Andrew A Radin Chief Executive Officer of Aria Pharmaceuticals, on the critical use of artificial intelligence in solving some of drug discovery’s most new and complex challenges.
Basic research and new technologies are uncovering insights every day that build on our understanding of human biology. At the same time, we are developing a better understanding of disease pathology, leading to a more complicated picture of how diseases manifest.
This creates two scenarios. The first is that in-roads made in decoding the complexity of human disease will make drug discovery more difficult and costly, limiting the potential to develop new life-saving therapies. The second scenario is that a better understanding of complex disease pathology will unlock new insights that lead to new, more effective and safer therapies. I believe this is where the future of drug research and development hinges.
One thing is for certain, as an industry we have always overcome new challenges that potentially stand in the way of finding breakthroughs. However, there is a pervasive problem that I would argue is central to the dilemma of solving for a growing complexity of disease. The pharmaceutical industry has traditionally focused on the lowest hanging fruit that the latest technologies can address. While the industry has benefited from that low hanging fruit, including high throughput screening and genomics, leading to substantial discoveries, there is more to be done.
I’ll preface this by saying the reasons to shy away from the latest technology are justified. While the industry has drawn criticism for lagging in this area, developing new pharmaceuticals carry much higher risks than new product development in other industries. It also carries a much higher price tag for failure. On top of that, pharmaceutical development is extremely complex in itself.
We are entrenched in the era of big data and are amassing an insurmountable quantity of human-derived data to better inform our efforts in finding treatments. Fortunately, we are also in the age of artificial intelligence (AI), allowing us to not only uncover novel biological insights, but also seek out treatments against very complex biology more efficiently and with more certainty. We have the tools at our disposal to meet this challenge.
How and where we apply that advantage is crucial to finding new breakthroughs. Scientists are already using AI to accelerate the traditional drug screening and drug discovery process. But most of that work is effectively speeding up a process that analyses a potential treatment against one or two types of data for a disease that is exponentially more complex and more intertwined than what this limited data represents.
The problem becomes that we are targeting diseases through one or two understood mechanisms of that disease, whether it be a single receptor or another target. We’re quickly learning that for every action there is an equal and opposite reaction in human biology that extends far beyond a single hypothesis. We need to reexamine our discovery approach with the acknowledgment that the most difficult diseases are not singular in nature. They are complex and interconnected with millions of signals and cellular interactions.
Therefore, the biggest challenge we need to solve for is screening methods that address the multi-dimensional nature of complex disease pathology. This is exactly where technology and AI gives us the advantage. Scientists have traditionally formed one hypothesis in which to seek treatments against, and through time consuming tests, screened compounds against that hypothesis.
Given what we know about the interconnected nature of diseases today and as our understanding of disease pathology grows, we need to rethink that approach. One such method that is proving out today is examining potential libraries of treatments against multiple, simultaneous orthogonal screens that could offer new insights into the potential of new treatments against any given disease. It’s akin to looking at the whole puzzle and how each piece of that puzzle is connected, rather than looking at only one puzzle piece and expecting profound insights.
Utilising AI, it is possible to build in silico models of any disease using dozens of heterogenous, human-derived data sources, then integrating potential treatments against that data in one process, breaking that data silo we face in traditional drug discovery avenues. Effectively, we can more accurately test any given molecule against a more real-world representation of a disease. This can have a profound impact bringing new insights of how treatments operate in the body against a disease.
My team and I recently presented early preclinical data from a promising compound in idiopathic pulmonary fibrosis (IPF) that has proven the utility of this approach. We built an in silico model of IPF using all exclusively human-derived data available on the disease in which to screen a large library of molecules against. Using an AI-driven approach, we analysed thousands of potential compounds against that model.
The results were twofold. First, we identified a potential treatment that had never been considered for IPF, a very complex disease where few treatments exist despite massive amounts of efforts thrown against trying to find new therapies. When we looked at those treatments against just one or two sets of data, they did not rank high for efficacy because the analysis only considered one dimension of the disease. What made the difference was integrating and simultaneously analysing those multiple heterogeneous data sources in one process. This gave us a far more accurate analysis of how those compounds would perform against the reality of the complexity of the disease.
Second is that because we conducted multiple orthogonal screens, we examined efficacy against a much more complete picture of IPF than what is traditionally accomplished in drug screening. This gave us confidence in the efficacy of the compounds, justifying the resources needed to move them through preclinical research. Three separate preclinical studies have confirmed the efficacy and safety of the compound so far.
Needless to say, we’re excited by those results but at the same time, we expected them. A benefit of this approach is that we’re able to also predict efficacy because we ensured the model was fully traceable, meaning we understood the rationale behind those efficacy signals. A tangible benefit of looking at a disease through multiple data lenses is that we can better understand how they will perform against the full picture of the disease.
Another benefit of this approach is that as new data becomes available, we can utilise that to build upon our approach because we aren’t confining our analysis to one or two types of data. This is important because we can use the problem of too much data and the growing complexity of any given disease to our advantage. It becomes an asset rather than a problem we need to solve.
These results by themselves are exciting, but what is most important is that we now have a potential therapy against an increasingly prevalent disease with few options today. A potential therapy that by all accounts would not be up for consideration in traditional drug discovery models that are based on examining drugs against a limited view of disease biology. And we have replicated this more than a dozen times with the same findings and same results. It’s an exciting time to be in drug discovery.
In summary, while in-roads decoding the complexity of human biology have made drug discovery more difficult, and this will continue, we have the tools within AI to uncover treatments against this new reality.
DDW Volume 24 – Issue 1, Winter 2022/2023
About the author
Andrew A Radin is the co-founder and Chief Executive Officer of Aria Pharmaceuticals. Radin created the company’s first drug development algorithms as part of his studies in biomedical informatics at Stanford University in 2014. Since co-founding Aria, Radin was named an Emerging Pharma Leader by Pharma Executive magazine, was invited to give a TEDMED talk, and was named a Top 100 AI Leader by Deep Knowledge Analytics.