Benchmarking study of parameter variation when using signature fingerprints together with support vector machines

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) modeling is crucial for virtual screening in drug discovery.
• Support Vector Machines (SVMs) with radial basis function (RBF) kernels are commonly used in QSAR.
• Key parameters (C, gamma, signature height) require optimization, which is computationally intensive.

Purpose of the Study:

• To identify optimal default parameter values for QSAR modeling using molecular signatures and SVMs.
• To reduce the computational cost associated with parameter optimization.
• To provide practical recommendations for improving virtual screening efficiency.

Main Methods:

• Utilized seven public QSAR datasets covering diverse endpoints.
• Employed both bit and count versions of molecular signatures.
• Investigated the impact of varying SVM parameters (C, gamma) and signature heights.

Main Results:

• Recommended signature heights: 0-2 for count version, 0-3 for bit version.
• Suggested SVM parameter ranges: C (1-100) and gamma (0.001-0.1).
• Found that minor improvements may occur with extended parameter searches (e.g., height 3 for count) on small datasets.

Conclusions:

• Established effective default parameter ranges for QSAR modeling with molecular signatures and SVMs.
• The recommended parameters can streamline virtual screening processes.
• Further exploration of parameter space is generally not warranted for significant gains.