Lei Wei
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Lei Wei has a total of 200 co-authors. Here are the top co-authors by number of shared publications:
- Xuegong Zhang
MOE Key Laboratory of Bioinformatics and Bioinformatics Division of BNRIST, Department of Automation, Tsinghua University , Beijing 100084
35 shared publications - Xiaowo Wang
Tsinghua University
18 shared publications Haiyang Bian
MOE Key Lab of Bioinformatics and Bioinformatics Division of BNRIST, Department of Automation, Tsinghua University, Beijing 100084
11 shared publications- Minsheng Hao
MOE Key Laboratory of Bioinformatics and Bioinformatics Division of BNRIST, Department of Automation, Tsinghua University, Beijing, China
11 shared publications Yixin Chen
MOE Key Laboratory of Bioinformatics and Bioinformatics Division of BNRIST, Department of Automation, Tsinghua University, Beijing, China
11 shared publicationsSijie Chen
MOE Key Laboratory of Bioinformatics and Bioinformatics Division, BNRIST, Department of Automation, Tsinghua University, Beijing, China
9 shared publications
Latest Publications
- Open Access
- 2 October 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (10)
DNA-BFN: Bayesian flow networks for generative modeling of regulatory DNA sequences
Motivation: Designing regulatory DNA sequences is challenging due to their discrete nucleotide alphabet and complex higher-order dependencies underlying regulatory function. Many existing generative models, particularly diffusion-based approaches, learn by injecting stochastic noise into sequences and training denoising processes, which can distort discrete sequence structure and weaken biological constraints. Results: Here, we introduce DNA Bayesian Flow Networks (DNA-BFN), a generative framework that models regulatory DNA sequences through iterative Bayesian updates of per-position nucleotide distributions. By directly modeling per-position categorical beliefs, DNA-BFN realizes a diffusion-like generative process without explicitly perturbing discrete nucleotide sequences, enabling fully differentiable training while preserving discrete sequence structure. Across eight regulatory sequence datasets spanning diverse species, scales, and sequence lengths, DNA-BFN consistently outperforms state-of-the-art baselines, faithfully preserves sequence feature statistics and regulatory patterns, enables few-shot sequence design, and achieves faster sampling efficiency. These results show that framing sequence generation as progressive uncertainty reduction in probability space provides an effective and generalizable approach for nucleotide sequence design. Availability and implementation: The software and its source code are publicly available at https://github.com/WangLabTHU/DNA-BFN. The dataset, source code and model weights are available in the Zenodo repository: https://doi.org/10.5281/zenodo.20177972.
- 2 October 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (10)
Open AccessDNA-BFN: Bayesian flow networks for generative modeling of regulatory DNA sequences
Motivation: Designing regulatory DNA sequences is challenging due to their discrete nucleotide alphabet and complex higher-order dependencies underlying regulatory function. Many existing generative models, particularly diffusion-based approaches, learn by injecting stochastic noise into sequences and training denoising processes, which can distort discrete sequence structure and weaken biological constraints. Results: Here, we introduce DNA Bayesian Flow Networks (DNA-BFN), a generative framework that models regulatory DNA sequences through iterative Bayesian updates of per-position nucleotide distributions. By directly modeling per-position categorical beliefs, DNA-BFN realizes a diffusion-like generative process without explicitly perturbing discrete nucleotide sequences, enabling fully differentiable training while preserving discrete sequence structure. Across eight regulatory sequence datasets spanning diverse species, scales, and sequence lengths, DNA-BFN consistently outperforms state-of-the-art baselines, faithfully preserves sequence feature statistics and regulatory patterns, enables few-shot sequence design, and achieves faster sampling efficiency. These results show that framing sequence generation as progressive uncertainty reduction in probability space provides an effective and generalizable approach for nucleotide sequence design. Availability and implementation: The software and its source code are publicly available at https://github.com/WangLabTHU/DNA-BFN. The dataset, source code and model weights are available in the Zenodo repository: https://doi.org/10.5281/zenodo.20177972.
- Open Access
- 15 September 2026
- Published by Springer Nature in Cell Discovery
- Vol. 12 (1) , 66
The accessory protein ORF3a hijacks the CLCC1 chloride channel to disrupt ER homeostasis in betacoronavirus pathogenesis
Beta-coronavirus infection disrupts endoplasmic reticulum (ER) homeostasis; however, the mechanisms by which viral proteins manipulate ER-resident factors remain unclear. Here, we report that the SARS-CoV-2 accessory protein ORF3a binds to the ER chloride channel CLCC1, thereby impairing ER ion homeostasis, activating the unfolded protein response, and inducing endomembrane remodeling. Using newly developed ratiometric reporters, we showed that this interaction exacerbated the basal levels of ER-phagy and nucleophagy. Importantly, CLCC1 counteracted ORF3a by sequestering it within the ER, attenuating its toxicity and suppressing viral replication, which established CLCC1 as a host restriction factor. We further elucidated a spatiotemporal mechanism: during early infection, ORF3a accumulates in the ER, co-assembles with CLCC1 into puncta, and disrupts ER homeostasis; at the later stage, ORF3a overload enables its escape from CLCC1-mediated ER retention and subsequent translocation to lysosomes, facilitating viral egress. This ER-centric function is conserved across diverse beta-coronaviruses, including SARS-CoV-1 and bat- or pangolin-derived strains, revealing a unified pathogenic strategy. Furthermore, in mouse brains, the ORF3a-CLCC1 axis recapitulates key neuropathological features, including ER stress, autophagic dysfunction, neuronal death, and neuroinflammation, providing a mechanistic link to COVID-19-associated neurological symptoms. Our findings identify CLCC1 as an essential host defense protein and establish a conceptual framework to guide future research on antiviral responses at the organelle level and related disease mechanisms.
- 15 September 2026
- Published by Springer Nature in Cell Discovery
- Vol. 12 (1) , 66
Open AccessThe accessory protein ORF3a hijacks the CLCC1 chloride channel to disrupt ER homeostasis in betacoronavirus pathogenesis
Beta-coronavirus infection disrupts endoplasmic reticulum (ER) homeostasis; however, the mechanisms by which viral proteins manipulate ER-resident factors remain unclear. Here, we report that the SARS-CoV-2 accessory protein ORF3a binds to the ER chloride channel CLCC1, thereby impairing ER ion homeostasis, activating the unfolded protein response, and inducing endomembrane remodeling. Using newly developed ratiometric reporters, we showed that this interaction exacerbated the basal levels of ER-phagy and nucleophagy. Importantly, CLCC1 counteracted ORF3a by sequestering it within the ER, attenuating its toxicity and suppressing viral replication, which established CLCC1 as a host restriction factor. We further elucidated a spatiotemporal mechanism: during early infection, ORF3a accumulates in the ER, co-assembles with CLCC1 into puncta, and disrupts ER homeostasis; at the later stage, ORF3a overload enables its escape from CLCC1-mediated ER retention and subsequent translocation to lysosomes, facilitating viral egress. This ER-centric function is conserved across diverse beta-coronaviruses, including SARS-CoV-1 and bat- or pangolin-derived strains, revealing a unified pathogenic strategy. Furthermore, in mouse brains, the ORF3a-CLCC1 axis recapitulates key neuropathological features, including ER stress, autophagic dysfunction, neuronal death, and neuroinflammation, providing a mechanistic link to COVID-19-associated neurological symptoms. Our findings identify CLCC1 as an essential host defense protein and establish a conceptual framework to guide future research on antiviral responses at the organelle level and related disease mechanisms.
- Open Access
- 2 July 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (7)
Cross-dataset annotation harmonization for cell-type hierarchy construction
Motivation: Single-cell transcriptomic datasets annotate cell types with diverse schemes and varying resolution. This poses challenges in building unified hierarchical cell-type structures and hinders integration of large-scale datasets. To address this, several computational methods have been developed to harmonize cell type annotations across datasets and build data-driven hierarchies of cell types. Results: Here, we benchmarked three state-of-the-art methods: scHPL, treeArches, and CellHint. We evaluated these methods across five simulated scenarios and five real-world scenarios across cell types and organs. To assess harmonization results, we designed three metrics, Annotation Harmonization F1-score (AH-F1), Tree Edit Distance Similarity and Parent–Children Branches Similarity, comparing the constructed cell-type hierarchies and the knowledge-based ones. Based on the benchmarking results, we found that methods performed well in simulated scenarios but still have room for improvement in complex real-world data. Thus, we developed OTHarmonizer, a tool based on partial optimal transport (OT) for cell-type harmonization and hierarchy construction. OTHarmonizer excels in accurately capturing equivalent and hierarchical relationships between cell types, offering a more effective approach for the cell-type hierarchy construction across datasets. Availability and implementation: The simulated and real-world datasets in the benchmark are available on https://figshare.com/articles/dataset/OTHarmonizer/28243205. The source codes for the benchmark and OTHarmonizer are available online on GitHub at https://github.com/Duck-Boss/OTHarmonizer.
- 2 July 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (7)
Open AccessCross-dataset annotation harmonization for cell-type hierarchy construction
Motivation: Single-cell transcriptomic datasets annotate cell types with diverse schemes and varying resolution. This poses challenges in building unified hierarchical cell-type structures and hinders integration of large-scale datasets. To address this, several computational methods have been developed to harmonize cell type annotations across datasets and build data-driven hierarchies of cell types. Results: Here, we benchmarked three state-of-the-art methods: scHPL, treeArches, and CellHint. We evaluated these methods across five simulated scenarios and five real-world scenarios across cell types and organs. To assess harmonization results, we designed three metrics, Annotation Harmonization F1-score (AH-F1), Tree Edit Distance Similarity and Parent–Children Branches Similarity, comparing the constructed cell-type hierarchies and the knowledge-based ones. Based on the benchmarking results, we found that methods performed well in simulated scenarios but still have room for improvement in complex real-world data. Thus, we developed OTHarmonizer, a tool based on partial optimal transport (OT) for cell-type harmonization and hierarchy construction. OTHarmonizer excels in accurately capturing equivalent and hierarchical relationships between cell types, offering a more effective approach for the cell-type hierarchy construction across datasets. Availability and implementation: The simulated and real-world datasets in the benchmark are available on https://figshare.com/articles/dataset/OTHarmonizer/28243205. The source codes for the benchmark and OTHarmonizer are available online on GitHub at https://github.com/Duck-Boss/OTHarmonizer.
- Open Access
- 1 July 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (Supplement)
Deciphering key factors of active learning performance in biomolecular design
Motivation: Employing machine learning (ML) to efficiently design biomolecules has become an emerging trend in genetic engineering. Active learning (AL) algorithms, as scalable approaches for ML-guided discovery, can automatically identify promising samples for function (i.e. fitness) optimization, and have therefore attracted growing interest across scientific domains. However, applying AL in genetic engineering presents several challenges. The regulatory patterns between sequence and fitness are highly complex, noisy, and sparse, making the existing evaluation of AL algorithm efficiency unreliable. Therefore, a comprehensive benchmark and thorough investigation into the key determinants of AL performance are urgently required to resolve these challenges. Results: We created a benchmark across multiple large-scale libraries of proteins and DNA regulatory sequences, evaluating uncertainty quantification (UQ) algorithms on metrics including calibration and accuracy, demonstrating the robustness and generality of ensemble-based algorithms. Moreover, we systematically assessed the efficiency of existing sampling strategies for fitness optimization. Our results show that no single sampling strategy is universally optimal across datasets, although greedy iterative strategies perform well in many practical scenarios. Finally, we evaluated the factors influencing optimization efficiency, and found that optimization efficiency is mainly determined by the choice of initial settings, distribution sparsity, and sequence similarity in high-fitness regions, rather than by the specific AL algorithm. Based on this, we proposed two quantifiable metrics to interpret the strategy performance and provide a practical reference for strategy selection. These findings offer valuable insights for the implementation of AL pipelines in biomolecular sequence design scenarios. Availability and implementation: The source code and supporting datasets used in this work are openly available on GitHub at https://github.com/WangLabTHU/biomolecule-al-decipher and have been archived on Zenodo at https://doi.org/10.5281/zenodo.19661002.
- 1 July 2026
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 42 (Supplement)
Open AccessDeciphering key factors of active learning performance in biomolecular design
Motivation: Employing machine learning (ML) to efficiently design biomolecules has become an emerging trend in genetic engineering. Active learning (AL) algorithms, as scalable approaches for ML-guided discovery, can automatically identify promising samples for function (i.e. fitness) optimization, and have therefore attracted growing interest across scientific domains. However, applying AL in genetic engineering presents several challenges. The regulatory patterns between sequence and fitness are highly complex, noisy, and sparse, making the existing evaluation of AL algorithm efficiency unreliable. Therefore, a comprehensive benchmark and thorough investigation into the key determinants of AL performance are urgently required to resolve these challenges. Results: We created a benchmark across multiple large-scale libraries of proteins and DNA regulatory sequences, evaluating uncertainty quantification (UQ) algorithms on metrics including calibration and accuracy, demonstrating the robustness and generality of ensemble-based algorithms. Moreover, we systematically assessed the efficiency of existing sampling strategies for fitness optimization. Our results show that no single sampling strategy is universally optimal across datasets, although greedy iterative strategies perform well in many practical scenarios. Finally, we evaluated the factors influencing optimization efficiency, and found that optimization efficiency is mainly determined by the choice of initial settings, distribution sparsity, and sequence similarity in high-fitness regions, rather than by the specific AL algorithm. Based on this, we proposed two quantifiable metrics to interpret the strategy performance and provide a practical reference for strategy selection. These findings offer valuable insights for the implementation of AL pipelines in biomolecular sequence design scenarios. Availability and implementation: The source code and supporting datasets used in this work are openly available on GitHub at https://github.com/WangLabTHU/biomolecule-al-decipher and have been archived on Zenodo at https://doi.org/10.5281/zenodo.19661002.
- Open Access
- 14 April 2026
- Published by Elsevier in Cell Genomics
- Vol. 6 (7) , 101214
Cross-task interpretability through unified modeling reveals a universal shortcut bias in neoantigen prediction
- 14 April 2026
- Published by Elsevier in Cell Genomics
- Vol. 6 (7) , 101214
Open AccessCross-task interpretability through unified modeling reveals a universal shortcut bias in neoantigen prediction
- Open Access
Programming Biomolecular Interactions with All-Atom Generative Model
Biomolecular interactions lie at the core of cellular life, spanning diverse molecular modalities from small molecules to nucleic acids and proteins. Nevertheless, design strategies remain separated despite shared physicochemical principles of molecular recognition. Here we present AnewOmni, a unified generative framework trained on more than 5 million biomolecular complexes, that enables transferable molecular design across molecular scales by assembling chemically meaningful building blocks at atomic resolution. We further introduce programmable graph prompts to support user-defined chemical, topological, and geometric steering during generation, exploring hybrid and unconventional chemistries beyond canonical structures. We demonstrate that transferable learning of interaction patterns and physical constraints across molecular modalities is possible, via an atom-to-block latent space capturing both atomic details and structural priors. The framework successfully designed small molecules, peptides, and nanobodies targeting the challenging KRAS G12D switch II pocket, as well as orthosteric peptides and allosteric small-molecule inhibitors for PCSK9 in the absence of known binding site, achieving 23%-75% success with only low-throughput validation, bypassing modality-specific high-throughput screening. AnewOmni is the first to succeed in functional molecular design across all scales, from small chemical entities to large biologics, and represents a stepstone towards general molecular reasoning engines, advocating a generative foundation model for biomolecular interactions to enter regimes where data and human intuition remain limited.
Open AccessProgramming Biomolecular Interactions with All-Atom Generative Model
Biomolecular interactions lie at the core of cellular life, spanning diverse molecular modalities from small molecules to nucleic acids and proteins. Nevertheless, design strategies remain separated despite shared physicochemical principles of molecular recognition. Here we present AnewOmni, a unified generative framework trained on more than 5 million biomolecular complexes, that enables transferable molecular design across molecular scales by assembling chemically meaningful building blocks at atomic resolution. We further introduce programmable graph prompts to support user-defined chemical, topological, and geometric steering during generation, exploring hybrid and unconventional chemistries beyond canonical structures. We demonstrate that transferable learning of interaction patterns and physical constraints across molecular modalities is possible, via an atom-to-block latent space capturing both atomic details and structural priors. The framework successfully designed small molecules, peptides, and nanobodies targeting the challenging KRAS G12D switch II pocket, as well as orthosteric peptides and allosteric small-molecule inhibitors for PCSK9 in the absence of known binding site, achieving 23%-75% success with only low-throughput validation, bypassing modality-specific high-throughput screening. AnewOmni is the first to succeed in functional molecular design across all scales, from small chemical entities to large biologics, and represents a stepstone towards general molecular reasoning engines, advocating a generative foundation model for biomolecular interactions to enter regimes where data and human intuition remain limited.
- Open Access
Autoregressive forecasting of future single-cell state transitions
Existing methods for dynamic analysis of static single-cell RNA-sequencing data can reconstruct temporal structures covered by observed cells, but cannot forecast unobserved future state transitions. We propose a temporal generative AI model, CellTempo, to forecast future cellular dynamics by representing cells as learned semantic codes and training an autoregressive generation decoder to predict ordered code sequences. It can forecast long-range cell-state transition trajectories and landscapes from snapshot data. To train the model, we constructed a comprehensive single-cell trajectory dataset scBaseTraj by integrating RNA velocity, pseudotime, and inferred transition probabilities to compose multi-step cellular sequences. Experiments on multiple real datasets showed that CellTempo can forecast cell state evolutions from individual cells, and reconstruct nuanced cell-state potential landscapes and their varied progressions after genetic or chemical perturbations, all with high fidelity to biological truth. This work opens a route for forecasting unseen future dynamics of cell state transitions from static observations.
Open AccessAutoregressive forecasting of future single-cell state transitions
Existing methods for dynamic analysis of static single-cell RNA-sequencing data can reconstruct temporal structures covered by observed cells, but cannot forecast unobserved future state transitions. We propose a temporal generative AI model, CellTempo, to forecast future cellular dynamics by representing cells as learned semantic codes and training an autoregressive generation decoder to predict ordered code sequences. It can forecast long-range cell-state transition trajectories and landscapes from snapshot data. To train the model, we constructed a comprehensive single-cell trajectory dataset scBaseTraj by integrating RNA velocity, pseudotime, and inferred transition probabilities to compose multi-step cellular sequences. Experiments on multiple real datasets showed that CellTempo can forecast cell state evolutions from individual cells, and reconstruct nuanced cell-state potential landscapes and their varied progressions after genetic or chemical perturbations, all with high fidelity to biological truth. This work opens a route for forecasting unseen future dynamics of cell state transitions from static observations.
- Open Access
- 9 February 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
A generic reference defined by consensus peaks for single-cell ATAC-seq data analysis
The rapid advancement of transposase-accessible chromatin using sequencing (ATAC-seq) technology, particularly with the emergence of single-cell ATAC-seq (scATAC-seq), accelerates the studies of gene regulation. However, the absence of a generic feature reference for ATAC-seq data limits single-cell analyses and hinders the development of comprehensive cell atlases. To address this, we construct a generic chromatin accessibility reference by aggregating peaks from 624 high-quality bulk ATAC-seq datasets, defining about 1.4 million consensus peaks (cPeaks). Leveraging a deep neural network model, we expand cPeaks to include previously unobserved regions, enhancing their coverage across diverse tissues and cell types. cPeaks exhibit consistent shapes across tissue types, sequencing technologies, and peak-calling methods, indicating that they represent inherent genomic features. Compared to existing feature-defining methods and references, cPeaks show superior performance in scATAC-seq analyses, improving cell annotation and rare cell type identification. Additionally, cPeaks provide insights into chromatin dynamics during cellular differentiation and tumor progression. cPeaks can serve as a robust reference for chromatin accessibility studies to promote cross-dataset consistency and accelerate biological discoveries.
- 9 February 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessA generic reference defined by consensus peaks for single-cell ATAC-seq data analysis
The rapid advancement of transposase-accessible chromatin using sequencing (ATAC-seq) technology, particularly with the emergence of single-cell ATAC-seq (scATAC-seq), accelerates the studies of gene regulation. However, the absence of a generic feature reference for ATAC-seq data limits single-cell analyses and hinders the development of comprehensive cell atlases. To address this, we construct a generic chromatin accessibility reference by aggregating peaks from 624 high-quality bulk ATAC-seq datasets, defining about 1.4 million consensus peaks (cPeaks). Leveraging a deep neural network model, we expand cPeaks to include previously unobserved regions, enhancing their coverage across diverse tissues and cell types. cPeaks exhibit consistent shapes across tissue types, sequencing technologies, and peak-calling methods, indicating that they represent inherent genomic features. Compared to existing feature-defining methods and references, cPeaks show superior performance in scATAC-seq analyses, improving cell annotation and rare cell type identification. Additionally, cPeaks provide insights into chromatin dynamics during cellular differentiation and tumor progression. cPeaks can serve as a robust reference for chromatin accessibility studies to promote cross-dataset consistency and accelerate biological discoveries.
- Open Access
How different AI models understand cells differently
AI single-cell foundation models (scFMs) are believed to be able to learn essential relations in cell transcriptomics with the attention modules in Transformer, but there is no method to reveal what they actually learned. We observed that different models may grasp different aspects of relations. To unravel the mystery, we propose scGeneLens, a framework for dissecting how scFMs perceive cells. We employed a sparse block attention to replace the original attention mechanism to concentrate attentions into a few dominant gene–gene relations, used attention propagation to trace how the relations propagate across Transformer layers, and used integrated gradients to disentangle the relative contributions of gene identity and expression in cell representations. We applied it to scFoundation and scGPT and show that they exhibit pluralistic perceptions of cells: scFoundation emphasizes relations among cell-type marker genes, resulting in stronger cell-type separability, whereas scGPT focuses more on genes involved in shared cellular pathways and core biological activities, leading to representations that generalize across conditions. The framework provides a unified lens for probing what scFMs learn about cells and offers actionable insights for the design of future cellular foundation models. Our code can be seen in https://anonymous.4open.science/r/scGeneLens-B771/ .
Open AccessHow different AI models understand cells differently
AI single-cell foundation models (scFMs) are believed to be able to learn essential relations in cell transcriptomics with the attention modules in Transformer, but there is no method to reveal what they actually learned. We observed that different models may grasp different aspects of relations. To unravel the mystery, we propose scGeneLens, a framework for dissecting how scFMs perceive cells. We employed a sparse block attention to replace the original attention mechanism to concentrate attentions into a few dominant gene–gene relations, used attention propagation to trace how the relations propagate across Transformer layers, and used integrated gradients to disentangle the relative contributions of gene identity and expression in cell representations. We applied it to scFoundation and scGPT and show that they exhibit pluralistic perceptions of cells: scFoundation emphasizes relations among cell-type marker genes, resulting in stronger cell-type separability, whereas scGPT focuses more on genes involved in shared cellular pathways and core biological activities, leading to representations that generalize across conditions. The framework provides a unified lens for probing what scFMs learn about cells and offers actionable insights for the design of future cellular foundation models. Our code can be seen in https://anonymous.4open.science/r/scGeneLens-B771/ .