DeepFusion: A deep learning based multi-scale feature fusion method for predicting drug-target interactions.

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Accurate prediction of drug-target interactions (DTIs) is crucial for drug discovery and repositioning.
• Existing deep learning methods often require extensive approved DTI data, posing a significant challenge.
• The need for efficient DTI prediction models with reduced data dependency is evident.

Purpose of the Study:

• To introduce DeepFusion, a novel deep learning-based multi-scale feature fusion method for DTI prediction.
• To address the data scarcity issue in training DTI prediction models.
• To enhance the accuracy and efficiency of predicting drug-target interactions.

Main Methods:

• Developed DeepFusion, a deep learning model integrating multi-scale features for DTI prediction.
• Generated global structural similarity features using convolutional neural networks.
• Extracted local chemical sub-structure semantic features via transformer networks for drugs and proteins.

Main Results:

• DeepFusion demonstrated high performance on BIOSNAP dataset subsets (70% yielding ROC-AUC of 0.877 and PR-AUC of 0.888).
• Achieved comparable results to baseline methods trained on the entire dataset, using only a fraction of the data.
• Case studies showed promising predictive power for potential DTIs.

Conclusions:

• DeepFusion offers an effective solution for DTI prediction, overcoming data limitations.
• The multi-scale feature fusion approach enhances prediction accuracy and efficiency.
• This method holds significant potential for accelerating drug discovery and repositioning efforts.