Topological structure of radiation-induced DNA damage encodes coupled LET-oxygen signatures

TL;DR

This study uses topological data analysis and machine learning to classify radiation-induced DNA damage features across various particle types, energies, and oxygen levels, revealing mechanistic insights relevant for tumor radiotherapy.

Contribution

First application of persistent homology and Random Forests to analyze DNA damage topology across particle therapy conditions, uncovering topological signatures of particle type, SOBP position, and oxygen tension.

Findings

Particle identity and SOBP position are perfectly decodable from topology.

Oxygen level classification accuracy decreases with LET, influenced by atomic number.

Topological summaries effectively encode oxygen-dependent damage features.

Abstract

We present the first nuclear-scale persistent homology and Random Forest classification analysis of radiation-induced DNA double-strand break (DSB) topology across the clinical particle therapy range. Using TOPAS-nBio and the Voxel-Aware Oxygen model, we generated 2,450 simulated nuclei across 49 conditions (seven particle configurations, 0.2--70.7~keV/\textmu{}m; seven oxygen levels, 0.005--21\%~O$_2$) and extracted a 107-feature matrix across seven modalities. DSB topology encodes particle identity, Spread-Out Bragg Peak (SOBP) position, and oxygen tension in a three-tier hierarchy, with fidelity at each tier governed by the physical mechanism controlling it. Particle identity and SOBP position are exactly decodable (balanced accuracy = 1.000). Oxygen-level classification degrades monotonically with LET from 0.517 (electrons) to 0.189 (carbon distal SOBP), with a charge-driven non-monotonicity at the helium-to-carbon transition confirming that atomic number, not LET alone, governs topological discriminability. The joint 49-class task achieves balanced accuracy 0.346, seventeen times above chance. Per-class recall peaks universally at 0.5\%~O$_2$ (0.788--0.976 across all configurations), which is consistent with the OER curve inflection. Topological Summaries (persistent entropy, landscape integrals) dominate oxygen encoding at all LET ($\eta^2_{O_2} =\,$0.300--0.622). A partial-out test reveals two mechanistically separable channels: a count-mediated scale signal ($\eta^2_{O_2}$ survival ratio 0.062) and a count-independent shape signal preserved or enhanced in five of seven configurations (balanced accuracy survival ratio 1.011). Persistent entropy and landscape integrals, as novel radiobiological observables, provide a computational basis for characterizing oxygen-dependent damage topology in hypoxic tumor treatment planning.