Lu Zhang

PhageFlare predicts functions for phage proteins with no structural matches, outperforming baselines and annotating many previously unresolved proteins.

U-Chrom provides a unified framework for analyzing single-cell and spatial 3D genome data, showing consistent cell-type differences across assays.

The paper introduces DeepMM, a deep learning method to correct chimeric contigs in MAGs, improving gene annotation accuracy and reducing reference-discordant genes and virulence-factor calls.

MetAgeFormer is a transformer-based model pre-trained on metabolomic data to estimate biological age and identify metabolic subtypes associated with disease risk.

SILO improves detection of disease susceptible variants in short IBD segments, outperforming existing models in simulated and real data.

stDyer-image improves clustering of spatial transcriptomics and proteomics data by integrating morphological images directly into the clustering process.

This study evaluates long single-end barcoded reads for structural variant detection, showing improved accuracy with longer read lengths compared to short-read and pangenome-based methods.

FEAST generates synthetic spatial transcriptomics data with tunable spatial and transcriptional variation for benchmarking and 3D reconstruction.

This paper identifies microbial transcriptional reprogramming in tumors that generates tumor-specific epitopes, enhancing immunotherapy safety by enabling targeted T cell responses.

TRAFICA improves TF-DNA binding affinity prediction by integrating open chromatin features and in vitro binding profiles, outperforming existing tools in both in vitro and in vivo settings.

HarmoDecon mitigates multi-scale biases in spatial cell-type deconvolution by using a semi-supervised deep learning model, improving accuracy in spatial domain clustering and marker gene correlations.

The paper introduces SpaMV, a method for integrating spatial multi-omics data that preserves modality-specific information and improves clustering and dimensionality reduction.

This paper introduces KEScape, a deep learning model that predicts viral immune escape and evolution by integrating structural information and prioritizing impactful mutations, outperforming existing models.

MimicNeoAI identifies microbial epitopes and their tumor neoepitope mimics from sequencing data, revealing highly immunogenic microbial epitopes specific to tumors.

RAPID is an automated tool using large language models to generate medical reporting checklists, achieving high accuracy on CONSORT and CONSORT-AI datasets.

CompleteBin improves short contig binning in complex microbial communities, recovering more complete genomes and novel genes, with applications in environmental metagenomics.

This study compares gut microbiome differences between Han Chinese and non-Chinese populations, identifying population-specific genome clusters and SNPs linked to functional traits.

The paper introduces a bacteria-centric AMP predictor that identifies more bacterial antimicrobial peptides from real genomes than fake ones, outperforming existing predictors.

stDyer improves spatial domain clustering in SRT data by combining GMVAE and GATs with dynamic graphs and mini-batch sampling, outperforming existing methods.

DeepMM identifies and corrects metagenomic misassemblies using deep learning, improving downstream binning and increasing near-complete MAGs in a real dataset.

Pangaea improves metagenome assembly using short-reads with barcodes, achieving better contig continuity and more complete genomes than existing assemblers.

MaskGraphene integrates multi-slice, multi-condition spatial transcriptomics data using a graph neural network to generate interpretable joint embeddings with preserved geometric information, improving downstream analyses.

The paper assesses how genome topology and viral selection affect biological interpretation in the human gut virome, highlighting the need for standardized reporting of circular-to-linear genome ratios and dereplication thresholds.

LRTK is a platform-agnostic toolkit for linked-read analysis of human genomes and metagenomes, offering simulation, alignment, variant calling, and taxonomic classification.

Bio-Mol improves small molecule property prediction by integrating missing bioactivity profiles through pretraining and finetuning, achieving a 5.2% higher AUROC than previous models.

Deepurify uses a multi-modal deep language model with contrastive learning to improve MAG decontamination by leveraging genomic sequences and taxonomic lineages.

IGSimpute improves scRNA-seq data imputation with high accuracy and interpretability, outperforming existing methods on multiple tissues.

This study uses an explainable AI model to classify chronic inflammatory responses as prognostic markers for glioma using TCGA data, identifying five inflammatory profiles linked to survival outcomes.

This paper benchmarks 15 de novo metagenome assembly tools on 32 datasets from various sequencing technologies, finding that long-read assemblers produce the highest contig continuity but miss some MAGs, while hybrid assemblers offer a balance of continuity and MAG completeness.

This paper introduces PRSIMD, a machine learning model that integrates genetic and non-genetic factors to improve disease risk prediction, showing better performance than existing PRS methods.