Zhejiang University

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

Single-cell perturbation model that generates a transcriptome gene by gene, letting a regulatory-network policy choose which genes come first.

Cross-context protein binder design model that generates one sequence binding multiple targets or conformational states of the same target.

Antibody CDR design model post-trained by on-policy distillation, cutting RAbD CDR-H3 backbone RMSD from 2.37 Å to 1.95 Å.

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.

Diffusion-based generative RNA model for de novo sequence design, conditioned on function, RNA family, structure, or binding proteins.

Lung pathology foundation model adapted from Virchow2 on whole-slide images, validated across 32 tasks spanning the lung diagnostic workflow.

Autoregressive nucleotide-and-text foundation model generating DNA and RNA sequences from natural-language prompts that name species and function.

Ab initio heterogeneous cryo-EM reconstruction seeds its encoder with foundation-model priors, sorting 100 structures from one simulated mixture.

De novo peptide sequencing transformer that reads modified and unmodified peptides directly from tandem mass spectra without a reference database.

Compact protein fitness predictor that fuses within-family evolutionary profiles with inverse-folding logits for zero-shot variant effect prediction.

Protein function prediction via compressed in-context learning on a sequence-structure language model, cutting 751-token demonstrations to under 16.

Medical image grounding model that localizes text phrases in CT, MRI, X-ray, ultrasound, endoscopy, dermoscopy, and fundus images.

Host-pathogen protein interaction predictor scoring bacterial effector and human protein pairs from frozen ESM-2 embeddings with a transformer.

Scoring function for protein, nucleic acid, and small-molecule complexes that predicts binding affinity, ranks docked poses, and screens ligands.

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

De novo peptide sequencing from tandem mass spectra with a non-autoregressive Transformer trained on 100 million peptide-spectrum matches.

De novo peptide sequencing from tandem mass spectra, using curriculum learning and iterative self-refinement to stabilize non-autoregressive decoding.

Multimodal ECG model that reads raw 12-lead waveforms and ECG images in one shared space via a structure-aware discrete signal tokenizer.

Agent-based pathology model that navigates whole-slide images by zooming and panning like a pathologist, scoring 88.6% on the PathMMU-HR2 benchmark.

Reasoning LLM for single-cell type annotation, mapping per-cell expression to labels with marker-by-marker chains of thought on one GPU.

Protein-ligand binding affinity model that tokenizes quantum electron-cloud density into discrete codes, plus a distilled cloud-free variant.

Zero-shot tumor segmentation on CT and MRI that reads text-prompted anomaly attention maps out of a frozen medical foundation diffusion model.

Single-cell foundation model inferring context-specific protein-protein interactions from cancer transcriptomes via a variational graph autoencoder.

Topology-guided protein backbone generation that turns hand-drawn 3D curves into designable structures by steering a diffusion sampler with a sketch.

Protein sequence design by flow matching in a compressed language-model latent space, spanning peptides, antibodies, and antimicrobial peptides.

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

Slide-level pathology foundation model that vector-quantizes tile patch tokens at 64x compression, keeping spatial detail for whole-slide analysis.

Cross-domain molecular foundation model encoding small molecules, protein pockets, and their complexes in 2D and 3D on one Transformer backbone.

Ternary complex structure predictor for PROTACs and molecular glues, placing E3 ligase, degrader, and target protein in one SE(3)-equivariant pass.

Medical vision-language model that unifies image comprehension and generation in one autoregressive transformer via heterogeneous LoRA adapters.

Genomic foundation model for metagenomic annotation: a 500M-parameter bidirectional encoder calling coding regions at single-nucleotide resolution.

Multi-omics foundation model that folds DNA, RNA, and protein into one codon-level nucleotide representation following the central dogma.

Cyclic peptide structure prediction for sequences carrying unnatural amino acids, adding atom-level features and cyclization-aware position encoding.

Single-cell analysis model driven by plain-language instructions, covering cell type annotation, pseudo-cell generation, and drug response prediction.

RNA foundation model unifying sequence representation, 3D structure prediction, and de novo design. Ranks first on 11 of 13 BEACON tasks.

Protein-text foundation model aligning sequences with function descriptions through segment-wise objectives for static and dynamic functional sites.

Histopathology vision-language model handling image patches and gigapixel slides in one 15B checkpoint, across classification, VQA, and captioning.

Antibody CDR sequence and structure co-design from the whole antigen, using a relation-aware equivariant graph network with no specified epitope.

Pathology vision-language model for whole-slide diagnosis, adding lesion detection and segmentation to visual question answering on gigapixel images.

EEG foundation model for brain-computer interface decoding, factorizing self-attention into parallel spatial and temporal branches.

TCR-antigen binding affinity prediction for unseen peptides, pairing an ESM-2-initialized receptor encoder with a peptide cross-attention module.

Protein inverse folding for low-resource enzyme design, distilling a frozen protein language model into a structure encoder used alone at inference.

RNA-binding protein binding profiles predicted base by base across 800 bp windows, with an m6A signal channel that exposes methylation-RBP crosstalk.

Prompt-guided protein sequence design conditioned on 3D backbones, fold blueprints, and functional tags. 63.21% native sequence recovery on CATH.

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

Post-translational modification prediction from sequence and 3D structure, quantizing each residue's micro-environment into a per-PTM discrete token.

Protein inverse folding as a generative Markov bridge, refining a structure-derived sequence prior with a frozen protein language model.

Predicts binding free energy change (ΔΔG) at protein-protein interfaces by scoring bound and unbound states with an inverse folding model.

AAV capsid design platform for gene therapy that steers a peptide language model toward inserts combining receptor targeting and production fitness.

Multimodal LLM aligning natural language, small molecules and proteins in any direction, turning prose design goals into molecules or enzymes.

Conditional diffusion model that generates P-type ATPase backbone conformations in a specified E1, E1P, E2P, or E2 functional state.

Enzyme redesign framework built on a structure-to-sequence protein network, scoring mutants and generating sequences without retraining.

Histopathology vision-language model classifying H&E patches zero-shot from text prompts, trained on 1.6M captions written for whole-slide crops.

Histopathology multimodal assistant answering questions about H&E patches, pairing a pathology-trained CLIP tower with a 13B Vicuna language model.

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

Unified DNA, RNA, and protein foundation model with 1.8B parameters, pretrained across 169,861 species to learn the central dogma from sequence.

Lightweight mixture-of-experts medical vision-language model routing visual question answering and image classification to domain-specific experts.

Foundation model spanning invasive SEEG/iEEG and non-invasive EEG in one backbone, with zero- and few-shot transfer across neurological disorders.

Conversational T5-based framework that turns scRNA-seq data into cell sentences for cell type annotation and drug sensitivity prediction.

Inverse protein folding from backbone coordinates, chaining a pretrained structure encoder into a pretrained sequence autoencoder on small data.

500M-parameter transformer model pretrained on intracranial SEEG recordings for neural signal forecasting, imputation, and seizure detection.

Equivariant heterogeneous graph network that rescores docked protein-ligand poses for virtual screening and ranks structural analogs by activity.

Multimodal pathology assistant that answers questions about histology and cytology images, pairing the PathCLIP vision encoder with a Vicuna-13B LLM.

Protein structure encoder pretrained by contrastive alignment to a frozen protein language model, anchored by self-supervised contact-map prediction.

3D CT segmentation of abdominal organs and tumors, where a deformable-attention Transformer decodes organ embeddings into the segmentation kernels.

Protein language model that fuses Gene Ontology knowledge graphs with masked language modeling, improving protein function and interaction prediction.