Molecular Similarity Perception Based on Machine-Learning Models.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• The principle 'similar molecules have similar properties' is fundamental in chemistry.
• Molecular similarity assessments are crucial in pharmaceutical research (structure-activity relationships) and chemical legislation (e.g., orphan drug status).
• Human perception of molecular similarity is subjective, necessitating objective computational models.

Purpose of the Study:

• To develop and evaluate machine-learning models for predicting molecular similarity.
• To create models that can reproduce human expert judgments on molecular similarity.
• To provide tools and datasets for broader accessibility and application in molecular similarity estimation.

Main Methods:

• Utilized 2D molecular fingerprints for similarity calculations.
• Employed 3D molecular descriptors, including shape and pharmacophore features.
• Validated models using a dataset of molecule pairs judged by human experts.

Main Results:

• Developed machine-learning models capable of estimating molecular similarity.
• Demonstrated the models' ability to align with expert human judgments.
• Established a new dataset of molecules and an online tool for similarity estimation.

Conclusions:

• Machine-learning models can effectively estimate molecular similarity, complementing human expertise.
• The developed models offer potential to streamline evaluations in drug discovery and regulatory processes.
• Freely available datasets and tools promote wider adoption and further research in molecular similarity.