Abraham Heifets,

Co-Founder & CEO,

Atomwise

"By predicting the binding of billions of small molecules to a protein of interest having a known disease association in a matter of days, Atomwise can accelerate the earliest stages of drug design by several orders of magnitude."

What aspects of the drug discovery process does Atomwise seek to improve?

AI for drug discovery can be broken down into three categories: AI for biology, AI for chemistry and AI for clinical trial design and execution. Atomwise works on AI for chemistry. We were the first group to use the modern machine learning technique of convolutional neural networks for drug discovery.

There are a number of fundamental challenges in the traditional, early-stage drug discovery process. Traditional approaches are notoriously lengthy and complex, often involving years of iterative and complex medicinal chemistry. Atomwise leverages artificial intelligence (AI) for structure-based small molecule drug discovery, removing the barriers of physical screening that have limited the success of traditional methods for drug discovery. By predicting the binding of billions of small molecules to a protein of interest having a known disease association in a matter of days, Atomwise can accelerate the earliest stages of drug design by several orders of magnitude. Atomwise also goes beyond other CADD and AI-based approaches that require structural information about a target before virtual screening can take place, or that use simplified docking models that make certain complex targets inaccessible to AI-based approaches.

Our technology is based on convolutional neural networks – the same AI technology that is used for image and speech recognition. If you have ever talked to Siri or Alexa, or uploaded a photo to Facebook and had it prompt you to tag certain friends, those are examples of convolutional neural networks at play. Atomwise was the first group to take what works in image and speech recognition and apply it to molecular recognition. We built an AI system called AtomNet® and ran it in the biggest application of machine learning in drug discovery in history. At Atomwise, we partner with top-100 pharmaceutical and emerging biotechnology companies, a rapidly growing market estimated to reach $729B in global market value by 2025, and academic researchers at universities, institutes, and hospitals around the world. We now have over 750 projects, across every major therapeutic area, addressing over 600 unique targets. We have 285 active drug discovery partnerships with researchers at top universities and we’ve maintained an over 75% success rate on these projects.

How does using AI for drug discovery compare with traditional high throughput screening methods?

At Atomwise, we routinely screen libraries of over 16 billion molecules. Atomwise drastically reduces physical screening efforts, helping our partners identify leads without having to synthesize or buy large libraries of compounds.

Chemical vendors are offering about a billion new molecules a month, each of which you can purchase and have shipped to you in four weeks. In contrast, if you screened the entirety of a big pharma corporate collection, you’d have tested only 3-5 million molecules. In other words, if you put together all the big pharma corporate collections, and multiplied that by 10, that is how many new molecules are being added to purchasable chemical space every month. In our experience, these ultra-large libraries yield more, better, and more varied starting points (“scaffolds”) than you can with physical techniques like high throughput screening. The challenge is that even a modest false positive rate swamps any right answers, so success requires predictive models that are highly performant and deliver exquisite accuracy.

Atomwise recently announced a substantial round of fundraising. How will that money help grow the company?

With this latest investment, we will continue to scale our AI technology platform and continue to help researchers accelerate their drug discovery projects. We have hundreds of projects with a great diversity of proteins, especially proteins which have historically been incredibly challenging. Often the protein will have no x-ray crystal structure and requires the use of distant homology models. Around half of our projects have no known drug-like inhibitors. This precludes traditional computational methods like QSAR, which require a training set of known active and inactive molecules for each protein. Even traditional human chemists can struggle when they have no starting point! And yet, across our hundreds of projects, we are able to deliver success repeatedly.

Our previous funding round let us challenge, refine and test our system, and prove that it works. This Series B round is about taking that system, using it to discover real medicines, and get them to patients. We will build our own internal pipeline and continue to grow our portfolio of joint ventures with leading researchers using AtomNet® for drug discovery, like those we have launched with X-37, Atropos Therapeutics, Theia Biosciences and vAIrus. We will also expand our work with corporate partners, which currently include major players in the biopharma space including Eli Lilly and Company, Bayer, Hansoh Pharmaceuticals, and Bridge Biotherapeutics, as well as emerging biotechnology companies like StemoniX and SEngine Precision Medicine.

Half of your projects have no training data or crystal structure. Can you elaborate on how Atomwise’s approach overcomes these problems?

Many of the early stage drug discovery tasks are very similar. Researchers often have a protein, maybe an exciting new assay for it and they must find a novel chemical that will interact with the protein. At Atomwise, we do that work across a broad range of different diseases and protein classes. Some of the most interesting targets come from academia because they are always pushing into new biology and new hitherto unexplored disease mechanisms. The corollary of working at the edge of human knowledge is that you do not usually have the chemistry or the structural biology completely fleshed out. You do not necessarily know what is going to work for that protein. Therefore, these are the most challenging discovery projects, wherein you must figure out how to begin. What we have done at Atomwise is instead of building a protein specific model based on known structural chemistry for that specific target, we built a single global model, which learns the fundamental principles of chemistry. The same way that a human medical chemist might, our technology looks for patterns that it can translate across different proteins and across different diseases.

What is the key to improving machine learning models over time?

One of the benefits of using a single global system is that every project you work on improves the system for every other project. As you refine and train the system, you can improve accuracy for every protein. It is key to be able to generalize previously unexplored, unlocked proteins. At Atomwise, we call that drugging the undruggable. Of the roughly 20,000 proteins encoded in the human genome, only about 750 have FDA-approved drugs, and only about 2000-3000 have a small molecule drug or biologic under development. About 80% of human gene targets are uncharted territory and represent the most challenging and most promising future for pharma and human health, however the majority of these genes do not yet have enough structural data or are protein classes that have not yet been unlocked by traditional drug design. There is an opportunity to unlock and drug proteins that have never been drugged before.

Has Covid-19 expedited adoption of advanced technologies and what are your views on how to move even faster toward medical progress?

Covid-19 has brought awareness to the pressing need for advanced life science technologies. The race to develop treatments for Covid-19 has also put the drug discovery process under a global microscope - revealing that it takes far too long, costs way too much, and stunts potential positive outcomes due to the physical limitations of testing and screening viable drug candidates.

We currently have 15 Covid-19 research projects in flight tackling different proteins in the virus and the host. Right now, nobody knows what the best approach should be for tackling Covid-19. Part of the challenge is that people have a very clean view of science: Newton invented physics and then Einstein fixed those problems and each one builds on the other as if the path to success was direct. However, in reality, science is much more nuanced because we get to see in hindsight what works. Whereas when you go forward, you don't know which hypothesis is going to work. Therefore, to make fast progress for cancer, infectious disease, neurodegeneration, I think being able to simultaneously support a number of different hypotheses and casting a wide net is key.

What key milestone does Atomwise hope to reach moving forward?

We will be scaling the company in a number of different ways. First, we will scale the capabilities of our technology. We have built an impressive capability, but there is always more that you can do. Over the past couple of years, we have gained a valuable breadth of experimental data, including the largest diversity of drug target sites, homology models, protein classes, and disease areas of any AI platform. Moving forward, we must continue to scale our algorithms to keep up with the growth of ultra large libraries of 100 billion molecules and up. This task is not easy and it doesn’t happen overnight, so we need to keep doing that because the chemical space continues to expand. Along with that, the accuracy we need to deliver continues to increase.

Furthermore, to scale the technology, we are going to need to scale the team. We will continue to scale the collaborations we have been doing, whether that is with big pharma, small pharma or joint ventures. For us, “success” is being able to make the discovery of effective drugs faster, less expensive and make the process more democratic by opening up possibilities for effective therapies to be discovered by a wider group of researchers.

Overall, the future of drug discovery is bright. At no time in history have we had more information, more insight and better tools to work in drug discovery. It is an exciting time for all of us. At Atomwise, we have validated our claims with proven results. Our technology is covered by 19 issued patents, and to date, research partnerships have generated 17 pending patent applications and several peer-reviewed publications. Our 285 active drug discovery partnerships and our projects have yielded significant advancements across many disease areas, from cancer, neurology and immunity to infectious disease, inherited disorders, and others. In the AI space, it is crucial to have results at that scale and not just predictions. Anybody can make predictions, but real results are what matters. That is what we have been able to do over and over again.