Drug-Target Interactions Prediction at Scale: The Komet Algorithm with the LCIdb Dataset.

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and computational biology
• Drug discovery and development

Background:

• Drug-target interactions (DTIs) prediction is crucial for drug discovery, aiding in target deorphanization and identification.
• Existing DTI prediction methods often struggle with scalability and broad applicability across diverse molecular and protein spaces.
• Developing large, high-quality datasets and efficient, scalable prediction algorithms are key challenges.

Purpose of the Study:

• To address the challenges of building large DTI datasets and developing scalable prediction methods.
• To introduce LCIdb, a comprehensive dataset for DTI prediction, expanding molecule space coverage.
• To propose Komet, a novel, scalable DTI prediction pipeline designed for high performance on large datasets.

Main Methods:

• Creation of LCIdb, a large, curated dataset of DTIs with extensive molecule and protein coverage.
• Development of Komet, a DTI prediction pipeline utilizing a Kronecker interaction module and Nyström approximation.
• Implementation of Komet with efficient computation, GPU parallelization, and quasi-Newton optimization for scalability.

Main Results:

• Komet demonstrates superior scalability and prediction performance compared to state-of-the-art deep learning approaches.
• The pipeline shows strong generalization capabilities on external datasets and scaffold hopping benchmarks.
• LCIdb provides significantly broader molecule space coverage than existing public benchmarks.

Conclusions:

• Komet offers an efficient and high-performing solution for large-scale DTI prediction in drug discovery.
• The developed methods and datasets facilitate advancements in identifying novel drug-target relationships.
• Open-source availability of Komet and datasets promotes further research and application in the field.