Haiguang Liu

Carbon-A annotates protein-coding genes across eukaryotic genomes using a large encoder model, achieving high accuracy and enabling the creation of a comprehensive annotation database.

This paper introduces GO-CRE, a method for designing cell-type-specific cis-regulatory elements by reconstructing sequence-grammar trajectories and using reinforcement learning with predictor guidance.

HoloCell is a generative foundation model that integrates epigenomics, transcriptomics, and proteomics data across single-cell multi-omics modalities, demonstrating superior performance in representation learning and cross-modal generation.

This paper introduces improved genomic language models that balance long-range context with single-nucleotide resolution, achieving better performance and efficiency than prior methods.

Ophiuchus-Ab is a generative foundation model for antibody design that captures inter-chain dependencies and predicts antibody properties like developability and binding affinity.

GEMGen generates small molecules from transcriptomic data to phenocopy gene knockdown effects and identify novel inhibitors.

FlexRibbon is a pretrained protein model that jointly learns from sequences and structures, achieving state-of-the-art performance in mutation-rich settings.

CellNavi predicts driver genes for cell state transitions using a deep learning framework that integrates gene graphs with transcriptomics data across diverse cell types and perturbations.

DeepVirus identifies and classifies viral dark matter in metagenomic data using a hierarchical transformer framework, revealing novel viral lineages and expanding known RNA virus diversity.

The paper introduces {micro}Protein, a framework combining deep learning and reinforcement learning to predict and discover high-gain protein mutants using single mutation data.

This study investigates how aromatic amino acids in the orthosteric region of CB1 regulate conformational transitions using molecular dynamics simulations, revealing their roles in stabilizing active states and modulating receptor dynamics.

This paper introduces TamGen, a chemical language model for target-aware molecule generation, and shows it produces high-quality compounds with inhibitory activity against Tuberculosis ClpP protease.

This paper develops a pre-trained antibody language model to improve CDR sequence and structure generation, outperforming previous methods in antigen-specific antibody design.

The paper models the structure of Omicron's RBD and human ACE2, showing mutations enhance binding strength through hydrogen bonding and increased buried surface area.

The paper analyzes ACE2 sequences from 16 mammals and predicts ACE2-RBD complex structures to compare binding strengths across species.

The paper presents a deep learning algorithm using an auto-encoder to reconstruct 3D protein models from SAXS data, demonstrating accuracy without prior size information.