jCompoundMapper: An open source Java library and command-line tool for chemical fingerprints.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Chemical graph decomposition is crucial for encoding organic compound information.
• Limited open-source toolkits exist for molecular fingerprint generation.
• Need for flexible tools for machine learning and data mining applications.

Purpose of the Study:

• Introduce jCompoundMapper, an open-source Java library for chemical graph fingerprints.
• Provide customizable options for molecular decomposition and feature generation.
• Facilitate applications in machine learning and data mining.

Main Methods:

• Developed a Java 1.6 library utilizing the Chemistry Development Kit.
• Reimplemented and introduced various fingerprinting algorithms (e.g., ECFP, CATS2D, Molprint2D, atom pairs, pharmacophores).
• Included custom fingerprints like the all-shortest path fingerprint.
• Provided a command-line executable binary for practical application.
• Evaluated conversion speed, feature composition, and data mining performance.

Main Results:

• jCompoundMapper successfully reimplemented popular fingerprinting algorithms.
• Achieved competitive performance on QSAR (Sutherland datasets) and toxicity prediction (Ames benchmark) benchmarks.
• Demonstrated AUC ROC of 0.87 on the Ames dataset using extended connectivity fingerprints.
• Outperformed or matched literature results on several Sutherland QSAR benchmarks.

Conclusions:

• jCompoundMapper is a versatile library for chemical graph fingerprints with extensive customization and export options.
• The library's performance, speed, and open-source nature (LGPL) are valuable for cheminformatics applications.
• Facilitates tasks such as benchmarking, algorithm comparison, and similarity searching.