Zhen Li

Single-cell proteomics provides a direct functional readout of cellular states, but targeted assays measure limited and non-identical protein panels, fragmenting information across experiments. Here, we present ProtCell, a foundation model developed using approximately 380 million cells from 98 datasets for pretraining and evaluation. ProtCell combines protein-identity and continuous-abundance embeddings in a transformer encoder and learns cell- and protein-level representations through masked abundance modeling. ProtCell first connects fragmented profiles through transferable annotation and batch correction while preserving biological organization. It then extends limited profiles through abundance reconstruction and inference of an unmeasured cytotoxic protein program enriched in natural killer cells and memory CD8 ⁺ T cells. The learned protein dependencies further guide the selection of compact panels that retain information for profile reconstruction and biological classification. ProtCell thus provides a framework for connecting fragmented proteomic profiles, extending information beyond measured panels and guiding more informative measurements.

Spatially resolved sequencing technologies have revolutionized our understanding of biological regulatory processes within tissue microenvironments by simultaneously capturing the states of genomic regions, genes, and proteins alongside the spatial organization of cells. However, inherent heterogeneity across modalities and samples poses substantial challenges for the integrative analysis of spatial omics data, underscoring the urgent need for advanced computational methods. In this study, we propose PRESENT, a contrastive learning-based integrative framework for cross-modality representation of spatial multi-omics data. PRESENT employs omics-specific encoders consisting of graph attention networks and Bayesian neural networks coupled with distribution-aware decoders to model distinct modalities, and an inter-omics alignment module for multi-omics integration. By effectively incorporating spatial dependencies with multi-omics information across diverse species and technologies, PRESENT facilitates the accurate identification of spatial domains and the elucidation of underlying regulatory mechanisms. Furthermore, PRESENT can be extended to multi-sample integration via a two-stage training workflow, which incorporates inter-batch alignment loss, intra-batch preserving loss, batch-adversarial learning, and cyclic graph refinement strategies to eliminate batch effects while retaining biological signals. Extensive experiments on tissue samples across different anatomical regions and developmental stages demonstrate that PRESENT enables the characterization of hierarchical tissue structures from a spatiotemporal perspective.

The single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) has emerged as a core technology for dissecting cellular epigenomic heterogeneity and gene regulatory programs. With the emergence of atlas-level scATAC-seq datasets, cell type annotation increasingly faces challenges arising from unprecedented data scale and increased cell-type diversity, which together place stringent demands on model reliability and robustness. Here, we introduce HitAnno, a hierarchical language model capable of accurate and scalable cell type annotation in atlas-level scATAC-seq data. Leveraging selected cell-type-specific peaks to construct “cell sentences”, HitAnno employs a two-level attention mechanism that captures accessibility profiles hierarchically. Extensive evaluations show that HitAnno robustly annotates both major and rare cell types across multiple settings, including intra-dataset, cross-donor and inter-dataset annotation. The hierarchical attention mechanisms of the model reveal co-accessibility patterns among peaks and dependencies across higher-order peak sets, ensuring an interpretable annotation process. Training on a 31-cell-type human atlas, HitAnno can directly annotate new query datasets without retraining and is accessible through an online interface. Our model identifies heterogeneous subgroups within mixed labeled cells from unseen datasets, demonstrating its potential to assist researchers in refining existing cell atlases.

With the growing accumulation of scattered single-cell data and the rapid advancement of artificial intelligence (AI), there is a pressing need for a high-quality, well-organized, and AI-ready single-cell data resources to support large-scale model. Here, we present version 2.0 of human Ensemble Cell Atlas (hECA), a cell atlas incorporating both single-cell RNA sequencing (scRNA-seq) and single-cell ATAC sequencing (scATAC-seq) data. It expands the scRNA-seq data collection to 10,831,024 human cells with unified labels, and adds the new modality of scATAC-seq profiles with 1,450,511 cells. The data cover 42 human organs and tissues. To ensure cross-dataset consistency and quality, we standardized gene expression and chromatin accessibility matrices, harmonized cellular metadata, and manually re-annotated cell types based on the unified Hierarchical Annotation Framework (uHAF). The strength of the dataset has been shown in pre-training the large generative cellular AI model scMulan. hECA2.0 provides a well-structured and ready-to-use data resource, serving as a robust data foundation for AI-driven single-cell research.

Single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) profiles chromatin accessibility at cellular resolution, making it possible to reveal epigenomic landscapes that govern gene regulation in a variety of cells. Nevertheless, heterogeneous feature spaces and complex processing pipelines have impeded the construction of an ensemble resource capable of supporting diverse downstream analytical scenarios. To address this gap, we present Human-scATAC-Corpus (https://health.tsinghua.edu.cn/human-scatac-corpus/), a comprehensive database of human scATAC-seq comprising 5 407 621 cells from 35 datasets across 37 tissues or cell lines. To support complementary use cases, each dataset is distributed in three aligned formats: cell-by-candidate cis-regulatory element matrices for cross-dataset integration, raw fragment files for flexible processing, and cell-by-peak matrices for within-dataset analyses. This resource spans diverse biological contexts and includes rich metadata, enabling method benchmarking and development, as well as pretraining of foundation models. The website offers searchable browsing, detailed dataset pages, on-demand downloads, and tutorials. EpiAgent, a foundation model pretrained on Human-scATAC-Corpus, is further integrated to provide online analyses, including reference mapping, embedding extraction, and cell type annotation. Human-scATAC-Corpus establishes a unified and scalable substrate for single-cell epigenomics and is intended to accelerate discovery while standardizing evaluation across tasks.

Although single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) enables the exploration of the epigenomic landscape that governs transcription at the cellular level, the complicated characteristics of the sequencing data and the broad scope of downstream tasks mean that a sophisticated and versatile computational method is urgently needed. Here we introduce EpiAgent, a foundation model pretrained on our manually curated large-scale Human-scATAC-Corpus. EpiAgent encodes chromatin accessibility patterns of cells as concise ‘cell sentences’ and captures cellular heterogeneity behind regulatory networks via bidirectional attention. Comprehensive benchmarks show that EpiAgent excels in typical downstream tasks, including unsupervised feature extraction, supervised cell type annotation and data imputation. By incorporating external embeddings, EpiAgent enables effective cellular response prediction for both out-of-sample stimulated and unseen genetic perturbations, reference data integration and query data mapping. Through in silico knockout of cis-regulatory elements, EpiAgent demonstrates the potential to model cell state changes. EpiAgent is further extended to directly annotate cell types in a zero-shot manner.

Cis-regulatory elements (CREs), including enhancers, silencers, promoters and insulators, play pivotal roles in orchestrating gene regulatory mechanisms that drive complex biological traits. However, current approaches for CRE identification are predominantly sequence-based and typically focus on individual CRE types, limiting insights into their cell-type-specific functions and regulatory dynamics. Here, we present CREATE, a multimodal deep learning framework based on Vector Quantized Variational AutoEncoder, tailored for comprehensive CRE identification and characterization. CREATE integrates genomic sequences, chromatin accessibility, and chromatin interaction data to generate discrete CRE embeddings, enabling accurate multi-class classification and robust characterization of CREs. CREATE excels in identifying cell-type-specific CREs, and provides quantitative and interpretable insights into CRE-specific features, uncovering the underlying regulatory codes. By facilitating large-scale prediction of CREs in specific cell types, CREATE enhances the recognition of disease- or phenotype-associated biological variabilities of CREs, thus advancing our understanding of gene regulatory landscapes and their roles in health and disease.

Recent advances in spatial epigenomic techniques have given rise to spatial assay for transposase-accessible chromatin using sequencing (spATAC-seq) data, enabling the characterization of epigenomic heterogeneity and spatial information simultaneously. Integrative analysis of multiple spATAC-seq samples, for which no method has been developed, allows for effective identification and elimination of unwanted non-biological factors within the data, enabling comprehensive exploration of tissue structures and providing a holistic epigenomic landscape, thereby facilitating the discovery of biological implications and the study of regulatory processes. In this article, we present INSTINCT, a method for multi-sample INtegration of Spatial chromaTIN accessibility sequencing data via stochastiC domain Translation. INSTINCT can efficiently handle the high dimensionality of spATAC-seq data and eliminate the complex noise and batch effects of samples through a stochastic domain translation procedure. We demonstrate the superiority and robustness of INSTINCT in integrating spATAC-seq data across multiple simulated scenarios and real datasets. Additionally, we highlight the advantages of INSTINCT in spatial domain identification, visualization, spot-type annotation, and various downstream analyses, including motif enrichment analysis, expression enrichment analysis, and partitioned heritability analysis.

Large-scale foundation models have recently opened a new avenue to artificial general intelligence for life sciences, showing great promise in the analysis of single-cell transcriptomic data. Nevertheless, such challenges as the tremendous number of signaling regions, extreme data sparsity, and the nearly binary nature of single-cell epigenomic data have prevented the construction of a foundation model for epigenomics thus far, though it is evident that abundant epigenomic properties such as chromatin accessibility provide more decisive insights into cell states than transcriptomics, shaping the chromatin regulatory landscapes that control transcription in distinct cell types. Here, we introduce EpiAgent, the first foundation model for single-cell chromatin accessibility data, pretrained on a manually curated large-scale Human-scATAC-Corpus that is comprised of approximately 5 million cells and 35 billion tokens. EpiAgent encodes chromatin accessibility patterns of cells as concise “cell sentences,” and employs a bidirectional attention mechanism to capture cellular heterogeneity behind regulatory networks. With comprehensive benchmarks, we demonstrate that EpiAgent excels in typical downstream tasks, including unsupervised feature extraction, supervised cell type annotation, and data imputation. By incorporating external embeddings, EpiAgent facilitates the prediction of cellular responses to both out-of-sample stimulated and unseen genetic perturbations, as well as reference data integration and query data mapping. By simulating the knockout of key cis-regulatory elements, EpiAgent enables in-silico treatment for cancer analysis. We further extended zero-shot capabilities of EpiAgent, allowing direct cell type annotation on newly sequenced datasets without additional training.