DTI-BERT: Identifying Drug-Target Interactions in Cellular Networking Based on BERT and Deep Learning Method.

Area of Science:

• Genomics
• Computational Biology
• Drug Discovery

Background:

• Drug-target interactions (DTIs) are crucial for genomic drug discovery.
• Existing computational methods for DTI prediction have limitations.
• Wet-lab techniques for DTI identification are time-consuming and expensive.

Purpose of the Study:

• To develop a novel, sequence-based computational method for accurate DTI identification.
• To improve upon existing DTI prediction techniques by integrating advanced feature extraction methods.
• To accelerate the identification of lead drugs for specific protein targets.

Main Methods:

• Utilized pre-trained Bidirectional Encoder Representations from Transformers (BERT) for protein sequence feature extraction.
• Employed Discrete Wavelet Transform (DWT) for generating features from drug molecular fingerprints.
• Integrated feature vectors through a BRL block and Convolutional Neural Networks (CNNs) for further feature extraction.
• Optimized the prediction model using contrastive loss and cross-entropy loss.

Main Results:

• Achieved high prediction accuracies for target families: G Protein-coupled receptors (90.1%), ion channels (94.7%), enzymes (94.9%), and nuclear receptors (89%).
• Demonstrated superior performance compared to existing DTI predictors.
• Developed a freely accessible web server for the new predictor.

Conclusions:

• The proposed sequence-based method significantly enhances the accuracy of DTI prediction.
• This approach offers a valuable tool for accelerating drug discovery and development.
• The method shows potential applicability to other DTI prediction tasks.