Mechanistic learning to predict and understand minimal residual disease

Area of Science:

• Computational Biology
• Systems Biology
• Oncology

Background:

• Mechanistic models offer interpretability in biological systems, particularly in cancer treatment response.
• Machine learning excels at prediction with high-dimensional data common in oncology.

Purpose of the Study:

• To develop a Mechanistic Learning framework combining interpretability and predictive power.
• To identify critical mechanistic parameters for predicting clinical outcomes in B-cell precursor acute lymphoblastic leukemia (BCP-ALL).

Main Methods:

• Employed a Markov chain model to infer 16 mechanistic parameters from patient data.
• Trained a ridge logistic regression machine learning model on these parameters.
• Iteratively assessed parameter importance for predicting BCR::ABL1 fusion gene and minimal residual disease (MRD) status.

Main Results:

• The stem-like cell state was the most predictive feature for BCR::ABL1-positive (score 0.80) and MRD-positive (score 0.67) disease.
• Mechanistic Learning achieved comparable or improved scores (0.81 and 0.71, respectively) using fewer parameters.
• Identified specific cell fate transitions (de-differentiation, primitive-state persistence, differentiation-directed exit) as key predictors.

Conclusions:

• The Mechanistic Learning approach preserves and can improve predictive performance over standard machine learning.
• This framework provides biological hypotheses linking cell states (stemness) to clinical outcomes in BCP-ALL.
• Highlights the potential of integrating mechanistic insights with machine learning for precision oncology.