Atom Spotlight: Q&A with Jeff Warrington, Senior Scientist

Atomwise is comprised of a world-class team of diverse disciplines internally referred to as "Atoms." In this Atom Spotlight, we sit down for a Q&A session with Jeff Warrington, PhD, a member of our Computer-Aided Drug Design (CADD) Modelers Team.

Tell us about yourself. What interested you about AI and machine learning?

I’m Jeff Warrington and I’m a Senior Scientist at Atomwise. My Ph.D. training is in Synthetic Organic chemistry and Natural Product Synthesis where I studied under Louis Barriault at the University of Ottawa, Canada. From there I did a postdoc at Stanford University with Paul Wender where we completed the interconversion of an abundant natural product, Phorbol, to a rare one, Prostratin - which offers unique medicinal properties as a possible treatment for HIV infection.

Following my academic career, I worked as a medicinal chemist for over a decade and participated in all phases of drug discovery from hit identification, hit to lead, through to lead op and candidate selection. I was part of the team that developed Reldesemtiv at Cytokinetics, which is in clinical trials for neuromuscular disorders such as spinal muscular atrophy (SMA) and ALS.

During my time as a laboratory-based medicinal chemist, I became increasingly interested in the computational side to designing new ligands to bind to proteins, and became intimately familiar with the challenges of drug discovery - which certainly do not end with designing ligands for proteins.

The complexities involved with not only getting the drug into the body and to the target, but having drug-like properties of being orally bioavailable and long half-life within the body is a tremendous challenge for even the most skilled in the art. Walking through these complexities can take years for a medicinal chemistry team using traditional, iterative, empirical methods.

Having spent many years semi-empirically designing new molecules as drug candidates using traditional means, I felt that recent advances in machine learning would be able to provide us with the tools for processing multidimensional data in new ways to come up with better solutions, and better design ideas. I continue to believe that machine learning and Artificial intelligence will allow us to look at small molecule design ideas in new and creative ways, and come up with elegant solutions to complex problems.

Today, I’m a member of the Atomwise Production Team where I get to apply AI and ML to early stages of drug discovery for our company. I am excited at the prospect of what Atomwise can bring to drug discovery and what the future holds for the new discoveries our technology can make.

What’s a typical day at Atomwise?

We have a number of ongoing projects with external collaborators both industrial and academic. The company has scaled so that each scientist has a lot of different projects on their plate, so any given day is about prioritizing work for projects and making sure that they all get proper amounts of attention. Typical projects involve a lot of researching biology and some history of drug discovery at the target or pathway, and how to best implement our technology to find small molecule modulators of pathways to fight disease. There’s a lot of cheminformatics work, ensuring that the compounds that are eventually selected fit the right criterion for the project. There’s also communication aspects to the work in making sure that the work is communicated in such a way that it makes it understandable by researchers with from many different backgrounds. This involves a lot of generating graphical representations of the data, and writing plotting scripts. Interpretation of results is very important.

How does Atomwise’s technology work?

Atomwise patented the first deep learning technology for structure based small molecule drug discovery. This AI technology harnesses millions of data points and thousands of protein structures to solve problems that are difficult to solve with traditional methods. AtomNet®, uses a similar approach to that which is used in advanced image recognition technologies today, but applied to the problem of finding new small molecule drug-like ligands for proteins.

Chemical space is vast - it has been estimated that using modern synthetic methods you could feasibly make 10^60 different molecules consisting only of carbon, hydrogen, nitrogen and oxygen. This is on the order of the number of atoms in the observable universe. As you can imagine, this number really stretches human ability to comprehend how many synthesizable compounds there truly are.

While physical screening libraries have certainly increased over the years, and corporate compound collections can be in the millions of compounds to test in physical screening assays, they’ve been rapidly outpaced by enumerated virtual libraries from commercial vendors of compounds that could be made in a matter of weeks. These make-on-demand libraries today consist of libraries of greater than 10 billion compounds, which is thousands of times larger than even the screening libraries for the largest pharmaceutical firms.

Using AtomNet® models we are able to rapidly traverse this vast commercially available chemical space and find precious needles in the haystack which could offer new starting points for discovery of small molecule solutions for human problems.

How do you see AI impacting the drug discovery process?

Drug discovery is a long, costly process that involves many disciplines and tens to hundreds millions of dollars in preclinical research. Unfortunately we still don’t fully understand all the rules that govern the interaction between biology and chemistry, and we must produce new materials and drug candidates and physically test how they behave when assayed in biological systems. We then hope to extrapolate these model systems to more complex systems. For example, what happens to an isolated protein in a test tube may or may not be reflective of what happens in a cell, which may or may not be reflective of what happens in the human body.

While you can think that more complete systems have the advantage of being closer to the natural world - the flip side of this is that because they have so many moving parts and interrelated variables at play that it can be really difficult to see how design changes affect small parts of the whole system. We do have models on how individual experimental results translate into more complex systems, but this is still an active area of research.

The iterative process of design, test, redesign process in drug discovery typically requires hundreds to thousands of experimental iterations, and that can collectively take many years to complete. By taking advantage of tools such as machine learning and AI in early design cycles we hope to dramatically shortcut this process, by coming to better designs in fewer iteration cycles and much faster - thereby dramatically reducing the cost and time needed for the development of new medicines.

Learn More

Jeff will be speaking at CDD Vault's Demystifying Machine Learning (AI) in Drug Discovery webinar on September 10th. To learn more about machine learning methods applied to drug discovery and register here for the event. 

Join our team