Shanghai Jiao Tong University

Histopathology foundation model for whole-slide cancer diagnosis, covering 19 common cancer types and 205 clinical diagnostic tasks.

Fragment-protein interaction model trained on cellular chemoproteomics, pairing ESM-2 residue embeddings with bilinear attention over ligand atoms.

Protein structure prediction from multiple sequence alignments, trained across MSA depths so one model spans deep alignments and orphan proteins.

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

Long-context RNA foundation model reading whole mRNA transcripts at single-nucleotide resolution, pretrained at a native 10,240 nt context.

Sequence-conditioned generative framework that reconstructs missing prostate MRI contrasts and restores artefact-degraded acquisitions.

Single-cell metabolome inference from scRNA-seq, learned from spatially paired Visium and MALDI-MSI sections by multiple-instance learning.

Virtual-cell model that compresses a transcriptome into eight discrete tokens in a reasoning LLM's vocabulary, predicting module-level drug response.

Protein stability predictor scoring ΔΔG for substitutions, multi-point mutations and indels from a folding model's latent structure representations.

Antisense oligonucleotide activity prediction from sequence, position-specific chemistry, dose, and cell context. Spearman 0.5970 on ASO Atlas.

Dual-target structure-based drug design that fuses two pocket-conditioned Bayesian flow distributions to generate 3D ligands binding both proteins.

Genomic foundation model on a Mamba, attention, and mixture-of-experts backbone, with 1M-token context for variant scoring and regulatory DNA design.

Antigen-specific antibody design model that conditions an ESM3 backbone on epitope geometry, then aligns CDR generation with calibrated DPO.

All-atom foundation model for immune-receptor design that predicts structures and co-designs CDR sequences for antibodies, nanobodies, and TCRs.

Cis-regulatory network reconstruction from DNA sequence, epigenomic tracks, and Hi-C priors, constrained by gene expression. 0.84 zero-shot auROC.

Autoregressive generative model that uses reinforcement learning to optimize mRNA codon sequences for MFE, CAI, and GC content.

Reinforcement-learning generative framework for multi-objective RNA codon optimization that generalizes across six species and five RNA types.

Motion artifact reduction for cardiac CT, folding a learned cardiac motion field into model-based reconstruction to image the heart at any phase.

Latent diffusion model for PTM-aware protein sequence design, using ControlNet-style conditioning to steer generation toward chosen PTM sites.

Knowledge-graph-grounded model that predicts single-cell transcriptomic responses to small molecules, with zero-shot prediction for unprofiled drugs.

Pan-cancer single-cell foundation model with a hybrid Transformer-Mamba architecture, released with the PanFoMaBench cancer evaluation benchmark.

Protein structure autoencoder compressing backbone coordinates into a latent space, paired with a latent diffusion model for generative design.

Brain MRI foundation model pairing DenseNet and Vision Transformer backbones with mixture of experts for disease diagnosis and brain age prediction.

Protein conformational ensemble and dynamics generator using latent diffusion to sample all-atom MD trajectories and transition pathways.

Generative antibody and nanobody design model that co-designs CDR sequences and antigen-bound structures for de novo design and affinity maturation.

Structure-based drug design model generating 3D ligands in a protein pocket with interaction-guided flow matching and a learned atom-count predictor.

Vision-language foundation model for kidney cancer CT, covering zero-shot malignancy diagnosis, report generation, and recurrence risk prediction.

Single-cell foundation model for yeast that injects regulatory network priors into transformer attention for zero-shot and fine-tuned analysis.

Protein stability predictor scoring ΔΔG for point mutations by fusing ESM-2 embeddings with ProteinMPNN backbone geometry. Wet-lab validated.

Structure-based virtual screening model that jointly predicts protein-ligand complex structures and binding fitness from sequence and SMILES.

Molecular reasoning language model for molecule captioning and text-to-molecule generation, trained by chain-of-thought distillation then reward RL.

Enzyme thermostability prediction from sequence, using segment-level attention over ESM-2 embeddings to rank mutation sites for protein engineering.

Structure-based drug design model that generates 3D ligands inside a protein pocket by interpolating distribution parameters instead of samples.

Cross-species brain spatial transcriptomics foundation model pretrained on 133M cells from human, macaque, marmoset, and mouse whole brains.

Backmapping model that rebuilds all-atom protein and nucleic acid structures from coarse-grained beads and inpaints unresolved residues.

Single-cell perturbation response prediction using dual conditional diffusion bridges that link unpaired control and perturbed populations.

Alignment-free taxonomic classification of eukaryotic DNA in metagenomes. Reaches an F1 of 0.871 on 500 bp contigs, where k-mer tools falter.

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

Medical multimodal LLM (2B and 8B) trained for generalizable, step-by-step clinical reasoning via Mentor-Intern Collaborative Search.

Multimodal drug-response model coupling cell and molecule foundation models, pretrained on 1.8M perturbation RNA-seq profiles over 22,000 compounds.

Inverse folding framework combining a Markov bridge generator with direct preference optimization to design low-energy sequences and predict ΔΔG.

Cyclic peptide design conditioned on target protein structure, generating all four cyclization types via all-atom, all-bond harmonic SDE modeling.

RNA conformational ensemble generation with a diffusion model, sampling excited states and folding pathways from one structure without MSA input.

RNA-binding protein predictor trained on eCLIP data that scores binding intensity along transcripts and recovers motifs via integrated gradients.

Cytogenetics foundation model detecting numerical and structural chromosome abnormalities from metaphase images, pretrained on 84,000 specimens.

Multimodal physiological foundation model spanning EEG, ECG, EOG, and EMG that keeps working when arbitrary modalities are missing at inference time.

Pocket-based molecular docking with an SE(3)-equivariant diffusion transformer. Places 77.65% of top-1 poses within 2 Å RMSD on PoseBusters.

Zero-shot RNA design pipeline that ranks variants by genomic language model likelihood combined with inverse-folding structural compatibility.

Protein language model pretrained on over nine billion sequences, giving residue embeddings and zero-shot single-site mutant scores from its logits.

Cyclic peptide binder design by Monte Carlo tree search over sequence space, scored by predicted confidence of the peptide-target complex fold.

Whole-body CT segmentation covering 235 fine-grained anatomies: 193 organs, 33 lymph node stations, and 9 lesion types from one unified network.

Pathology image restoration recovering all-in-focus histology from single defocused focal planes, guided by semantic, defocus, and edge prompts.

Protein inverse folding ensemble that fuses five pretrained sequence designers through a self-attention encoder, reaching 63.1% recovery on CATH4.2.

Histopathology foundation model pretrained with BEiT masked image modeling on 11M+ tissue image tiles for cancer diagnosis and survival prediction.

Dual-target protein sequence design conditioned on two receptor structures at once, combining a heterogeneous graph network with ESM-2 features.

Histopathology foundation model pretrained on 300K whole-slide images across 20 tissue types and validated on 112 clinical-grade downstream tasks.

Multimodal sequence model spanning proteins, coding DNA, and regulatory DNA for zero-shot fitness scoring and conditional sequence generation.

Flow matching model for de novo 3D peptide design that converges peptide position before conformation, mirroring the physical order of docking.

Pan-tumour CT foundation model pretrained on 30,000 synthetic 3D scans carrying lesion masks and structured reports across ten organ systems.

Yeast gene regulatory network model with one pretrained subnetwork per gene, simulating target-gene response to transcription factor perturbation.

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

Protein expression prediction that pinpoints expression-governing residues by matching a language model landscape against measured fitness data.

Multi-omics instruction-tuned LLM that reads DNA, RNA, protein, and multi-molecule sequences and answers natural-language questions about them.

Structure-based molecular generation guided by imputed ligand electron density, assembling drug-like compounds into a pocket fragment by fragment.

Knowledge-enhanced ECG foundation model aligning a ResNet encoder with LLM-generated disease descriptions for zero- and few-shot interpretation.

Histopathology vision-language foundation model that folds a disease knowledge graph into pretraining for zero-shot cancer detection and subtyping.

Geometric foundation model matching enzymes to the reactions they catalyze, trained on 1.5 million structure-informed enzyme-reaction pairs.

Pathology image embeddings supervised by spatial transcriptomics instead of text captions, aligned over 697K image-gene expression pairs.

Lung cancer histopathology model that predicts spread through air spaces from whole-slide images using a feature-interactive Siamese graph encoder.

Single-cell foundation model that compresses each expression profile into 64 cross-attention patch tokens for annotation and spatial transfer.

ECG foundation model that learns discrete rhythm tokens from noisy real-world recordings for arrhythmia classification and anomaly detection.

De novo enzyme design conditioned on the reaction to be catalysed: substrate and product SMILES in, catalytic pocket, enzyme, and docked complex out.

Chemical language model for molecular property prediction, pretrained on PubChem SMILES by masking whole functional groups rather than random tokens.

RNA language model built from bidirectional Mamba2 blocks with a flash-attention head, pretrained on 100 million sequences up to 2,048 nucleotides.

Phase separation prediction from sequence alone, pairing a protein language model with MD-trained conformational features to score every residue.

Retrieval-augmented diffusion model that designs antibody CDR sequences by conditioning on structurally homologous CDR-like motifs from the PDB.

Multimodal protein model pairing a sequence encoder with a Gene Ontology branch, trained in recursive cycles that clean their own noisy labels.

Immunogenicity classifier for reverse vaccinology, fusing frozen protein language model embeddings with Foldseek and ESM3 structure tokens.

Enzyme catalytic pocket design conditioned on a reaction: substrate and product in, pocket backbone, sequence, and EC class out.

Spatial transcriptomics foundation model using cross-attention over niche ligand genes, pretrained on 4.1M deconvolved human Visium samples.

Multi-task EEG foundation model that treats brain signals as a foreign language, pairing a text-aligned neural tokenizer with a GPT-2 backbone.

Structure-based molecular design that samples a quantum electron cloud in the protein pocket, then decodes it into ligands with a Llama-style model.

EEG foundation model that learns transferable brain-signal representations with a vector-quantized tokenizer and masked transformer pretraining.

RNA language model pretrained on 23 million non-coding sequences with motif-level masking, then steered by predicted RNA type at fine-tuning.

Protein language model pairing sequence with quantized local-structure tokens via disentangled attention, for zero-shot variant effect prediction.

Separable 4D CNN that strips rotational streak artifacts from respiration-resolved cone-beam CT, processing all ten breathing phases in one pass.

Zero-shot variant effect prediction that fuses a frozen protein language model with an equivariant graph network over residue contact graphs.

Protein motif-scaffolding driven by a protein language model instead of MSA pretraining, solving 22 of 24 benchmark problems with diverse backbones.

Vision-language pre-training framework for universal brain MRI diagnosis, learning from imaging-report pairs to cover more than ten brain diseases.

Cardiac CT motion artifact reduction that treats the phase series as video, using self-attention along time to deblur the whole heart at any phase.

Radiology foundation model that reads interleaved 2D and 3D scans with text for diagnosis, visual question answering, and report generation.

Protein language model ranking stabilizing mutations with no assay data, trained jointly to predict the growth temperature of each sequence's host.

Endoscopy video foundation model that learns spatial-temporal representations from unlabeled clips for classification, segmentation, and detection.

3D CT localization foundation model that pairs MedLAM with SAM to segment any anatomical structure at a fixed, dataset-independent annotation cost.

Vision-language framework for 3D medical image diagnosis and visual question answering, bridging frozen image encoders and LLMs, shown on brain MRI.

Generative medical visual question answering model that pairs a vision encoder with a language model, trained on the 227k-pair PMC-VQA dataset.

Biomedical vision-language model trained contrastively on 1.6M figure-caption pairs mined from PubMed Central open-access articles.