By utilizing one of the fastest AI supercomputers in the world, a DGX A 100 housed at LSU, the DeepDrug™ Artificial Intelligence Platform for COVID-19 discovered a combination therapy that began human clinical trials in just 13 months after its inception — one of the fastest discoveries for a combination therapy in the pharmaceutical world. It is currently undergoing human trials in the Ukraine and at the Riverside University Health System Medical Center in California. The DeepDrug™ technology has been licensed by Skymount Medical.

Dr. Mukhopadhyay led the DeepDrug™ team to compete in the IBM Watson Artificial Intelligence XPrize semifinals. He has garnered more than $8 million in research grants at LSU from NASA, U.S. Department of Energy, National Science Foundation, Army Research Laboratory, Defense Advanced Research Projects Agency, U.S. Department of Transportation, National Geospatial-Intelligence Agency, state agencies and industry. He has published about 125 refereed papers in journals and proceedings of reputed conferences which includes the best paper award at the 20^th International Conference on Runtime Verification, 2020. He has three U.S. patents awarded and eight pending. He cofounded Ailectric LLC, an AI startup focusing on sound, text and image analytics. He serves as associate editor for IEEE Transactions on Artificial Intelligence and Remote Sensing Letters as well as on the program committee of AAAI 2022. In the last 12 years, he has graduated 11 Ph.D. students and has advised three postdoctoral research associates. He developed a new online course for LSU Online on Blockchain and Cryptocurrency that is offered in collaboration with IBM. 

The following is a Q&A with Dr. Mukhopadhyay adapted from an article in Pharmafile.

What have been the biggest milestones over the past eight years since the launch of DeepDrug™?

I had the first idea of using artificial intelligence, or AI, to accelerate drug discovery around 2007. However, at that time, AI was not really as mature as it is today. The AI revolution had not yet started. Also, I didn’t have access to high-performance supercomputers. So, I let the idea germinate. In 2009, I joined the LSU faculty and had access to LSU supercomputers, which was one of the reasons why I came to LSU. By 2012, the AI revolution had started. I also found a fantastic collaborator, LSU Department of Biological Sciences Associate Professor Michal Brylinski, who had just arrived at LSU. That’s when I decided to start developing the DeepDrug project.  

Since 2012, we have made several accomplishments. For example, we broke up four adenosine receptor antagonists into their molecular fragments and then recombined them to create an adenosine receptor molecule. Adenosine receptor paired with various G proteins can affect inflammation, pain and immune responses.

By 2016, we entered the IBM Watson Artificial Intelligence XPRIZE; and by 2020, we had reached the semifinal of the Artificial Intelligence XPRIZE. We were among 142 teams selected to compete worldwide. In February 2020, I gave the semifinal presentation at the TED World Headquarters in New York City.  After we came back from New York, we saw the first COVID-19 wave hit the U.S. Soon, there was a lockdown in Baton Rouge. At this point, we decided to pivot our engine towards therapeutics discovery for COVID-19. Today, a combination therapy is undergoing human trials both at the Riverside University Health System Medical Center in California and in the Ukraine.

What is the potential of AI in drug discovery and research? What will it change?

I will provide an analogy to explain this. If you think about the time period between 200 BC and 1500 AD, science and technology progressed at a very slow pace. For example, throughout this period, our mode of transportation was horse-drawn carriages and sail-driven ships. However, in just the last 300 years, suddenly we have had rapid progress in science and technology. We went from horse-drawn carriages and sail-driven ships to high-speed locomotives, cars, airplanes and rockets. What changed?

Let us look at physics. Before the 17th century, physics was mostly descriptive. It was called ‘natural philosophy.’  All of that changed when Sir Isaac Newton invented calculus, which brought in the quantitative aspect to physics. Now, we could compute and predict.

I believe that biology and pharmaceutical sciences, have been, for a long time, like physics before the 17th century, largely descriptive. What AI brings is the quantitative aspect. We can now compute and predict. This can change the nature of drug discovery.

Today, the time and cost needed to take a new drug to the market from scratch is 10 years and $2.5 billion. A lot of this time and cost required is due to the fact that only 12 percent of the drugs developed by pharmaceutical companies succeed. As a result, pharmaceutical companies have been reluctant to develop new drugs for sectors that have low profit margins such as infectious diseases. In fact, the last new antibiotic was developed by the pharmaceutical industry in 1987. AI can not only accelerate the identification of hits but also improve the success rate dramatically. So, now we will be able to take new drugs to the market at a fraction of the time and cost.

What developments in drug discovery do you anticipate in the next five years?

I think AI will be tightly integrated into the drug discovery pipeline. There will be a renewed focus on drug discovery for infectious diseases and other neglected global diseases. AI-enabled drug discovery pipelines should be able to handle pandemics in a much better way. Also, I think there will be lot of interest in nutraceuticals. In addition, given the ubiquity of genomics datasets as well as electronic health record data, I see the rise of personalized therapeutics.

How can AI help us improve global health and access to therapeutics?

AI can help repurpose generic drugs and their combinations towards emerging diseases thereby providing affordable therapeutics for the world as opposed to expensive treatments like monoclonal antibodies. AI can help develop environmentally friendly drug manufacturing processes. AI can also help identify nutraceuticals that can act as prophylactics and boost the immune system. In addition, AI can help discover lipid nanoparticles with certain properties that make them suitable for drug delivery.

How can AI speed up the response to outbreaks of previously unknown diseases?

By speeding up drug discovery, reducing the chance of failure and discovering affordable therapeutics, AI can help control outbreaks of previously unknown diseases. There is also the use of AI for diagnostic purposes. When you combine that, you get an end-to-end pipeline for handling such outbreaks.

DeepDrug™ can reduce the time and cost of drug discovery by as much as 90 percent. How will this huge reduction in resource allocation change things, and how will it allow the focus of drug development to shift?

For new drug development, DeepDrug™ can reduce the amount of time for new drug development by about 30 percent. However, through drug repurposing, DeepDrug™ can save up to 90 percent of time. Because DeepDrug™ can already filter out candidates that are likely to fail, you will need less in vitro and in vivo testing. It helps rapidly take a drug to human trials. Because of the drastic reduction in time and cost, pharmaceutical companies can now attend to neglected global diseases that have so far received very little attention.

The potential applications of AI are limitless. What are some of the most promising applications that you’re exploring currently?  

Outside of healthcare, we are doing fundamental research to make AI more reliable wherein it provides warning if it thinks that there is a possibility that it made a wrong decision. We call these confidence-aware AI. This is very important in biomedical applications like drug discovery or medical diagnosis. We have also performed fundamental research to prevent catastrophic forgetting in AI systems. For example, suppose you have an AI that has been trained to identify dogs. If you start training the same system to recognize horses, as its skill in recognizing horses improves, its skill for recognizing dogs fades away. We have developed new architectures, like the deep branching neural network, CactusNets, to remedy this situation.

In terms of applications, we are working on several environmental and geospatial uses. One of them is to predict wildfires before they take place and identify them when the first signs show up. This would allow emergency management agencies to deploy resources efficiently as well as respond to incidents in real time.

Another application is related to bird conservation. Migratory birds can land on toxic ponds that contain leftover water, minerals and other materials from mining called tailing ponds, which can kill them. This threatens several endangered species such as the Aleutian Canada Goose. We are using AI to track birds in flight and prevent them from landing on tailing ponds.

In previous work, we have computed the tree cover for the entire state of California at 1 meter resolution, where each pixel represents 1 square meter. This has applications in computing Above Ground Biomass, or AGB, density and Forest Carbon Mapping.

We have also developed AI techniques that can identify rooftops in high-resolution satellite imagery from small training datasets. The applications include post-disaster damage assessment, solar capture, mapping and planning urban and rural extent, urban heat island effects, etc.   

In another project with a graduate student, we are working on mapping dirt roads in the Saudi Arabian desert. In the Saudi Arabian desert, dirt roads often merge with the sand only to re-emerge later. The idea is to join these dirt road segments together to create a coherent map. This has several applications such as generating addresses and directions, computing land cover estimates, mapping natural resources, etc.  

In collaboration with LSU Bert S. Turner Department of Construction Management Professor Yimin Zhu, we are combining virtual reality and AI to optimize the design of energy-efficient buildings. Funded by the Federal Motor Carrier Safety Administration, we are trying to use AI to understand driver behavior, especially commercial vehicles, that leads to crashes. Funded by the U.S. Department of Transportation, we have developed AI-based models for understanding route-choice behavior of drivers in an urban transportation system.