Over the years, many aspects of Cyclica’s business and technology have changed, sometimes dramatically. In business, we started as a service provider, then created a Software as a Service (SaaS) offering, and most recently grew into a neo-biotech company with a vibrant drug discovery pipeline . In technology, we started out with a structure based surface matching algorithm that took weeks to compute potential targets for a single molecule, and ended up with our current deep learning model that can screen billions of molecules in just minutes, with industry leading prediction accuracy
But, underneath all this transformation, one constant vision has remained unchanged: we set out to look at molecules not only as they relate to a single protein target, but as they relate to the universe of all proteins. This constancy of vision has led us from what was initially a relatively clunky reverse docking algorithm for proteome screening to an ultrafast and highly predictive deep learning model, MatchMaker™, that has enabled us to discover novel molecules for dozens of different targets, some of them notoriously difficult to drug (often referred to as ‘undruggable’).
The original motivation for our proteome-wide approach was to understand the polypharmacology of a molecule, and to find new targets for an existing drug. But, as we developed our deep learning approach, we realized another key advantage of the proteome-wide view: by training a deep neural network on the observed chemical interactions of tens of thousands of proteins and millions of molecules, we provide an opportunity for the model to learn general principles and become good at predicting new active molecules for proteins that have no chemical data associated with them, at all. We end up with a single model that works for any target, even those with low or no data, out of the box.
This is a very unique approach to AI in drug discovery. Almost all other companies in the space focus on ligand-based models, where generalization occurs only in chemical space, with a single, fixed target that needs to have some chemical data already available. MatchMaker™ uniquely enables us to address low data targets and to quickly generate new chemical matter for first in class drug programs. At the same time, we always look at the full polypharmacological profile of molecules, with the aim of improving their selectivity and safety profile.
As we look back, we see that where we are now is the result of a long series of constant innovation. It is particularly dramatic for MatchMaker™, our hit finding engine, as described above. However, a similar process has led to several other ground-breaking technologies, including our ligand-based ADMET prediction engine POEM useful for lead optimization. We also recently released a preprint and open-source software for NodeCoder, a graph convolutional network developed to predict a wide range of functional sites on 3D protein structures. This project was co-developed between members of the Cyclica and Bo Wang’s team at the Vector Institute and previously described in this blog series and NeurIPS MLSB 2021. The modeling framework was designed for use with predictive protein structures that otherwise don’t have the biological context, such as ligand binding sites, that are typically available through experimental approaches.
As companies grow, it becomes harder and harder to keep up the pace of innovation, and progress can slow down as more and more work is devoted to maintenance and legacy systems.
At Cyclica, we have created a dedicated innovation team that is tasked to identify opportunities for innovative new tools to fill out or expand our platform, and explore these opportunities to eventually lead to prototypes that can be integrated into our drug discovery workflow to boost its efficiency, scale and scope. This way, we hope to keep alive the innovative spirit that has allowed us to build our platform so far, and harness it to keep us at the cutting edge as we move forward into the future.
Authored by Naheed Kurji and Dr. Andreas Windemuth, Chief Innovation Officer
