A machine learning algorithm can predict whether people with major depression will respond to a common antidepressant after only 1 week.
Credit: Yuuji/Getty Images

Researchers based at the Gwangju Institute of Science and Technology in the Republic of Korea have developed a deep learning artificial intelligence model to predict how likely drugs are to interact with other drugs and produce adverse side effects.

In current medicine it is very common for patients, particularly older individuals, to be taking multiple medicines at the same time. “This can be problematic because drug interactions can alter the intended responses,” write Eunyoung Kim and Hojung Nam, a PhD student and senior researcher at the Gwangju Institute of Science and Technology, in their Journal of Cheminformatics article describing the study.

Drug-drug interactions can have a variety of effects ranging from decreased clinical efficacy of the drugs involved to serious adverse events. It is therefore important to identify if these effects might occur when developing new therapeutics.

These interactions have mostly been found through experimentation and clinical records in the past. “However, in vitro and in vivo identification of drug-drug interactions are largely infeasible, owing to patient safety and ethical considerations that increase time and costs,” write Kim and Nam. “Furthermore, major polypharmacy side effects are difficult to identify from small trials and cohorts.”

The researchers decided to apply deep learning to the problem, as it has previously shown promise in this area. Deep learning is a type of machine learning using artificial neural networks with three or more layers to simulate the behavior of the human brain.

Most previous attempts to understand and predict drug-drug interactions have focused on drug compound interactions rather than looking at how they interact with the body via genes. In this study, the model was designed to predict drug interactions based on drug-induced gene expression.

Their model works in two parts, firstly predicting a drug’s effect on gene expression by considering many factors including pharmacological structure and properties. The second part assesses how a given drug will interact with other drugs if given at the same time.

The model was trained using the entire LINCS L1000 dataset, a resource linking genes, drugs, and disease states through common gene-expression signatures. They also used two datasets on drug-drug interaction: TWOSIDES and DrugBank.

Initial accuracy of the model was good, although the researchers say more work is needed before it can be applied more widely. “First, the datasets in this case were very sparse in terms of side-effect type. The small proportion of interactions between drugs were known, but the numbers were extremely sparse and imbalanced when considered with relation type. For this reason, rare side effects were excluded for the current model.”

They add that identifying side effect mechanisms remains challenging and say more information on drug-linked gene expression is needed to boost accuracy, as current datasets remain limited.

Also of Interest