Flexible Data Trimming Improves Performance of Global Machine Learning Methods in Omics-Based Personalized Oncology

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Machine learning (ML) applications in omics-based cancer drug prescription are limited by insufficient clinical outcome data, leading to overtraining and vulnerability.
• A hybrid global-local ML approach, floating window projective separator (FloWPS), was developed to mitigate these issues through data trimming (sample-specific feature removal).

Purpose of the Study:

• To evaluate the effectiveness of the FloWPS approach in improving the performance and robustness of various ML methods for predicting cancer treatment response.
• To assess the impact of FloWPS on classifier quality and overtraining across multiple gene expression datasets.

Main Methods:

• Applied FloWPS to seven popular ML algorithms: linear SVM, kNN, RF, RR, BNB, ADA, and MLP.
• Conducted computational experiments on 21 high-throughput gene expression datasets from 1778 cancer patients with known chemotherapy responses.
• Evaluated classifier performance using ROC AUC and tested for overtraining by analyzing feature importance correlations.

Main Results:

• FloWPS significantly improved classifier performance for global ML methods (SVM, RF, BNB, ADA, MLP), increasing ROC AUC from 0.61-0.88 to 0.70-0.94.
• The FloWPS approach demonstrated robustness against overtraining, evidenced by increased feature importance correlation between different ML methods.
• Binomial Naive Bayes (BNB) achieved the best performance with FloWPS data trimming across all tested datasets.

Conclusions:

• FloWPS enhances the accuracy and robustness of ML models for cancer drug response prediction.
• The method's ability to reduce overtraining makes it a valuable tool for developing reliable ML classifiers in personalized oncology.
• Binomial Naive Bayes, when combined with FloWPS, shows particular promise for future applications in precision medicine.