Block-wise Exploration of Molecular Descriptors with Multi-block Orthogonal Component Analysis (MOCA)

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Machine learning and Quantitative Structure-Activity Relationship (QSAR) models often utilize data organized in blocks, representing diverse molecular features or descriptors.
• Exploring relationships within and between these data blocks is crucial for effective modeling.

Purpose of the Study:

• To introduce and apply Multi-block Orthogonal Component Analysis (MOCA) as a novel tool for analyzing multi-block data structures.
• To identify principal components unique to single blocks or shared across multiple blocks.
• To develop a quantitative metric for assessing block redundancy and evaluate predictive modeling potential.

Main Methods:

• Application of MOCA to two distinct datasets comprising 550 and 300 molecules with up to 9213 molecular descriptors across 11 blocks.
• Analysis of MOCA models to reveal inter-block relationships and overarching dataset trends.
• Utilizing MOCA joint components to derive a quantitative redundancy metric.
• Application to a dataset of 7 ecotoxicological endpoints for crop protection chemicals.

Main Results:

• MOCA models successfully identified relationships between data blocks and global trends within the datasets.
• A quantitative metric for block redundancy was proposed, useful for feature selection and evaluating molecular representations.
• General trends in ecotoxicological data were rediscovered and linked to specific molecular properties.
• The predictive potential of individual blocks and the overall modelability of target blocks were estimated using MOCA.

Conclusions:

• MOCA is an effective analytical tool for exploring complex, multi-block chemical data, revealing both unique and shared information.
• The developed redundancy metric provides a valuable approach for optimizing feature selection and assessing data block relevance in QSAR and machine learning.
• MOCA facilitates the linkage of molecular properties to biological endpoints, enhancing predictive modeling capabilities in areas like ecotoxicology.