The influence of negative training set size on machine learning-based virtual screening.

Area of Science:

• Computational Chemistry
• Bioinformatics
• Machine Learning

Background:

• The performance of machine learning (ML) methods is often influenced by the number of negative training examples, a factor frequently overlooked.
• This study investigates the impact of varying negative training data on ML model efficacy.

Purpose of the Study:

• To analyze how the ratio of positive to negative training instances affects ML performance in virtual screening.
• To identify optimal training data compositions for improved ML-based virtual screening.

Main Methods:

• Evaluated ML methods with fixed positive and variable negative examples from the ZINC database.
• Tested five ML algorithms (SMO, Naïve Bayes, Ibk, J48, Random Forest) using MACCS and CDK molecular fingerprints.
• Simulated virtual screening experiments across multiple protein targets.

Main Results:

• Increased positive-to-negative training ratios significantly influenced ML performance parameters.
• Observed improvements in precision and Matthews Correlation Coefficient (MCC), with minor decreases in hit recall.
• CDK FP with SMO or Random Forest yielded the most effective classification; Naïve Bayes showed low sensitivity to negative instance variations.

Conclusions:

• The ratio of positive to negative training instances is a critical factor in ML experiments and virtual screening.
• Optimizing negative training set size can serve as a boosting approach to enhance ML-based virtual screening efficacy.