Komet

Komet predicts drug-target interactions (DTIs) through a scalable three-step framework that leverages Kronecker product structure and quasi-Newton optimization to enable large-scale training without compromising performance.

Key Features:

• Kronecker interaction module: Captures determinants of DTIs for (molecule, protein) pairs with reduced computational complexity.
• Nyström approximation: Enables efficient computation tailored for large datasets.
• Quasi-Newton optimization: Facilitates scalable model training on large datasets.
• GPU parallel computation: Accelerates computational efficiency during training and inference.
• Integration with LCIdb: Leverages a curated, large-scale DTI dataset with extensive coverage of molecule and druggable protein spaces.

Scientific Applications:

• Deorphanization of therapeutic targets: Enables large-scale prediction of interactions for newly identified targets across the protein space.
• Target identification from phenotypic screens: Supports identification of protein targets for drug candidates arising from phenotypic screening.
• Scaffold hopping: Demonstrated performance on the LH benchmark for scaffold hopping problems.
• Generalization assessment: Evaluated on external datasets to validate broad applicability across sparse regions of chemical and protein space.

Methodology:

Komet employs a three-step framework incorporating a Kronecker interaction module for (molecule, protein) pairs, Nyström approximation for efficient computation on large datasets, and quasi-Newton optimization, with GPU parallel computation for efficiency.