BMC GenomicsBioinformaticsPeer-reviewed journal article
Slingshot: Cell lineage and pseudotime inference for single-cell transcriptomics
Kelly StreetiD, Davide RissoiD, Russell B. FletcheriD, Diya DasiD, John NgaiiD, Nir YosefiD, Elizabeth A. PurdomiD, Sandrine DudoitiD
Abstract
Single-cell transcriptomics allows researchers to investigate complex communities of heterogeneous cells. These methods can be applied to stem cells and their descendants in order to chart the progression from multipotent progenitors to fully differentiated cells. While a number of statistical and computational methods have been proposed for analyzing cell lineages, the problem of accurately characterizing multiple branching lineages remains difficult to solve. Here, we introduce a novel method, Slingshot, for inferring multiple developmental lineages from single-cell gene expression data. Slingshot is a uniquely robust and flexible tool for inferring developmental lineages and ordering cells to reflect continuous, branching processes.
- organism: human
- organism: mouse
- assay: scRNA-seq
- research focus: trajectory inference
- research focus: software / tool
- approach: statistical model
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (19 Apr 2017) · CC BY 4.0
Slingshot builds a cluster-level minimum spanning tree to find branching lineages without fixing their number, then fits simultaneous principal curves for stable per-lineage pseudotimes.
Earlier pseudotime tools trade stability for flexibility: Monocle's cell-level tree is noise-sensitive and needs the lineage count fixed in advance, Embeddr fits one lineage, and Wishbone fits at most two. Slingshot separates finding the lineage structure from fitting the curves, staying robust to noise and clustering choices while finding an unknown number of branches, with optional cluster-level endpoint constraints.
Key results
ShownUnder resampling, principal-curve orderings were the most stable and Monocle's cell-level MST diameter path was the least stable. This motivated extending principal curves to branching lineages.
HSMM subset of 212 cells in ICA space with 50 bootstrap-like subsamples (each about 63% of cells on average). Monocle's path varied widely. A k-means (k = 5) cluster MST gave tied pseudotimes at vertices and occasional spurious branches. Principal curves gave very similar orderings across subsets. No numeric stability metric was reported. · Robustness to Noise; Figure legends
SuggestedSimultaneous principal curves keep pseudotime estimates stable across clusterings, even when the underlying cluster MSTs change a lot. MST-only orderings (Waterfall/TSCAN style) shift with the cluster choice.
On a simulated 2D dataset with two endpoints and k-means for k = 3 to 14, Slingshot curves were nearly identical while the trees varied. Clustering still matters at the extremes: k = 2 found no branch, and k = 15 gave spurious branches and overfit curves. · Robustness to Cluster Assignments; Figure legends
ShownSupervising only the endpoints, together with a covariance-scaled distance between clusters, recovered the three-lineage structure of the olfactory epithelium (neuronal, sustentacular, and microvillous branching off the neuronal path). The unsupervised and Euclidean variants did not.
616 mouse olfactory epithelium cells, first five PCs, horizontal basal cell (HBC) cluster as root. Without supervision the sustentacular cluster was not an endpoint. Marking the mature sustentacular cluster as an endpoint gave the lineages validated in the source study. Euclidean distances between cluster centers created a spurious early HBC branch. · 2.2 Multiple Lineage Identification; 4.3 Datasets; Figure legends
ShownMonocle and Wishbone did not reproduce the known olfactory epithelium lineages, even with favourable settings. Slingshot's lineages matched prior biology on both the full dataset and the two-lineage subset.
Monocle was given the correct count of three lineages and HBC orientation. Its neuronal path skipped most of the GBC intermediates, treated GBCs as a terminal branch, and absorbed microvillous and immature sustentacular cells. On the two-lineage subset, Wishbone's bifurcation did not separate neurons from sustentacular cells with either diffusion maps or PCA. · 2.2 Multiple Lineage Identification; Figure legends
SuggestedSlingshot recovers a single lineage and a two-branch bifurcation in two other published datasets, without being told the number of lineages and without subsetting the data.
The datasets are the 212-cell HSMM subset (single lineage) and the Shin et al. mouse hippocampal qNSC data (two bifurcating fates). The authors say the pseudotimes were highly similar to those in the original papers. The comparisons appear only in Supplemental Figure S3, and no concordance values are given. · 2.2 Multiple Lineage Identification; 3 Discussion; 4.3 Datasets
Methods, data and reproducibility
Starting from clustered cells in a reduced space, Slingshot builds a cluster MST with a covariance-scaled (Mahalanobis-like) distance, rooted at a user-chosen cluster with optional terminal-cluster constraints, then fits simultaneous principal curves that shrink lineage curves toward their average where cells are shared; pseudotime is the arclength of each cell's orthogonal projection.
- Human skeletal muscle myoblast scRNA-seq (Trapnell et al. 2014) with mesenchymal contaminants removed; Monocle or k-means labels in a 2D ICA space · human · n = 212 of 271 cells · GSE52529
- Adult mouse hippocampal quiescent neural stem cells and their progeny, scRNA-seq (Shin et al. 2015); hierarchical clustering labels, first two PCs · mouse · n = 132 cells (101-cell main lineage, 23-cell side cluster) · GSE71485
- Adult mouse olfactory epithelium scRNA-seq from a companion study (in press); RSEC cluster labels, first five PCs, three terminal fates · mouse · n = 616 cells · GSE95601
- Simulated two-dimensional dataset with two endpoints, clustered with k-means for the robustness tests
- Methods & tools
- slingshot (R package), Monocle, TSCAN, Waterfall, Wishbone, Embeddr, principal curves (Hastie and Stuetzle 1989), Mahalanobis-type covariance-scaled cluster distance, k-means clustering, RSEC clustering, PCA and ICA dimensionality reduction, diffusion maps
- Software & environment
- Open-source R package slingshot on GitHub, with a Bioconductor release planned; no R or package versions stated.
- Reproducibility gaps
- Data are given as GEO accessions (GSE52529, GSE71485, GSE95601) rather than URLs. The Bioconductor release was pending, the olfactory epithelium study was in press, and no simulation or figure scripts are linked.
Limitations (5)
- Lineages are defined by the clusters, so poor clustering carries through. In the simulation, k = 2 missed the branch entirely and k = 15 produced spurious branches and overfit curves. · stated by the authors
- Curve fitting uses Euclidean distances, which behave poorly in high dimensions, so the authors strongly recommend reducing dimensions first. They also note that the choice of reduction method can change results substantially (Figure S1). · stated by the authors
- With small clusters the summed covariance matrix may not be invertible, so the distance between clusters falls back to a diagonal covariance approximation. · stated by the authors
- Stability and accuracy are judged visually from plots. No quantitative concordance or ground-truth accuracy metric is reported, and the claim of robustness to clustering rests on a single 2D simulation. · noted in this summary
- The evaluation covers three real datasets and a handful of competitor methods. The three-lineage benchmark depends on endpoint supervision, and its biological validation comes from an in-press companion study rather than from experiments reported here. · noted in this summary
Read this if you need branching lineage and pseudotime inference for single-cell data with an unknown number of lineages, or want the original description of Slingshot's cluster MST and simultaneous principal curves.
- Citations
- 3,148
- FWCI
- 50.31
- Field percentile
- top 1%
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
- CALISTA: Clustering And Lineage Inference in Single-Cell Transcriptional Analysis
Nan Papili Gao, Thomas Hartmann, … Rudiyanto Gunawan·Frontiers in Bioengineering and BiotechnologyPeer-reviewed·Feb 3, 2020
- Cell Lineage and Communication Network Inference via Optimization for Single-Cell Transcriptomics
Shuxiong Wang, Matthew K. Karikomi, … Qing Hua Nie·Nucleic Acids ResearchPeer-reviewed·Mar 27, 2019
- Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning
Mehrshad Sadria, Allen A. Zhang, … Gary D. Bader·Cell SystemsPeer-reviewed·Oct 1, 2026
- scTrace+: enhance the cell fate inference by integrating the lineage-tracing and multi-faceted transcriptomic similarity information
Wenbo Guo, Zeyu Chen, … Jin Gu·Cell SystemsPeer-reviewed·Sep 1, 2025
- Reconstructing complex lineage trees from scRNA-seq data using MERLoT
R. Gonzalo Parra, Νικόλαος Παπαδόπουλος, … Johannes Söding·Nucleic Acids ResearchPeer-reviewed·Aug 1, 2019
- Robust Lineage Reconstruction from High-Dimensional Single-Cell Data
Gregory Giecold, Eugenio Marco, … Guo-Cheng Yuan·Nucleic Acids ResearchPeer-reviewed·May 20, 2016