Chinese Academy of Sciences

A direction-aware foundation model trained on ~23M bulk RNA-seq differential-expression profiles that simulates coordinated gene dynamics in viral infection.

Self-supervised generative foundation model jointly trained on 91.7M nucleotide sequences and structured annotations spanning 1.076 trillion bases, achieving SOTA on 23 nucleotide-language benchmarks.

Autoregressive language model trained on 37 million intrinsically disordered region sequences from the AlphaFold Database, generating IDR sequences conditioned on surrounding structured context.

Autoregressive generative model pretrained on 30 million full-length natural mRNA sequences that jointly optimizes 5' UTR, CDS, and 3' UTR for therapeutic mRNA stability and translation efficiency.

Subspecialty-specific computational pathology foundation model pretrained on 353 million multi-scale patches from 210,000 H&E slides for gastrointestinal pathology, achieving SOTA on 32 of 33 systematic downstream tasks.

Billion-parameter single-cell foundation model performing full self-attention across all 28,000 human genes, integrating Gene Ontology priors via GCN for long-range gene context capture in transcriptomics.

GRN-informed single-cell foundation model combining gene regulatory hierarchy priors with long-sequence Mamba modeling for clustering, batch integration, perturbation modeling, and drug response prediction.

Transformer-based protein language diffusion model generating all-atom intrinsically disordered protein conformational ensembles, validated against experimental NMR and SAXS data.

Long-context generative RNA foundation model trained on 114 million full-length RNA sequences, supporting de novo design of tRNAs, aptamers, CRISPR guide RNAs, mRNAs, and circular RNAs.

A structure-aware generative DNA language model pretrained on influenza genomes that forecasts future antigenic variants across regions and subtypes.

A deep learning framework that predicts locus-specific DNA methylation across 39 human tissues from genomic sequence, with a scRNA-seq-augmented variant for unseen cell types.

Grounded multimodal large language model for endoscopic surgery, supporting visual dialogue, region-based question answering, and bounding-box grounding across surgical scene understanding tasks.

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

Knowledge-informed cross-species foundation model pre-trained on 101M human and mouse single-cell transcriptomes to decipher universal gene regulatory mechanisms.