Lei Wei

This paper introduces CellTempo, a framework that forecasts long-range cell state trajectories from static transcriptomic data, demonstrating its ability to predict lineage progression and distinguish perturbation responses from long-term fate changes.

This paper introduces AnewOmni, a generative model that designs biomolecules across scales by learning interaction patterns and physical constraints from 5 million complexes.

scDiffusion-X generates high-fidelity multi-omics data and enables modality translation with uncertainty quantification and gene regulatory network inference.

This paper introduces scGeneLens, a framework to analyze how different AI models perceive cells, revealing that scFoundation focuses on cell-type markers while scGPT emphasizes shared pathways.

X-Pert is a unified framework for predicting cellular responses to diverse perturbations, outperforming existing methods across genetic and chemical perturbations and enabling generalization across cell types.

This paper introduces CellTok, a method to convert single-cell transcriptomes into tokens for LLMs, enabling joint processing of cellular data and text.

GeST is a generative transformer model pretrained on spatial transcriptomics data to learn spatial context, showing improved generation and zero-shot spatial niche embedding extraction.

This study benchmarks 10 in silico gene perturbation methods across four scenarios, identifying strengths and limitations in predicting perturbation effects and zero-shot transfer.

This study introduces a mathematical model for Molecular Competition Networks and shows they outperform non-competitive networks in computational tasks.

The paper introduces UniCoord, a joint-VAE model that learns a universal coordinate system for single-cell transcriptomic data to capture heterogeneity and enable data generation and analysis.

This paper introduces a transformer-based framework for neoantigen identification, identifies a bias in immunogenicity prediction, and proposes a debiasing strategy to improve model fairness.

The paper introduces scMulan, a large AI model trained to generate cell transcriptomics data with specific attributes, showing it can generate pseudo data, predict missing attributes, and identify driving regulators.

This study benchmarks 12 multi-omics integration methods on scRNA-seq and scATAC-seq data across three tasks, identifying method-specific strengths and providing selection guidelines.

scDecouple separates true cellular response from gRNA infection bias in scCRISPR-seq data, improving gene perturbation analysis in heterogeneous samples.

scDirect identifies key transcription factors driving cell state transitions using single-cell multi-omics data, showing robust performance in benchmarking and real-world applications.

Deconer evaluates reference-based cell type deconvolution methods using simulated and real gene expression data, comparing 14 methods across multiple criteria and providing guidance for users.

This paper introduces HoloNet, a method to decode functional cell-cell communication events from spatial transcriptomic data by integrating ligand-receptor pairs, cell-type spatial distribution, and downstream gene expression.

The paper constructs a generic chromatin accessibility reference (cPeaks) from bulk ATAC-seq data to improve scATAC-seq analysis and cell type identification.

STEM maps single-cell and spatial transcriptomics data using transfer learning to create spatially aware embeddings and predict pseudo-spatial adjacencies.

The paper introduces hECA, a cell-centric human Ensemble Cell Atlas, which integrates scRNA-seq data from multiple studies to enable targeted data retrieval, multi-view presentations, and customizable reference creation for cell type annotations.

PccGEO designs functional genetic elements using a knowledge-constrained deep learning model, achieving high success rates in promoter design for E. coli.

ARIC estimates cell type proportions from bulk gene expression or DNA methylation data, showing accuracy in both data types.

This study quantifies how microRNAs modulate gene expression noise and shows how this affects synthetic gene circuits and cell type classification.

DISMIR detects cancer using cfDNA methylation and sequence data with high accuracy at low sequencing depth.

NeuronMotif decodes cis-regulatory motifs from DCNNs by analyzing neuron activations, revealing motif combinations and improving model performance.

This study uses a generative adversarial network to design new functional promoters in E. coli, achieving higher success rates and activity than natural promoters.

This study introduces a competition motif model to analyze molecular competition's role in gene regulatory networks, revealing its impact on dose-response curves, response speed, noise, and target correlations.

The paper shows miRNAs reduce gene expression noise in mouse cells using single-cell RNA-sequencing data analysis.