ShanghaiTech University

EEG foundation model pairing masked contextual modelling with cross-view invariance learning over 11,000 hours of routine clinical recordings.

Atomic protein model building from cryo-EM density maps, resolving conformational heterogeneity through atom-centric sampling and diffusion.

RNA language model transfer-trained from ESM-2 via a pseudo-protein alphabet mapping, outperforming 12 RNA language models in zero-shot evaluation.

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

RNA-protein contact prediction from sequence, built on ERNIE-RNA and ESM-2 embeddings. Reaches 0.77 auROC where AlphaFold 3 reaches 0.61.

Relative protein-ligand binding affinity prediction from docked complexes, matching Schrodinger FEP+ ranking accuracy zero-shot on the FEP benchmark.

Physics-guided all-atom diffusion model for protein-ligand complex prediction, reaching 95.3% success on PoseBusters redocking with a known pocket.

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

Vision-language model for open-vocabulary mouse behavior analysis, describing multi-view video and pose kinematics in natural language.

PET/CT foundation model pretrained by cross-modal masked autoencoding on whole-body scans, for tumor lesion segmentation and lymphoma staging.

CT segmentation foundation model that uses task prompts to segment 83 anatomical structures and lesions across whole-body scans in a single network.

Text-guided MRI synthesis model that generates brain MR sequences and resolutions on demand from routine scans using imaging-metadata prompts.

EEG foundation model pretrained with vector-quantized self-supervision, yielding interpretable discrete codes that transfer to seizure detection.

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

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

Self-supervised pretraining framework for medical imaging that unifies pixel restoration with contrastive learning across 2D and 3D image backbones.