MIT researchers are hoping their new generative AI model, DiffDock, can unlock the future of faster drug discoveries with few adverse side effects. The future of AI and healthcare is one that many hope not only to reduce care-related costs but also to improve all patient outcomes. And this is what a new paper that introduces a new molecular docking model, called DiffDock, hopes to do.
DiffDock will be presented at the 11th International Conference on Learning Representations where it will show off its approach to computational drug design. If proven successful, it could be a major breakthrough and could potentially rock the foundations of the traditional drug development pipeline. Overall, the goal is to create a new process of development that can accelerate the speed at which new drugs are created while also reducing the likelihood of adverse side effects through generative AI.
Right now, the most common molecular docking tools used for in-silico drug design take a “sampling and scoring” approach. This works by searching for a ligand “pose” that best fits the protein pocket. The issue is that it’s a time-consuming process to evaluate a large number of different poses. Once evaluated the tool scores them based on how well the ligand binds to the protein.
Though the issue is that even though this process has led to the identification and development of successful drugs, it takes over a decade of development with up to 90% of drug candidates failing clinical trials. This is quite costly as most studies, according to the CBO, estimate the price tag of the entire process as around $1 to $2 billion dollars per drug.
Gabriele Corso, co-author and second-year MIT PhD student in electrical engineering and computer science explained how previous deep learning methods treated molecule docking as a regression problem and how generative modeling excels, “it assumes that you have a single target that you’re trying to optimize for and there’s a single right answer…With generative modeling, you assume that there is a distribution of possible answers — this is critical in the presence of uncertainty.”
Hannes Stärk, co-author of the paper added, “Instead of a single prediction as previously, you now allow multiple poses to be predicted, and each one with a different probability.” The result of which is the model doesn’t need to compromise in attempting to arrive at a single conclusion. Doing so can be a recipe for failure during the development pipeline.
But it isn’t just DiffDock’s ability to unitize generative modeling that is showcasing its uniqueness. According to MIT News, the model is significantly more accurate than previous methods because of its ability to reason at a higher scale. Where other models began to fail, DiffDoc instead is able to maintain performance standards.
According to the same report, DiffDock places 22 percent of its predictions within 2 angstroms. This is widely considered to be the threshold for an accurate pose, 1Å corresponds to one over 10 billion meters. This result makes it more than double of the best current docking models, which barely hover over 10 percent.
As Tim Peterson, an assistant professor at the University of Washington St. Louis School of Medicine states of DiffDock’s potential, “DiffDock makes drug target identification much more possible. Before, one had to do laborious and costly experiments (months to years) with each protein to define the drug docking. But now, one can screen many proteins and do the triaging virtually in a day.”