Our client is developing a metabolite-based diagnostic test to screen for one of the most common cancer types in men and women. After getting promising results from a smaller scale phase 1 clinical study, the client needed machine learning expertise to begin evaluating a clinical dataset comprised of raw biomarker concentrations for hundreds of clinical study participants. In addition to validating the results from the first study, the client needed fast but robust analyses to narrow in on a small panel of predictive biomarkers that could best drive ML model performance and achieve the sensitivity and specificity needed for a commercial diagnostic.

Our team customized our biomarker discovery codebase to rapidly focus on a core set of predictive biomarkers. The first step in this process was applying a comprehensive set of transformations to expand our feature set and maximize potential signal. We then implemented our feature importance ranking algorithms to hone in on an initial subset of biomarkers. Using our approach, we were able to provide a biological interpretation for why these specific biomarkers had the highest performance by linking them to known signaling and homeostatic pathways from tumor biology. These features were then evaluated across a wide array of ML models through many-fold iterations of cross validation to ensure performance generalizability. Finally, we conducted subpopulation analysis to construct an ensemble model which was able to demonstrate high accuracy while maintaining interpretability. Our client is using these insights to plan their pivotal clinical trial.