Joint fusion of sequences and structures of drugs and targets for identifying targets based on intra and inter cross-attention mechanisms.

Abstract

Background: Accurately identifying targets not only guides treatments for diseases with unclear pathogenic mechanisms, but also reduces pharmaceutical costs and accelerates drug development timelines. However, the primary challenge in targets identification currently lies in the low accuracy of existing computational methods.

Results: We propose MM-IDTarget, a novel deep learning framework that employ a multimodal fusion strategy based on the intra and inter cross-attention mechanisms. MM-IDTarget integrates some cutting-edge deep learning techniques such as graph transformer, multi-scale convolutional neural networks (MCNN), and residual edge-weighted graph convolutional network (EW-GCN) to extract sequence and structure modal features of drugs and targets. This framework enhances the complementary of multimodal features by employing the intra and inter cross-attention mechanisms, facilitating effective fusion of multimodal features within drug and target and between drug and target. Furthermore, MM-IDTarget incorporates the physicochemical features of drug and target, utilizing fully connected networks to predict drug-target interactions (DTI).

Conclusions: Experimental results show that despite our benchmark dataset being one-third or the same size of those used by current state-of-the-art methods, MM-IDTarget achieves the performance on par with or superior to these methods across most Top-K evaluation metrics based on the same test set for targets identification. Moreover, MM-IDTarget exhibits the strong application capability on two generalization datasets and one dataset constructed from approved drugs, establishing it as a robust tool for targets identification.