DeepSide

DeepSide predicts potential drug side effects by applying deep learning to chemical structure (SMILES) and LINCS L1000 gene expression profiles to improve adverse-event prediction during drug development.

Key Features:

• Utilization of LINCS L1000 Dataset: Uses LINCS L1000 gene expression data from drug-perturbed cell lines as context-specific features for prediction.
• Multi-modal Deep Learning Architectures: Implements and evaluates five deep learning architectures, including a multi-modal model that integrates chemical structure (CS) and gene expression (GEX).
• SMILES-based Convolutional Neural Network: Employs a convolutional neural network trained on SMILES string representations of drugs, reflecting that chemical structure is more informative than gene expression for this task.
• Multi-layer Perceptron Models: Includes multi-layer perceptron (MLP)-based models among the evaluated architectures.
• Performance Improvements: Reports improvements of 13.0% macro-AUC and 3.1% micro-AUC over prior state-of-the-art methods.
• Evaluation Metrics: Assesses predictive performance using AUC metrics (macro-AUC and micro-AUC).

Scientific Applications:

• Drug design guidance: Provides early predictions of potential adverse effects to inform candidate selection in drug design.
• Clinical safety assessment: Supports safety profiling during clinical development by predicting side effects that may lead to trial failures.
• Recovery of missing associations: Identifies side effect-drug pairs documented in the literature but absent from ground-truth datasets.

Methodology:

Combines chemical structure (CS) and gene expression profiles (GEX) from the LINCS L1000 dataset; implements and compares five deep learning architectures including multi-layer perceptrons and convolutional neural networks trained on SMILES representations; evaluates models using macro-AUC and micro-AUC metrics.