Drug-Target Binding Affinity Prediction in a Continuous Latent Space Using Variational Autoencoders.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Accurate drug-target binding affinity (DTA) prediction is crucial for efficient drug discovery.
• Existing deep learning models for DTA often represent molecules and proteins in vector spaces.
• There is a need for more nuanced modeling approaches to capture input sample diversity.

Purpose of the Study:

• To propose a novel deep learning model for DTA prediction that operates in a continuous space.
• To enhance the accuracy and efficiency of drug discovery processes through improved DTA prediction.
• To jointly learn hidden representations of drugs and targets using Gaussian distributions.

Main Methods:

• Drug encoding using Simplified Molecular Input Line Entry System (SMILES).
• Target sequence characterization via a pretrained language model.
• Extraction of correlative information using residual gated convolutional neural networks.
• Joint learning of drug and target hidden representations as Gaussian distributions.

Main Results:

• The proposed continuous space model demonstrated superior performance compared to state-of-the-art vectorial representation methods.
• Experimental evaluations on benchmark datasets validated the effectiveness of the Gaussian distribution approach.
• The method achieved higher precision in DTA predictions.

Conclusions:

• Modeling drug-target interactions in continuous space offers advantages over traditional vector-based methods.
• The joint learning of drug and target representations as Gaussian distributions improves DTA prediction accuracy.
• This approach has the potential to significantly contribute to more efficient and precise drug discovery pipelines.