Identification of high-dimensional omics-derived predictors for tumor growth dynamics using machine learning and...

Area of Science:

• Pharmacology
• Bioinformatics
• Computational Biology

Background:

• Pharmacometric modeling captures tumor growth inhibition (TGI) dynamics but often lacks high-dimensional covariate analysis.
• High-dimensional molecular profiling ('omics') is increasingly used for predicting anticancer drug response.
• Machine learning (ML) methods can identify omics predictors for treatment outcomes.

Purpose of the Study:

• To combine TGI modeling and ML for omics-based prediction of tumor growth profiles.
• To identify biological pathways associated with treatment response and resistance using genomic data.
• To develop a two-step approach integrating ML (LASSO regression) with pharmacometric modeling.

Main Methods:

• A two-step workflow combining ML (LASSO regression) and pharmacometric modeling was proposed.
• Pharmacometric TGI models were fitted to 4706 patient-derived xenograft (PDX) tumor growth profiles.
• LASSO regression was used to link empirical Bayes estimates of TGI parameters to high-dimensional genomic copy number variation data (>20,000 variables).

Main Results:

• The integrated model reduced median prediction error by 4% compared to models without genomic information.
• LASSO identified 74 pathways related to treatment response or resistance.
• A portion of the identified pathways were validated through existing literature.

Conclusions:

• The combined ML and pharmacometric modeling approach enhances pharmacological understanding of genomic factors influencing treatment response variation.
• This workflow provides a powerful tool for integrating complex omics data into predictive models for anticancer drug efficacy.
• The study highlights the potential of leveraging genomic insights for personalized cancer therapy prediction.