MSDSE: Predicting drug-side effects based on multi-scale features and deep multi-structure neural network.

Area of Science:

• Pharmacology
• Computational Chemistry
• Artificial Intelligence

Background:

• Drug development faces challenges from unexpected side effects, impacting patient safety and leading to research failures.
• Current in silico methods often overlook intrinsic drug attributes, limiting their effectiveness to later drug development stages.

Purpose of the Study:

• To develop a novel deep learning framework, MSDSE, for early and accurate prediction of drug side effects.
• To address the limitations of existing methods by incorporating multi-scale, intrinsic drug features.

Main Methods:

• MSDSE employs a multi-structural deep learning approach, integrating SMILES sequence embeddings, molecular fingerprints, and graph embeddings.
• Features are projected into a common abstract space and processed through a bi-level channel strategy using CNNs with Inception and multi-head Self-Attention modules.
• Drug side effect prediction is framed as pair-wise learning, with probabilities outputted via inner product operations.

Main Results:

• MSDSE demonstrated optimal performance compared to baseline models on benchmark datasets.
• Ablation studies validated the effectiveness of the chosen feature representations and model architecture.
• Case studies confirmed the practical capability of MSDSE in predicting drug side effects.

Conclusions:

• MSDSE offers a robust framework for early drug side effect screening by leveraging multi-modal intrinsic drug features.
• The model's ability to integrate diverse data types enhances prediction accuracy and supports safer drug development.
• The findings highlight the potential of advanced deep learning for identifying potential adverse drug reactions proactively.