City University of Hong Kong

Pocket-conditioned 3D ligand generator that steers flow matching with LLM-written chemical priors, for de novo design and scaffold hopping.

Fragment-protein interaction model trained on cellular chemoproteomics, pairing ESM-2 residue embeddings with bilinear attention over ligand atoms.

Single-cell perturbation model that generates a transcriptome gene by gene, letting a regulatory-network policy choose which genes come first.

Cis-regulatory network reconstruction from DNA sequence, epigenomic tracks, and Hi-C priors, constrained by gene expression. 0.84 zero-shot auROC.

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

Cross-species RBP-RNA binding site predictor that turns RNA-binding protein conservation into label smoothing, reaching 0.85 AUC from human to mouse.

Antibody-antigen binding free energy predictor fusing ESM-2 sequence embeddings with persistent homology and interface geometry via cross-attention.

Blind protein-ligand docking model adding Ollivier-Ricci curvature descriptors and degree-aware message passing, predicting poses in 0.09 seconds.

Protein-DNA binding free energy change prediction for missense mutations, with double- and single-stranded DNA binders modeled separately.

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

Multimodal foundation model for cardiac biosignals, pretrained by masked modeling on ECG, PPG, and clinical text from ~1.7 million individuals.

Conditional GAN generating HLA class I pseudo-sequences from a peptide, then resolving them to candidate alleles by nearest-neighbor lookup.

Allele-free HLA class I epitope classification from peptide sequence alone, via LoRA-adapted ESM-2 with parallel CNN and Transformer branches.

Plasmid characterization and retrieval model aligning DNA sequences with property text across ten facets, from antimicrobial resistance to host range.

Multimodal 3D genome foundation model pairing Hi-C contact maps with epigenomic tracks, pretrained on over one million paired samples.

Protein function prediction fusing five Gene Ontology pipelines, two of them deep models over protein and DNA language model embeddings.

Epitope prediction model scoring whether a peptide is presented by HLA class I or II, with no allele input needed. Built on ESM-2 embeddings.

Phosphorylation site prediction from protein sequence. A LoRA-adapted ESM-2 encoder feeds a conformer, reaching 79.5% AUC at serine sites.

Whole-slide pathology assistant that states the morphological findings behind each diagnosis, trained on 180k VQA pairs from 9,850 gigapixel slides.

RNA language model pretrained on 23 million non-coding sequences with motif-level masking, then steered by predicted RNA type at fine-tuning.

Unified DNA, RNA, and protein foundation model with 1.8B parameters, pretrained across 169,861 species to learn the central dogma from sequence.

Abdominal CT segmentation model driven by CLIP text embeddings, covering 25 organs and 6 tumor types with zero-shot extension to new categories.

Open large language models for natural science, fine-tuned on physics, chemistry, and materials science literature with automated instruction tuning.