“Ensemble of uniformally distributed initial positions of a candidate repurposed drug prior to its interaction with the COVID-19 protein target ADRP, performed over many long-time molecular dynamics simulations run on Summit at OLCF.”
Credit: Art Hoti, Aghastya Bhati, and Peter Coveney, University College London
Coupling machine learning and physics-based methods, with this work researchers aim to accelerate the slow process of drug discovery, which typically lasts many years and costs billions of dollars—a major weakness in public health emergencies.
Coupling machine learning and physics-based methods, with this work researchers aim to accelerate the slow process of drug discovery, which typically lasts many years and costs billions of dollars—a major weakness in public health emergencies. Furthermore, virtual screening methods employed in drug discovery are currently hampered by their reliance on human intelligence in the application of chemical knowledge.
To overcome this, the researchers have developed a method called “IMPECCABLE” that involves sampling candidate compounds from both a billion-compound, synthetically accessible space, as well as from the output of a deep learning generative algorithm. The selected compounds are scored based on calculated binding free energies, and fed back into the deep learning algorithm to iteratively refine predictive capability.
This approach will have direct applicability in the pharmaceutical industry for quick identification of potent binders for a given target protein and binding pocket. Additionally, a machine learning-based method will enable assessment of the resistance of COVID-19 protein variants and their impact on existing vaccines and drugs.
