Different molecular enumeration influences in deep learning: an example using aqueous solubility.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning

Background:

• Aqueous solubility is a fundamental property influencing chemical and biological processes.
• Accurate solubility prediction is essential but challenging, even with advanced computational methods.
• Molecular representations significantly impact deep learning model performance and interpretability.

Purpose of the Study:

• To review molecular representations for aqueous solubility prediction.
• To propose and evaluate the use of full Simplified Molecular Input Line Entry System (SMILES) enumerations for improved accuracy.
• To integrate attention mechanisms for model explainability.

Main Methods:

• Review of existing molecular representations (fingerprint-based, feature-based, graph-based).
• Implementation of a convolutional neural network (CNN) utilizing full SMILES enumerations.
• Application of attention mechanisms within the decoding network.

Main Results:

• Full SMILES enumerations offer a more comprehensive molecular representation, improving prediction accuracy.
• The CNN model demonstrates robustness with large datasets, requiring no explicit chemistry knowledge.
• Attention mechanisms successfully identify molecular substructures relevant to solubility, enhancing model interpretability.

Conclusions:

• Full SMILES enumerations represent a significant advancement for accurate aqueous solubility prediction using deep learning.
• CNNs combined with attention provide a robust and interpretable approach for solubility modeling.
• This method facilitates understanding the contribution of specific molecular features to solubility.