Journal Club: Building the Single-Cell Atlas Era — From Pediatric Tumors to Mouse Pituitaries to Spatial Epigenomes

· 10 min read · Paper Feed

Journal Club: Building the Single-Cell Atlas Era — From Pediatric Tumors to Mouse Pituitaries to Spatial Epigenomes

Conceptual illustration: Building the Single-Cell Atlas Era — From Pediatric Tumors to Mouse Pituitaries to Spatial Epigenomes

Introduction

Ten years ago, a single-cell RNA-seq paper might profile a few thousand cells from one tissue and feel like a milestone. Today, the bottleneck has flipped: we are drowning in single-cell data, but most of it sits in fragmented lab-specific formats with bespoke annotations that make cross-study comparisons almost meaningless. Three recent papers — the Single-Cell Pediatric Cancer Atlas (ScPCA), the ISON framework for inferring spatial chromatin accessibility, and the Consensus Pituitary Atlas (CPA) — each tackle this problem from a different angle, yet together point to a single theme: single-cell biology is becoming an engineering discipline as much as a discovery science.

What unites these works is the conviction that infrastructure — pipelines, portals, and reference resources — now matters as much as the next flashy dataset. All three emphasize uniform preprocessing, harmonized metadata, open-source tooling, and reproducible downstream analysis. They span a wide range of biological contexts: rare pediatric cancers (ScPCA), the central endocrine hub of the mouse pituitary (CPA), and complex neurodegenerative tissue in Alzheimer's disease (ISON). Read together, they sketch how the field is moving from "one paper, one dataset" to a modular ecosystem where atlases, methods, and platforms can be mixed and matched.

In this post, we walk through what each paper actually does, the tools they introduce, and where they leave open questions. If you have ever stared at a dozen incompatible .h5 files wondering how to even start a meta-analysis, these are the kinds of resources — and the kinds of methodological thinking — you'll want to know about.

Paper 1: ScPCA — A Pediatric Cancer Data Portal Built for Reuse

Key Findings

  • The ScPCA Portal hosts uniformly processed single-cell and single-nucleus RNA-seq data from over 700 pediatric tumor samples across 55 cancer types, expanding well beyond its original 10 Alex's Lemonade Stand Foundation (ALSF)-funded projects to include community-contributed datasets.
  • Downloads are distributed as ready-to-use SingleCellExperiment (R) and AnnData (Python) objects, including raw and normalized counts, PCA/UMAP coordinates, automated cell-type annotations, and per-sample copy-number variation (CNV) estimates.
  • A subset of samples is paired with bulk RNA-seq, spatial transcriptomics, or feature-barcoded data (e.g., CITE-seq), making multimodal analysis feasible without re-processing from raw reads.
  • Every download comes with a summary report, allowing users to triage samples without writing a single line of code before committing to deeper analysis.

Methodology & Study Design

The authors built scpca-nf, an open-source Nextflow workflow that wraps alevin-fry for fast pseudoalignment-based quantification and runs the rest of a standard sc/snRNA-seq QC, normalization, dimensionality-reduction, and clustering pipeline. Crucially, the same pipeline runs across every sample in the portal, which removes pipeline-induced batch effects that have derailed past meta-analyses of pediatric tumors. Cell-type annotation is automated rather than left to the user, and CNV inference is built in — important for distinguishing malignant from non-malignant cells in tumor scRNA-seq data.

Significance

Pediatric cancers are biologically distinct from adult malignancies, but cohorts are small and scattered across institutions. ScPCA effectively turns dozens of underpowered studies into one queryable resource. For the bioinformatics community, the most important takeaway is not any specific biological finding but the demonstration that a pediatric oncology atlas can be built, versioned, and consumed entirely with open tooling. This is the kind of infrastructure that lowers the barrier for trainees entering cancer single-cell analysis, and it sets a benchmark for what reusable data should look like: standardized formats, harmonized metadata, and downloadable processed objects.

Paper 2: ISON — Inferring Spatial Chromatin Accessibility Without Spatial ATAC

Key Findings

  • ISON (Integrated Spatial Omics Network) is a unified computational framework that predicts chromatin accessibility at the spatial spot level by integrating single-cell multiome (RNA + ATAC) data with matched spatial transcriptomics.
  • Predicted accessibility profiles capture cis- and trans-regulatory information and enable estimation of transcription factor (TF) activity at each spot.
  • ISON can distinguish TFs within the same family — a granularity that bulk or even scATAC-only approaches typically miss.
  • Applied to Alzheimer's disease spatial transcriptomics, the method revealed disease- and age-specific spatially variable gene regulatory modules, demonstrating biological interpretability on a real disease dataset.

Methodology & Study Design

Commercial kits for simultaneous spatial RNA + ATAC profiling do not yet exist at scale, so the authors took a different route: they assumed that joint RNA + ATAC measurements from a single-cell multiome experiment, combined with spatial transcriptomics from the same tissue, contain enough shared structure to impute what chromatin accessibility would look like at each spatial spot. ISON uses this shared structure to learn a mapping between expression and regulatory landscapes, then reconstructs spatially resolved gene regulatory networks (GRNs). The authors emphasize scalability in both time and memory — a practical concern when applying any spatial method to whole-transcriptome atlases.

Significance

This is a methods paper in the truest sense: it does not generate a new atlas, it adds a new dimension to atlases others have already built. The clever bit is recognizing that when you cannot measure something directly (spatial ATAC), you can sometimes infer it well enough to do biologically meaningful downstream analysis. The ability to resolve TF activity within the same family is genuinely novel and opens the door to asking mechanistic questions — which member of, say, the AP-1 or ETS family is active where — in spatial contexts. For anyone building or using a single-cell atlas, ISON is a reminder that the most valuable methodological contributions are often those that retroactively enrich existing data.

Paper 3: CPA — Crowdsourcing the Mouse Pituitary

Key Findings

  • The Consensus Pituitary Atlas (CPA) aggregates 283 existing mouse pituitary single-cell datasets — roughly 1.3 million high-quality cells — under one uniform preprocessing workflow.
  • At this scale, the authors identify consensus cell-type and lineage markers, including low-expression transcripts that were statistically invisible in the original individual studies.
  • Using curated metadata, they characterize sex-biased and age-dependent gene expression patterns at cell-type resolution.
  • They train machine-learning models for automated cell-type annotation and doublet detection, both released as reusable tools.
  • A complementary platform, epitome, offers a user-friendly web interface for streamlined analyses.

Methodology & Study Design

Every one of the 283 datasets was reprocessed through the same pipeline — a critical design decision that converts a noisy meta-analysis into a coherent atlas. The authors then went beyond cell typing: they inferred consensus cell–cell communication patterns and integrated RNA-seq with chromatin accessibility data to associate transcription factors with cell-fate programs across modalities. The ML models for annotation and doublet detection are explicitly designed to be applied prospectively by other labs generating new pituitary data, not just to retrospectively annotate the atlas itself.

Significance

The pituitary is small, understudied relative to organs like the liver or brain, and has been profiled piecemeal for over a decade — exactly the kind of field that benefits most from atlas-level integration. CPA is conceptually the closest analog to ScPCA in our trio: both are uniformly processed, openly shared references that turn a fragmented literature into a queryable resource. The biological payoff is concrete — new low-expression lineage markers and robust sex/age effects — but the broader methodological message is that ML-based annotation trained on harmonized atlases is becoming a default output of major data resources, not an afterthought.

Synthesis & Discussion

Read together, these three papers trace a clear arc for single-cell biology in 2026.

Complementarity. ScPCA and CPA are resource papers — they consolidate data. ISON is a method paper — it shows how to add value to such consolidated data by inferring a missing modality. In principle, a future ScPCA-style pediatric tumor atlas paired with spatial transcriptomics could feed directly into an ISON-style pipeline to add a chromatin-accessibility layer to spatial tumor maps. Similarly, ML-based annotation tools of the kind CPA trains are exactly what ScPCA's automated annotations hint at — a sign that this workflow (harmonize → annotate with ML → expose via portal) is becoming a template rather than an exception.

Contrasts. The biological scope could hardly be wider: pediatric human cancers, normal mouse pituitary, and Alzheimer's disease human brain. The modalities also differ: ScPCA and CPA are transcriptomics-first, while ISON explicitly tackles multi-omics integration. One useful tension to flag in lab meeting: ScPCA and CPA are descriptive resources with curated ground truth, whereas ISON is predictive and depends on assumptions about the relationship between RNA and chromatin. Predictions can be powerful, but they require validation — and ISON's evaluation is largely computational rather than experimentally confirmed at scale.

Emerging patterns. Several themes stand out: (1) uniform preprocessing is no longer optional — it is the central contribution of every resource paper we discussed; (2) ML-based annotation is becoming a first-class deliverable, alongside the data itself; (3) portals with web interfaces (ScPCA, epitome) are replacing supplementary tables as the primary user-facing artifact; and (4) methods papers are increasingly atlas-aware, designed to plug into existing resources rather than to compete with them.

Open questions. How generalizable are annotation models trained on one atlas to datasets generated by very different chemistries? Can ISON-style imputed spatial ATAC ever fully replace experimental spatial epigenomics, and how should users report uncertainty? And perhaps most importantly: how do we keep these atlases alive — versioned, updated, and financially supported — as data types evolve?

For Your Lab Meeting

  1. ScPCA standardizes on alevin-fry and Nextflow. Do you think pseudoalignment-based quantification is the right default for all single-cell contexts, or are there cases (e.g., viral integration, custom spike-ins) where full alignment is still necessary?
  2. ISON predicts chromatin accessibility at spatial spots using single-cell multiome data as a bridge. What assumptions about RNA–ATAC coupling does this rest on, and where might those assumptions break down — for example, in post-mitotic neurons or in cells with very different RNA/DNA ratios?
  3. CPA trains ML models on 1.3M cells across 283 datasets. How would you design a fair benchmark to test whether those models generalize to a new dataset generated with a chemistry or species the atlas has never seen?
  4. All three papers invest heavily in infrastructure (workflows, portals, ML tools). As a community, are we rewarding infrastructure contributions appropriately relative to biological "discovery" papers — through citations, funding, or hiring?
  5. Pediatric cancers (ScPCA), pituitary biology (CPA), and Alzheimer's (ISON) are very different fields. Do you see a future in which one unified cross-tissue, cross-disease single-cell atlas replaces field-specific ones, or will the field remain federated by biology?

Key Terms

  • scpca-nf: An open-source Nextflow workflow developed by the ScPCA team that uniformly processes scRNA-seq and snRNA-seq data using alevin-fry for quantification.
  • SingleCellExperiment / AnnData: Standardized R and Python data containers for single-cell omics — SingleCellExperiment from the Bioconductor ecosystem, AnnData popular in scanpy/scverse tooling.
  • Single-cell multiome: Simultaneous profiling of multiple modalities (typically RNA + ATAC) from the same single cell, enabling direct linkage of gene expression and chromatin accessibility.
  • Spatial transcriptomics: Technologies (e.g., 10x Visium, Slide-seq, MERFISH) that preserve the tissue location of gene expression measurements, generating "spot-level" data with positional information.
  • Gene regulatory network (GRN): A model of how transcription factors and cis-regulatory elements interact to control gene expression, often inferred from RNA and chromatin accessibility data.

References

  1. Hawkins AG, Shapiro JA, Spielman SJ, Mejia DS, Venkatesh Prasad D, Ichihara N, et al. (2026). The Single-Cell Pediatric Cancer Atlas: Data portal and open-source tools for single-cell transcriptomics of pediatric tumors.. Cell genomics. https://doi.org/10.1016/j.xgen.2026.101283

  2. Debnath I & Duren Z (2026). Inference of spatial chromatin accessibility via integration of spatial transcriptomics and single-cell multi-omics data.. Nature communications. https://doi.org/10.1038/s41467-026-73948-7

  3. Kövér B, Willis TL, Sherwin O, Kaufman-Cook J, Kemkem Y, Vazquez Segoviano M, et al. (2026). Consensus Pituitary Atlas, a scalable resource for annotation, novel marker discovery, and analyses in mouse pituitary gland research.. Cell reports. https://doi.org/10.1016/j.celrep.2026.117407

Single-Cell Transcriptomics scRNA-seq Genomics