Westlake University

Virtual-cell model that compresses a transcriptome into eight discrete tokens in a reasoning LLM's vocabulary, predicting module-level drug response.

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

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

Hallucination framework for de novo nucleic acid design, pairing NA-MPNN sequence proposals with a frozen AlphaFold3 or Protenix structure oracle.

Single-cell perturbation prediction model using conditional flow matching to map control cells to perturbed expression distributions.

Generative language model for phenotype-driven drug discovery, proposing small-molecule structures from up- and down-regulated gene signatures.

Ab initio heterogeneous cryo-EM reconstruction seeds its encoder with foundation-model priors, sorting 100 structures from one simulated mixture.

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

Hierarchical DNA foundation model that co-trains a dynamic token-merging tokenizer with latent Transformers to match genomic information density.

Long-sequence DNA foundation model with groove-aware convolutions and reverse-complement gating over 100kb contexts. Averages 0.708 MCC on GUE.

Nanomaterial-protein interaction prediction from protein sequence, structure, and experimental context that generalizes to unseen materials.

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

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

Agent-based pathology model that navigates whole-slide images by zooming and panning like a pathologist, scoring 88.6% on the PathMMU-HR2 benchmark.

Protein-ligand binding affinity model that tokenizes quantum electron-cloud density into discrete codes, plus a distilled cloud-free variant.

De novo protein design from natural language: a 16B-parameter framework turning text descriptions into sequences via structure-conditioned generation.

Slide-level pathology foundation model that vector-quantizes tile patch tokens at 64x compression, keeping spatial detail for whole-slide analysis.

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

Multi-omics foundation model that folds DNA, RNA, and protein into one codon-level nucleotide representation following the central dogma.

Binding free energy change prediction for antibody mutations, from a lightweight transformer that scores 10,000 variants in under five minutes.

Multimodal medical imaging foundation model for zero-shot clinical diagnosis and report generation from chest X-ray and CT in English and Chinese.

Spatial genomics foundation model that predicts a cell's transcriptome from its neighbours and forward-simulates tissue response to perturbation.

Biomedical vision-language model aligning image regions to UMLS clinical concepts, for zero-shot diagnosis across 10 imaging modalities.

Multimodal 80B-parameter protein-language model that answers natural language questions about protein function from sequence and structure.

Multiomic foundation model for zero-shot in silico perturbation, predicting gene regulation and cell fate transitions from DNA and ATAC signal.

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

Antibody CDR sequence and structure co-design from the whole antigen, using a relation-aware equivariant graph network with no specified epitope.

Protein inverse folding for low-resource enzyme design, distilling a frozen protein language model into a structure encoder used alone at inference.

TCR-antigen binding affinity prediction for unseen peptides, pairing an ESM-2-initialized receptor encoder with a peptide cross-attention module.

Peptide-spectrum match rescoring for DDA proteomics, learned end to end from raw MS2 spectra and peptide sequence across 271 million PSMs.

De novo atomic model building from cryo-EM density maps, adapting AlphaFold2 with local attention and a 3D rotary position embedding.

DIA proteomics scoring model that identifies and quantifies peptide precursors, pretrained across 952 mass spectrometry runs instead of one.

Cryo-EM foundation model pre-trained on 65 million particle images, enabling zero-shot classification, pose clustering, and quality assessment.

Post-translational modification prediction from sequence and 3D structure, quantizing each residue's micro-environment into a per-PTM discrete token.

De novo peptide sequencing model that retrieves a similar peptide-spectrum match from a database and fuses it into transformer decoding.

Enzyme redesign framework built on a structure-to-sequence protein network, scoring mutants and generating sequences without retraining.

Histopathology vision-language model classifying H&E patches zero-shot from text prompts, trained on 1.6M captions written for whole-slide crops.

Histopathology multimodal assistant answering questions about H&E patches, pairing a pathology-trained CLIP tower with a 13B Vicuna language model.

Tri-modal protein language model aligning sequence, structure, and text in one embedding space for natural-language search over billions of proteins.

Inverse protein folding from backbone coordinates, chaining a pretrained structure encoder into a pretrained sequence autoencoder on small data.

Structure-aware protein language model pairing amino acid tokens with Foldseek 3Di structural states, outperforming ESM-2 across 10 downstream tasks.

Multimodal pathology assistant that answers questions about histology and cytology images, pairing the PathCLIP vision encoder with a Vicuna-13B LLM.

Protein structure encoder pretrained by contrastive alignment to a frozen protein language model, anchored by self-supervised contact-map prediction.

Remote-homolog template recognition that threads a sequence against clustered PDB and AlphaFold DB structures to improve AlphaFold2 modelling.