Peking University

Multimodal foundation model pretrained on 1.76 billion paired histology-spatial transcriptomics spots, linking whole-slide images to spatial molecular programs.

A CLIP-style dual-encoder model that learns a shared peptide-spectrum representation for zero-shot peptide-spectrum-match inference in DIA proteomics.

A probabilistic autoregressive model for protein molecular-dynamics trajectories that generates flexible-length paths frame-by-frame with an anti-drifting sampling strategy.

Property-driven protein inverse folding: a ProteinMPNN checkpoint aligned via multi-objective preference optimization to improve developability while preserving structural fidelity.

A pan-cancer pretrained latent diffusion model that enhances spatial transcriptomics, imputing genome-wide expression from sparse panels with zero-shot generalization.

A medical vision-language model that accepts visual-referring multimodal input and produces pixel-grounded multimodal output, jointly answering and segmenting medical images.

A spatiotemporal foundation model that learns generalizable representations directly from 4D functional MRI volumes for diverse brain-imaging tasks.

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

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

Self-supervised ECG foundation model that treats heartbeats as words and rhythms as sentences, using a QRS-Tokenizer and dual-level pretraining on MIMIC-IV-ECG.

A self-improving text-to-image diffusion foundation model that generates synthetic medical images across multiple modalities and organs to augment downstream clinical AI tasks.

Biomedical multimodal LLM with pixel-level insight, combining visual question answering, pixel-grounded prompts, and segmentation via a mixture-of-experts design.

A GPT-style language model trained on whole-night sleep stage sequences that enhances automated sleep staging and enables sleep disorder diagnosis.

A 1D convolutional foundation model for electrocardiogram analysis, trained on over 10 million expert-annotated recordings across 150 diagnostic categories.

A lightweight mixture-of-experts medical vision-language model that routes between domain-specific experts for VQA and image classification while activating only 30-50% of parameters.

A 7B-parameter protein language model built on LLaMA-2 that performs both protein sequence generation and superfamily classification in a unified framework.

Unsupervised RNA language model using multiple sequence alignments to predict secondary structure and solvent accessibility from evolutionary information.

Text-informed self-supervised vision-language pretraining for 3D CT volumes, enabling zero-shot classification, retrieval, report generation, and segmentation.

A BERT-based RNA foundation model trained on 23.7 million non-coding RNA sequences, producing embeddings for structure prediction, functional annotation, and RNA design.