Liang Hong

VenusGenome improves microbial genome representation learning by using a genome-aware masked autoencoder, achieving high performance in phenotype prediction and gene-level tasks.

MAXWELL calibrates protein language models to predict mutation-induced stability changes, outperforming existing methods and successfully designing stabilizing mutations in DhaA dehalogenase.

This study investigates the structure-activity relationship of mRNA-LNPs, finding that a more ordered internal structure correlates with higher in vitro activity.

This paper designs an engineered PET hydrolase, VenusPETase, using deep learning to enhance PET depolymerization efficiency and thermostability, achieving high TPA yields and complete degradation of commercial PET.

This paper introduces VenusRXN, a deep learning framework for reaction-conditioned enzyme discovery, achieving high accuracy in identifying enzymes for novel reactions.

This paper introduces OpenPheno, a multimodal model that predicts bioactivity using phenotypic profiles and natural language descriptions, achieving strong zero-shot performance on unseen assays.

The paper introduces SaprotHub, a platform that simplifies training and sharing of protein language models, enabling non-experts to build and collaborate on customized models.

This study identifies and quantifies drug-free lipid nanoparticles in mRNA-LNPs using SANS and nano flow cytometry, leading to more accurate structural models.

This paper uses a protein language model and structural representation tree to discover improved PET hydrolases, identifying KbPETase with higher thermal stability and degradation activity.

ProDMM is a multimodal sequence modeling framework that captures interdependencies between protein motifs, coding DNA, and regulatory elements to predict phenotypes and generate optimized regulatory sequences.

GRATCR generates epitope-specific TCR sequences with fewer training data than state-of-the-art models, showing higher specificity and biological functionality.

The paper introduces Venus-MAXWELL, a framework for predicting protein mutant stability using sequence-to-landscape learning, achieving higher accuracy and faster inference than existing methods.

RiboDiffusion generates RNA sequences from 3D structures using a diffusion model, outperforming baselines in sequence recovery and showing consistent performance across RNA types and lengths.

ProSST integrates protein sequences and structures into a Transformer-based model, achieving state-of-the-art performance in mutation effect prediction and downstream tasks.

A BERT-style model pretrained on 23.7 million unlabelled ncRNA sequences gives embeddings that improve secondary structure, contact, distance-map and some function predictions over task-specific and MSA-based baselines.

fastMSA accelerates multiple sequence alignment by using a dual encoder based on a protein language model, achieving a 34-fold speed-up while maintaining homolog recall and compatibility with downstream tasks like AlphaFold.