AIM-DDI: Model-Agnostic Multimodal Integration for Drug-Drug Interaction Prediction
AIM-DDI introduces an innovative multimodal integration module that is independent of architecture, aimed at predicting drug-drug interactions (DDIs) while tackling the issue of generalizing to unseen drugs. It encodes diverse modality information as tokens within a unified latent space, facilitating model-agnostic integration of structural, chemical, and semantic signals related to drugs across various DDI prediction frameworks. This module addresses the constraints of current multimodal DDI models, which are often linked to specific architectures, hindering their adaptability. Comprehensive assessments highlight its efficacy across a range of DDI models. This research has been published on arXiv with the identifier 2605.14327.
Key facts
- AIM-DDI is an architecture-independent multimodal integration module for DDI prediction.
- It represents heterogeneous modality information as tokens in a shared latent space.
- It enables model-agnostic integration of structural, chemical, and semantic drug signals.
- The module addresses unseen-drug generalization for drugs not observed during training.
- Existing multimodal DDI models have fusion mechanisms tied to specific architectures.
- Extensive evaluations were conducted across diverse DDI prediction architectures.
- The research is published on arXiv with ID 2605.14327.
- The paper is categorized as a cross-type announcement.
Entities
Institutions
- arXiv