Longitudinal single cell RNA-sequencing reveals evolution of micro- and macro-states in chronic myeloid leukemia

Background

Single-cell RNA sequencing (scRNA-seq) has transformed the field of cancer research by allowing scientists to study gene expression of cancerous cells at the individual cell level [1][2][3]. However, identifying disease states is more challenging with scRNA-seq than with bulk RNA sequencing (RNA-seq). Bulk RNA-seq reveals different disease stages by providing a gene expression profile from a large population of cells [4]. In contrast, scRNA-seq data are derived from single cells and contain a lot of variability, making it much harder to identify disease stages [4].
In this preprint, Frankhouser and colleagues sought to understand how the single-cell transcriptome translates into the distinct disease states detected by bulk RNA-seq. Their approach was based on state-transition theory which suggests that while the individual cells exhibit variability, their disease states will become clear at the cell aggregate level. They focused on chronic myeloid leukemia (CML), a type of cancer that has two phases: it begins at the chronic phase (CP) and develops into the blast crisis (BC) phase [5]. Ultimately, they showed that scRNA-seq data can be used to reveal different disease states at the pseudobulk scale.

Key Highlights

Discrete CML stages are only visible at the cell aggregate level
Frankhouser and team first investigated CML with bulk RNA-seq using time-series gene expression data. Using Principal Component Analysis via Singular Value Decomposition (SVD), they identified that the biggest change in gene expression is associated with the shift from healthy to diseased states.
The researchers then examined a time-series scRNA-seq dataset, collected weekly from mice as they developed CP CML. Their initial goal was to map the disease’s progression from a healthy state to a diseased state through analyzing individual cells. However, their attempts showed that the patterns of change were due to differences in cell type rather than the stage of leukemia. When investigating the cells within each cell type, they were still unable to detect a unique state-transition between the healthy and diseased cells.
The researchers decided to combine the gene counts of all the single cells together into a pseudobulk sample to see if state-transition could be detected in aggregate. Using the SVD method, they were able to map the trajectory of the single-cell dataset from healthy to leukemic. They identified three distinct states within the trajectory: the Early state, the Transition state, and the Late state.

Each cell type contributes to CML state-transition
The authors investigated state-transition within each of the four cell types (B cells, T cells, myeloid cells, and stem cells) of the scRNA-seq data. Using SVD at the pseudobulk stage, they found that the trajectory states are clear when cells are aggregated together in each cell type. They then conducted a computer simulation to see how much each cell type influences the overall disease trajectory. They fixed one cell type at the healthy state while allowing the other cell types to naturally evolve. They compared the fixed simulation with the natural trajectory and saw that the simulated myeloid population caused the biggest information loss (a proxy for cell type importance), suggesting it plays the most important role in the CML state-transition. They also found that, although myeloid cells made the largest contribution, each cell type contributed to the overall leukemic state.

 

Figure 2C from the preprint – The computer simulation of the mouse CML model in the chronic phase (top) and blast crisis phase (bottom). The researchers used a principal component analysis (PCA) and time-ordering to fit a trajectory line through the pseudobulk (PsB) data points from healthy to sick. The graphs on the left show the simulated trajectory compared to the real trajectory, and the graphs on the right show the cell type contributions through information loss. Figure made available under a CC-BY-NC-ND 4.0 International license.

The researchers then confirmed these findings with a different CML mouse model that mimics the BC phase. Once again, they identified that state-transition can only be found at the cell aggregate level for single-cell data. They then separated the cells into their respective cell types and investigated the trajectory within each cell type. This time, when implementing the simulation, they found that the B cells, along with the myeloid cells, were the biggest drivers of state-transition.

State-transition approach is validated in human CML model
The authors decided to test their approach on human CML samples due to potential limitations of their mouse models. They analyzed scRNA-seq data from bone marrow stem and progenitor cells. They were unable to detect state-transitions when investigating the cells individually. However, when aggregated together, they were able to differentiate between the healthy state and the leukemic state.
The authors then adapted their computational simulation to work with the patient data. Using this simulation, they fixed each cell type to measure the information loss of the overall trajectory. They identified that common myeloid progenitors (CMP) and Pro B-cells contributed the most information to state-transition.

A mathematical model demonstrates cell-type contribution to state transitions
After demonstrating that the overall disease state is made up of each cell type’s contribution, the researchers built a mathematical model to formalize this relationship. The model is based on a concept of linear combination, where the state of the disease is the sum of the individual state of each cell type. For validation, the researchers implemented the model on the time-series CP CML mouse scRNA-seq dataset and found that the model perfectly recreated the previously measured disease trajectory.

Figure 4A from the preprint – A visual representation of the mathematical model developed for pseudobulk state-transition. Figure made available under a CC-BY-NC-ND 4.0 International license.

Conclusion
By tracking CML progression with longitudinal single-cell RNA-seq, Frankhouser and team showed that the trajectory from healthy to diseased states can only be visible when single-cell microstates are aggregated into pseudobulk macrostates. By investigating the separate cell types of the cell aggregate, they found that each cell type contributes to state-transition. They built and validated a mathematical model representing the cell-type contribution. In the future, this model can pave the way for clinical applications, such as predicting disease development and treatment response.

Why this preprint is important
A notable strength of this paper is the authors’ ability to connect single-cell data directly to clinically relevant phenotypes. Typically, scRNA-seq data does not show a direct clinical relevance due to high variability at the individual cell level. However, when aggregated, the more relevant information of the disease was clearly shown. Additionally, as the authors mention, their framework showed that complex systems can be built from simple parts. Through their mathematical model they demonstrated that a complex, three-state disease trajectory can emerge from the combination of simple, single-state cells. The disease trajectory and systems-level approach discovered in this paper holds promise for future clinical applications, such as predicting a patient’s risk of a disease becoming more aggressive.

Questions for the authors:

1. You define the macrostate by aggregating gene expression. Do you think other biological layers, like protein levels or cell-to-cell communication signals, are also part of this macrostate, and would incorporating them make the model even more accurate?
2. How many single cells do you think are needed per sample to create a stable and reliable macrostate?
3. In the computational simulation you conducted, is it possible that when you fix one cell type, the other cell types might over- or under-compensate for their absence? If so, how might these dynamic interactions affect the information loss you calculated for each cell type?

References:
[1] Li, L., Xiong, F., Wang, Y., Zhang, S., Gong, Z., Li, X., He, Y., Shi, L., Wang, F., Liao, Q., Xiang, B., Zhou, M., Li, X., Li, Y., Li, G., Zeng, Z., Xiong, W., & Guo, C. (2021). What are the applications of single-cell RNA sequencing in cancer research: a systematic review. Journal of experimental & clinical cancer research : CR, 40(1), 163. https://doi.org/10.1186/s13046-021-01955-1
[2] Zhang, Y., Wang, D., Peng, M., Tang, L., Ouyang, J., Xiong, F., Guo, C., Tang, Y., Zhou, Y., Liao, Q., Wu, X., Wang, H., Yu, J., Li, Y., Li, X., Li, G., Zeng, Z., Tan, Y., & Xiong, W. (2021). Single-cell RNA sequencing in cancer research. Journal of experimental & clinical cancer research : CR, 40(1), 81. https://doi.org/10.1186/s13046-021-01874-1
[3] Chang, X., Zheng, Y., & Xu, K. (2024). Single-Cell RNA Sequencing: Technological Progress and Biomedical Application in Cancer Research. Molecular biotechnology, 66(7), 1497–1519. https://doi.org/10.1007/s12033-023-00777-0
[4] Tzec‐Interián, J. A., González‐Padilla, D., & Góngora‐Castillo, E. B. (2025). Bioinformatics perspectives on transcriptomics: A comprehensive review of bulk and single‐cell RNA sequencing analyses. Quantitative Biology, e78. https://doi.org/10.1002/qub2.78
[5] Michor F. (2007). Chronic myeloid leukemia blast crisis arises from progenitors. Stem cells (Dayton, Ohio), 25(5), 1114–1118. https://doi.org/10.1634/stemcells.2006-0638

Tags: chronic myeloid leukemia, mathematical modeling, scrna-seq

doi: https://doi.org/10.1242/prelights.41929

Read preprint (No Ratings Yet)