Artificial intelligence (AI) has the potential to dramatically improve many aspects of drug development. A particularly pressing challenge in drug development is the translation of results from preclinical studies into clinical trials. In this Q&A, Jo Varshney, Founder and Chief Executive Officer of VeriSIM Life, discusses her company’s unique AI platform and its Translational Index score, which it uses to help clients identify opportunities for improvement of their preclinical and clinical candidates, thereby derisking programs and increasing the likelihood of success, in conversation with Pharma’s Almanac Editor in Chief David Alvaro, Ph.D.
David Alvaro (DA): To start, can you please introduce me to VeriSIM Life and the vision on which the company was founded?
Jo Varshney (JV): I founded VeriSIM Life in 2017 with a mission to address the translational gap using AI and to alleviate some of the most pressing frustrations within drug development.
My obsession with translation goes way back. My father worked in pharma since I was two, and I was always fascinated by the many different types of animals used to evaluate developmental drugs and make extrapolations about the effects those candidates will have when administered to humans. That to me seemed a little nonlinear — we are not rats and rats are not us. The question then was: What else could be done to develop more valuable data?
Another fundamentally frustrating aspect of drug development is the fact that 95% or more of candidates fail in clinical trials, essentially rendering all that early nonclinical work useless. We spend millions of dollars and significant time, yet the overwhelming majority of candidates do not ultimately work in patients. Those are the problems VeriSIM was created to solve. I was able to bring a diverse set of skills to bear on these problems, combining my training as a veterinarian, my Ph.D. studies in comparative genomics, specifically oncology, and an interest in computer programming that goes back to my childhood.
DA: Did you found the company and then create the platform, or was the company founded to commercialize an existing idea?
JV: The technology was the foundation for everything else. Initially, I was not interested in starting a company; things just moved in that direction serendipitously. With our early technology, we participated in a hackathon organized by Google in San Francisco exploring data about a cancer patient. We won that hackathon, but more importantly, the three days we spent working on solutions for that patient motivated me to create a scalable platform. The opportunity to have an impact on thousands of patients rather than just one and really improve drug development was very attractive to me, and I knew that launching a new company was the best way to do that.
DA: How did you assemble a team with the varied expertise necessary to get this technology to a point where it can have that kind of impact?
JV: I firmly believe that if you want to create innovative technology, you first have to really learn the fundamentals. I took it upon myself during my Ph.D. and postdoctoral studies to learn the fundamentals of artificial intelligence: supervised learning, unsupervised learning, deep learning, etc. I also learned to code. That helped me to understand the type of team needed to establish this platform technology, which drove our initial recruiting. Today, we have folks ranging from software engineers to artificial intelligence specialists, space engineers, and translational scientists who understand different elements of drug development. Each has their area of expertise that fits perfectly with the others, helping develop solutions to address the translational gap.
In addition to bringing together all these people with diverse knowledge, we created a culture and environment that promotes cross talk among them. One of the first things we did was create the concept of a Translational Index score, which is inspired by credit scores. In the manner that credit scores provide an indication of financial health, the Translational Index score provides an indication of a client’s translational needs. It helps guide everyone in the company working to address that client's specific drug development needs. As a result, no one has to become an expert in all of the technologies involved in the platform. By looking at the Translational Index score, each team member — regardless of their area of expertise — can understand what solutions might be needed to improve the score.
DA: Can you explain the general technology and features of the BIOiSIM® platform that your team developed?
JV: The platform was designed to be analogous to the way that a human mind learns and creates insights from individual elements. With different applications, we can address a host of drug development–related questions, such as “What’s the right formulation strategy?” or “What would be a good dosing regimen?” The foundation is in the brain of the platform, which we call BIOiSIM. It contains the core fundamentals based on an understanding of how these different concepts are interconnected. These models are then fine-tuned to the specific application, which involves model optimization to address a specific question. The generalized platform contains the knowledge base that makes it possible to accurately predict potential specific outcomes.
People learn different things and then apply them and solve complex problems. Machines have the capacity to do this exponentially faster than human minds. What is really important, though, is that our team is constantly learning from the outcomes generated by the AI platform, as does the platform. And the platform is also constantly learning from the team because both are enabling a commingling of human brains and machine intelligence.
DA: How do you see your platform’s potential to transform drug development?
JV: VeriSIM Life’s platform is not the answer to all problems with drug development. Going back to the credit score analogy: knowing your credit score on its own doesn’t make someone rich but it does enable informed decision making and provides guidance to establish a path to better financial health. Similarly, the BIOiSIM platform provides guidance to drug development experts. With this guidance, it is possible to reduce waste in unnecessary research by 20% to 30% and boost both asset and company valuation. Most importantly, patients in need get access to new medicines sooner.
It is important to understand that AI is not suitable for all aspects of drug development. Our mantra is that we don’t try to express a solution when there isn’t a problem. We are a big proponent of “explainable AI.” Returning to the credit score analogy, this would be akin to a detailed credit report that shows missed payments, hard credit pulls, and possible fraudulent activity — risks to be avoided. VeriSIM Life generates detailed reports for clients and for our internal use, being as transparent as possible about the validation of our models, the accuracy of the system, and the propensity for potential bias. We provide these reports because we want people to understand how these systems come together to generate the Translational Index score and how that can be used to optimize things.
DA: How would you describe your initial data lake and its evolution to today?
JV: Before even starting the company, I wanted to understand what the minimum viable product (MVP) would look like. That involved determining the parameters reflecting the physio logical system of humans and animals, creating a system that took them into account, and finding the data to populate the system in the public domain.
Once that was established, the next step was to take an existing, commercial drug and run simulations to see if the outcomes predicted by the system matched the real-world evidence. The results were promising, so I hired my first software engineer.
We then started seeking public databases — many of which had been released by the U.S. Food and Drug Administration (FDA)— and hired the appropriate experts to understand all of those data. We have since also formed partnerships with data providers, such as the one we have with Clarivate, one of the largest data providers in the life science industry. All these different types of data are combined into a data lake to help address myriad challenges in drug development.
Where no data are available for a specific problem, we pursue one of two approaches. One is to leverage the data we have generated on previous projects, using outcomes and working backward to fill gaps. That involves leveraging synthetic data derived from the real-world data. The second method involves transfer learning, in which we leverage the fact that many diseases have common mechanisms of action. Knowledge from one disease, such as cancer, can often be applied to others, such as infectious diseases, and vice versa. This approach requires deep knowledge and understanding about the diseases involved, and VeriSIM Life has access to experts who can help us tackle the data diversity that we have in our industry.
DA: How does your AI platform shorten drug development timelines, and what do you see as its biggest impact in this regard?
JV: I think the biggest gap we are addressing is in the area of human intelligence, more specifically by enabling better decision making. People in the industry spend years trying to understand their programs well, but they still have difficulty seeing where those programs should go. We bring the superpowers of machine intelligence to enhance their ability to make the decision to move a program forward and invest more or to cut it down.
That is just one example. There are so many problems our platform can help solve by enabling human expertise. The tangible results are seen in the time, cost, and risk reductions that result from better decision making.
VeriSIM Life is also different from other AI companies because most of them are focused on the initial problem of candidate discovery but not the other activities that can occur during the many additional years it takes to bring a new drug to market. We also help with questions around off-target effects, potential toxicities, optimal dosage, and other problems that aren’t usually considered at the earliest development stages. They are typically addressed once a lead candidate (or candidates) is selected. We are helping to address the gap during all stages of development, trying to reduce the number of iterations and physical experimentation needed to get to the optimal drug product.
It can be challenging, however, because people in the pharma industry are very hesitant to adopt new technologies or change ways of working because of ingrained risk aversion. But if the odds can be improved, we could open up opportunities for new areas of drug development that otherwise would not be pursued given the time and cost involved. The lack of new antibiotics is a good example. Hundreds of thousands of people die each year from infections by pathogens that are resistant to current antibiotics, but there has been practically no investment in this area by pharma companies because the margins on antibiotic drugs are minimal and the cost and risk of development preclude investment. A technology like ours can help to identify successful candidates and reduce the risk and cost of development.
DA: By reducing risk at key points in the drug development cycle, are you changing the dynamics of where candidates fail — early vs. late phases, for instance?
JV: If you ask a drug developer, they will say phase I is not really an issue because it involves healthy volunteers testing out the dose. Preclinical models can always be improved, and work should constantly be focused in that area, which is expensive, but an important first step. Having said that, most failures now are occurring in phases II and III, when safety and efficacy are being evaluated. That includes off-target effects and statistical significance for response rates, whether compared to placebo or an existing product on the market.
That is where our technology can be very useful. If one patient responds, we can help identify the key attributes of that patient so additional, similar patients with the highest likelihood of responding to the drug candidate can be recruited. Our technology can also create the opportunity for drug developers to have conversations with regulatory agencies about the potential to expand beyond a specific population to others. Often, regulators are more open-minded because proposals can be backed by better trends data.
DA: Going forward, do you anticipate further development of your core engine, or is it primarily a matter of training from now on?
JV: Artificial intelligence involves lifelong learning. We use the Translational Index score to identify weaknesses and strengths, and we will continue to do that. For example, we looked at random schizophrenia drugs that failed clinical trials for different reasons and those that received FDA approval and are on the market. The Translational Index scores for the approved drugs just based on the structure of the drug substance and the target were 90% higher than those that failed. Schizophrenia is incredibly complex — and we did not train the model on schizophrenia or a specific outcome — but still got these amazing results. Imagine if every company used this system during early-stage development!
To get to your question, I see the evolution of our system coming down to how rare and complex the problems are that it is tackling. And, of course, whether the Translational Index score is really sufficient. If not, then we will continue to invest. It is going to be a constant, lifelong learning experience for us and the system we’ve built.
DA: What is the potential of BIOiSIM for drug repurposing?
JV: Drug repurposing is an amazing application for AI. We have been exploring this internally to generate our own assets. For instance, we looked at several approved drugs for their potential to combat pulmonary arterial hypertension (PAH), a rare condition with high mortality rates, as was seen in COVID-19 cases. We identified one that could be repurposed for this application, and within six months we received an Orphan Drug designation from the FDA. In less than two years, we are already preparing an investigational new drug (IND) package for the asset. This rapid pace was possible because there were already extensive data on the drug. Many companies hesitate to fully embrace drug repurposing because they are concerned about intellectual property issues, particularly composition of matter. But there are several ways to overcome those challenges for repurposing applications, and I expect people will get more and more creative about how to manage those applications.
DA: Is this work on the new drug tied to your subsidiary PulmoSIM Therapeutics?
JV: Yes. We created a subsidiary because this asset represents a huge opportunity for which we did not have to invest a lot of time or money. It is also a way to practice what we are preaching to others. Once this asset gets into the clinic, we will be able to share how we got there more openly. We are very close and very confident — we already know how humans are going to respond to this candidate because they have taken it before. We also already know where the challenges lie and have addressed them using our AI platform, such as changing the route of administration to avoid systemic side effects. It is a win–win for both patients with these rare conditions and our company because there was no need for massive investments. It is also a win for the companies that will co-develop the asset with us because they get access to a program with a lot of derisked value.
DA: How do your partnerships with drug development companies work?
JV: We work with clients in two different ways. For those with a molecule or target at a very early stage and data suggesting there may be problems for whatever reasons, we enter into a co-development partnership and jointly share the intellectual property in one form or another. We use our platform to identify one or more new candidates, which the client then takes further into a clinical program.
The second approach supports clients who are looking to avoid unnecessary risk at much later stages in the drug development process. They look to us to give them an indication of the probability of clinical success. Often, they have some experimental data that suggest potential safety, dosing, or other issues. These clients sign a fee-for-service, tech-enabled services contract.
For instance, we had a client with a small molecule that did not show a good profile in animal models, and they wanted to assess whether any of its metabolites would be more appropriate for development. In two months, VeriSIM Life evaluated close to 4,000 different metabolites and developed Translational Index scores for each of them and the actual candidate molecule. The molecule itself scored 3.5 out of 10, but two or three metabolites had scores of 9 — a huge difference. We are working with the client to improve the score of the molecule and significantly reduce the risks.
Both types of services are available to all companies, from small to large. We have a co-development partnership with the Mayo Clinic to enhance their programs, which leverages their significant expertise in many areas and direct learning from patients, physicians, and scientists using our system. Many of our smaller clients have single molecules that are in development, while some Big Pharma companies have high-throughput challenges with respect to lead candidate identification. Essentially, we work with many different types of clients on different protocols and approaches to address a variety of translational challenges.
DA: How does AI address and manage biases in data?
JV: On one hand, AI can help avoid some of the biases humans might have when interpreting the data. On the other hand, there are biases in how the data are captured, which is inherent to the way drug development is performed these days. In animal testing, for instance, the animals are from inbred species with the same genetic strains and are kept in controlled environments. That does not reflect the diversity of the human race nor the environments in which they live. AI can potentially address these gaps, because incorporating diversity is much more feasible in an in silico system than in a real animal setting.
Of course, we cannot completely eliminate biases because in general the bias in a data system is driven by the largest source of data. We need to look at how to tackle these unnecessary biases and reduce their impacts, which will involve strong guidelines around what type of data is used, ensuring that these data are appropriate for the particular application.
DA: To close, is there anything else you would like to add about your long-term vision and plans for VeriSIM Life?
JV: I think the next five years will be truly fundamental and game-changing for not just the company but for the AI field in general. We expect to have more drug partnerships and far more companies reach the market with drugs that have used our technology. In the next 10 years, I see the Translational Index score becoming the gold standard for companies to evaluate assets and M&A opportunities. There are many small companies that are founded with the intention of being acquired by a large pharma firm, and having a way to score the potential of their assets would help probable purchasers make comparisons.
I would also stress that AI is here to stay. The world we are living in today is not going to be the same world in the next five years. The concept of massive investment in R&D by large pharma companies is going to vanish — it is already happening if you consider all the layoffs making the news. There will be fewer people in R&D departments and more investment in sales and marketing, and the emphasis will be on finding partners who can help bridge these gaps by using AI to reduce the cost of drug development.
I also believe that the really successful companies of the future won’t be the same Big Pharma firms from today but companies that achieve more in shorter timeframes and thus are able to capture more of the market. I also see companies like Google, which are just entering the market today, becoming important players because they know how to find the right experts, and they can be agile. They have the potential to change the way drug development is going to be viewed. Once they figure out the process, it is a matter of repeating those steps thousands of times.
That raises questions about how the current workforce is being trained to manage these changes, which have already started and will only occur more rapidly in the coming years. They need to be prepared because these changes won’t take place slowly like in past industrial revolutions. Fortunately for VeriSIM Life, AI will be one of the easiest changes because it has so much to offer and can solve complex problems quickly. We hope to assist in the transition of the healthcare industry into AI-enabled Pharma in a time- and cost-effective manner!