The human Ensemble Cell Atlas (hECA) v1.0 assembled 1,093,299 labeled single-cell RNA-seq cells from 116 published datasets into one cell-centric repository and introduced 'in data' cell sorting, quantitative portraiture, and customizable annotation references. Its authors enumerated the unresolved issues standing between such draft ensemble atlases and a comprehensive, high-fidelity human cell atlas. Coverage scale-up and one new modality have since been partly discharged by the authors' own hECA v2.0 (10,831,024 scRNA-seq cells plus 1,450,511 scATAC-seq cells across 42 organs) and by community platforms hosting tens of millions of cells. The surviving open core is methodological: (i) a general-purpose normalization and batch-effect-correction scheme that yields adjusted, biologically interpretable expression values applicable simultaneously across studies and all organs, rather than latent-space embeddings per downstream task; (ii) correction of non-uniform sampling of cells and genes so that global quantitative portraits estimate biology rather than study design; and (iii) automated, ontology-consistent cross-study re-annotation replacing manually curated marker references. A 2026 review confirms that existing integration methods still struggle with over- or under-correction, and hECA v2.0 (2025) still uses only log-normalization and manual re-annotation. These methodological obstacles remain open.

Background

Single-cell RNA sequencing (scRNA-seq) measures gene expression in individual cells, and large projects such as the Human Cell Atlas (HCA) and the US Human BioMolecular Atlas Program (HuBMAP) aim to build molecular reference maps of all human cell types in health and disease. Most public collections are 'file-centric': each study's data are archived as separate files, indexed by metadata, and downloaded one dataset at a time. The human Ensemble Cell Atlas (hECA) instead proposed 'cell-centric' assembly: cells from many studies are merged into one orchestrated repository, a unified giant table (uGT), a NoSQL column-based store holding each cell's gene expression values together with descriptive attributes such as organ, donor, study, and cell type, with all cell types renamed onto a single tree of standardized names called the unified Hierarchical Annotation Framework (uHAF). On this substrate hECA v1.0 (1,093,299 labeled cells from 116 datasets, 38 organs) introduced three applications: 'in data' cell sorting, meaning retrieval of cells from the whole assembled atlas via logic expressions that combine metadata labels and gene-expression thresholds; quantitative portraiture, meaning multi-view quantitative summaries of organs, cell types, and genes; and customizable reference creation for automatic annotation (label transfer) of user-supplied query data.

Three notions from single-cell genomics are needed to state the obstruction. Batch effects are systematic technical differences between measurements from different laboratories, protocols, or runs that can swamp biological signal; most correction methods align cells from different batches in a latent representation space tailored to one downstream task such as clustering, rather than converting the measured expression values themselves into adjusted values that remain directly interpretable and comparable across studies. Dropout is the failure of scRNA-seq to detect transcripts that are present, producing false zeros. Non-uniform sampling bias arises because tissue dissociation and capture efficiencies differ by cell type and protocol, so the proportion of each cell type in assembled data reflects study design at least as much as true anatomy.

The hECA authors located the gap precisely. They applied only library-size normalization (dividing each cell's counts by its total, scaling, and log-transforming) precisely because, as their discussion states, low-level processing issues such as normalization and batch-effect correction are 'far from perfectly solved yet, either in our own work or in the community', and because available batch-correction methods were designed for specific downstream tasks, typically within one organ, and mostly align cells in latent space rather than producing adjusted expression values with preserved biological meaning. The community benchmark of Luecken et al. (2022), which evaluated 68 method and preprocessing combinations on 13 atlas-level integration tasks spanning more than 1.2 million cells, confirmed that the strong performers (scANVI, Scanorama, scVI, scGen) deliver latent embeddings and involve an explicit trade-off between batch removal and conservation of biological variation, not a general-purpose expression-scale correction. Meanwhile, every cell in hECA was re-annotated by comparing cluster marker genes against manually curated marker references mapped onto the uHAF, and labels were then mapped to the Cell Ontology, a community-controlled vocabulary of cell types, partly by exact string matching and partly by hand. The published discussion therefore lists the unresolved issues: coverage is a tiny fraction of all human cells; normalization, batch-effect correction, and non-uniform sampling remain unsolved and 'may poison any global analysis of atlas data'; current portraits 'can only reveal properties of the collected data but cannot be expected of full fidelity for the underlying biology'; future versions should add modalities such as scATAC-seq (single-cell assay for transposase-accessible chromatin, which measures chromatin accessibility) and spatial information; and the scalability of in-data cell sorting and the NoSQL backend at very large data sizes is untested. In addition, the published methods document that re-annotation relies on manually curated marker references and manual mapping to the Cell Ontology — a practice that persists in hECA v2.0.

Problem Statement

The problem is the methodology layer separating current cell-centric ensemble atlases from a comprehensive human cell atlas whose summaries are quantitatively faithful to the underlying biology rather than to the collected data. The common objective is that a user retrieving cells or portrait-level statistics from the assembled atlas obtains values that estimate biology, without per-task manual reprocessing. The linked open questions are: (1) General-purpose preprocessing — construct a normalization and batch-effect-correction scheme that operates simultaneously on data from many studies and all organs of the body and outputs adjusted gene-expression values (not only latent embeddings) that remain biologically interpretable and comparable across studies and can serve as a general reference for diverse downstream tasks, avoiding over-correction of study-specific variation that is the biology of interest in other studies. (2) Sampling-bias-robust portraiture — handle non-uniform sampling of cells across cell types and of expressed genes, together with dropout and technical variability, so that global quantitative portraits, such as cell-type compositions of organs and gene-expression distributions, estimate the underlying biology rather than reflecting properties of the collected data and the sampling design of the assembled studies. (3) Automated, ontology-consistent re-annotation — replace the manually curated marker references and manual name mapping with automated methods that transfer labels across studies and map heterogeneous, abbreviated, and synonym-laden cell-type names onto a unified hierarchy (uHAF / Cell Ontology) at whole-body atlas scale. The three questions are independent open problems: a contribution may address any one of them, and resolving one does not require resolving the others. A resolving contribution must make the target of the question it addresses precise — the quantity to be estimated, or the measurable criterion by which its method is to be judged against the underlying biology rather than against the collected data — and provide evidence that its method attains that target. No particular benchmark or ground-truth dataset is presupposed; where the natural target is not directly observable, as for the portraiture question, the contribution must state and defend the operationalization it uses. The source's remaining roadmap items — assembling far more data across organs, developmental stages, and disease conditions; integrating further modalities (scATAC-seq has since been added; proteomics and spatial coordinates have not); and benchmarking in-data cell sorting and the NoSQL backend at the tens-of-millions-of-cells scale — are data-engineering milestones toward the same objective rather than bounded scientific questions; hECA v2.0 (about 12.3 million cells) and community platforms hosting tens of millions of cells have absorbed part of them; they fall outside the scope of the open problem.

References

  1. Chen S, Luo Y, Gao H, Li F, Li J, Chen Y, You R, Hao M, Bian H, Xi X, Li W, Li W, Ye M, Meng Q, Zou Z, Li C, Li H, Zhang Y, Cui Y, Wei L, Chen F, Wang X, Lv H, Hua K, Jiang R, Zhang X (2021). A human ensemble cell atlas (hECA) enables in data cell sorting (v1 title; later posting titled hECA: the cell-centric assembly of a cell atlas). bioRxiv. doi:10.1101/2021.07.21.453289. https://doi.org/10.1101/2021.07.21.453289
  2. Chen S, Luo Y, Gao H, Li F, Chen Y, Li J, You R, Hao M, Bian H, Xi X, Li W, Li W, Ye M, Meng Q, Zou Z, Li C, Li H, Zhang Y, Cui Y, Wei L, Chen F, Wang X, Lv H, Hua K, Jiang R, Zhang X (2022). hECA: The cell-centric assembly of a cell atlas. iScience 25:104318. doi:10.1016/j.isci.2022.104318. https://doi.org/10.1016/j.isci.2022.104318
  3. Xi X, Chen Y, Wu X, Hao M, Li J, Bian H, Meng Q, Li F, Li C, Xiao C, Dong X, You R, Xiong Y, Yang P, Gao Z, Cui X, Pan Y, Li Z, Li W, Li Z, Chen X, Cui Y, Lv H, Jiang R, Wei L, Zhang X (2025). hECA v2.0: an AI-ready ensemble cell atlas of single-cell RNA and ATAC sequencing data. Scientific Data 13:110. doi:10.1038/s41597-025-06426-2. https://doi.org/10.1038/s41597-025-06426-2
  4. Chen S, Luo Y, Gao H, Li F, Li J, Chen Y, You R, Lv H, Hua K, Jiang R, Zhang X (2022). Toward a unified information framework for cell atlas assembly. National Science Review 9:nwab179. doi:10.1093/nsr/nwab179. https://doi.org/10.1093/nsr/nwab179
  5. Chen Y, Hao M, Gao H, Li J, Chen S, Li F, Wei L, Zhang X (2022). Protocol for profiling cell-centric assembled single-cell human transcriptome data in hECA. STAR Protocols 3:101589. doi:10.1016/j.xpro.2022.101589. https://doi.org/10.1016/j.xpro.2022.101589
  6. Luecken MD, Büttner M, Chaichoompu K, Danese A, Interlandi M, Mueller MF, Strobl DC, Zappia L, Dugas M, Colomé-Tatché M, Theis FJ (2022). Benchmarking atlas-level data integration in single-cell genomics. Nature Methods 19(1):41-50. doi:10.1038/s41592-021-01336-8. https://doi.org/10.1038/s41592-021-01336-8
  7. Li S, Lücken M, Marioni JC, Teichmann SA, He P (2026). Toward informed batch correction for single-cell transcriptome integration. Nature Computational Science 6(2):123-133. doi:10.1038/s43588-025-00943-1. https://doi.org/10.1038/s43588-025-00943-1
  8. Ben-Kiki O, Bercovich A, Lifshitz A, Raz O, Brook D, Tanay A (2023). MCProj: metacell projection for interpretable and quantitative use of transcriptional atlases. Genome Biology 24:220. doi:10.1186/s13059-023-03069-7. https://doi.org/10.1186/s13059-023-03069-7 (preprint: bioRxiv 2022, doi:10.1101/2022.12.01.518678)
  9. Pasquini G, Rojo Arias JE, Schäfer P, Busskamp V (2021). Automated methods for cell type annotation on scRNA-seq data. Computational and Structural Biotechnology Journal 19:961-969. doi:10.1016/j.csbj.2021.01.015. https://doi.org/10.1016/j.csbj.2021.01.015
  10. Zeng Z, Wang X, Du H (2025). CellOntologyMapper: Consensus mapping of cell type annotation. bioRxiv. doi:10.1101/2025.06.10.658951. https://doi.org/10.1101/2025.06.10.658951
  11. CZI Cell Science Program, Abdulla S, Aevermann B, Assis P, Badajoz S, Bell SM, Bezzi E, Cakir B, et al. (2025). CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Research 53(D1):D886-D900. doi:10.1093/nar/gkae1142. https://doi.org/10.1093/nar/gkae1142