Technical University of Munich

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

RNA foundation model pretrained on 223 eCLIP experiments to predict base-resolution RBP binding, with frozen embeddings that transfer downstream.

Open-weights pathology foundation models pairing a 1.1B-parameter ViT-g/8 tile encoder with distilled 86M and 22M variants for H&E histology.

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

A robustified fine-tune of Phikon-v2 for H&E tiles, retrained to strip scanner and laboratory signatures out of its feature space.

A pathology tile encoder fine-tuned from Midnight-12k so its features track tissue morphology rather than the scanner and lab behind the slide.

Hierarchy-aware self-supervised model for single-cell microscopy that preserves morphological structure suppressed by imaging-modality confounders.

Variational autoencoder trained on scRNA-seq and applied frozen to impute unmeasured genes and denoise spatial transcriptomics profiles.

Virtual spatial transcriptomics foundation model predicting pan-cancer, spatially-resolved single-cell gene expression from H&E histology slides.

LoRA adapter on ProstT5 predicting per-residue distributions over Foldseek 3Di tokens, capturing conformational flexibility from MD trajectories.

Protein-family language model trained on unaligned homolog sets for zero-shot variant fitness prediction and design. ProFam-1 holds 251M parameters.

Spectral prediction model for modified peptides, forecasting fragment-ion intensities and retention time with zero-shot generalization to unseen PTMs.

Pathology foundation model that fuses global patch and cell-level tokens via joint-weighted attention pooling for H&E-based biomarker detection.

SE(3)-equivariant flow-matching model for pocket-aware 3D ligand generation, predicting binding affinity and confidence in the same network.

Predicts single-cell scRNA-seq coverage and scATAC-seq insertion profiles from DNA sequence, adapting the Borzoi trunk with a cell-specific decoder.

Linear B-cell epitope prediction from peptide sequence alone, pairing ProtT5 embeddings with an SVM trained on 222,030 curated IEDB peptides.

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

Sparse autoencoder for blood-cell microscopy that decomposes hematology foundation model embeddings into expert-validated sub-cellular concepts.

Decoder-only transformer that recasts ancestral recombination graph inference as next-token prediction, estimating coalescence times from variation.

Peptide-MHC structure prediction by SE(3)-equivariant diffusion, sampling 10 Cα conformations in under 6 seconds at 0.45 Å best-of-10 RMSD.

Whole-genome bacterial pathogenicity prediction from ProtT5 embeddings, alignment-free and taxonomy-agnostic, with per-protein attention scores.

Bacterial exotoxin classifier over frozen ProtT5 embeddings that separates secreted toxins from non-toxic secreted proteins at MCC 0.94.

Promptable 3D segmentation for particle picking in cryo-electron tomography, conditioned on a reference subtomogram to detect any target complex.

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

Grounded multimodal language model for endoscopic surgery, supporting visual dialogue, region-based question answering, and bounding-box grounding.

3D vision foundation model for computed tomography, contrastively pretrained on 148,000 scans for segmentation, triage, retrieval, and concept search.

Regulatory genomics model predicting cell-type-specific RNA-seq coverage from DNA sequence, unifying transcription, splicing, and polyadenylation.

Fragment-ion intensity prediction for cross-linked peptides, covering cleavable DSSO and DSBU chemistries alongside non-cleavable DSS and BS3.

Histopathology patch encoder turning 512x512 tiles into 768-dimensional features, trained with a CoCa objective on 1.26 million captioned images.

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.

Aligns structure, binding-pocket, text and molecular-dynamics encoders to a protein sequence anchor, giving frozen embeddings that transfer widely.

SMILES transformer pretrained to predict 113 RDKit molecular descriptors, giving embeddings that carry physicochemical properties into ADMET models.

Fine-tuned Prosit predictor of spectra and retention time for citrullinated peptides, separating them from isobaric deamidation in MS searches.

Cardiac MR vision foundation model self-supervised on 36 million images, fine-tuned for segmentation, view classification and pathology detection.

Transformer foundation model pretrained on 110M single-cell and spatial transcriptomics profiles, transferring spatial context to dissociated cells.

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

Bilingual protein language model that translates bidirectionally between amino acid sequences and the 3Di structural alphabet for inverse folding.

Rigid protein-protein docking by diffusion over the rigid-body pose, with a confidence model ranking sampled complexes. Median C-RMSD 4.85 on DIPS.

Masked DNA language model trained on over 800 vertebrate genomes and conditioned on species identity to learn conserved regulatory sequence features.

Masked DNA language model trained on 800+ species with explicit species conditioning, separating conserved regulatory motifs from background bias.

Parameter-efficient protein language model that matches larger models such as ESM-2 on protein prediction tasks using under 10% of the parameters.

Predicts CLIP-seq crosslink counts along an RNA sequence base by base, separating protein-specific signal from experimental background.

Suite of six protein language models, including ProtBERT and ProtT5, trained on up to 393 billion amino acids without multiple sequence alignments.