How do scientists discover drug therapies for diseases such as COVID-19? Many researchers identify potential therapies based on molecular binding affinity—the strength of pharmaceutical binding—to a protein that’s pertinent to disease progression. Historically, this has involved extensive lab work: experimental examinations of the binding affinity of each molecule.
Lab work is slow and expensive. There are many possible molecules to test. Such work can’t be avoided entirely. Nevertheless, researchers need methods to quickly determine which pharmaceutical candidates are the most promising. Otherwise, the job will basically take forever.
Computational modeling can help. This is a modern innovation that helps speed up drug discovery. Here, you predict—rather than experimentally determine—the binding of each molecule. This is much faster than lab work. Once you have a small number of pharmaceutical candidates, screened from a large body of possible molecules, you can reduce the required lab work.
How does computational modeling work? It can be either structure-based or ligand-based. Structure-based drug design predicts binding affinity based on the three-dimensional structure of the protein. This data may not be available. Ligand-based drug design predicts binding affinity based on the chemical characteristics of molecules that are known to bind to the protein.
As you may have expected, identifying possible drug molecules can’t be that easy. You’re right; it’s not. The results from computational modeling may be incorrect. Why?
One reason is that proteins are flexible. For example, upon binding a drug molecule, the protein may change its shape at the binding site. This protein shape change almost certainly affects the drug binding affinity. In other words, scientists may be designing computational models based on the wrong protein shape.
Another reason is that molecular binding is extremely complex in practice. Scientists have proposed many strategies for their binding affinity simulations. Unfortunately, none of these strategies work for all protein–drug combinations.
Machine learning may help solve these computer-aided drug design challenges. Researchers use machine learning to devise a mathematical model based on experimental data, and use the mathematical model to extrapolate predicted—yet otherwise unknown—binding affinities.
However, the results of machine learning are not always sufficient. For example, numerous computational tools poorly predict molecular binding to the protein known as beta-secretase—pertinent to Alzheimer’s disease, a common drug target.
Many researchers, such as Dr. Ho-Leung Ng’s research group, have helped improve the performance of machine learning. For example, the Ng research group recently studied drug binding to beta-secretase. In a future post, we’ll learn about the Drug Design Data Resource Grand Challenge 4, and why Dr. Ng used it as a benchmark for their research. These studies are highly pertinent to finding effective drug therapies for COVID-19.