Overfit deep neural network for predicting drug-target interactions.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Drug-target interactions (DTIs) prediction is crucial for efficient drug discovery.
• Traditional experimental methods for DTI identification are costly and time-consuming.
• Deep learning models offer a promising alternative but face challenges like overfitting.

Purpose of the Study:

• To introduce OverfitDTI, a novel framework for DTI prediction that intentionally overfits deep neural networks (DNNs).
• To leverage the implicit representation of drug-target relationships learned by overfit DNNs.
• To demonstrate the efficacy of OverfitDTI in accurately predicting DTIs.

Main Methods:

• A deep neural network (DNN) model is intentionally overfit to learn features from both drug chemical spaces and target biological spaces.
• The learned weights of the overfit DNN serve as an implicit representation of nonlinear drug-target relationships.
• The OverfitDTI framework was evaluated on three publicly available datasets.

Main Results:

• Overfit DNN models demonstrated high accuracy in fitting nonlinear drug-target relationships.
• The framework successfully identified fifteen compounds predicted to interact with TEK (TIE-2).
• Experimental validation confirmed AT9283 and dorsomorphin as TEK inhibitors in HUVECs.

Conclusions:

• Overfitting DNNs can be an effective strategy for capturing complex nonlinear relationships in DTI prediction.
• OverfitDTI provides a computationally efficient and accurate method for accelerating drug discovery.
• The study validates the potential of computational approaches in identifying novel therapeutic agents.