[C13] Cancer Biology

[C13-02] Machine learning-assisted mechanistic modeling to predict disease progression in acute myeloid leukemia patients

*Chenxu Zhu¹, Er Jin², Johannes Stegmaier², Thomas Stiehl¹ (1. Institute for Computational Biomedicine-Disease Modeling, RWTH Aachen University, Aachen, Germany. (Germany), 2. Institute of Imaging and Computer Vision, RWTH Aachen University, Aachen, Germany (Germany))

Keywords:

Acute myeloid leukemia,Stem cell niche,Mathematical model,Computational model,Machine Learning,Blood cell formation (Hematopoiesis),Cancer stem cell,Hematopoietic stem cell

Blood cell formation is a complex process which is driven by hematopoietic stem cells (HSCs). HSCs give rise to progenitors and precursors and eventually produce mature blood cells, such as white blood cells, red blood cells, and platelets. One of the aggressive blood cancers called acute myeloid leukemia (AML) originates from leukemic stem cells (LSCs) and is characterized by the accumulation of aberrant immature cells, referred to as leukemic blasts. Due to the impairment of healthy blood cell formation, many AML patients suffer from life-threatening complications, such as bleeding or infection. Although treated with high-dose chemotherapy, many patients relapse and need salvage therapy. To reveal the mechanisms of disease progression and relapse, we proposed a mathematical model that accounts for competition of HSCs and LSCs in the stem cell niche and physiological feedback regulations before, during, and after chemotherapy. We fit the model to data of 7 individual patients and simulate variations of the treatment protocol. Our simulation results can recapitulate the non-monotonic recovery of HSCs observed in relapsing patients. The model suggests using the decline of HSC counts during remission as an indication for salvage therapy in patients lacking minimal residual disease markers. To bring our model closer to clinical applications, we propose a machine learning assisted mechanistic model that ensures adherence to biological principles while learning from a larger clinical AML dataset. By embedding mechanistic constraints into machine learning, we aim to identify patient-specific predictors of relapse while preserving biological interpretability.