Continuous immune composition retains longitudinal information beyond coarse ecological-context labels in pediatric leukemia

Background Single-cell pediatric leukemia datasets capture substantial variation in cellular composition, but apparent ecological states can be confounded by leukemia lineage, project structure, annotation differences, malignant-cell composition, and compositional constraints. We developed a Bayesian framework to test whether harmonized single-cell-derived immune composition provides longitudinal information beyond coarse ecological-context labels in continuous-time Ornstein–Uhlenbeck (OU) models of a prespecified semiquantitative leukemia-state trajectory. Results We reprocessed two public Single-cell Pediatric Cancer Atlas (ScPCA) leukemia projects using common cell-type mapping and constructed a cross-project immune composition from T/natural-killer (TNK), B-lineage, and myeloid compartments. Among 104 compositionally eligible participants, coarse $$\:k=2$$ clustering was reproducible under bootstrap resampling and showed limited association with project or diagnosis. Within 99 B-cell acute lymphoblastic leukemia (B-ALL) participants, it separated 73 TNK-dominant and 26 B/Myeloid-shifted profiles with minimal project association. However, poor concordance between raw-composition and centered log-ratio-derived assignments indicated representation-sensitive context membership. We therefore evaluated continuous ecological composition separately from coarse context labels in a corrected longitudinal analysis of 34 participants contributing 241 observations and 207 transitions. Higher B-lineage composition was associated with a lower effective OU attractor \:μ\: in both B-only and B+myeloid models, whereas coarse context effects were unresolved. True leave-one-participant-out validation modestly favored B-only over OU-only in aggregate, but predictive differences were heterogeneous across participants; context-only and ecology-plus-context models provided no incremental advantage. A directionally consistent B-associated signal was observed in a secondary ordinal-state analysis. Simulations using observed irregular sampling schedules supported recovery of moderate ecological effects, while posterior-predictive analyses showed improved uncertainty calibration with continuous ecological composition. Conclusions Pediatric leukemia single-cell composition contains reproducible coarse immune structure, but stable clustering does not imply a longitudinally predictive ecological state. Continuous B-lineage composition retained more longitudinal information than discrete ecological-context labels, although its participant-level predictive contribution was modest. This framework distinguishes descriptive ecological coarse-graining from ecological covariates that add information to stochastic longitudinal models while avoiding interpretation of OU mean reversion as direct evidence of stabilizing selection or fixed leukemia ecotypes.