bioRxivBioinformaticsPreprint
Moana: A robust and scalable cell type classification framework for single-cell RNA-Seq data
Abstract
Single-cell RNA-Seq (scRNA-Seq) enables the systematic molecular characterization of heterogeneous tissues at an unprecedented resolution and scale. However, it is currently unclear how to establish formal cell type definitions, which impedes the systematic analysis of scRNA-Seq data across experiments and studies. To address this challenge, we have developed Moana, a hierarchical machine learning framework that enables the construction of robust cell type classifiers from heterogeneous scRNA-Seq datasets. To demonstrate Moanas capabilities, we construct cell type classifiers for human immune cells that accurately distinguish between closely related cell types in the presence of experimental perturbations and systematic differences between scRNA-Seq protocols. We show that Moana is generally applicable and scales to datasets with more than ten thousand cells, thus enabling the construction of tissue-specific cell type atlases that can be directly applied to analyze new scRNASeq datasets. A Python implementation of Moana can be found at https://github.com/yanailab/moana.
- organism: human
- assay: scRNA-seq
- research focus: cell type annotation
- research focus: clustering
- research focus: software / tool
- approach: machine learning (classical)
In-depth summary
Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (30 Oct 2018) · CC BY 4.0
Chained linear SVMs on kNN-smoothed, PCA-transformed scRNA-seq classify closely related PBMC subtypes across 10x chemistries, plus a 'mirror validation' check that estimates accuracy without ground truth.
Without formal cell type definitions, each scRNA-seq dataset is labelled by hand through study-specific clustering and marker choices; the paper notes that alignment methods still require that manual step and that scmap-cluster was validated mostly on pancreas, so its handling of closely related types was unclear. Moana instead builds explicit, reusable classifiers as a hierarchy whose levels can use different resolutions and genes, and adds a validation scheme that needs no labelled test data.
Key results
ShownA classifier trained on 10x v2 PBMC data mostly assigned purified 10x v1 PBMC populations correctly, although v1 and v2 cells separated by protocol rather than by cell type in t-SNE.
On purified v1 populations: B, T and NK cells identified with >97% accuracy; >99% of CD8+ T cells placed in the CD8+ lineage and 85% of CD4+ T cells in the CD4+ lineage; accuracy of 94-99% across three T subsets (naïve CD4+, memory CD4+, naïve CD8+). Weak spot: 19% of a purified CD14+ monocyte set called dendritic cells (possible contamination not examined). · Accurate prediction of PBMC cell types in data…
ShownWithin one 10x protocol, the nine-type PBMC hierarchy (eight binary SVMs trained on PBMC-8k) reproduced manual annotations of a dataset from the same donor, including naïve CD4+ and CD8+ T cells that overlapped in t-SNE.
PBMC-8k (n=8,381) for training; PBMC-4k (n=4,340) scored against clustering labels: precision above 98% for CD14+ monocytes, DCs, B and NK cells and 80-90% for T cell subtypes; about 20% of predicted CD16+ monocytes had been annotated CD14+. Smoothing was unnecessary for the top split (k=1) but essential for most finer ones. · Construction and initial validation of a cell…
Suggested'Mirror validation' (retrain on a test set's predicted labels, then check whether the new model reproduces the original training labels) agreed with the pattern seen against manual labels, giving an accuracy estimate without labelled test cells.
With PBMC-4k as mirror set, coherence (a combined precision/recall score) was at or above 90% for all PBMC types, with monocyte and T cell subtype models again lowest; lowering ν lifted imbalanced classes such as T and NK cells above 98%. A pancreas classifier (Baron16-3, n=4,591) also gave highly coherent mirror results. · Systematic assessment of classification…; Construction and validation of a classifier for…
SuggestedA reimplementation of scmap-cluster, a non-hierarchical correlation-based classifier, lost resolution on the cross-protocol PBMC test where Moana held up, especially for T cells.
After reproducing its published results, the authors trained scmap-cluster on PBMC-8k, tuned gene number on PBMC-4k and applied it to purified v1 populations: B and NK cells were accurate, other types much less so, and 28% to 70% of cells in each purified T cell subset were left 'unassigned' (supplementary figures). · Moana outperforms a previously described method…
SuggestedRobustness to IFN-β exposure needed a classifier retrained on matched control cells; it then recovered all major types in treated PBMCs, and per-type expression pointed to a response shared by all cell types.
PBMC-8k classifier confused monocyte and T cell subtypes in Kang18-Ctl (n=12,757) per mirror validation. A classifier retrained on Kang18-Ctl (no T subtypes) gave similar proportions in Kang18-Tx (n=13,551), high coherence except dendritic cells, and matched Butler et al. labels except treated DCs. Over 50 genes rose in all types, including CXCL11, CCL8 and CMPK2, earlier called cell type-specific. · Accurate prediction of PBMC cell types after an…; Expression analysis based on predicted PBMC cell…
Methods, data and reproducibility
Clustering uses heavy kNN-smoothing then DBSCAN in 2-D PC space, repeated within clusters; each split is learned by a linear SVM on lightly smoothed PCA scores, and models are chained into a tree, with large datasets handled via random subsets of about 2,000 cells. 'Mirror validation' retrains on a new dataset's predicted labels and scores agreement back on the training set.
- PBMC-8k, 10x Genomics v2 chemistry PBMC dataset (training) · human · n = 8,381 cells
- PBMC-4k, 10x Genomics PBMC dataset from the same donor (validation; 37 plasmacytoid DCs excluded) · human · n = 4,340 cells
- v1-PBMC-16k, 10x Genomics v1 chemistry PBMC dataset · human · n = 16,000
- 10x v1 datasets of experimentally purified PBMC subpopulations (Zheng et al.) · human
- Baron16-3 pancreas scRNA-seq from the authors' lab (classifier training; four biological samples shown) · human · n = 4,591
- Kang18-Ctl, unstimulated control PBMCs · human · n = 12,757
- Kang18-Tx, PBMCs exposed to interferon beta (IFN-β) · human · n = 13,551
- Methods & tools
- Moana, kNN-smoothing, DBSCAN, principal component analysis (PCA), linear support vector machine (SVM), t-SNE, scmap-cluster (reimplemented), variance-stabilizing transformation (VST), 10x Genomics v1 and v2 chemistry
- Reproducibility gaps
- All method details are deferred to a Supplement not included in this text; no dataset accessions or download links are given, and no software versions or hardware are stated.
Limitations (6)
- Doublets were left in the data rather than removed; the authors note they can confuse the clustering stage and that automated doublet removal is not yet part of the framework. · stated by the authors
- Rare populations are a weak point: dendritic cells (2% of the Kang control training data) had low coherence after IFN-β, and oversampling rare types only helps above some minimum cell count. · stated by the authors
- The PBMC-8k classifier did not separate monocyte or T cell subtypes in the Kang control PBMCs (which differed strongly, possibly from culturing), so a new classifier was trained, and it could not resolve T cell subtypes in that lower-count data. · stated by the authors
- Classifiers are built for robustness and do not flag previously unseen cell types (follow-up clustering or outlier detection is suggested); PCA also gives no sparse gene selection. · stated by the authors
- Same-protocol validation used a second dataset from the same donor scored against the authors' own clustering, and the scmap-cluster comparison used the authors' reimplementation on a benchmark they designed. · noted in this summary
- Training parameters, dataset sources and the synthetic oversampling scheme are described only in a Supplement absent from this text, so settings cannot be checked here. · noted in this summary
Read this if you need an interpretable PCA-plus-linear-SVM approach to transfer PBMC cell type labels across 10x chemistries, or a way to estimate classifier accuracy without a labelled test set.
- Citations
- 68
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- scClassify: hierarchical classification of cells
Yingxin Lin, Yue Cao, … Jean Yee Hwa Yang·Molecular Systems BiologyPeer-reviewed·Jun 1, 2020
- Knowledge-based classification of fine-grained immune cell types in single-cell RNA-Seq data with ImmClassifier
Xuan Liu, Sara J.C. Gosline, … Jeffrey T. Chang·Briefings in BioinformaticsPeer-reviewed·Feb 3, 2021
- scClassifR: Framework to accurately classify cell types in single-cell RNA-sequencing data
Vy Nguyen, Johannes Griss·bioRxiv·Dec 22, 2020
- CellO: Comprehensive and hierarchical cell type classification of human cells with the Cell Ontology
Matthew N. Bernstein, Zhongjie Ma, … Colin N. Dewey·iSciencePeer-reviewed·Dec 8, 2020
- SingleCellNet: a computational tool to classify single cell RNA-Seq data across platforms and across species
Yuqi Tan, Patrick Cahan·Cell SystemsPeer-reviewed·Jul 31, 2019
- Digital Cell Sorter (DCS): a cell type identification, anomaly detection, and Hopfield landscapes toolkit for single-cell transcriptomics
Sergii Domanskyi, Alex Hakansson, … Carlo Piermarocchi·PeerJPeer-reviewed·Jan 13, 2021