Drug Discovery Online published an article by Verseon’s Anirban Datta, PhD, explaining the shortcomings of current attempts to use AI in drug discovery which is constrained by the small pool of existing experimental data. He also illustrates how physics-based predictive modeling can surmount these limitations.
—
Fremont, Calif.—Drug Discovery Online recently published an article by Verseon’s Anirban Datta, PhD, explaining the shortcomings of current attempts to use AI in drug discovery. Dr. Datta provides examples illustrating the limitations of AI and explains how previously collected experimental data constrains the number and types of compounds AI can find. He also discusses how accurate physics-based predictive modeling can surmount the limitations imposed by the small pool of existing data.
As Dr. Datta remarks, “The success of machine learning, and in particular deep learning, depends heavily on the availability—and quality—of large data sets for training.” He cites real-world examples in which AI systems fail. These systems fail because they can only successfully analyze situations that resemble their set of prior data. When situations outside its data set require extrapolation, AI often fails to yield useful results.
Finding novel drugs that better treat diseases remain the pharma industry’s biggest challenge. However, it is impossible to synthesize and test the entire vast ocean of possible drug-like molecules. The sheer number of possibilities is too large. Without any prior experimental data for those molecules, AI-centric systems remain stuck in tiny tide pools of data with no means to extrapolate to the wider ocean. These systems are consequently biased toward producing “me-too” drugs.
AI systems also face other limitations. They need data on compounds that did not work. But no one publishes this sort of data. In addition, drug-like chemical properties change radically with even slight changes to a compound, adding another layer of complexity to the data required for correct predictions.
Dr. Datta explains that it is necessary to build “synthetic” data sets to surmount the limitations that have prevented AI systems from creating genuinely new drugs. Physics-based molecular modeling can efficiently generate these synthetic data sets and provide AI a way to navigate beyond the tidepools at the edge of the chemical ocean. Tools that correctly utilize the rules of physics and chemistry can determine the behavior of compounds that no one has ever synthesized before. Coupled with the interpolative power of AI, it is possible to find the novel compounds necessary to advance modern medicine. Dr. Datta concludes that “companies with deep expertise in both physics-based molecular modeling and AI may have the ultimate advantage.”
About Anirban Datta:
Anirban Datta, Ph.D., is the head of discovery biology at Verseon International Corporation. He has over 20 years of experience in biomedical research and pharmaceutical drug discovery. He is the driving force behind Verseon’s automated processes for biological characterization of compounds, teasing out their unique properties, and structuring drug candidate development pathways. He has led multiple drug discovery programs in diverse disease areas, including cardiometabolic disorders, ophthalmology, and oncology. Datta was previously a scientist and Susan B. Komen Breast Cancer Foundation Fellow at UCSF and the recipient of lung and breast cancer concept awards from the U.S. Department of Defense. His early research was spun out into a cancer diagnostics company. He received his B.S. in physics and biology from the University of Chicago and his Ph.D. in molecular biology from the University of Pennsylvania.
Contact Info:
Name: Walter Jones
Email: Send Email
Organization: Verseon
Address: 47000 Warm Springs Boulevard, Fremont, CA 94539, United States
Website: https://www.verseon.com
Release ID: 89074394
