Artificial intelligence models may help develop drugs that can help people overcome opioid use disorder, according to findings presented at the Biophysical Society Annual Meeting.

“Artificial intelligence has the advantage of being able to take huge amounts of information and learn to recognize patterns from it,” Leslie Salas-Estrada, PhD, a postdoctoral fellow in the department of pharmacological sciences at Icahn School of Medicine at Mount Sinai, said in a press release. “So, we believe that machine learning can help us to leverage the information that can be derived from large chemical databases to design new drugs from scratch. And in that way, we can potentially reduce the time and costs associated with drug discovery.”

Healio spoke with Salas-Estrada to learn more about her development of an artificial intelligence strategy that identifies compounds that may treat opioid use disorder.

Healio: What prompted this research?

Salas-Estrada: There is an unmet medical need to develop therapies that can treat individuals whose repeated exposure to opioids for chronic pain or recreational purposes has caused them to develop brain abnormalities, leading to a chronic disease known as opioid use disorder (OUD) or addiction. I work in a lab that studies opioid receptors, which are the main protein targets for clinically used opioid drugs, among others belonging to the family of G protein-coupled receptors (GPCRs). Opioid abuse can also have other concerning side effects, such as respiratory depression, which is responsible for the majority of opioid overdose deaths. It is known that approximately 1,000 people die from opioid overdoses in the United States every week, and that the number of deaths has increased substantially during the COVID-19 pandemic. Thus, the U.S. is facing a major public health issue which has serious costs to society due to increased health care spending, lost productivity and criminal justice issues.

One member of the opioid receptor subfamily, the kappa-opioid receptor, is known to play a role in reward processing and responses to stress, which are key elements in addiction. Inhibition of kappa-opioid receptor activity has shown promise in the treatment of opioid dependence in preclinical studies. However, there are very few known kappa-opioid receptor inhibitors and most of them present concerning safety issues, such as persistent pharmacological effects that can result in toxicity.

My research goal was to design a kappa-opioid receptor inhibitor with favorable drug-like properties, taking advantage of the large amount of structural and physicochemical information available about opioid receptors, opioid ligands and drug-like molecules. As new methods are needed to leverage this enormous amount of information more efficiently, I implemented an artificial intelligence (AI)-driven strategy to design putative optimal inhibitors of the kappa opioid receptor from scratch. The first step in my strategy was to teach the computer basic chemistry. I trained a computational model with information from millions of chemical compounds, such as what a drug-like molecule looks like, so that it could learn to recognize patterns in the data and create new compounds that were chemically sound. At the beginning, the model would produce very simple molecules or things that did not make sense. However, as it received more and more information, it started to improve. Once the model learned how to create compounds, I used a type of training technique called reinforcement learning to teach it how to create compounds with characteristics that are specific to kappa-opioid receptor inhibitors. This training involved positively rewarding the model for creating molecules with these characteristics, or giving it a penalty for those without them. In the end, the resulting model was able to produce a set of molecules predicted to inhibit the kappa-opioid receptor. Our collaborators are now helping us synthesize and experimentally test these molecules. So far, three compounds have been synthesized and they were able to bind the receptor in vitro. One of these compounds showed inhibitory effects in cells.

Healio: What is the take–home message?

Salas-Estrada: AI and other machine learning-based tools can be used to good effect in creating novel inhibitors of the kappa opioid receptor, using both structural information about the receptor and physicochemical information about known molecules. These tools hold great potential in the development of drugs to treat opioid addiction in the future. Their potential will continue to improve as new structural and functional information on opioid receptors becomes available.

Healio: What are the clinical implications?

Salas-Estrada: The compounds that we have identified with our trained models need to be further optimized and tested in cells before they can be used in preclinical trials with animal models. If they prove to be safe and effective in these tests, then they can be tested in a clinical setting.

Healio: What research on this topic would you like to conduct/see next?

Salas-Estrada: I would like to establish a broader impact of the tool I implemented by extending its application to other GPCRs and pharmacological targets. Additionally, I would like to improve its yield of valid molecules by testing the use of different physicochemical properties and structural evaluations for training.