Evaluating parameters for ligand-based modeling with random forest on sparse data sets.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Ligand-based predictive modeling is crucial for drug discovery decision-making.
• Increasing dataset sizes necessitate efficient data analysis for rapid and robust modeling.

Purpose of the Study:

• To evaluate the efficiency of machine learning methods on sparse data structures.
• To compare the impact of Morgan fingerprints (radii, hash sizes) and molecular signatures on modeling time, predictive performance, and memory usage.
• To assess Scikit-learn and FEST implementations of random forest, alongside a support vector machine.

Main Methods:

• Analysis of four datasets using Morgan fingerprints (varying radii and hash sizes) and molecular signatures.
• Comparison of modeling time, predictive performance, and memory requirements.
• Utilized Scikit-learn and FEST implementations of random forest, and a support vector machine.

Main Results:

• Unhashed fingerprints significantly outperformed hashed fingerprints in accuracy, with comparable modeling time and memory usage.
• The FEST algorithm demonstrated fast execution and low memory usage, suitable for large, high-dimensional sparse data.
• Support vector machines and random forests performed comparably, with support vector machines better utilizing information from larger Morgan fingerprint radii.

Conclusions:

• Unhashed Morgan fingerprints are recommended for improved accuracy in ligand-based predictive modeling.
• The FEST algorithm is a viable and efficient option for handling large, sparse chemical datasets.
• Both random forest and support vector machines are effective, with nuances in descriptor utilization.