DL-SMILES#: A Novel Encoding Scheme for Predicting Compound Protein Affinity Using Deep Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Drug repositioning accelerates development by repurposing safe compounds.
• Machine learning, particularly deep learning, is increasingly used for predicting drug-target interactions.
• Current methods often use simplified molecular representations like SMILES strings.

Purpose of the Study:

• To introduce a novel compound representation, SMILES#, incorporating compound properties.
• To develop a deep learning model for predicting binding affinity using SMILES#.
• To improve the accuracy and efficiency of drug repositioning.

Main Methods:

• Developed SMILES#, a new method for encoding SMILES strings with compound properties.
• Proposed a deep learning architecture combining recurrent neural networks (RNNs), convolutional neural networks (CNNs), and attention mechanisms.
• Utilized both labeled and unlabeled data for joint molecular encoding and binding affinity prediction.

Main Results:

• SMILES# significantly improved model accuracy and reduced Root Mean Square (RMS) error across datasets.
• The deep learning model demonstrated effectiveness in predicting binding affinity.
• Validation confirmed the method's ability to distinguish related and unrelated compounds for the same target.

Conclusions:

• SMILES# offers a superior approach to molecular representation in drug discovery.
• The proposed deep learning model enhances the prediction of drug-target interactions for repositioning.
• This work advances computational methods for accelerating drug development.