Problem

Current eligibility biomarkers for Immune checkpoint inhibitors (ICIs) poorly predict patient’s response. Drug clearance, both at baseline and over time, is a stronger predictor of ICI effectiveness, however the current non-linear mixed effects modeling (NLME) approach for estimating clearance of ICIs requires administering at least one dose of the drug and collecting plasma samples, which is time-consuming and costly. Previous studies demonstrated that patient demographic and clinical covariates were effective in explaining variability in clearance, yet there is no established machine learning (ML) framework that leverages these variables to predict ICIs clearance before treatment.

Solution

Researchers at The Ohio State University have developed a virtual population of cancer patients using clinical trial data to reflect realistic relationships between patient characteristics. Using an existing population pharmacokinetic model, ICI pembrolizumab drug levels were simulated after one dose and baseline clearance (CL₀) was calculated. Machine learning models were then trained to predict CL₀ using patient features such as body weight, sex, serum albumin, and kidney function, under two scenarios: using the true simulated clearance or clearance estimated from limited timepoint data. Several ML methods were tested, with Gradient Boosting and Random Forest performing best. Key predictors identified by the models matched known clinical factors affecting drug clearance. Adding even one or two drug concentration measurements (peak and trough) further improved predictions, suggesting baseline data alone may not fully capture real-world variability. Overall, the results show that optimized ML models could help predict ICI clearance before treatment and support better patient selection for immunotherapy.

Applications

• Pre-treatment patient stratification to identify individuals less likely to benefit from ICI therapy
• Supports oncologists in optimizing therapy selection and potentially redirecting non-responders to alternative treatments earlier in the care pathway
• May complement existing biomarker-based eligibility criteria

Advantages

• Leverages routinely available clinical and demographic variables already known to influence ICI pharmacokinetics, with optional incorporation of limited plasma concentration data to improve predictive performance

• Non-invasive, faster and more cost-effective compared to traditional NLME clearance estimation
• Enables earlier decision-making before initiation of expensive and potentially ineffective ICI therapy
• Flexible ML approaches (Gradient Boosting, Random Forest)

Seeking opportunities for out-licensing and collaboration