Can Empirical Rbe Models Predict Variations In Intrinsic Radiosensitivity?
Abstract
Purpose
To assess the accuracy of empirical relative biological effectiveness (RBE) models in predicting variations in intrinsic radiosensitivity using a comprehensive dataset of clonogenic cell survival.
Methods
A dataset composed of 46 pancreatic cancer cell lines, irradiated with a constant dose-averaged linear energy transfer (LET) of 3.85 keV/μm, was used to extract radiosensitivity metrics, including D10%, D50%, mean inactivation dose and survival fraction at 2 Gy with their corresponding RBEs. Eleven RBE models were retrained on this dataset using the leave-one-out cross-validation method, with minimizing the root mean squared error (RMSE) for either α and β values, or RBE at a specific endpoint. The goodness-of-fit for each model before and after retraining was reported in terms of the RMSE, reduced chi-squared statistic (χ2red), and Bayesian information criterion (BIC). Additionally, the dataset was categorized into four groups based on cell line radiosensitivity to assess at which level of radiosensitivity the models perform more accurately.
Results
Most models showed similar RMSE and χ2red values with a similar trend for all endpoints. However, poor RBE predictions were observed for three models, likely attributed to how these models were first derived and their functional expressions, which limit their use to conditions (LET and α and β values) outside the conditions they were derived. Although models with more complex functional forms outperformed the others, those with simpler expressions achieved the best BIC scores. Model retraining improved prediction accuracy, with some models showing substantial improvements. Overall, the models exhibited better accuracy for radioresistant cell lines.
Conclusion
Our findings suggest that it is best to retrain RBE models to a specific dataset that closely represents the range of LET, and α and β values in which they will be utilized. However, inaccuracies in RBE predictions are still present, highlighting incorporating intrinsic biological/genomic characteristics into future RBE models.