Topological Distance-Based Electron Interaction Tensor to Apply a Convolutional Neural Network on Drug-like Compounds

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Deep learning (DL) models in quantitative structure-activity relationship (QSAR) traditionally use simplified molecular representations like graphs or SMILES codes.
• These representations often oversimplify complex chemical properties, limiting the application of diverse DL architectures.
• A novel molecular representation is needed to unlock the potential of advanced DL models in drug discovery and chemical property prediction.

Purpose of the Study:

• To develop a novel molecular representation that captures richer chemical information for DL applications.
• To apply this new representation to predict key physicochemical properties of drug-like compounds.
• To explore the effectiveness of convolutional neural network (CNN) models with this novel representation.

Main Methods:

• Developed a topological distance-based electron interaction (TDEi) tensor, inspired by quantum mechanics.
• Represented atomic orbitals as electron configuration (EC) vectors (bit strings) indicating electron presence and spin.
• Calculated inter-EC vector interactions based on topological distances within the molecule.
• Applied modified VGGNet CNN models to the TDEi tensor (3D array) for property prediction.

Main Results:

• Successfully predicted four physicochemical properties (melting point, lipophilicity, water solubility) using the TDEi tensor and CNN models.
• Achieved good prediction accuracy across multiple drug-like compound datasets.
• Principal Component Analysis (PCA) indicated stronger feature-endpoint correlations from deeper CNN layers, suggesting hierarchical feature extraction.

Conclusions:

• The TDEi tensor offers a powerful, chemically informative molecular representation for DL.
• This approach enables the application of advanced CNN architectures for accurate physicochemical property prediction.
• The findings pave the way for more sophisticated DL models in cheminformatics and drug design.