzyme-substrate specificity model that scores catalytic pairs from sequence with a physics-derived dual-encoder and a contrastive objective.

CRISPR off-target prediction model that scores gRNA-DNA specificity from sequence, framing guide-target recognition as cross-modal retrieval.

Specificity foundation model predicting small-molecule drug-target binding from sequence, scored as cross-modal retrieval without docking or assays.

Peptide-MHC binding specificity model that frames presentation as cross-modal retrieval, aligning peptide and MHC encoders by contrastive learning.

Diffusion transformer for virtual tissue synthesis, generating H&E histopathology patches conditioned on spatial gene expression and morphology.

SE(3)-invariant masked autoencoder that learns protein fold representations from AlphaFold-DB structures, supporting zero-shot fold classification.

Contrastive antibody language model predicting antibody-antigen binding specificity from sequence with a dual-encoder, cross-attentive architecture.

Causal 309M-parameter protein language model that scores variant fitness zero-shot and generates sequences, reaching 0.390 Spearman on ProteinGym.

Latent flow-matching method that repurposes protein language model embeddings to generate high-fitness protein variants without predictor guidance.

EEG foundation model whose learned queries map any electrode montage into a fixed latent space, scaling linearly in the number of channels.

Autoregressive protein language model for antibody Fc domains, reinforcement-tuned to design variants with programmable Fc-receptor binding profiles.

Predicts virtual single-cell spatial transcriptomics from H&E histology using frozen pathology foundation models and spot-level supervision.

Cas9 PAM preference prediction from protein sequence with an ESM-2 backbone, extending PAM annotation to 50,308 metagenome-mined orthologs.

Contrastive alignment framework that projects H&E histology and single-cell transcriptomic foundation model embeddings into one shared latent space.

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

Physiological signal foundation model for ECG, EMG, and EEG pairing learnable multi-scale wavelet decomposition with masked transformer pretraining.

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

Patch-based 3D diffusion model that generates teravoxel-scale virtual mouse brain volumes conditioned on spatially resolved gene expression.

EEG foundation model built on bidirectional Mamba blocks that scale linearly with recording length, pretrained on 21,000 hours of clinical EEG.

Compact EEG foundation model whose alternating attention separates within-channel time from across-channel space, cutting attention memory sixfold.

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

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

3D blood vessel segmentation across CT, MRI, light-sheet microscopy and OCTA volumes, generalizing zero-shot to imaging domains absent from training.

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

Single-cell RNA-seq foundation model pretrained only on malignant cells, for zero-shot batch integration and drug response prediction in tumors.

Vision-language chat model for 3D chest CT volumes, answering free-form questions and drafting radiology report findings from a frozen 3D encoder.

Vision-language model for 3D chest CT that aligns whole volumes with radiology reports for zero-shot abnormality detection and case retrieval.

Virtual staining model that generates 11-marker spatially resolved protein multiplexes from routine H&E histopathology whole-slide images.

Cryo-electron tomography membrane analysis pipeline pairing generalizable U-Net membrane segmentation with mesh-based particle localization.

Generative toolkit that synthesizes six aligned immunohistochemistry markers from one H&E histopathology image, trained on unpaired stains.

Rigid protein-protein docking model that predicts a complex from two unbound structures in one pass, with no candidate sampling or refinement.