Yu Li

Trimole-Hybrid improves ADMET prediction by combining multiple molecular representations and task-specific models, achieving top performance on 22 benchmarks.

BioKinema generates physically accurate all-atom biomolecular trajectories at lower computational cost than traditional MD simulations, enabling dynamic analysis of protein and protein-ligand systems.

Navigo models mouse embryogenesis to predict cell fates and disease by learning developmental trajectories from scRNA-seq data.

The paper shows that biological meaning in protein embeddings depends on resolution, with different levels of organisation preserving distinct biological relationships.

GFETM improves scATAC-seq modeling by integrating genome foundation model embeddings, revealing cell-state specific TF activity and kidney diabetes epigenomic signatures.

This paper introduces MultiAMP, a framework for predicting and designing antimicrobial peptides by integrating sequence and structural information, achieving state-of-the-art performance and discovering novel peptides.

GiMat identifies gene-gene interactions across multiple dichotomous phenotypes using a statistical framework, finding four novel interaction pairs in type 2 diabetes and hypertension comorbidity.

LigUnity is a foundation model that jointly embeds ligands and protein pockets into a shared space to improve affinity prediction in virtual screening and hit-to-lead optimization.

The paper introduces MoRNiNG, a database of RNA modification sites associated with RNA secondary structure dynamics, and validates the impact of modifications on G-quadruplex formation.

This paper introduces MorphDiff, a transcriptome-guided diffusion model that predicts cell morphology changes under perturbations and improves MOA identification.

cfDecon improves cfDNA deconvolution accuracy and interpretability using deep learning, outperforming existing methods in simulation and clinical datasets.

VIRALpre improves virus identification by combining genomic foundation model embeddings with K-mer features, achieving higher accuracy and better performance on short contigs and cross-dataset tests.

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.

scBalance is a sparse neural network framework for cell type annotation in scRNA-seq data, showing strong performance on both common and rare cell types across large datasets.

AcrNET predicts anti-CRISPR proteins using a deep learning model, achieving significant performance improvements and class-level predictions.

TAPE deconvolves bulk RNA-seq data into cell-type fractions and gene expression using a deep autoencoder, showing better performance and robustness compared to existing methods.

CLEAR is a self-supervised contrastive learning framework for scRNA-seq that improves clustering, batch correction, and pseudo-time inference by reducing technical noise and heterogeneity.

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.

scMinerva is an unsupervised graph learning framework for single-cell multi-omics analysis that outperforms existing methods with few labels and provides biomarkers for COVID-19 cell differentiation.

This paper introduces RHSNet, a deep learning method that identifies and quantifies recombination hotspot determinants across species, sexes, and populations.

HiC-LDNet is a deep learning framework for chromatin loop detection in Hi-C data, achieving high accuracy in both bulk and sparse single-cell Hi-C maps.

conST learns embeddings from gene expression, spatial neighbors, and histology via three-level contrastive learning, reaching adjusted Rand index 0.65 on spatialLIBD slice 151673, with GNNExplainer-based explanations.

This study identifies conserved RNA-binding motifs in coronavirus nucleocapsid proteins that enable broad-spectrum antiviral drug development.

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.

This paper introduces a framework to align and integrate single-cell RNA-seq and ATAC-seq data, achieving higher accuracy than existing methods on both simulated and real data.

CLMB improves metagenomic binning by using deep contrastive learning to handle noise, outperforming existing methods and recovering more high-quality genomes.

This study introduces HMD-AMP, a deep learning framework for multi-label classification of antimicrobial peptides, improving performance on minor classes and handling multiple targets.

USPNet is a signal peptide predictor using a deep protein language model that outperforms existing methods, especially in metagenomic data without additional group information.

ReFeaFi predicts regulatory elements genome-wide using a two-model approach, achieving good performance across species and identifying disease-associated variants in human regulatory regions.

DeeReCT-APA quantitatively predicts alternative polyadenylation site usage across genes using a deep learning model, outperforming existing methods in multiple tasks.