The use of artificial intelligence (AI) has been increasing in scientific research and clinical applications. From determining best treatment options for cancer patients, to identifying novel protein-ligand interactions, and in drug discovery, AI has been revolutionary.
Much of this work has been retrospective, where significant and specific data with known interactions are fed into the AI software to answer specific questions. However, prospective studies utilize AI as a predictive tool. Bryan Roth, MD, PhD, professor at the University of North Carolina School of Medicine, and his colleagues at UCSF, Stanford, and Harvard aimed to test the prospective functionality of the AI program, AlphaFold2.
The team found that AlphaFold2 was effective at producing 3D models of proteins as well as predicting ligand binding ability. “Our results suggest that AlphaFold2 structures can be useful for drug discovery,” shared Roth.
Their recently published study, “AlphaFold2 structures template ligand discovery,” was published in Science. The authors wrote, “AlphaFold2 models may sample conformations that are relevant for ligand discovery, much extending the domain of applicability of structure-based ligand discovery.”
AlphaFold2 is an AI system that predicts the 3D structures of proteins and can be used to help identify a multitude of drug candidates. “AlphaFold2 [has] greatly expanded the number of structures available for structure-based ligand discovery, even though retrospective studies have cast doubt on their direct usefulness for that goal,” the authors wrote.
Their work aimed at determining if the system could be used not only to determine the likelihood of binding interaction of potential drugs to known ligand binding sites, but if the system can determine ligand binding sites in an unknown protein and predict interactions with possible ligands (and potential drug candidates). The authors wrote, “It is possible that the relatively poor performance of AlphaFold2 models in retrospective simulation of structure-based ligand discovery may underestimate the ability of AlphaFoldF2 structures to template new ligand discovery prospectively.”
Two proteins, sigma-2 and 5-HT2A, have not been previously used for AlphaFold2 training. Both proteins are useful in cell communication and may be targets for treating neurological conditions including Alzheimer’s disease and schizophrenia. Thus, using these proteins in this study is both useful to test the functionality of the AI and may well also provide actionable data for drug candidate studies to treat these and other neurological conditions.
The team used AlphaFold2 first in a prospective study using both proteins. The AI system presented 3D structures with a variety of possible binding sites for both proteins that the researchers confirmed experimentally with microscopy and x-ray crystallography, finding that the sites were fairly well predicted.
Using the experimental data fed into AlphaFold2, they were then able to conduct a second retrospective study on ligand binding compared to the fully prospective study they began with.
Both prospective and retrospective models were used to predict as many as 1.6 billion drug candidates based on the experimental and AI models. “Our results suggest that AlphaFold2 structures may be more relevant to prospective structure-based ligand discovery than thought based on retrospective simulation,” wrote the authors.
Though the models had some differences in their output, there weren’t significant differences in success between the models. For the sigma-2 protein, the prospective study resulted in 54% of the protein-ligand interactions that were successful, compared to 51% in the retrospective model. Similarly, in 5-HT2A, the successful combinations were 26% in the prospective model versus 23% in the retrospective model.
“With a nearly infinite number of possibilities to create drugs that hit their intended target to treat a disease, this sort of AI tool can be invaluable,” concluded Roth. “Going forward we will test whether these results might be applicable to other therapeutic targets and target classes.”
