henhaobio
Sign in

bioRxivBioinformaticsPreprint

SCORPIUS improves trajectory inference and identifies novel modules in dendritic cell development

Robrecht CannoodtiD, Wouter SaelensiD, Dorine SichieniD, Simon Jan TavernieriD, Sophie JanssensiD, Martin GuilliamsiD, Bart N. LambrechtiD, Katleen De PreteriD, Yvan SaeysiD

PDFDOICode

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 2 (7 Oct 2016)

Unsupervised trajectory inference by SCORPIUS outperformed Wanderlust, Monocle and Waterfall across ten scRNA-seq datasets and predicted falling protein synthesis during dendritic cell development, validated in vivo.

Pioneering trajectory inference tools were each tested on just one or two datasets and mostly rely on prior knowledge (Wanderlust needs a start cell; Monocle picks genes using given cell states), and in a head-to-head benchmark none performed consistently well. SCORPIUS orders cells with no user input, and its ten-dataset benchmark scoring label-based accuracy and ordering consistency gives a basis for objectively comparing future methods.

Key results

  1. ShownAcross the ten-dataset benchmark, SCORPIUS scored significantly higher than Wanderlust, Monocle and Waterfall on both dimensionality-reduction accuracy and cell-ordering consistency, and beat all of them on every dataset except 5c, where Monocle ordered cells more consistently.

    Embedding scored by five-nearest-neighbor (5-NN) cross-validation accuracy (CVA), ordering by the share of cell pairs ordered consistently with known stages (COS); each averaged over 100 runs per method and dataset; one-sided paired Wilcoxon tests, significance flagged at p < 0.05 and p < 0.005 (Figure 2b); numeric scores appear only in figures. · 3.2 Extensive benchmarking shows SCORPIUS…; Evaluation metrics; Performance comparison; Figure legends

  2. ShownOn dendritic cell (DC) progenitor scRNA-seq data, SCORPIUS recovered the known differentiation order without supervision and split the most trajectory-predictive genes into six modules sharing expression patterns and functions, with module 4 holding established DC genes.

    57 monocyte-DC progenitors (MDPs), 95 common DC progenitors (CDPs) and 96 pre-dendritic cells (PreDCs); Random Forest importance with permutation testing retained the top 120 genes (p < 10^-4); Gaussian mixtures with the Bayesian information criterion gave six modules; module 4 includes Itgax, Cd209a, Lgals3; no ordering score is reported for these data. · 3.3 SCORPIUS highlights different functional…; Feature selection; Figure legends

  3. ShownIn vivo labeling of nascent protein in mouse bone marrow supported the hypothesis from module 3 (translation genes): protein synthesis was high in MDPs and declined progressively through later DC progenitor stages.

    O-propargyl-puromycin (OP-Puro; 50 mg/kg intraperitoneally, bone marrow collected 1 hour later and analyzed by flow cytometry) in five mice; absolute fluorescence varied by mouse but relative levels agreed across replicates; hematopoietic stem cells and multipotent progenitors reportedly incorporated significantly less than common myeloid progenitors; no p-values or test statistics are printed. · 3.4 A decrease in protein synthesis rate during…; 4 Discussion; Measurement of protein synthesis; Figure legends

  4. SuggestedDimensionality reduction scored relatively high for every method, but trajectory ordering varied widely; SCORPIUS's ordering was comparatively weak on datasets 3c and 5d, attributed to within-time-point subpopulations and a possible systematic error in the data.

    Per visual inspection, dataset 3c showed strong heterogeneity among cells from the same time point, and on dataset 5d every method misordered the short-term and long-term HSC (ST-HSC, LT-HSC) stages; Waterfall and Wanderlust were highly variable yet correlated, and Monocle's longest minimum-spanning-tree path was deemed noise-sensitive. · 3.2 Extensive benchmarking shows SCORPIUS…

  5. SuggestedModule dynamics outline functional waves: early-hematopoiesis/alternative-lineage and translation modules decline during DC differentiation, actin-polymerization genes rise at CDP and PreDC stages, and major histocompatibility complex class II genes are highly expressed in late PreDCs.

    Modules 1 and 2 include Cd27, Cd34, Mpo, Prtn3 and Nkg7; module 3 Ncl and Cdk4; module 5 Tmsb4x, Crip1 and Actb; module 6 Cd74, H2-Aa, H2-Ab1 and H2-Eb1; bulk microarray data are said to closely match but no similarity metric is given, and functional roles rest on prior literature. · 3.3 SCORPIUS highlights different functional…; 3.4 A decrease in protein synthesis rate during…; Figure legends

Methods, data and reproducibility

SCORPIUS applies classical multidimensional scaling to Spearman correlation distances, seeds a path through k-means centers, refines it with principal curves into pseudotime, and ranks genes by Random Forest importance, grouping them with Gaussian mixtures. Evaluation used 5-NN accuracy and ordering consistency, plus OP-Puro translation labeling in mice.

  • Ten benchmark scRNA-seq datasets from five studies (differentiation, cell cycle, stimulus response) with stage labels from cell sorting or sampling time points · n = 10 datasets from 5 studies; at least 50 cells per stage
  • scRNA-seq profiles of dendritic cell progenitors: monocyte-DC progenitors (MDPs), common DC progenitors (CDPs) and pre-dendritic cells (PreDCs) · n = 57 MDPs, 95 CDPs, 96 PreDCs
  • Bulk microarray expression covering further DC developmental stages and other cell types, used to cross-check the selected module genes
  • In vivo OP-Puro incorporation into nascent protein of bone marrow cells, quantified per cell by flow cytometry across stem, myeloid and DC progenitor stages · mouse · n = five mice
Methods & tools
SCORPIUS (all results: n = 3 scaling dimensions, k = 4 clusters), Wanderlust (defaults: num_landmarks = 20, num_graphs = 100, k = 30, l = 8), Monocle (defaults: num_genes = 1000, num_paths = 1), Waterfall (default: k = 5), Classical Torgerson multidimensional scaling of Spearman rank correlation distances, k-means clustering with a density-aware shortest path through cluster centers, Principal curves (Hastie and Stuetzle; princurve R package), Random Forest pseudotime regression with permutation-based gene p-values, mclust R package (Gaussian mixtures; cluster number by Bayesian information criterion), fitdistrplus R package (normal fit for outlier removal), One-sided paired Wilcoxon rank test, O-propargyl-puromycin (OP-Puro) labeling, BD Fortessa X20 flow cytometry, FlowJo 10
Software & environment
R packages fitdistrplus, princurve and mclust underpin outlier filtering, principal curves and gene-module clustering (no versions given); flow cytometry analyzed in FlowJo 10.
Experimental conditions
Hardware and run times are not reported; Table S2 rates scalability by how many cells a method handled before runtime exceeded 10 seconds, but the ratings themselves are not in the text.
Reproducibility gaps
Accession codes for the ten benchmark datasets sit in supplementary Table S1; the dendritic cell scRNA-seq and bulk microarray sources are not identified in the text, and R package versions are not stated.
Limitations (5)
  • SCORPIUS assumes cells sampled uniformly along a linear process; the authors call linearity a shortcoming, noting branching or more general models need many more cells and that benchmarking them requires new metrics and datasets. · stated by the authors
  • Concurrent processes such as maturation and cell cycle complicate a single ordering; the authors note that regressing one out can harm downstream analyses and that separating them would require larger datasets. · stated by the authors
  • Benchmarking was performed by the method's own developers; the authors acknowledge that self-assessment can overstate performance and used ten different datasets to curb selective-reporting bias. · stated by the authors
  • Competing methods ran with default parameters and without the prior knowledge the authors say they were designed to exploit, and the text does not explain how Wanderlust's required start cell was chosen, so the benchmark may disadvantage them. · noted in this summary
  • Support for the DC biology is largely qualitative in the text: no ordering score is given for the DC data, bulk microarray agreement lacks a metric, and the OP-Puro decline across five mice is reported without statistics. · noted in this summary

Read this if you develop or benchmark single-cell trajectory inference methods, or study gene programs and protein synthesis changes during dendritic cell development.

Citations
166

Abstract

1Recent advances in RNA sequencing enable the generation of genome-wide expression data at the single-cell level, opening up new avenues for transcriptomics and systems biology. A new application of single-cell whole-transcriptomics is the unbiased ordering of cells according to their progression along a dynamic process of interest. We introduce SCORPIUS, a method which can effectively reconstruct an ordering of individual cells without any prior information about the dynamic process. Comprehensive evaluation using ten scRNA-seq datasets shows that SCORPIUS consistently outperforms state-of-the-art techniques. We used SCORPIUS to generate novel hypotheses regarding dendritic cell development, which were subsequently validated in vivo. This work enables data-driven investigation and characterization of dynamic processes and lays the foundation for objective benchmarking of future trajectory inference methods.

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

Similarity Map
SelectedSame fieldCross-field
THISTran '20Weiler '22Zhang '21Li '21Zang '26Lange '22
68% semantic matchBioinformatics
Thinh Ngoc Tran, Gary D. Bader·PLoS Computational Biology·Sep 9, 2020· 87 citations