IBM unveils new AI model to predict potentially harmful drug-to-drug interactions

The work is being highlighted at this week's AAAI-20 Conference in New York City.

Medical technology concept. Electronic medical record.

Image: iStockphoto/metamorworks

IBM this week unveiled new research on how an artificial intelligence (AI) model can help better predict drug-to-drug interactions (DDIs).

Researchers from the MIT-IBM Watson AI Lab, Harvard School of Public Health, Georgia Institute of Technology, and IQVIA have created a new AI tool called CASTER that they claim can more accurately predict potentially harmful and unsafe adverse interactions for drugs in the market, as well as ones in the early development phase. 
 
The findings were released at the AAAI-20 Conference taking place all week in New York City. CASTER stands for chemical substructure representation.

"Think of [CASTER] as an AI copilot that helps flag potential drug interactions doctors need to be careful of," said David Cox, director of the MIT-IBM Watson AI Lab. "The potential problem when you have great therapies is they'll interact in ways you don't expect. You can have an adverse reaction that could be dangerous."

SEE: Telemedicine, AI, and deep learning are revolutionizing healthcare (TechRepublic download)

This is made especially difficult because new drugs are constantly coming into the market, Cox said.  

Every year, more than one million people in the US are hospitalized as a result of adverse drug events. Unlike current methods to check drug-to-drug interactions, this new AI tool develops a specialized drug representation to predict the likelihood of adverse reactions between drugs based on their frequent chemical substructures, IBM said.

There are databases of known drug interactions and the system is trained on what they are and then is asked to predict others, Cox said. "The hope is as new drugs become available you can put [them] in the system." The CASTER tool provides an early heads up of what might be a problem, he said.

CASTER was shown to help achieve higher accuracy than previous methods in predicting DDI, IBM said.

The work is just one example of how AI can be used on new kinds of data like chemical structures--and not just pictures and audio and other areas where it has traditionally been used to make decisions, Cox said. The researchers are also using AI to suggest potential new drugs with therapeutic benefits, so it can discover new materials, such as new molecules and properties in existing molecules, he said.

The researchers tested the model on two common drug databases, DrugBank and BioSNAP, and it performed better than state-of-the-art results from existing AI systems, according to IBM. In a paper highlighting their work, the researchers examined the predictions the framework made between the known interaction between sildenafil, an effective treatment for erectile dysfunction and pulmonary hypertension, and nitrate-based drugs.

Unlike previous methods, which take into account only a few substructures of a drug's molecular structure, the predictive analysis capabilities in CASTER focus on what's important and ignores what's not, IBM said.

Researchers were motivated to "devise a specialized representation that automatically allows the predictive learning to focus only on the most relevant functional substructures, which are more likely to be responsible for the interaction," IBM said.

"We demonstrated empirically that CASTER can provide more accurate and interpretable DDI predictions than the previous approaches that use generic drug representations,'' the researchers wrote. "For future works, we plan to extend it to chemical sub-graph embedding and incorporate metric learning for further improvement."

"This is an early glimpse of what is possible with technology, and you can expect [CASTER] to find its way into products down the road,'' Cox said.

Also see