DeepSide: A Deep Learning Approach for Drug Side Effect Prediction | JoVE Visualize

Area of Science:

Computational biology
Pharmacology
Machine learning

Background:

Drug development faces significant challenges due to unforeseen adverse effects discovered during clinical trials, leading to participant risks and financial losses.
Predictive models for drug side effects can streamline drug design, but current methods often discard valuable data from large datasets like LINCS L1000.
The LINCS L1000 dataset offers extensive cell line gene expression data perturbed by various compounds, serving as a rich resource for context-specific drug features.

Purpose of the Study:

To improve drug side effect prediction accuracy by leveraging the complete LINCS L1000 dataset, including previously discarded experimental data.
To evaluate the predictive performance of five different deep learning architectures for side effect prediction.
To compare the informativeness of drug chemical structures (CS) versus gene expression profiles (GEX) for predicting drug side effects.

Main Methods:

Experimented with five deep learning architectures, including multi-modal approaches combining chemical structure and gene expression data.
Utilized the full LINCS L1000 dataset, overcoming the limitations of state-of-the-art methods that only use high-quality experiments.
Developed and tested a convolutional neural network (CNN) model using SMILES string representations of drugs.

Main Results:

A multi-modal architecture integrating chemical structure and gene expression data showed the best performance among multi-layer perceptron-based models.
Drug chemical structure (CS) was found to be more informative for prediction than gene expression profiles (GEX).
The CNN model using only SMILES strings achieved state-of-the-art performance, with 13.0% macro-AUC and 3.1% micro-AUC improvements. It also identified novel side effect-drug pairs.

Conclusions:

Deep learning models, particularly CNNs utilizing chemical structure, can significantly enhance the accuracy of drug side effect prediction.
The full utilization of datasets like LINCS L1000, including less curated data, is crucial for advancing predictive modeling in pharmacology.
The developed model, DeepSide, demonstrates the potential to identify known and novel drug-side effect relationships, aiding drug discovery and safety assessment.