Symbolic Regression: A Versatile Approach for Constructing Phenomenological Models of Radiobiological Effects.

Abstract

The development of quantitative models that correlate physical, chemical, and biological parameters with radiobiological effects is imperative in the domains of radiotherapy and radiation protection. Due to the challenges associated with quantifying underlying mechanisms, phenomenological models are frequently established in preference to mechanistic models. However, the lack of a universal methodology for constructing phenomenological models presents a significant challenge in the field. We employ symbolic regression as a method for constructing phenomenological models. We attempt to develop models for the survival fraction, microdosimetric parameters, the radiation oxygen effect, and the FLASH effect. Additionally, we compare the results obtained from our symbolic regression approach with existing formulas in the scientific literature to assess the efficacy and validity of our method. Symbolic regression yields multiple simple formulas for each modeling task undertaken. These formulas demonstrate a comparable ability to predict radiobiological effects as the formulas presented in previous scientific publications. Our findings propose that symbolic regression is an automated and flexible strategy for constructing phenomenological models of radiobiological effects. Additionally, they underscore that the interpretability of a model is as crucial as its goodness of fit, as symbolic regression can identify various distinct formulas that adequately fit the provided data points.