PREFER: A New Predictive Modeling Framework for Molecular Discovery.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Machine learning and deep learning are vital in cheminformatics for predicting molecular properties and prioritizing compounds.
• Evaluating and comparing different machine learning models and molecular representations is challenging due to diverse frameworks and setups.

Purpose of the Study:

• To introduce PREdictive modeling FramEwoRk for molecular discovery (PREFER), a Python-based framework.
• To facilitate the comparison of various molecular representations and machine learning models.
• To address the need for reproducible and comparable evaluations in cheminformatics.

Main Methods:

• Developed PREFER using Python (3.7.7) and AutoSklearn (0.14.7).
• Designed PREFER to enable direct comparison between different molecular representations and machine learning models.
• Tested framework performance on diverse public and in-house datasets.

Main Results:

• Demonstrated the framework's capability to compare representation-model combinations.
• Showcased exemplary use cases and results on various datasets.
• Provided insights into applying PREFER to small datasets.

Conclusions:

• PREFER offers a standardized approach for evaluating machine learning models in molecular discovery.
• The framework enhances reproducibility and comparability in cheminformatics research.
• PREFER is freely available on GitHub, promoting open science and collaboration.