Many radiotherapy (RT) treatments carry a significant risk of serious complications to normal tissues. Previously, efforts to predict outcome have used either physical factors (e.g., volume irradiated to high dose) or biological factors (e.g., inherent radiosensitivity), but not both. We hypothesize that combining physical treatment data with patient-specific biomarkers will increase the predictive power of RT outcomes models. We will test this hypothesis specifically for prediction of radiation pneumonitis (RP) for lung cancer patients. Because no appropriate dataset exists to test this hypothesis, under Specific Aim 1, we will conduct a prospective clinical trial for non-small-cell lung cancer patients and collect candidate physical and biological data for modeling. Biomarker families selected based on peer-reviewed literature will mainly include: (a) the levels of DNA-end binding complexes of DNA damage detection and repair proteins, which has been well correlated with radiosensitivity;and (b) pretreatment blood levels of the interleukin family of inflammatory cytokines (IL-1 alpha and IL-6) and ACE enzymes, which have also been correlated to RP. Physical data to be collected includes volumes irradiated to varying doses of the normal lung (dose-volume histogram data, collected using 4-D methods to avoid breathing artifacts), and the spatial location of the high-dose regions, both of which have been correlated with risk of RP. We will also image tumor regression over the course of RT (mid-course and at the end of RT) in order to better characterize the high doses received by normal lung tissue after tumor regression. In SA2, we test the hypothesized improvement in outcome prediction by combining biological and physical data. To select the best mathematical model, we will adapt and validate a new form of statistical model-building known as kernel-based learning. Based on our preliminary results, the kernel-based methods will likely provide a natural framework for understanding the interactions among the different physical and biological variables resulting in an effectively optimal predictive model. In summary, we propose to test the combination of biological/biomarker data and dose data to improve our ability to predict radiotherapy complications. In particular, we will use detectable blood-based biomarkers as well as treatment dose distribution characteristics to potentially improve our ability to predict (or avoid) radiation pneumonitis.
