1195 - A DNA Damage-Based Multiscale Framework for Biological Effect Modeling in Proton Therapy Treatment Planning

Purpose/Objective(s): Variable relative biological effectiveness (RBE) is increasingly recognized as important in proton therapy treatment planning. Voxel-wise prediction of survival fraction (SF), a clinically relevant endpoint for RBE, remains challenging due to the lack of mechanistic models linking macroscopic dose to underlying biological damage and the computational burden of microscopic simulations. We present a mechanistic, multiscale framework that links voxel-level proton dose and energy spectra to DNA damage of varying complexities and ultimately to SF and RBE, enabling biologically grounded evaluation of clinical proton therapy plans.

Materials/Methods: The multiscale framework comprises the GPU-Accelerated Proton Spectrum Model (GPSM) for voxel-wise proton spectrum calculation, the Spectrum-to-Damage Neural Network (SDNN) for predicting DNA damage complexity from spectra, and the Damage-to-Survival Polynomial Model (DSPM) for converting DNA damage to SF and RBE. In GPSM, voxel-wise proton energy spectra and dose were computed using an in-house GPU-based Monte Carlo (MC) simulation tool. SDNN was trained using paired proton spectra and DNA damage data generated across 37 monoenergetic proton energies, 12 high-throughput experimental conditions, and 48 irradiation conditions spanning multiple depths, with spectra computed by GPSM and corresponding DNA double-strand break (DSB) yields (DSB, DSB₃, DSB₄₊ with 2, 3, 4+ damage sites) calculated using microscopic MC simulations of radiation–DNA interactions across physical and chemical stages. To develop DSPM, experimental SF data were compiled from the literature, and corresponding DSB yields were computed using microscopic MC under matched irradiation conditions. DSPM was formulated as a second-order multivariable polynomial to map DNA damage to SF while minimizing overfitting, and RBE was derived at SF=0.1 relative to Cs-137 reference radiation. The model was constructed for the H460 cell line, though the framework is generalizable. We applied the developed framework a conventional IMPT pelvis irradiation plan (1.8 GyE/fraction) and a proton SBRT kidney plan (14 GyE/fraction).

Results: The SDNN and DSPM models achieved relative root mean square errors 9.6% and 9.5%, respectively. The integrated framework successfully generated voxel-wise dose, SF, and RBE maps for clinical treatment plans. In both IMPT and SBRT cases, spatially heterogeneous SF and RBE distributions were observed despite relatively uniform physical dose. Mean PTV RBE values were 1.09 ± 0.08 (IMPT) and 1.11 ± 0.08 (SBRT), while RBE increased to 1.27 and 1.29 at the distal end, respectively, with further increase observed in the distal fall-off region.

Conclusion: We developed a mechanistic, DNA damage–based framework for rapid voxel-level prediction of SF and RBE in proton therapy treatment planning. This framework enables biologically informed plan evaluation and provides a foundation for future biology-guided treatment optimization.