Machine learning model to predict oncologic outcomes for drugs in randomized clinical trials

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Predicting cancer patient outcomes is complex due to tumor heterogeneity and biological variability.
• Machine learning (ML) offers a potential approach to identify associations between clinical data and patient outcomes.

Purpose of the Study:

• To determine if ML can effectively predict oncologic outcomes using clinical trial, drug-related biomarker, and molecular profile data.
• To develop and validate an ML model for predicting progression-free survival (PFS) and overall survival (OS) in cancer patients.

Main Methods:

• Analyzed data from 104,758 patients across 1102 therapeutic clinical trials for colorectal, pancreatic, melanoma, and non-small cell lung cancers.
• Curated datasets included treatment lines, drug types, molecular alterations, and probability of drug sensitivity (PDS).
• A random forest ML model was trained on PFS and OS data, with performance validated using cross-validation and independent test sets.

Main Results:

• The ML model achieved high correlation coefficients in cross-validation (PFS: r=0.82, OS: r=0.70).
• Spearman correlation on test sets showed significant accuracy (PFS: rs=0.879, OS: rs=0.878).
• The model correctly predicted the better outcome arm in 81% of PFS and 71% of OS randomized trials.

Conclusions:

• Machine learning can successfully predict oncologic outcomes by integrating diverse clinical and molecular data.
• The developed algorithm demonstrates potential for optimizing clinical trial design and pharmaceutical agent development.
• Accurate prediction of patient outcomes can guide therapeutic strategies and improve drug discovery pipelines.