CPSign: conformal prediction for cheminformatics modeling.

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Conformal prediction calibrates machine learning models, providing valid prediction intervals crucial for pharmaceutical science.
• Existing methods often lack comprehensive tools for direct chemical structure analysis and prediction.

Purpose of the Study:

• Introduce CPSign, an open-source software for conformal prediction in cheminformatics.
• Enable users to perform data preprocessing, modeling, and predictions directly on chemical structures.
• Evaluate CPSign's performance against contemporary modeling approaches.

Main Methods:

• Implemented inductive and transductive conformal prediction for classification and regression.
• Utilized Venn-ABERS methodology for probabilistic prediction.
• Supported chemical signatures and other descriptors, with Support Vector Machines (SVM) as the primary modeling method, extensible to other models like DeepLearning4J.
• Included features for result visualization and publishing models as REST services.

Main Results:

• CPSign demonstrated robust predictive performance and efficiency comparable to other methods.
• Outperformed neural network-based models in runtime and hardware requirements.
• Showcased on par performance with state-of-the-art deep learning models in evaluations.
• Validated through use in several studies and production environments.

Conclusions:

• CPSign offers a convenient, flexible, and efficient software package for cheminformatics modeling.
• Its ability to handle chemical input, descriptor calculation, and SVM modeling within a conformal prediction framework is a significant advantage.
• The software provides a high level of abstraction for model building and evaluation without compromising flexibility or performance.