Large-scale learning of structure-activity relationships using a linear support vector machine and problem-specific...

Area of Science:

• Computational Chemistry
• Machine Learning
• Bioinformatics

Background:

• Large-scale binary classification is crucial in cheminformatics, particularly for virtual screening.
• Existing methods struggle with large, imbalanced datasets common in high-throughput screening.
• Linear Support Vector Machines (SVMs) offer computational efficiency for high-dimensional data.

Purpose of the Study:

• To adapt and evaluate a linear large-scale SVM (LIBLINEAR) for cheminformatics classification tasks.
• To assess LIBLINEAR's performance on large and imbalanced datasets using virtual screening metrics.
• To compare LIBLINEAR against established classification and ranking approaches.

Main Methods:

• Extended LIBLINEAR with virtual high-throughput screening metrics for unbalanced datasets.
• Employed Feature Trees for chemotype clustering and weighted AUC-based performance measures.
• Utilized nested cross-validation and nested leave-cluster-out cross-validation for robust evaluation.
• Compared LIBLINEAR against Naïve Bayes, random decision forest, and maximum similarity ranking.

Main Results:

• LIBLINEAR demonstrated superior performance compared to Naïve Bayes, random decision forest, and maximum similarity ranking.
• The method achieved competitive results against literature benchmarks, though nonlinear machines performed better.
• LIBLINEAR showed excellent performance on large-scale problems up to 175,000 samples.
• The approach offers linear complexity in prediction, making it computationally efficient.

Conclusions:

• Adapted LIBLINEAR provides a computationally efficient and effective approach for large-scale cheminformatics classification.
• It outperforms several established methods on large and imbalanced datasets.
• LIBLINEAR is a strong alternative to existing large-scale classification methods, especially when computational time is a concern.