Jeffrey M. Trent

Cell types contribute different information to leukemia dynamics. A, Cell type PsB transcriptomes were created for each sample by combining gene counts from the cells of each cell type. Using the same procedure to identify the PsB state-space, a CML state-space was identified in either PC1 or PC2 for each cell type. The ctPsB state-spaces showed different leukemia development dynamics when compared with the PsB state-space. B, The number of cells in each cell type is shown as a proportion over the scaled time of leukemia development (left). The cell type proportions are grouped by disease state to test cell type population changes over leukemia development (*, P < 0.05; **, P £0.001). C, Computational simulation strategy to quantify each cell type’s contribution to CML dynamics is depicted in the diagram. Each cell type was fixed at the healthy (T0) time point by replacing all cells of that type with the healthy time point cells for all subsequent time points. Those cells were then used to construct a simulated PsB sample for each time point sample to simulate the effect of each cell type being fixed at T0 cell population count and gene expression state. D, Full PsB trajectories for the CP CML mice were compared with the simulated trajectories where each cell type is fixed at the healthy (T0) time point (left). For each sample, the distance between the full and simulated trajectories was calculated as a percentage and then summarized by each cell type to show how much less state distance the fixed cell type simulations traveled in the state-space (middle). To quantify the information loss from fixing each cell type, KLD was calculated between the full PsB and simulated trajectories for each cell type to show that myeloid cells contributed the most information to the state-transition if not allowed to evolve. E, The same simulations were performed for the BC CML mice which showed different contribution from the cell types with B cells contributing the most information to BC CML state-transition. KLD, Kullback–Leibler divergence.

Validation on human CD34+ BM cells. A, Using a cohort of four CML and three healthy BM samples, CD34+ cells were isolated, and scRNA-seq was performed. Cell types were labeled from the data, and PCA was performed to determine whether a unique transcription disease could be identified from the data. However, similar to the mouse model, cell type and not the disease status of the samples was the primary source of variance. B, PCA was also performed on each cell type separately to identify whether disease state emerged at the single-cell level, but the healthy and CML cells were observed to be in a continuum of transcriptional states in all PCs. C, PsB samples were made for all cells in each sample, and PCA was performed. When PC1 vs. PC2 were plotted, PC1 separated CML from healthy samples (percent of total variance explained reported below each axis label). D, Fixed simulations were used on all pairwise combinations of the healthy and CML samples to quantify how each cell type contributed to CML. For each pairwise healthy–CML sample pairing, the KLD was calculated to quantify how much information was lost for each cell type and the mean from all pairwise pairings was reported. The CMP contributed the most information whereas the “other” cell type group that combined all remaining cell types accounting for less than 5% of the total cells in both healthy and CML samples contributed almost no information to CML. CMP, common myeloid progenitor; KLD, Kullback–Leibler divergence.

Leukemia phenotype is a macrostate observable in the pseudobulk data. A, Time-series PBMC samples from a CP CML mouse model were collected weekly, and scRNA-seq was performed to determine the cell contribution to leukemia state-transition (top left). Because the mice developed leukemia at different rates, a scaled time was used to align the samples from the healthy (T0) to leukemia (Tf) time points. PCA was performed using all time points and all cells to determine whether a unique transcriptional state associated with leukemia (Tf) could be identified (i.e., a SC-level CML state-space; bottom). After labeling cells by cell type and visualizing cell type in PCA space, the primary source of variance (PC1 and PC2, which accounted for 35.2% and 19.6% of the variance in the data, respectively) separated cells by cell type and not leukemic from healthy time points. B, PCA was performed on each cell type separately to determine whether a CML state-space could be identified within each cell type. In all PCs, the cells were observed to be in a continuum of transcriptional states that included states from both health and leukemic time points. C, PsB transcriptomes for each sample were created using all cells and, following the same procedure used previously to identify a CML state-space using bulk RNA-seq, a PsB CML state-space was identified in PC1 which explained 34.6% of the total variance. When plotted vs. time, each mouse’s trajectory in PC1 shows the CML development as the mice move from the healthy to leukemic state. The sample density in the bulk state-space was used to identify boundaries between stable critical points (c1, c3, and c5) and the unstable boundary critical points (c2 and c4) in the PsB space. D, Using the critical points of the PsB state-space, a potential landscape was constructed that describes the sample’s CML dynamics. Each mouse is modeled as a particle undergoing Brownian motion in the potential landscape. The stable critical points are used to define the early, transition, and late phenotypic states of CML. E, Association between covariates and the SC- and PsB-level data were quantified using the PC structure produced by each data type. Partial regression, performed for each covariate while adjusting for all others, was used to produce R2 values for each PC. Time was used as a proxy for CML evolution as the mice developed leukemia over the course of the time-series sampling. In PsB data, time showed strong association with the leading PCs which capture the dominant sources of variation, whereas in the SC-data, time exhibited uniformly weak associations across PCs. F, Variance attributable to each covariate was summarized by weighting partial R2 by each PC’s variance fraction and summing across PCs. Time dominates variation in PsB but contributes little in SC data.

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cancer. However, identifying meaningful disease states from single-cell data remains challenging due to the complex continuum of transcriptional alterations, which may obscure clear phenotype boundaries and complicate biological interpretation and clinical relevance. In this study, we systematically explored the chronic myeloid leukemia (CML)-specific information content encoded in scRNA-seq versus bulk transcriptomics to resolve this paradox and clarify how discrete disease-defining states emerge from inherently noisy single-cell data. Although CML single-cell transcriptomes existed along continuous transcriptional microstates, clinically relevant leukemia phenotypes clearly manifested only at the pseudobulk (macrostate) level. State-transition theory was leveraged to reveal how disease phenotype state transitions are governed by cell type–specific contributions. Together, these results establish a theoretical framework explaining why discrete disease phenotypes remain hidden at the single-cell scale but emerge clearly at the aggregated macrostate level, enabling previously inaccessible biological insights into leukemia evolution. By resolving how single-cell variation aggregates into macroscopic disease states, this framework provides insights into CML progression and offers a broadly applicable strategy for exploring disease dynamics across cancers and other complex conditions. Significance:Linking transcriptional single-cell microstates to emergent phenotypic macrostates provides insights into CML progression and establishes a generalizable paradigm for interpreting cancer dynamics from single-cell data.

Cell type counts and categories.

Cell type proportion testing table.

Model of CML dynamics from cell type transcriptomes.

Identifying a CML state-space in human CD34+ BM scRNA-seq.

Parameter values for the model.