Methodological challenges in translational drug response modeling in cancer: A systematic analysis with FORESEE

Area of Science:

• Oncology
• Pharmacology
• Bioinformatics

Background:

• Translational models are vital for personalized medicine, linking in vitro drug responses to in vivo patient outcomes.
• Current research often optimizes isolated model features, neglecting workflow integration and leading to irreproducibility.
• Existing models are frequently dataset-specific, limiting their generalizability and transferability.

Purpose of the Study:

• To thoroughly investigate translational models for predicting cancer patient drug responses across diverse datasets.
• To identify optimal modeling strategies and develop guidelines for robust and simple predictive models.
• To understand factors affecting model robustness and the correlation between cell line and patient drug response predictions.

Main Methods:

• Utilized the R-package FORESEE for systematic investigation of translational models.
• Scanned the modeling space to find optimal combinations of different model settings.
• Analyzed diverse datasets to assess model predictivity and robustness.

Main Results:

• Identified models with extremely high predictivity and established modeling guidelines promoting simplicity.
• Determined that data noise, sample size, and drug unspecificity negatively impact model robustness.
• Demonstrated that accurate cell line models do not always translate to accurate patient drug response predictions.

Conclusions:

• Emphasizes the need for a holistic approach to translational modeling, considering the entire workflow.
• Highlights critical factors that challenge the robustness and generalizability of predictive models.
• Recommends increased focus on in vivo aspects to advance precision medicine and improve patient outcomes.