Chronic Myeloid Leukemia (CML) is driven by expression of the BCR-ABL fusion gene. Despite the success of treatment with tyrosine kinase inhibitors (TKIs), the disease-originating leukemic stem cells frequently persist through therapy, and disease cannot be eradicated. Thus, new biomarkers beyond BCR/ABL may be necessary to assess disease state and treatment response, and to identify novel targets and therapeutics that could lead to a cure. Herein, we used time-series RNA-seq to molecularly characterize the transcriptional states of CML and predict disease evolution and treatment response. Using weekly blood draws from a murine tetracycline (Tet)-inducible BCR/ABL expression model of chronic phase CML and non-BCR/ABL expressing controls ( Fig 1A), we applied state-transition theory to mathematically model CML development. We constructed a CML state-space using dimensionality reduction to the full transcriptomic data of the time-series CML and control mice. From the sample distribution in the state-space, we empirically derived healthy (no BCR/ABL) and perturbed (BCR/ABL-induced) potentials ( Fig 1B). Using these potentials, our model described the transcriptome evolution during CML development as a particle undergoing Brownian motion. We used the Fokker-Planck (FP) solution of the Langevin equation of motion to both accurately model the transcriptome and predict the future dynamics from a single sample used as an initial condition. While control mice remain in the region of the state-space corresponding to health, each of the BCR/ABL expressing mice moved toward CML state as early as one week after BCR/ABL induction and before detection of BCR/ABL mRNA expression. Of note, no significant difference was observed between the experimental and the FP predicted trajectories and survivals, supporting the accuracy of our approach in predicting disease evolution and outcome at the earliest time points. The CML potential was described as a three well potential with three stable (Early c1, Transition c3, Late c5) and two unstable (Early Transition c2 and Late Transition c4). Differentially expressed genes (DEGs) corresponding to Early c1, Transition c3, Late c5 states were identified by comparing with control mice. Importantly, the state-space model also allowed assessment of whether each DEG had a pro- or anti-CML effect. At the Early state c1, we identified 78 DEGs; 55% of them were anti-CML, indicating an early attempt of the biological system to prevent disease development. By GSEA, we showed that the anti-CML Hallmark gene sets included EMT, angiogenesis, hypoxia and the only pro-CML gene set was IL2/Stat5. In contrast, at the Transition c3 and Late c5 states, we identify 366 and 1858 DEGs, respectively; nearly all these DEGs had a pro-CML effect and were enriched for Hallmark gene sets involving inflammation, angiogenesis and metabolism. Using CML state-space as a pseudotime ordering of the CML samples, we also studied modules of non-linear gene expression dynamics. This approach revealed changes in gene expression rates (inflection points) coinciding with the unstable critical points (c2 and c4) that separated the otherwise stable states. The genes in these transition modules included immune/inflammation processes and likely represented those genes that drove CML state-transition. To investigate the evolution of disease “in reverse” (i.e., after the CML phenotype developed), we treated two cohorts of CML mice, one with a TKI (nilotinib) for 4 weeks and one with Tet to turn off the expression of BCR/ABL (Tet-off, Tet-on [TOTO]). We adapted the state-transition model to a “treatment state-transition” model by introducing a specific parameter, g, which both accurately described and predicted individual treatment efficacy. We showed that although the circulating leukemic cells were suppressed by TKI, the disease state recovered only from c5 to c3, but never reached the health (control) state, and the treated mice rapidly relapsed once treatment was completed. Conversely, the TOTO cohort, which had the BCR/ABL expression completed turn off, returned stably to the healthy state. In conclusion, our state-transition model provides a theory-based analytic framework for investigating disease dynamics and predicting treatment response and outcome. Importantly, can be translated into the clinic for novel transcriptomic-based diagnostic approaches.