GANsDTA: Predicting Drug-Target Binding Affinity Using GANs.

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Predicting drug-target interactions is crucial but challenging due to the high cost of acquiring labeled training data.
• Current supervised methods struggle with limited labeled biomedical data.
• Generative Adversarial Networks (GANs) offer potential for feature learning in complex biological systems.

Purpose of the Study:

• To develop a semi-supervised generative adversarial network (GAN)-based approach for predicting drug-target binding affinity.
• To leverage both labeled and unlabeled data for improved feature extraction in drug-target interaction prediction.
• To address the limitations of data scarcity in supervised machine learning for drug discovery.

Main Methods:

• A novel semi-supervised GAN framework was designed, consisting of two GANs for feature extraction and a regression network for prediction.
• The model utilizes a semi-supervised learning mechanism to learn features from both labeled and unlabeled protein and drug data.
• Performance was evaluated using multiple public datasets for binding affinity prediction.

Main Results:

• The proposed semi-supervised GAN method achieved competitive performance in predicting binding affinity.
• Utilizing freely available unlabeled data significantly improved prediction accuracy.
• The approach demonstrated the utility of unlabeled data in enhancing biomedical relation extraction tasks, including drug-target and protein-protein interactions.

Conclusions:

• This work presents the first semi-supervised GAN-based method for binding affinity prediction.
• Leveraging unlabeled data is a viable strategy to enhance drug-target interaction prediction, particularly with limited labeled datasets.
• The findings suggest broader applicability of this approach to other biomedical relation extraction challenges.