Researchers Find Innovative AI in Drug Development

2023 saw a surge in interest in generative AI systems, such as ChatGPT and Midjourney. But in addition to making collages and assisting with email writing, GenAI may also be used to develop novel medications for medical conditions. 

To create novel synthetic medication molecules with the ideal qualities and traits, scientists today employ cutting-edge technology in a process known as “de novo drug design.” But the present approaches can be expensive, time-consuming, and labor-intensive. 

Scientists at Chapman University’s Schmid College of Science and Technology in Orange, California, were inspired by ChatGPT’s widespread use and wondered if this method could expedite the drug design process. As a result, they developed their own GenAI model, which is described in a recent paper titled “De Novo Drug Design using Transformer-based Machine Translation and Reinforcement Learning of Adaptive Monte-Carlo Tree Search,” which was published in the journal Pharmaceuticals. 

In order to understand a vast dataset of known compounds, how they bind to target proteins, and the general principles and syntax of chemical structure and properties, Dony Ang, Cyril Rakovski, and Hagop Atamian programmed a model. 

The final product has the potential to produce an enormous number of distinct molecular structures that adhere to fundamental chemical and biological restrictions and successfully attach to their targets. This might significantly speed up the process of finding effective treatment candidates for a variety of diseases at a far lower expense. 

For the first time in the domains of cheminformatics and bioinformatics, researchers combined two state-of-the-art AI approaches—the renowned “Encoder-Decoder Transformer architecture” and “Reinforcement Learning via Monte Carlo Tree Search” (RL-MCTS)—to produce the ground-breaking model. Users of the platform, appropriately called “drugAI,” can enter a target protein sequence (such as a protein that is commonly involved in the course of cancer). 

DrugAI can create original chemical structures from scratch and then iteratively refine candidates, guaranteeing finalists demonstrate strong binding affinities to respective therapeutic targets—critical for the efficacy of proposed medications. DrugAI was trained on data from the extensive public database BindingDB. The algorithm finds 50–100 novel compounds that are probably going to block these specific proteins. 

“This approach allows us to generate a potential drug that has never been conceived of,” stated Dr. Atamian. It’s been verified and put to the test. We’re now witnessing amazing outcomes.” 

Researchers evaluated the compounds that drugAI produced based on a number of factors, and discovered that the outcomes were on par with, or occasionally even superior to, those from two other popular approaches. The researchers discovered that all of the potential medications produced by drugAI had a 100% validity rate, which means that none of them were included in the training set. 

medicine-likeness, or how closely a chemical resembles the qualities of an oral medicine, was another metric used to evaluate DrugAI’s candidate drugs. The candidate drugs scored at least 42% and 75% higher than previous models. Furthermore, every drugAI-generated molecule showed substantial binding affinities to its target, like those found using conventional virtual screening techniques. 

Additionally, Ang, Rakovski, and Atamian sought to evaluate the efficaciousness of drugAI for a given ailment to that of currently approved medications for that condition. In a different experiment, drugAI produced a list of unique medications that target the same protein in order to compare their properties, based on screening methods that produced a list of natural chemicals that inhibited COVID-19 proteins. When they examined the drug-likeness and binding affinity of the natural compounds and drugAIs, they discovered that the metrics were identical in both cases. However, drugAI was able to detect these far more quickly and affordably. 

Furthermore, the algorithm’s flexible structure was built by the scientists to enable the addition of new functionalities by future researchers. “That means you’re going to end up with more refined drug candidates with an even higher probability of ending up as a real drug,” Dr. Atamian added. “We’re excited for the possibilities moving forward.” 

Journal Reference  

Dony Ang et al, De Novo Drug Design Using Transformer-Based Machine Translation and Reinforcement Learning of an Adaptive Monte Carlo Tree Search, Pharmaceuticals (2024). DOI: 10.3390/ph17020161.