DNA computing has emerged as a transformative paradigm for tackling computational problems at the molecular level, yet existing approaches remain constrained in algorithmic interpretability, efficiency, and scalability. Here we present a DNA-based decision tree system that modularly embeds classification rules into DNA strand displacement reaction cascades for interpretable decision-making across various configurations. It supports cascaded networks exceeding 10 layers, parallel computation of 13 decision trees in a Random Forest involving 333 strands, and multimode operation (linear/nonlinear, binary/multi-class, single/tandem trees), while maintaining low leakage, rapid signal propagation, and minimal computational elements. Coupled with a DNA-methylation sensing module, it translates biomarker profiles into molecular instructions for tree traversal, reproduces in-silico predictions and enables accurate disease subtype classification. The decision tree system represents an interpretable, scalable, and memory-efficient DNA computing approach and will open new avenues for programming intelligent molecular machines with broad applicability.
© 2025. The Author(s).