DTI-Voodoo: machine learning over interaction networks and ontology-based background knowledge predicts drug-target...

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• In silico drug-target interaction (DTI) prediction is crucial for identifying new drug candidates and repurposing existing drugs.
• Current DTI prediction methods often rely on either indirect phenotypic effects or direct molecular information, with varying degrees of success.
• Integrating diverse data sources and network information can potentially improve the accuracy and scope of DTI predictions.

Purpose of the Study:

• To develop a novel computational method, DTI-Voodoo, for predicting drug-target interactions (DTIs).
• To integrate molecular features, ontology-encoded drug phenotypic effects, and protein-protein interaction networks for enhanced DTI prediction.
• To address and mitigate intrinsic biases present in common DTI datasets that affect performance evaluation.

Main Methods:

• Developed DTI-Voodoo, a computational approach utilizing a graph convolutional neural network (GCN).
• Combined molecular drug features and ontology-encoded phenotypic drug effects with protein-protein interaction networks.
• Implemented a modified evaluation scheme to account for dataset biases and accurately assess prediction performance.

Main Results:

• DTI-Voodoo effectively leverages drug effect features within the interaction network, outperforming methods relying solely on molecular features.
• Analysis revealed significant intrinsic biases in standard DTI datasets, impacting the evaluation of prediction methods.
• The proposed modified evaluation scheme demonstrated that DTI-Voodoo significantly improves upon state-of-the-art DTI prediction methods.

Conclusions:

• DTI-Voodoo offers a robust and improved approach for in silico drug-target interaction prediction.
• The method's ability to integrate diverse data types and account for network information enhances its predictive power.
• Addressing dataset biases is critical for reliable benchmarking and advancement of DTI prediction methodologies.