Clonal embeddings allow exploratory analysis of lineage-resolved single-cell data

 

Background

Lineage tracing – labelling a progenitor cell and following the cells it gives rise to over time – allows us to understand the origin of different cell types. Classical cell fate studies, using dye injection or sparse genetic recombination, are limited in the number of clones tracked simultaneously and in the molecular characterization of the progeny.

Over the past decade, a number of DNA barcoding tools have been developed that address these limitations by allowing thousands of clones to be analyzed at once from one organism (GESTALT, CARLIN/DARLIN, PE Tracer, TREX). Upon barcoding induction, each cell acquires a unique DNA sequence (a ‘barcode’). As the cell divides, that sequence is replicated and inherited by all daughter cells. The barcode can be read out by RNA sequencing alongside the whole transcriptome information of the cell, allowing molecularly-phenotyped cells to be grouped into individual clones. Barcoding tools have been applied to the study of many biological systems including hematopoiesis, the early mammalian embryo and the developing brain.

 

Figure 1. Conceptual summary. Traditional clonal analysis represents a clone by the combination of cell types it gives rise to. Clone2vec instead represents a clone by the combination of its nearest neighboring clones, overcoming key limitations of barcoding data analysis, including undersampling and annotation biases. Created with Affinity Designer.

As illustrated in Figure 1, analysis methods are needed to interpret and group the individual clonal patterns. Furthermore, barcoding experiments typically have two limitations. First, the datasets are undersampled because not all cells in a tissue are sequenced, and a barcode may not be retrieved from every cell. Second, the analysis usually relies on cell annotations as an input, and these can have insufficient resolution. As a result, “manual” clonal analysis or simple clustering algorithms may group clones into distinct patterns inaccurately. The new computational method presented in the preprint highlighted here addresses both limitations and enables the analysis of complex clonal patterns from DNA barcoding data (Figure1).

Study approach

Studying thousands of clones is challenging, so the method described in the preprint grouped clones with similar cellular compositions, thereby facilitating analysis and interpretation of the data. This approach leveraged transcriptomic neighborhoods: for a given barcoded cell, its 15 closest barcoded cells in a low-dimensional embedding (e.g. PCA) were considered its neighbors. This neighborhood information was summarized in a clone-by-clone co-occurrence matrix which captured both the similarities between clones and the cell state proportions within each clone. A neural network was then trained on clone pairs to produce a clonal embedding relating clones to each other that can be clustered using standard algorithms. The architecture was inspired by word2vec which predicts the meaning of a target word, as a vectorized representation, based on the context of surrounding words.

To validate their approach, the authors compared it with several alternatives: cluster-based PCA, cluster-based Poisson GLM-PCA or kNN graphs constructed directly from inter-clone distances (Laplacian Maximum Mean Discrepancy (MMD), Gaussian MMD, Energy distance or Sinkhorn divergence). They used simulated data to test agreement of the clustering of clones from subsampled datasets with the full dataset and showed performance superior to other methods (Figure S4).

Why I highlight this preprint

The new clone2vec method has many useful features. Using the transcriptomic nearest-neighbor identities rather than coarse-grained cell type labels (table Figure 1) as inputs to the first step of the embedding is an advance for the field, because it captures finer transcriptomic differences between cells. Clonal patterns can be challenging to classify when clones differ across multiple cell fates, therefore an unbiased clustering approach is valuable. Combining a neighbor-based description of each clone with the skip-gram neural network also handles sparse data more robustly than previous methods: even when some cells are missing, the transcriptional similarity between the remaining cells of clones X and Y will still place them close together in the embedding. The accompanying package is well documented and user-friendly. Follow-up differential gene expression analysis can then address whether cells of the same terminal fate inherit distinct transcriptional signatures depending on their clonal origin.

Key biological findings

When applied to a central nervous system dataset containing progenitors and mature cells (De Haan and He et al., 2025), the clone2vec method recapitulated the original finding that distinct clonal patterns generate different anatomical structures. Differential gene expression analysis between clonal patterns, averaging expression at the clone level for specific cell types of interest, identified dorso-ventral markers that distinguish two clonal behaviors in both neuronal progenitors and their progeny.

In the small-cell lung cancer dataset by Ireland et al. (2025), the method resolved finer clonal patterns, with one cluster related to an epithelial cell state and another to a neuronal cell state.

Re-analysis of non-small cell lung cancer data (Caushi et al., 2021) revealed fuzzy clonal boundaries amongst CD8 T cells, reflecting continuous transcriptomic variation between subpopulations rather than clones giving rise to discrete cell types. The authors therefore applied archetype analysis to the clonal embedding: extreme patterns were identified, and each clone was represented as a combination of them. The resulting grouping was consistent with distinct histories of antigen exposure and activation, and similar clonal patterns were observed across cancer datasets. To integrate datasets, the authors used average clone expression profiles to build a mutual nearest-neighbors graph and then applied a linear transformation to align the clonal spaces of individual datasets while preserving their structure. Interestingly, anti-PD-1 therapy changed clonal frequencies but not the phenotypic properties of the clones.

Future directions

The clustering step is supervised, whereas recent single-cell RNA sequencing methods select clustering resolution in an unbiased way by testing for statistical significance between candidate clusters (Sant et al., 2025). Incorporating such an approach into the clonal clustering step could be an improvement.

The method makes it possible to look at transcriptional differences between progenitors with different clonal behaviors, but progenitor cells are often no longer present by the time the progeny is sequenced. Protocols that allow repeated sampling of the same clones will therefore need to be developed.

Finally, clonal dynamics result from both cell-intrinsic programming and exposure to the environment. The next frontier will be acquiring spatial information at the same time, so that stronger conclusions can be drawn about the influence of the environment and the factors that regulate clonal dynamics.

Question to the authors

What percentage of subsampling is commonly used in the field?

How does clonal heterogeneity, in terms of both the diversity and number of fate patterns, affect dropout stability? Do you see the same trends when subsampling a biological dataset rather than simulating dropout?

References

de Haan,  S. et al. et al. Ectoderm barcoding reveals neural and cochlear compartmentalization. Science 388, 60–68 https://doi.org/10.1126/science.adq9248 (2025)

Ireland,  A. S. et al. et al. Basal cell of origin resolves neuroendocrine-tuft lineage plasticity in cancer. Nature 647, 257–267 https://doi.org/10.1038/s41586-025-09503-z (2025).

Caushi,  J. X. et al. et al. Transcriptional programs of neoantigen-specific TIL in anti-PD-1-treated lung cancers. Nature 596, 126–132 https://doi.org/10.1038/s41586-021-03752-4 (2021).

Sant, C., Mucke, L. & Corces, M. R. CHOIR improves significance-based detection of cell types and states from single-cell data. Nat. Genet. 57, 1309-1319, doi:10.1038/s41588-025-02148-8 (2025).

 

Read preprint 1 Star (No Ratings Yet)