A Multimodal Deep Learning Framework for Predicting PPI-Modulator Interactions.

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Protein-protein interactions (PPIs) are crucial for biological functions and disease development.
• Current computational methods for identifying PPI modulators often require target structures or known modulators, limiting their use for novel targets.
• There is a need for versatile computational tools to predict PPI modulators applicable to new targets.

Purpose of the Study:

• To develop a novel sequence-based deep learning framework, MultiPPIMI, for predicting interactions between protein targets and modulators.
• To address the limitations of existing methods in identifying modulators for novel PPI targets.
• To provide a computational tool that can assist in the screening of potential therapeutic agents targeting PPIs.

Main Methods:

• Developed MultiPPIMI, a sequence-based deep learning framework.
• Integrated multimodal representations of PPI targets and modulators.
• Employed a bilinear attention network to capture intermolecular interactions.
• Evaluated performance on a curated benchmark dataset using AUROC metrics in cold-start and random-split scenarios.

Main Results:

• MultiPPIMI achieved an average AUROC of 0.837 in three cold-start scenarios, demonstrating effectiveness for novel targets.
• Achieved an AUROC of 0.994 in the random-split scenario, indicating high accuracy.
• A case study demonstrated MultiPPIMI's utility in screening inhibitors for Keap1/Nrf2 PPI interactions, aiding molecular docking simulations.

Conclusions:

• MultiPPIMI offers a promising sequence-based deep learning approach for predicting PPI modulators.
• The framework effectively handles novel PPI targets without requiring structural information or reference modulators.
• MultiPPIMI can serve as a valuable tool to accelerate the discovery and screening of modulators for therapeutic PPIs.