Sichuan University

Gene regulatory network inference from scRNA-seq that returns directed TF-to-target edges in one forward pass, with no per-dataset refitting.

Breast-specialized multimodal pathology foundation model for core needle biopsy diagnosis, with conformal risk control gating report release.

Vision-language model for neuroblastoma pathology that reads H&E slides with their reports to grade tumors, infer biomarkers and stratify risk.

Spatial transcriptomics foundation model for the tumor microenvironment, giving TME-aware embeddings and in silico perturbation from one checkpoint.

Protein representation model adding global fold-similarity and local substructure signals to masked pretraining, reaching 79.2 long-range contact P@L.

Structure-based drug design that schedules noise separately for 3D coordinates and 2D topology, reaching a 95.9% PoseBusters valid rate on CrossDock.

Pocket-conditioned 3D diffusion model for scaffold decoration, guided by evolutionary residue conservation and a protein-ligand interaction prior.

Sparse-view CBCT reconstruction foundation model pretrained on 8,407 CT volumes, recovering full 3D anatomy from as few as six X-ray projections.

PROTAC degradation prediction from molecular graphs of the target, linker, and E3 ligase, combining cross-attention with contrastive learning.

SMILES language model pretrained by editing: substructures are dropped and restored, giving fragment-level supervision for property prediction.

Pocket-conditioned 3D ligand generator trained on its own predicted conformations, closing the train-inference gap that degrades diffusion sampling.

Direct RNA nanopore basecaller pairing a densely connected 1D CNN and CTC decoder with a Random Forest classifier that demultiplexes 24 barcodes.

Self-supervised foundation model for 3D medical image segmentation, pretrained on roughly 110,000 unannotated CT volumes via Volume Fusion.

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

Histopathology tile encoder pairing a CNN stem with a multi-scale Swin Transformer, pretrained on 15.6 million unlabeled H&E patches.

Histopathology patch encoder self-supervised with BYOL on 2.7M unlabeled H&E tiles, using attention that gates feature channels with pooled context.