M-Tune: imbalanced data handling in machine learning by tuning the decision threshold.

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Chemistry

Background:

• Class imbalance in machine learning leads to majority class overprediction and poor predictive performance, particularly in biological datasets.
• Existing methods like Random Under Sampling (RUS), Synthetic Minority Oversampling Technique (SMOTE), and Generalized tHreshOld ShifTing (GHOST) have limitations, including sample manipulation or majority class bias.

Purpose of the Study:

• To develop and evaluate a novel technique, M-Tune, for effectively addressing class imbalance in machine learning classifiers.
• To assess M-Tune's performance on imbalanced drug discovery datasets and compare it with established methods.

Main Methods:

• Developed M-Tune, a novel ensemble strategy incorporating threshold shifting and majority voting to handle class imbalance.
• Tested M-Tune on 138 drug discovery datasets using two classifiers and eleven different fingerprints.
• Evaluated performance against RUS, SMOTE, and GHOST for minority class identification.

Main Results:

• M-Tune demonstrated superior performance in predicting the minority class compared to GHOST and SMOTE.
• M-Tune achieved performance comparable to RUS in identifying minority classes.
• Classifiers benefited from M-Tune in effectively classifying the minority class across various fingerprints.

Conclusions:

• M-Tune is a robust and effective method for handling class imbalance in real-world imbalanced datasets, especially in drug discovery.
• The method aids in identifying the minority class in highly imbalanced datasets, offering practical advantages for active compound recovery.
• While M-Tune may increase false positives, its benefit in maximizing the recovery of active compounds is significant.