Mechanistic Learning Reveals the Reciprocal Cell Fate Transitions That Drive Disease Progression in B-Cell Acute Lymphoblastic Leukemia

Mechanistic Learning, an emerging field with growing applications in oncology, is a novel approach for integrating mechanistic mathematical modeling with machine learning. Mechanistic Learning increases the predictive accuracy of mechanistic models by incorporating techniques from machine learning, while also maintaining the mechanistic insight that is difficult to obtain from otherwise complex, highly nonlinear, black-box prediction algorithms.

We analyzed samples from bone marrow (N=63) and peripheral blood (N=46) from patients diagnosed with B-ALL. Patients received induction chemotherapy with three-year follow-up. We also obtained bone marrow measurements for a cohort of individuals in molecular remission from B-ALL. Data provides flow cytometry characterization of four B-ALL cell states with their normal counterpart appearing in parentheses: CD34+/CD38- (hematopoietic stem cells), CD34+/CD38+ (stage 1 hematogones), CD34-/CD38+ (stage 2 and 3 hematogones), and CD34-/CD38- (naïve B-cells). We compare the utility of 1) traditional machine learning methods, 2) traditional mechanistic modeling, and 3) Mechanistic Learning approaches to make accurate predictions in classifying patient 1) disease status (cancer patients at diagnosis vs molecular remission from B-ALL) 2) BCR::ABL1 status and 3) post-induction minimal residual disease (MRD) status based on flow cytometry, PCR for BCR::ABL1, or ClonoSeq.

First, we used a traditional supervised machine learning approach know as Linear Discriminant Analysis (LDA) that reduces data dimensionality while maximizing the separability of classification data. LDA achieves an accuracy of 80% for predicting BCR::ABL1 status, 84% (bone marrow) and 89% (peripheral blood) for disease status classification, and 71% for MRD prediction. However, LDA alone provides little-to-no mechanistic insight into the underlying biology.

Second, we employ a mechanistic mathematical modeling approach to estimate cell state transition rates between four cell fates described by CD34 and CD38 flow cytometry markers. We hypothesized that disease progression in B-ALL relies on high rates of cell state transitions between stem cell-like, hematogone-like, and naïve B-cell-like leukemia subpopulations. We parameterized a model of cell state transitions using a mechanistic modeling approach known as a Markov chain. However, none of the Markov rate parameters in isolation correlate with clinical outcomes.

Third, we combine mechanistic modeling and machine learning by repeating LDA on parameters found using mechanistic modeling approach above. Similar to LDA strategy, we achieve an accuracy of 76% for BCR::ABL1 status, 83% (bone marrow) and 89% (peripheral blood) disease state classification, and 64% for MRD prediction.

Finally, we search for a reduced list of the most informative parameters such that a high degree of accuracy is maintained. Principle component analysis indicates that the most informative parameters are structural parameters known as a “reciprocity” index which is the rate of incoming and outgoing transitions of each cell state. Re-training the LDA on only the stem-like state reciprocity indices, we maintain a high predictive accuracy for BCR::ABL1 status (80%), disease status (86.5% peripheral blood and 76% bone marrow) and MRD status (70%).

In conclusion, our Mechanistic Learning approach underscores the critical role of cell state transitions that are reciprocating to and from the stem-like CD34+/CD38- state, toward cell fates in the hematogone-like and naïve B-cell-like leukemia states. Stem reciprocity is predictive of disease state, BCR::ABL1 status and MRD status. Our findings highlight the potential of mechanistic learning in enhancing both the understanding and predictive accuracy of disease progression.