My primary research interests is currently virutal screening with deep learning.
Medicinal chemists and biologists use virtual screening to scan large chemical databases to find molecules of interest to perform expensive downstream analysis. The standard practice uses classical techniques such as structural docking or traditional machine learning to screen libraries on the order of 100k. The recent introduction of deep learning means we can score massive libraries (order tens of billions). And if that isn't enough, there have been recent successes in de-novo molecule generations. In my testing, these generators can output valid molecules at a rate of 25,000 unique to all known databases per second per GPU. In this regime, downstream analysis such as MD on even 0.001% would be intractable.
My work focuses on solving this problem in two ways:
Creating workflows and analysis that can intelligently sample these enormous libraries always to provide downstream analysis the best possible candidates
Rethinking the way lead discovery is performed, so we don't have to think of these steps as all unique. If we reframe the problem with the specific goal in mind, it may be the case the workflow arises different (postmodernism can teach you a bit about this concept)
In many ways, I am skeptical of typical supervised deep learning. The workflow of collecting data, generalizing over it, and predicting off it has its uses--but in the regime discussed, it seems aimless. And yes, I am very inspired by 20th-century French philosophy.
AACR "Virtual Screening with Deep Learning using Cancer Cell Line Dose Response Data", Austin Clyde, Arvind Ramanathan, Rick Stevens. Advancing Precision Medicine Drug Development: Incorporation of Real-World Data and Other Novel Strategies, San Deigo CA. 10 January 2020. (poster)
Janssen Pharmaceutica "Accelerating Virtual Docking Screens with Deep Learning", Austin Clyde. 5 December 2019. (talk)