Peking University

Designs RNA and DNA aptamers against protein targets by backpropagating binding and anti-binding objectives through a frozen all-atom predictor.

Histopathology foundation model for uterine malignancies that orders whole-slide morphology into continuous, progression-associated tumor states.

De novo siRNA design conditioned on target mRNA context, generating 21-nt duplexes by masked discrete diffusion under efficacy guidance.

EEG foundation model pretrained to predict structured latent states rather than masked waveforms, reaching 52.94% frozen macro balanced accuracy.

Protein language model adding long-range contact supervision to ESM2 via LoRA, improving all eight protein-level tasks with no structural input.

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

Single-cell perturbation-response model predicting transcriptomic and cell-number changes for unseen perturbations plus inverse design.

Pan-fungal circRNA prediction from genome sequence and gene annotation alone, ranking candidate backsplice junctions without requiring RNA-seq.

Multimodal molecular generation model for drug design, conditioned on properties, pharmacophores, protein sequences, or protein binding pockets.

Multimodal foundation model pretrained on 1.76B histology and spatial transcriptomics spots, inferring molecular state from whole-slide images.

Multiphoton pathology vision-language system turning one label-free breast section into virtual H&E, a margin heatmap, and a written report.

Contrastive dual-encoder model for DIA proteomics, embedding peptides and spectra in a shared space for zero-shot peptide-spectrum matching.

Autoregressive generative model for protein molecular dynamics that emits flexible-length trajectories frame by frame with anti-drifting sampling.

Protein inverse folding model aligning ProteinMPNN by multi-objective preference optimization to improve developability without losing fold fidelity.

Pan-cancer pretrained diffusion model imputing genome-wide expression from sparse spatial transcriptomics panels, zero-shot and reference-free.

Mixture-of-Experts generative model turning DNA sequence plus cell-type ATAC-seq into unified epigenomic, transcriptomic, and 3D chromatin profiles.

Medical vision-language model that takes visual prompts on an image and returns answers grounded in pixel-level segmentation masks.

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

Protein structure prediction model that weights residues by protein-language-model importance scores to improve accuracy on hard AlphaFold2 targets.

Scientific multitask language model whose byte-level genome modeling beats Evo 7B on DNA perplexity and mutation-effect prediction at 1.5B params.

Cross-species RBP-RNA binding site predictor that turns RNA-binding protein conservation into label smoothing, reaching 0.85 AUC from human to mouse.

SE(3)-equivariant chemical language model for pocket-based 3D molecule generation, used to design an HPK1 inhibitor with in vivo anti-tumor efficacy.

Hierarchical transformer for virus discovery in metagenomes, classifying viral genomes across taxonomic ranks and flagging candidate novel lineages.

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

Dermatology foundation model pretrained on 432,776 skin images, covering malignancy classification, severity grading, and lesion segmentation.

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

Organic reaction foundation model that tokenizes 3D molecular structure to predict products, retrosynthetic routes, conditions, and yields.

Spatial transcriptomics foundation model pairing gene-scale cell embeddings with an SE(2) Transformer over cell coordinates, pretrained on 88M cells.

Fungal genome mining framework that detects biosynthetic gene clusters and identifies their core enzymes from a pretrained Pfam-domain transformer.

Proteomics foundation model for peptide-spectrum scoring and open de novo sequencing, reading over 1,300 modifications from tandem mass spectra.

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

Bulk transcriptome foundation model, 150M parameters over ~20,000 protein-coding genes. Imputes masked expression at Pearson r = 0.954.

Graph-level self-supervised pretraining for 3D molecules, reconstructing whole-molecule geometry to sharpen quantum property and force prediction.

Spatiotemporal foundation model that learns representations directly from 4D functional MRI volumes for disease diagnosis and phenotype prediction.

Tri-modal pathology foundation model aligning whole-slide images, transcriptomes, and diagnostic reports, and running on any subset of the three.

End-to-end protein structure alignment from 3D coordinates, matching TM-align quality at LDDT 0.56 and beating Foldseek on SCOPe40 search.

Unified electron microscopy image analysis toolkit built on EM-DINO, a vision foundation model pretrained on 5 million diverse EM images.

Autoregressive 3D structure model built on an octree tokenizer, spanning molecule generation, molecular docking, and protein pocket prediction.

Binding affinity scoring for protein-ligand and lipid-protein pairs without a docked pose, used to rank the protein corona on candidate liposomes.

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

Protein language model that tokenizes sequence, backbone structure, and text into one vocabulary for function prediction, design, and fold editing.

Generative foundation model that co-generates sequence and 3D coordinates for proteins, small molecules, and crystals under functional objectives.

Multimodal LLM unifying 12-lead ECG time series, ECG images, and text for grounded, clinician-aligned electrocardiogram interpretation.

Full-atom flow matching model that generates a ligand and the induced-fit holo pocket together, starting from an apo binding site.

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

Broad-spectrum antiviral screening framework pairing a pretrained molecular encoder with ESM-2 embeddings for phenotype- and target-based prediction.

ECG foundation model that treats heartbeats as words and rhythm strips as sentences, using heartbeat-level tokenization for diagnostic classification.

Neural ODE model of protein network dynamics, pretrained on 38 million perturbed protein measurements for drug efficacy and synergy prediction.

Cervical cytology screening system pretrained on 127,471 whole-slide images from 48 centers, with test-time adaptation for new clinical sites.

Protein question-answering model that fuses sequence and structure into an LLM prompt as virtual tokens, answering free-form questions about function.

Molecular docking framework that poses several ligands sharing one protein pocket at once, using their consistency to sharpen each prediction.

Per-residue membrane contact and solvent accessibility prediction from sequence alone, replacing MSA input with language model embeddings.

Peptide binder design by mimicking the binding interface of a known receptor or antibody, generating all-atom peptides through latent diffusion.

Breast ultrasound generative foundation model that synthesizes conditioned images to train screening, diagnosis, and prognosis models.

DNA methylation foundation model that encodes 5mC as a fifth base, pretrained on 568 million BS-seq reads for tissue-of-origin and expression tasks.

Biomedical multimodal LLM that answers questions about medical images and returns pixel-level segmentation masks, using a mixture-of-experts design.

Text-guided medical image synthesis across OCT, fundus, X-ray, CT and MRI. Synthetic data lifts downstream clinical tasks by 12-17%.

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

Virtual staining diffusion model that converts kidney histology between H&E, Masson's trichrome, PAS, and PASM in any direction from one checkpoint.

Structure-based molecule optimization that steers a Bayesian flow network with property gradients over 3D coordinates and atom types at once.

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

Language model over whole-night sleep stage sequences that corrects automated sleep staging and supplies features for sleep disorder diagnosis.

Reprograms a frozen single-target diffusion model for dual-target drug design by composing SE(3)-equivariant messages across two aligned pockets.

Molecular docking and design foundation model that unifies structure-based drug design and peptide design at the atom level in one checkpoint.

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

Protein-ligand docking framework that picks the binding pocket by contrastive alignment, then refines the pose with bi-level iterative refinement.

Convolutional ECG foundation model trained on expert annotations spanning 150 diagnostic categories, with 12-lead and single-lead wearable variants.

Single-cell foundation model built on MAMBA2 state-space blocks, reading all 60,530 genes of a transcriptome without variable-gene selection.

Cell-free RNA language model for multi-cancer detection, classifying plasma samples straight from raw sequencing reads without gene annotation.

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

Protein large language model adapted from LLaMA-2 that unifies sequence generation and superfamily classification in one 7B-parameter framework.

Structure-based drug design diffusion model that re-extracts the essential binding subcomplex from a pocket at every step of 3D ligand generation.

RNA language model trained on multiple sequence alignments of Rfam families, predicting secondary structure and solvent accessibility from homology.

Peptide language model that generates antimicrobial, anticancer, and target-binding sequences, adapted per task by Mixture-of-Experts plugins.

Vision-language pretraining for 3D CT volumes, aligning scans with their radiology reports for zero-shot classification, retrieval, and segmentation.

Protein structure tokenizer that encodes all-atom folds as artificial amino acids and decodes them back to coordinates at TM-score above 0.96.

Structure-based drug design by SE(3)-equivariant diffusion over 3D atom coordinates and types, with the same frozen network scoring binding affinity.

Drug repurposing model that predicts a compound's L1000 transcriptional signature from SMILES and ranks it against a disease gene signature.