Virtual screening by a new Clustering-based Weighted Similarity Extreme Learning Machine approach

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Machine learning, particularly Extreme Learning Machine (ELM), is increasingly used in virtual screening.
• Conventional ELM's performance can be limited by random weight selection in the hidden layer, affecting robustness.

Purpose of the Study:

• To propose and evaluate a novel Clustering-based Weighted Similarity Extreme Learning Machine (CWS-ELM) for enhanced virtual screening.
• To address the robustness issues of conventional ELM by introducing deterministic weight assignment methods.

Main Methods:

• Developed Weighted Similarity ELM (WS-ELM) using a single-layer feed-forward neural network with 16 similarity coefficients as activation functions.
• Introduced CWS-ELM by integrating k-means clustering and support vector clustering for deterministic weight assignment.
• Evaluated algorithms on the Maximum Unbiased Validation Dataset, comparing against support vector machine, random forest, and similarity searching.

Main Results:

• CWS-ELM combined with support vector clustering and the Sokal/Sneath(1) coefficient demonstrated superior performance.
• ECFP_6 molecular fingerprints outperformed ECFP_4, FCFP_4, and FCFP_6 within the CWS-ELM framework.
• The proposed CWS-ELM significantly improved virtual screening accuracy compared to other methods.

Conclusions:

• CWS-ELM offers a robust and accurate approach for virtual screening tasks.
• The optimal combination for high performance involves CWS-ELM with support vector clustering, Sokal/Sneath(1) coefficient, and ECFP_6 fingerprints.