Cycle-configuration descriptors: a novel graph-theoretic approach to enhancing molecular inference

Area of Science:

• Computational chemistry
• Cheminformatics
• Bioinformatics

Background:

• Molecular inference is crucial for discovering molecules with specific activities and properties.
• Existing frameworks like mol-infer rely on machine learning and mixed-integer linear programming.
• The accuracy of molecular inference is limited by the representational power of chemical descriptors.

Purpose of the Study:

• To address the limitations of the standard two-layered (2L) model in distinguishing chemical graphs with similar structures but different properties.
• To introduce a novel family of cycle-configuration (CC) descriptors capable of capturing nuanced structural patterns, specifically ortho/meta/para relationships in aromatic rings.
• To enhance the accuracy and efficiency of the mol-infer framework for molecular property prediction and inference.

Main Methods:

• Development of a novel family of cycle-configuration (CC) descriptors designed to capture aromatic ring patterns.
• Integration of CC descriptors into the mol-infer framework, formulating them as linear constraints for mixed-integer linear programming.
• Extensive computational experiments on 44 chemical properties (27 regression, 17 classification) to evaluate descriptor performance.

Main Results:

• The new CC descriptors enable the distinction of non-isomorphic chemical graphs that were previously indistinguishable by the 2L model.
• Prediction functions constructed using CC descriptors achieved similar or improved performance across all 44 tested chemical properties compared to previous methods.
• Molecular inference for chemical graphs up to 50 non-hydrogen vertices was demonstrated to be feasible within practical time limits using the CC descriptors.

Conclusions:

• The proposed cycle-configuration (CC) descriptors significantly enhance the distinguishability and predictive power for molecular properties.
• CC descriptors represent a substantial improvement over standard descriptors, overcoming limitations in capturing specific aromatic ring configurations.
• The enhanced mol-infer framework with CC descriptors offers a more accurate and efficient tool for cheminformatics and drug discovery.