University of Chinese Academy of Sciences

All-atom de novo protein design that denoises sequence and structure as a single Cartesian process, reading residue identity from atomic occupancy.

3' UTR foundation model pretrained on 7 million eukaryotic untranslated regions for mRNA stability prediction and de novo UTR sequence design.

Protein inverse folding that fuses sequence and structure priors from two frozen language models per residue. Reaches 63.11% recovery on CATH 4.2.

Protein structure prediction from multiple sequence alignments, trained across MSA depths so one model spans deep alignments and orphan proteins.

Antifreeze protein classifier over frozen ESM-2 embeddings, trained only on sequences whose antifreeze activity was measured in the lab.

Long-context RNA foundation model reading whole mRNA transcripts at single-nucleotide resolution, pretrained at a native 10,240 nt context.

Codon language model for mRNA prediction and coding sequence design, steering synonymous codon choice with a swappable host usage prior at inference.

Molecular glue degrader activity prediction from SMILES and protein sequence, with two-stage cross-attention that follows E3-then-substrate binding.

Single-cell foundation model that predicts latent representations of graph-connected gene blocks instead of reconstructing individual gene counts.

RNA and single-stranded DNA 3D structure prediction from sequence alone, with no MSA or language-model inputs and roughly 100x cheaper inference.

Vision-guided model that builds virtual 3D organoid surrogates from brightfield microscopy to predict chemical perturbation responses without omics.

Multi-view single-cell foundation model at 12B parameters, aligning expression and perception views with an LLM-JEPA joint-embedding objective.

Diffusion-based generative RNA model for de novo sequence design, conditioned on function, RNA family, structure, or binding proteins.

Phylogeny-aware genome annotation model predicting exons, introns, UTRs and repeats directly from eukaryotic DNA, with no RNA or protein evidence.

Phosphopeptide detectability prediction for mass spectrometry, rescoring DDA identifications and pruning DIA spectral libraries to cut search time.

Latent diffusion model that backmaps coarse-grained protein structures to all-atom detail, with bond geometry learned into the latent space.

Generative antibody and nanobody design model that co-designs CDR sequences and antigen-bound structures for de novo design and affinity maturation.

Influenza genomic language model adapting DNABERT-2 to ~900,000 viral genomes, identifying subtypes, segments, and pathogenicity from sequence.

Species-conditioned codon language model that jointly reads 5'UTRs, coding sequence, and RNA secondary structure to design native-like genes.

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

Cross-species brain spatial transcriptomics foundation model pretrained on 133M cells from human, macaque, marmoset, and mouse whole brains.

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.

Relative protein-ligand binding affinity prediction from docked complexes, matching Schrodinger FEP+ ranking accuracy zero-shot on the FEP benchmark.

Cyclic peptide design conditioned on target protein structure, generating all four cyclization types via all-atom, all-bond harmonic SDE modeling.

Compact 1.1M-parameter DNA language model distilled from Nucleotide Transformer v2, outperforming its 500M teacher on 11 of 18 benchmark tasks.

Molecular conformation description language encoding 3D geometry as SMILES plus internal-coordinate tokens, turning 3D modeling into a sequence task.

All-atom generative model for protein complexes that designs multi-chain binders from scratch and performs multimer folding and inverse folding.

Single-cell multi-omics foundation model with a Mamba backbone, pretrained on 2.7 million paired scRNA-seq and scATAC-seq profiles.

De novo peptide sequencing from mirror-protease mass spectra, reading paired complementary spectra to recover near-complete fragment ion coverage.

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

Secondary metabolite structure prediction from microbial biosynthetic gene clusters, generating SMILES strings from Pfam functional-domain tokens.

Ternary complex structure predictor for PROTACs and molecular glues, placing E3 ligase, degrader, and target protein in one SE(3)-equivariant pass.

Translates whole-brain imaging phenotypes between humans and mice through a shared latent space built from transcriptomics and connectivity.

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

Structure-based molecular generation guided by imputed ligand electron density, assembling drug-like compounds into a pocket fragment by fragment.

Histopathology vision-language model handling image patches and gigapixel slides in one 15B checkpoint, across classification, VQA, and captioning.

Protein conformational ensemble generator using SE(3) flow matching from a perturbed ESMFold prior, sampling MD-like dynamics from sequence alone.

Immune protein structure prediction for TCRs, antibodies, and nanobodies. Adapts ESMFold with LoRA, reaching 1.31 Å RMSD on the CDR3-beta loop.

Prokaryotic promoter design and zero-shot promoter-strength ranking from a nucleotide language model pretrained on 17,806 microbial genomes.

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

Chemical perturbation model generating post-treatment transcriptomes for compounds and cell lines never screened, from SMILES structure and dose.

EEG foundation model that makes each electrode its own token stream, pretrained by causal next-signal prediction over 138 electrode positions.

Self-supervised 3D molecular pretraining that corrupts and rebuilds tetrahedral local structures, then transfers to toxicity and affinity prediction.

Discrete diffusion model that generates RNA secondary structure contact maps as pixel-wise segmentation, conditioned on RNA-FM and UFold features.

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

Vision-language model for annotation-free pathology localization, marking the finding a text prompt names in X-ray, histology and fundus images.

Conditional GAN that generates small molecules against a protein-protein interaction interface, encoding the complex with graph attention networks.

Equivariant heterogeneous graph network that rescores docked protein-ligand poses for virtual screening and ranks structural analogs by activity.