DDintensity: Addressing imbalanced drug-drug interaction risk levels using pre-trained deep learning model embeddings

Area of Science:

• Bioinformatics
• Computational Biology
• Pharmacogenomics

Background:

• Imbalanced datasets pose a significant challenge in bioinformatics, particularly for predicting drug-drug interaction (DDI) risk levels.
• Biased models resulting from imbalanced data lead to poor performance on underrepresented classes, hindering accurate DDI risk assessment.

Purpose of the Study:

• To introduce DDIntensity, a novel approach for handling imbalanced DDI risk level datasets.
• To leverage pre-trained deep learning embeddings and LSTM-attention models to improve DDI prediction accuracy.

Main Methods:

• Utilized pre-trained deep learning models (including BioGPT) as embedding generators.
• Integrated embeddings from diverse data types (images, graphs, text) with LSTM-attention networks.
• Trained and validated the DDIntensity model on DDinter and MecDDI datasets.

Main Results:

• BioGPT embeddings yielded superior performance, achieving an Area Under the Curve (AUC) of 0.97 and an Area Under the Precision-Recall curve (AUPR) of 0.92.
• Demonstrated high scalability across different DDI data modalities.
• Successfully identified novel drug-drug interactions through case studies on oncology drugs (Sorafenib, Mitoxantrone).

Conclusions:

• DDIntensity offers a robust solution for imbalanced bioinformatics datasets, specifically in DDI risk prediction.
• The approach enhances model performance and facilitates the discovery of new DDI.
• Pre-trained deep learning embeddings are crucial for improving DDI risk level classification.