I recently invited Naheed Kurji, CEO of Cyclica, to an interview on the Artificial Intelligence in Drug Discovery podcast. Unfortunately, he came down with a cold, and we had to cancel. Fortunately, he agreed to provide written answers to my questions.
Cyclica focuses on predicting a drug’s polypharmacology to provide insights into adverse effects, repurposing, and more. The Toronto-based startup launched a cloud-based tool called Ligand Express last November to empower researchers with the technology. Now the company is expanding it to create a comprehensive platform for drug discovery and design.
Below, you’ll learn about challenges and opportunities of polypharmacology, how AI is necessary—but not sufficient—to address them, and emerging opportunities in drug discovery driven by advances in computing and genomics.
SS: I’d like to start with a discussion of polypharmacology. For those who don’t know, how were things done before, and could you explain polypharmacology, and why it’s so important to drug discovery and development?
NK: In the prevailing paradigm of drug discovery, scientists first study the biological mechanism of the condition to be addressed, then identify a specific molecular component that is involved, the target (usually a protein), and then embark on finding a drug that will interact with the target and improve the condition. This is known as target-based drug discovery. However, at the end of the day, target-based drugs, whether found through virtual screening or augmented with artificial intelligence (AI), has led to some disappointments because focusing on a single target is insufficient.
“It is becoming increasingly clear that polypharmacology plays a major role in the success of drugs.”
Polypharmacology refers to a drug’s ability to bind multiple molecular targets and thereby affect multiple biological processes. Many studies have indicated that a limited number of distinct binding sites exist, a fact that, when coupled with the large number of proteins found in the human body, implies a drug will interact with many proteins besides its intended target. Those off-target interactions may cause unexpected and undesirable adverse effects or may alternatively provide attractive repurposing opportunities. Also, complex diseases such as central nervous system diseases, cancer and inflammatory diseases may require more complex therapeutic approaches. Typically this was accomplished by using multiple drugs or drug cocktails; however, the concept of designing a drug to affect multiple pathways has gained popularity. It is becoming increasingly clear that polypharmacology plays a major role in the success of drugs. By keeping this concept at the forefront of drug development, scientists can explore novel targets for new drug programs, expand the indications for developing drugs, repurpose existing drug for new indications, and computationally design multi-targeted drugs. In an effort to capitalize those opportunities and bring innovative drugs to the market, we have spent the past four years building an integrated collection of enabling technologies specifically for polypharmacology.
SS: You initially created your technology, Ligand Express, to focus on polypharmacology. Can you describe how it works?
NK: While we initially embarked on building Ligand Express, we are focused on much more than just that. We do not believe that offering pharma with a point solution to one part of the value chain is what they are looking for. We believe they are looking for a more end-to-end enabling platform that augments their existing workflow. So, our winning aspiration is to drive drug discovery by empowering scientists in pharma with an integrated cloud-based and AI-augmented platform that enhances how they design, screen, and personalize medicines. While this is a lofty goal, we have a focused strategy, which, as a young company, is crucial. So we put our heads down to solve the problem of uncovering the polypharmacology of a small molecule, in silico.
Taking a polypharmacology view of small molecules is a highly non-trivial problem, due largely to the availability of sufficient protein data, and the computational intensity screening a small molecule across the entire known proteome. Ligand Express is a platform that consists of several integrated technologies. Currently, there are three major components, but in the future, we will be adding other components to facilitate drug discovery. Ligand Express starts with Proteome Screening, which takes a small molecule drug—hypothetical, preclinical, clinical or FDA-approved—and uncovers putative protein targets that are likely to bind the molecule. There are a few of important considerations in how we have approached the build of our tech: i) only a small molecule is required as an input to Ligand Express (i.e. one does not require a target to go alongside the small molecule); ii) Ligand Express moves beyond just human proteins as we built a proprietary database of surfaces of all known protein structures across all species; iii) by using a biophysics-based approach, the identification of potential binders is supported by protein and chemical structures and is resolved at the molecular level to be presented to our customers. Computational biophysics can produce new hypotheses about protein-ligand interactions, instead of inferences from existing knowledge like an AI approach would do. This enables the discovery of new targets as well as new sites on targets (e.g. allosteric sites).
Once Proteome Screening identifies a potential protein-molecule complex, another component of Ligand Express, Effect Prediction, predicts the modulatory effect that a molecule has on each predicted protein using machine learning on large amounts of experimental data. The last component of Ligand Express, Network Analysis, leverages bioinformatics and systems biology information to link the predicted protein-ligand interactions to diseases, pathways, and other biological functions. Graphs can be generated and tuned in real time to help scientists probe results and action on interesting hypotheses.
When you look at the fully integrated platform, in essence we use molecular modeling to uncover novel information, and then leverage AI to link that with existing information. By understanding how a small-molecule drug will interact with all known proteins, Ligand Express provides value in elucidating mechanism-of-action (e.g. target deconvolution after phenotypic screening), prioritizing lead candidates, understanding side effects, as well as repurposing existing drugs.
SS: You often talk about how AI is not a silver bullet. It’s good for specific things, but not for others. Can you talk a bit about how your thinking on this is reflected in Ligand Express? What is AI good for, in your opinion, and what is better done through other technologies and approaches?
NK: AI has been around for a very long time, probably longer than a lot of people realize. Alan Turing first described AI back in the 1950s, and computational chemists have been utilizing it for decades. However, it is understandable as to why there is a lot of excitement about AI nowadays given the invention of new approaches and the creation of powerful hardware capable of performing massive calculations. Most of the practical successes of modern AI are from machine learning approaches such as deep neural networks. Machine learning is great when there is a lot of data available to train on. However, it is less effective in “thinking outside the box,” i.e. coming up with novel hypotheses that add to rather than interpret existing data. We only employ AI within our platform for problems where it is best suited, and use alternative, mechanistic approaches where they are more appropriate. Force-fitting an AI technology to a problem is tempting given all the hype and buzz, but ultimately it can negatively impact performance and utility. A good example of this was recently described by Ragoza et al. in 2017, who attempted to model detailed protein/ligand interactions using convolutional neural networks instead of biophysically motivated force fields. The results were mixed, with one of the problems identified the insufficient amount of good data available to learn complicated physical interaction from the ground up.
“Force-fitting an AI technology to a problem is tempting given all the hype and buzz, but ultimately it can negatively impact performance and utility.”
In Ligand Express, AI and biophysical modeling go hand in hand. Driving the components of Ligand Express (Proteome Screening, Effect Analysis, and Network Analysis) is a combination of approaches involving big data, statistics, biophysical modeling, and machine learning. Together, these technologies synergize to give scientists a better understanding of a drug’s polypharmacology and its activity in patients. Along the way to building this technology, we identified multiple opportunities to employ AI in improving our platform, but before we integrate these ideas into Ligand Express, we perform exacting due diligence to ensure they work. If the AI doesn’t add substantial value to our platform, we won’t hesitate to shut the effort down and move on to the next opportunity.
SS: You’ve also talked about the importance of using AI to augment what scientists do, not try to change their workflow, or force technology upon them. Do you see that happening a lot? In what ways are you consciously working to avoid it? How do you get people on your team to do so as well?
NK: The successful implementation of AI in drug discovery requires a patient strategy and a long term view. Many scientists, especially bench scientists at the forefront of drug discovery, have spent over a decade of their lives working on a PhD in chemistry or biochemistry. If we simply put an AI technology in their hands and approach them with the attitude that “you’re going to have use it because it’s the next big thing, trust us,” we will not succeed. We have to show them how adopting and leveraging our technologies can augment their capabilities, and that comes from upfront validation work, evaluation through testing, and integration that augments, not disrupts existing workflow.
“If we simply put an AI technology in their hands and approach them with the attitude that ‘you’re going to have use it because it’s the next big thing, trust us,’ we will not succeed.”
When we first embarked on the creation of Ligand Express, we knew that we would be doing something that departed from the norm, and that it was natural to expect some skepticism and push back for those adopting the platform. In this case, we hoped that this departure from the ordinary would not unnecessarily change workflows or feel forced, but rather would be integrated and augment current practices. That said, there have been cases where technology has radically changed the industry, computational examples being virtual screening and QSAR modeling, which have led to increased efficiency in drug discovery. Perhaps one day, we will foster more sweeping changes. I don’t think we are there yet but with some of the innovations we have planned I am confident that we will catalyze important changes to the way things are done.
Our philosophy is that we are the technologists—we are very good at what we do—but we must rely on industry scientists to provide clarity on the challenges they face. Working together we can solve problems in a streamlined fashion. That is what led to Ligand Express and the philosophy behind how it is implemented. The platform is designed to allow users to bring their scientific acumen to the data, interact with the data, and drive novel hypothesis directly from the platform into the lab. In other words, we want scientists to autonomously navigate through our platform to augment their research, and our job is to make their experience as smooth as possible.
We’ve built a cross-functional team consisting of biologists, pharmacologists, medicinal chemists, and bioinformaticians who test our platform on a daily basis to gain multifaceted perspectives and ensure we are building the correct tools for them and their peers. We regularly host user testing sessions to improve the intuitive UI design for scientists in their day to day work. We also put a tremendous amount of weight on customer feedback. In fact, we have recently doubled down on our customer focus, implementing a new development processes that prioritizes all feedback from our users so that we can orient ourselves towards the near, short, and long term goals of delivering the best technologies and features.
SS: You’ve published a number of papers and case studies related to use of your technology. Are there any that stand out for you in illustrating its value?
NK: This is a great question. We have made it a priority to disclose our approach and methodology (within reason), and have published five peer-reviewed papers and five validations notes that describe our core technologies including—all of these can be found on our website). One study I’d like to point out is a drug repositioning and phenotypic screening target deconvolution study that was done in collaboration with Dr. Sanchez and colleagues from Tulane University.
Dr. Sanchez was testing FDA approved therapeutics as possible treatments for treating systemic sclerosis (SSc) associated with fibrosis in organ systems. SSc is an orphan disease whose pathogenesis is poorly understood, with a 9-year survival rate of 30%. Currently, SSc is treated by anti-inflammatory therapies with modest benefits, motivating Dr. Sanchez’s research into alternative therapies. It was interesting that Nelfinavir (NFV), an FDA approved drug to treat HIV type 1 infection, demonstrated activity in her preliminary phenotypic screen. NFV has a well established safety profile, and has also demonstrated anti-inflammatory activity, making it a potential drug repositioning candidate. While this was promising, Dr. Sanchez was hoping to understand the underlying target driving NFV’s activity in her early phenotypic models. We used Ligand Express to screen NFV against the structurally characterized proteome, revealing putative beneficial mechanisms of action consistent with canonical pathways in fibrogenesis. More importantly, this novel finding by Ligand Express was subsequently validated by a series of in vitro experiments.
This success has not only led to a publication, but also provided valuable insights that support the idea of repositioning NFV for treatment of fibrogenesis. This study is a highlight for Cyclica because it constitutes prospective validation of Proteome Screening. It also showcases Ligand Express’ ability to support critical research into rare and orphan diseases, and its utility in discovering the underlying target-of-action coming out of phenotypic screens.
SS: You recently announced an expansion of the platform. Can you tell me a bit about these new technologies, and what you learned that inspired them? What feedback were you hearing from users? What opportunities were you seeing in the market?
NK: We recently launched a novel ADMET-Prediction technology that is being integrated into Ligand Express to predict the pharmacokinetic properties of small molecules. As mentioned above, QSAR models have been used for a while, but following our evaluation of the marketplace, discussions with our partners, and exploration with our advisors and team, it became clear that i) there was significant opportunity to out-innovate existing approaches, and ii) integrating an ADMET-Prediction technology into a more holistic end-to-end platform is what pharma is looking for to drive more efficient discovery strategies. Our validation note on ADMET-Prediction has been benchmarked to leading open source models and out performs them all.
“An important theme to consider when thinking about strategy is ‘technology unification’—building technologies that can be applied in many ways is how we will drive not only differentiation but also defensibility.”
With Ligand Express as described above, we were repeatedly asked, “where does the small molecule come from that is being run through the Ligand Express platform?” The answer was simple, the molecule was supplied by our clients or partners, but that question led us to ask ourselves, “is there a way that we can start providing the structures that would be run through Ligand Express, and can we do it in a way that is uniquely Cyclica?” This is what prompted us to develop our state-of-the-art Differential Drug Design (DDD) technology. DDD is inspired by Ligand Express in that it takes a polypharmacology view to drug design. An important theme to consider when thinking about strategy is “technology unification”—building technologies that can be applied in many ways is how we will drive not only differentiation but also defensibility. So, our ADMET-Prediction technology will also play an important role in the DDD workflow so that the molecules we design have the favourable pharmacokinetic profiles to treat the disease of interest. Currently, DDD is is being developed and tested in collaboration with multiple pharma partners, and will be broadly released by the end of 2018. For now, we will leave it at that, but leave it say that we are very excited with initial results.
The other feature we are developing, Structural Pharmacogenomics (SPGx), addresses a major challenge faced by drug developers, namely differential drug responses in sub-populations. SPGx helps navigate this tricky aspect of drug development and strengthens the data generated by our platform. The market for structural pharmacogenomics’ application to personalized medicine is massive. We are obtaining solid validation on our SPGx technology as well. We have a publication co-authored by our collaborators at the University of Toronto published in Proteins that showcases the applications of our SPGx technology to cystic fibrosis.
You may be asking, why are we tackling this problem? Well, we are very optimistic about the future of personalized medicine given where the market is going. You’ve probably heard about the most recent $300 million collaboration between GSK and 23andMe to drive genetics-driven drug research. These types of collaborations are really interesting and applicable to where we are trying to position ourselves. Our goal is to provide layer of information on top of the geno-pheno data that pharma are looking to exploit. This would help with target discovery, DDD, and personalized medicine (via better understanding the role of single nucleotide polymorphisms (SNPs) / variants on drug response). I think this should help crystallize the intuition behind our winning aspiration: design → screen → personalize, all within an integrated end-to-end enabling platform.
SS: How will adding this functionality change the product? How will it improve users’ workflow?
NK: We are excited to bring these changes to the platform because they enable our users to span a greater portion of the drug discovery/development pipeline. In terms of workflow, as mentioned, DDD precedes Proteome Screening, allowing scientists to develop their own molecules. SPGx picks up after proteome screening by exploring how variations in proteins within the population may impact drug activity.
Scientists face many challenges when starting a drug discovery program with lead identification being a pivotal step. DDD can be applied to identify scaffolds/molecules with desirable properties that would serve as a solid foundation for a drug discovery project. Arming scientists with a tool that generates and evaluates small molecules for polypharmacology and drug-likeness will allow them to explore chemical space more effectively and lead to increased success at the bench.
Scientists can run molecules generated by DDD through Ligand Express and, with the inclusion of SNP data through SPGx, they will be able to identify protein mutations that may alter the activity of these potential drugs. Identifying these mutations during the drug development process supplements drug discovery scientist workflow and enables them to select experiments (both in vitro and in vivo), and design drugs that are more tolerant to mutation or are specific for a certain sub-population.
What’s your ultimate vision for Cyclica?
NK: We envision ourselves to offer an integrated cloud-based and AI-augmented platform that enhances how scientists design, screen, and personalize medicines. We firmly believe that biophysical and AI-augmented computational methods will accelerate drug discovery, and we are confident that our technologies will eventually be used throughout the pharmaceutical industry. By putting relevant information at researcher’s fingertips, whether from public sources or in-house proprietary data, biophysically computed or experimentally acquired data, Ligand Express will enable them to discover insights that would otherwise stay hidden, and streamline or even eliminate many of the time consuming and capital intensive laboratory experiments currently used for drug discovery.
SS: What are some of the challenges you face in achieving that vision?
NK: Creating a comprehensive software platform as we envision poses many challenges. There are scientific, technical, and behavioural and cultural challenges. Scientifically, we need to ensure that any new methods we introduce are well-tested and validated, as well as presented to and accepted by the scientific community via presentations and peer-reviewed publications. Technically, we need to stay at the forefront of a rapidly evolving environment in high-performance computing, and software development methodology. We also need to maintain the scientific computing expertise necessary to bridge the gap between science and software. Behaviourally and culturally we need to focus on user interface design that is aligned to the needs of the scientific community, and maintain and grow our agile and effective team of developers that can execute on our designs. To achieve this, we continue to to work very closely with our pharma partners who use our platform, to guide them along the path to the effective use of AI in drug discovery. As it turns out, during this process our partners guide us as much as we guide them, and we take away the experience needed to adapt our software to fit the workflow rather than the other way around. This way, we see ourselves achieving our vision gradually in a continuous cycle of product development and customer interaction. This isn’t meant to be easy, and it isn’t. But we accept that challenge as we believe that at the end of the day not only will we drive significant benefit to scientists in pharma and therefore patients, but become a highly valued and valuable company.
SS: What’s your prediction for how this space will continue to evolve? (Noting that predictions are always risky, of course, so I won’t hold you to them!)
NK: A widely accepted answer to this question would be with more technology, more science, and more data. That’s important and probably true, but I believe answering that question with only that approach is insufficient. For AI to best be adopted in any industry, it is really important to think about the psychology of the people who are using it, and how they are going to adopt it. So I think the next five to ten years are going to be a learning exercise on how computer scientists and technologists can best interact with scientists in pharma, so they together can address the challenges faced by the industry.
“I think the next five to ten years are going to be a learning exercise on how computer scientists and technologists can best interact with scientists in pharma, so they together can address the challenges faced by the industry.”
As for the evolution of AI we know there is a plethora of data, but a great portion of the data is inaccessible. Historically, pharma companies gain their market positions based on their own proprietary information and intellectual properties, which are often kept in silos or hoarded by particular research groups or within research groups by particular scientists. That’s a limitation for the evolution of AI because high-quality AI needs access to high-quality data.
The good news is that there have been salient changes happening in the industry in the past two to three years. There is this trend of opening up with data and sharing amongst institutions. Even better, there are efforts where data are being shared openly across the ecosystem. A great example of such is the Structural Genomics Consortium (SGC). It’s a pre-competitive environment where pharma companies have come together to support the SGC’s mandate of an open concept in structural biology. Gaining access to more data will be a determining factor to the success of AI; and we will see more of that, whether it’s genomic data, proteomic data, drug binding data, or even electronic medical records. I believe this trend of data sharing will continue to rise in the next five to ten years.
Lastly, as BenchSci has shown, there are now close to 100 companies in this space. If you rewind 5-7 years, there were probably only around 15-20. So there has been a mass acceleration in companies entering the market. I believe that many of the companies that offer only one solution in the value chain will eventually be integrated into a more holistic offering, and that those that can add value in more than one place will come out ahead. I think that captures the essence of our strategy.
SS: Where can people learn more about Cyclica and connect with you? Any upcoming papers or conferences?
NK: Our website, cyclicarx.com, is always a great starting point to get to know our company’s story, technologies, validations, team, and so on. I can also be reached through email at naheed.kurji@cyclicarx.com.
We are in the process of filing a provisional patent on one of the core technologies used in our Differential Drug Design. After that patent is published, we hope to generate publications based on novel insights generated by our DDD technologies. In particular, you might want to be on the lookout for a publication about our ADMET prediction method, the outlines of which are already available as a note on our documentation website. As for our SPGx technology, we have a publication co-authored by our collaborators of University of Toronto now that was published in Proteins that showcases the applications of this technology to cystic fibrosis.
Another exciting initiative I hope to share is that we have designed Cyclica Academic Partnership Program (CAPP), with the goal of increasing the impact of existing research efforts and bolstering publication(s) with novel in silico insights. The inaugural program under CAPP is the Ligand Express™ Academic Polypharmacology (LEAP) program, which is now calling for applications from principal investigators at academic institutions and non-profit organizations. We will award successful applicants with access to Ligand Express™ for a period of 90 days to screen up to five small molecules. This is an excellent opportunity for academics to use our cutting edge technologies at no cost. With the launch of LEAP, we hope to join forces with the brightest minds from research institutions around the globe to push in silico drug discovery forward. The application form to LEAP can be found on our CAPP website. We encourage researchers who are interested to submit their application by September 30, 2018.