- Probabilistically map cells across time points in lineage-traced single-cell RNA-sequencing (scRNA-seq) studies.
- Infer ancestors and descendants of rare or transient cell types or states.
- Combine with CellRank to compute putative driver genes, expression trends, activation cascades, and much more.
Our manuscript is available as a preprint on bioRxiv.
High-diversity lineage relationships can be recorded using evolving barcoding systems (Wagner and Klein, Nat Rev Genet 2020); when applied in-vivo, these record independent lineage relationships in each individual. To infer the molecular identity of putative ancestor states, samples need to be related from early to late time points.
In our setting, each individual corresponds to a different time point, and we wish to relate cells across time to infer precise differentiation trajectories ( Forrow and Schiebinger, Nature Comms 2021). While gene expression is directly comparable across time points, lineage information is not: individual lineage trees may be reconstructed at each time point (Alemany et al., Nature 2018, Raj et al., Nature Biotech 2018, Spanjaard et al., Nature Biotech 2018, Jones et al., Genome Biology 2020), but these do not uncover the molecular identity of putative ancestors or descendants.
Critically, moslin uses two sources of information to map cells across time in an optimal transport (OT) formulation (Peyré and Cuturi, arXiv 2019):
- gene expression: directly comparable across time points, included in a Wasserstein (W)-term (Schiebinger et al., Cell 2019). The W-term compares individual early and late cells and seeks to minimize the distance cells travel in phenotypic space.
- lineage information: not directly comparable across time points, included in a Gromov-Wasserstein (GW)-term (Nitzan et al., Nature 2019, Peyré et al., PMLR 2016). The GW-term compares pairwise early and late cells and seeks to maximize lineage concordance.
We combine both sources of information in a Fused Gromov-Wasserstein (FGW) problem (Vayer et al., Algorithms 2020), a type of OT-problem. Additionally, we use entropic regularization (Cuturi 2013) to speed up computations and to improve the statistical properties of the solution (Peyré and Cuturi, arXiv 2019).
Under the hood, moslin is based on moscot to solve the optimal transport problem of mapping lineage-traced cells across time points. Specifically, we implement moslin via the LineageClass , we demonstrate a use case in our tutorial and we showcase how to work with tree distances in an example. Downstream analysis, like visualizing the inferred cell-cell transitions, is available via moscot's API.
Raw published data is available from the Gene Expression Omnibus (GEO) under accession codes:
Additionally, we simulated data using LineageOT and TedSim. Processed data is available on figshare. To ease reproducibility, our data examples can also be accessed through moscot's dataset interface.
To ease reproducibility of our preprint results, we've organized this repository along the categories below. Each folder contains notebooks and scripts necessary to reproduce the corresponding analysis. We read data from data and write figures to figures. Please open an issue should you experience difficulties reproducing any result.
| Application | Folder path |
|---|---|
| Simulated data (Fig. 2) | analysis/simulations/ |
| C elegans embryogenesis (Fig. 3) | analysis/packer_c_elegans/ |
| Zebrafish heart regeneration (Fig. 4) | analysis/hu_zebrafish_linnaeus/ |
The concept figures in this README have been created with BioRender.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent