Files
Abstract
Immuno-oncology (IO) is an innovative treatment approach to cancer with substantial benefits on a subgroup of patients. Predictive biomarkers can be used to predict treatment effects for patients and find biomarkers that will help identify suitable subgroups of patients for clinical trials of IO therapeutics. Finding predictive biomarkers is complicated by non-proportional hazard patterns, which are common in IO clinical trials data. The effects of the IO interventions are often seen to vary over time, violating proportional hazards. By using more flexible tools that don’t require proportional hazards to identify predictive biomarkers we can better select the patients who will benefit from IO treatment.
Detecting interaction effects between biomarkers and treatment is an intuitive way to find candidate predictive biomarkers for IO. The LASSO is a well-developed method for finding possible predictive biomarkers, especially useful when there are more candidate biomarkers than samples. However, the LASSO requires the proportional hazards assumption, and that the interaction terms are predefined and included in the model fitting process. Machine learning tools like Random Forest and Gradient Boosting for Additive Models (GAMBoost) provide more flexible interaction term patterns, and Random Forests do not require proportional hazard assumption. However, these models are considered “black-box models”, and they are difficult to make inference about the interaction effects between the biomarkers and treatment.
In this study, we were inspired by variable importance (VIMP) measures originally from Random Forests and have modified and generalized them to a unified framework. These measures detect interaction effects in a wide variety of statistical models, including black box models. The interaction VIMP is the paired VIMP minus two independent VIMPs for a pair of variables of interest; in our case, the variables are a candidate biomarker variable and treatment indicator variable. Hypothesis tests are based on a normality assumption on VIMP resampling distribution. We tested the method on Random Forest and GAMBoost. The method showed robust results in terms of biomarker detection rates in different simulation settings, some of which the LASSO cannot handle. We also investigated delayed effects or diminishing effects settings on the treatment arm, to mimic IO therapeutic effects.