Shanghai AI Laboratory

Infrared spectroscopy foundation model pretrained on 60 million simulated spectra, then adapted to real FTIR measurements of molecules and mixtures.

Genomic foundation model on a Mamba, attention, and mixture-of-experts backbone, with 1M-token context for variant scoring and regulatory DNA design.

Antibody CDR design model post-trained by on-policy distillation, cutting RAbD CDR-H3 backbone RMSD from 2.37 Å to 1.95 Å.

Decoder-only foundation model that unifies sequences, 3D structures, and natural language for small molecules and proteins in one shared token space.

NMR foundation model trained on 158 million simulated 1H and 13C spectra, transferring simulation-learned representations to real experimental data.

Protein-text foundation model placing amino acid sequences and natural language in one token space for protein understanding and de novo design.

Codon-level mRNA language model adapted from ESM-2 650M by swapping amino-acid tokens for codon tokens, transferring protein knowledge to mRNA tasks.

Virtual cell foundation model predicting single-cell responses to genetic, chemical, and cytokine perturbations with conditional flow matching.

De novo peptide sequencing transformer that reads modified and unmodified peptides directly from tandem mass spectra without a reference database.

Conversational single-cell and spatial multi-omics brain foundation model, with zero-shot cell annotation and disease prediction across species.

Spectroscopy-grounded molecular foundation model that reads NMR, IR, and mass spectra as text, elucidating structures and generating 3D conformers.

Multimodal scientific foundation model unifying protein, DNA/RNA, and small-molecule structure in one token vocabulary for cross-domain reasoning.

Structure-based virtual screening model that jointly predicts protein-ligand complex structures and binding fitness from sequence and SMILES.

Scientific multimodal foundation model, a 241B-parameter MoE with a tokenizer that reads molecular formulas and protein sequences natively.

Protein foundation model built on Bayesian Flow Networks, prompted with MSA profiles for structure- and function-preserving sequence design.

EEG foundation model with cross-scale spatiotemporal tokenization and sparse structured attention, evaluated on 11 decoding tasks across 16 datasets.

De novo peptide sequencing from tandem mass spectra with a non-autoregressive Transformer trained on 100 million peptide-spectrum matches.

Single-cell perturbation response prediction using dual conditional diffusion bridges that link unpaired control and perturbed populations.

Medical multimodal LLM (2B and 8B) trained for generalizable, step-by-step clinical reasoning via Mentor-Intern Collaborative Search.

De novo peptide sequencing from tandem mass spectra, using curriculum learning and iterative self-refinement to stabilize non-autoregressive decoding.

Brain foundation model unifying EEG and MEG in a single encoder via a shared discrete tokenizer that transfers across sensor layouts and montages.

Zero-shot tumor segmentation on CT and MRI that reads text-prompted anomaly attention maps out of a frozen medical foundation diffusion model.

Brain MRI segmentation foundation model trained on 66,000+ image-label pairs across 14 MRI sub-modalities, with a hypergraph dynamic adapter.

General medical vision-language model trained with reinforcement learning to reason step by step over medical images for diagnosis and visual QA.

Multimodal sequence model spanning proteins, coding DNA, and regulatory DNA for zero-shot fitness scoring and conditional sequence generation.

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

Pan-tumour CT foundation model pretrained on 30,000 synthetic 3D scans carrying lesion masks and structured reports across ten organ systems.

Protein expression prediction that pinpoints expression-governing residues by matching a language model landscape against measured fitness data.

Multi-omics instruction-tuned LLM that reads DNA, RNA, protein, and multi-molecule sequences and answers natural-language questions about them.

Protein and RNA sequence annotation that also names the residues driving each label, learned from sequence-level supervision alone.

Histopathology vision-language foundation model that folds a disease knowledge graph into pretraining for zero-shot cancer detection and subtyping.

Protein subcellular localization from sequence, returning both a text label and a synthetic fluorescence image of the protein inside a given nucleus.

Vision-language assistant that reads a whole gigapixel pathology slide, answering diagnostic questions and writing slide-level descriptions.

Masked-autoencoder foundation model for chest radiographs, self-supervised on 1.04 million unlabelled images for disease screening and localization.

Cryo-EM and cryo-ET heterogeneity analysis that separates subunit rigid-body motion from compositional change into distinct latent spaces.

RNA language model adapted from ESM-2 by cross-modality transfer learning, matching RNA-native baselines with 1/8 the trainable parameters.

Chest X-ray foundation model that pairs masked image modeling with image-report contrastive alignment for zero-shot diagnosis and phrase grounding.

Protein language model pairing sequence with quantized local-structure tokens via disentangled attention, for zero-shot variant effect prediction.

Zero-shot variant effect prediction that fuses a frozen protein language model with an equivariant graph network over residue contact graphs.

Fully 3D promptable segmentation foundation model for volumetric CT and MR, encoding whole volumes so anatomy can be segmented from one prompt point.

Vision-language pre-training framework for universal brain MRI diagnosis, learning from imaging-report pairs to cover more than ten brain diseases.

Medical imaging adaptation of the Segment Anything Model, fine-tuned on 4.6M images and 19.7M masks for promptable segmentation across 10 modalities.

Radiology foundation model that reads interleaved 2D and 3D scans with text for diagnosis, visual question answering, and report generation.

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

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.

Generative medical visual question answering model that pairs a vision encoder with a language model, trained on the 227k-pair PMC-VQA dataset.

Scalable and transferable U-Net family (14M–1.4B parameters) for 3D medical image segmentation, supervised-pretrained on TotalSegmentator.

RNA foundation model pretrained on 23.7 million non-coding RNA sequences, producing embeddings for structure prediction, annotation, and RNA design.