Reshaping drug discovery with deep learning and polypharmacology
Posted 23rd October 2019 by Joshua Sewell
Cyclica is a Toronto-based biotech that leverages artificial intelligence and computational biophysics to reshape the drug discovery process. We spoke to their Chief Scientific Officer Andreas Windemuth.
What is the main goal of Cyclica?
Most computational discovery companies focus on an individual protein as the target for the drug being designed. From the beginning our goal at Cyclica has been polypharmacology i.e. looking at the action of a drug on the whole proteome, covering unanticipated side effects and hopefully new indications for existing drugs.
These days drug design works by finding a target from one of the many thousands of proteins in the body. This one protein is identified based on significant research, but there are many reasons why interfering with the target might not have the desired effect. Any drug that is developed for this one target will interact with other targets as well.
The second most integral part of our philosophy is combining first-principle biophysics and machine learning approaches. There are a lot of systems where there isn’t very much data available, and so analysing the biophysical mechanisms is the only way to make predictions. In other cases, there is a lot of real-world information available, and a machine learning approach makes sense. Therefore, we take a combined approach: a hybrid between structural, or mechanistic, approaches, and data-driven AI approaches.
What are you currently most excited to be working on?
We started out doing everything in the old-fashioned biophysical way of molecular modelling. But we found that there is a limit in terms of accuracy with that approach, simply because we don’t know everything about the systems.
We started looking into deep learning as a way to leverage the large amounts of experimental data on drug/protein interactions for prediction. Over the last few years, we have come up with our own deep-learning model called MatchMaker, which combines biophysical protein structure information with copious experimental binding data available through public and private databases.
When deep learning was first applied by Google to machine translation, for example, there was a huge step up in improved fidelity. For us, the same thing happened with MatchMaker. The field of drug/protein binding prediction had long reached a plateau of accuracy using structure-based docking approaches, which we initially used. Once we tried out our first successful MatchMaker models, we found that we could predict far more accurately which protein a drug molecule would bind to.
So much so, that we could not initially believe our results. We checked and double-checked everything, and finally performed multiple blind tests with our partners in the Pharma industry. It all came out to the same thing: MatchMaker was far more predictive than docking. We now have confidence that MatchMaker is the most accurate available tool predicting the interactions between proteins and drugs.
Are there particular ways that you think AI benefits pharma?
AI has applications in almost every corner of Pharma. Drug development itself is a huge undertaking with multiple different aspects to it and there are hundreds of points at which one could apply AI. A lot of AI and Pharma companies are doing quite different things. For example, many companies focus on target discovery, on image processing for assay read-out, or on clinical trial analysis and design.
Together with a handful of others, we are focusing on drug-target interactions. The advantage of AI over older methods of computational chemistry is that the accuracy is significantly higher, so we can achieve things we couldn’t before. To us, this is the most exciting part about AI in Pharma.
We also aim to provide a comprehensive platform in the drug discovery area: anything to do with drugs and interactions with biology. We want to provide a platform that becomes valued in the pharmaceutical industry. For this purpose, we have developed Ligand Express, a cloud-based interactive analysis platform covering drug discovery in general.
Looking to the future, what are your predictions for the industry?
In the next 5-10 years, things will change in the Pharma industry, and not just because of AI.
There is already a trend at the big Pharma companies where research is increasingly less important. Smaller biotechs are doing more and more of the research. This will remove some of the inflexibility in the Pharma business, as the smaller biotechs are much more willing to innovate and partner with people like us. When it comes to R&D, we will see most of the action happening in small biotechs in the future.
Andreas Windemuth is Chief Scientific Officer at Cyclica.
Only a week away, the 3rd Global Pharma R&D Informatics & AI Congress promises to be a stimulating networking and knowledge sharing opportunity. See who will be joining Andreas Windemuth on the agenda.
Leave a Reply