A hallmark of artificial intelligence is its ability to learn from the past. As researchers advance and refine AI applications, it could also increasingly become part of routine research – the kind of work that can help and respond to progress in managing this pandemic and can support the response to a future one, too.
Finding meaning in a sea of messy or incomplete data is exactly what data scientists at the Pacific Northwest National Laboratory (PNNL) do, with experience using graph-based machine learning, detailed molecular modeling, and questionable artificial intelligence security and basic research. , PNNL researchers are now applying their artificial intelligence tools to investigate fundamental questions about treating COVID. What they learn sharpen the tools available in the Computational Toolbox to quickly respond to a future pandemic.
When a pandemic flu spread around the world about a century ago, scientists didn’t know that viruses existed; During the coronavirus pandemic, scientists had sequences of genetic information to track the spread of the virus and its variants, molecular details to develop rapid diagnostic tests, and tools to develop an entirely new class of vaccines. Some of these vaccines were approved in the US a year after the new coronavirus was first discovered.
Imagining individual treatment effects through counterfactual reasoning
Every time COVID19 cases occur elsewhere in the world, access to treatment becomes a problem. When there were more sick patients than treatment options, doctors have made difficult decisions about how to get the most out of the medical resources available.
One mindset that can be part of these decisions is counterfactual thinking.It compares the results of patients who have received treatment with their imaginary results when, contrary to the facts, they have not been treated, based on knowledge of similar situations with previous patients turned out.
AI algorithms can also use counterfactual reasoning, provided they have enough basic knowledge to draw on. The amount of COVID-related research over the past year has provided computer scientist Jeremy Zucker and his colleagues with a number of biochemical details about the novel coronavirus and how our immune system responds to it.
Taken together, these details can be represented using a data science approach called a knowledge graph. From this knowledge graph, the team derived a counterfactual model to answer a specific scientific question about the results of COVID19 treatment.
“With data science leveraging experimental biomedical knowledge about COVID disease progression and response to treatment, artificial intelligence can learn to more accurately predict the effects of treatments on individual patient outcomes,” Zucker said.
The team used such an artificial intelligence framework to simulate certain biochemical data from hypothetical patients seriously ill with COVID19. Each patient had different viral loads, received a different dose of a drug, and either recovered or died.
In each case, the team wanted to predict whether a surviving patient would have died without treatment with the drug or if they would have died if they had survived had been given a higher dose of the drug.
The analysis provided more accurate information about the potential benefit of the treatment for individual patients compared to algorithms that only predicted average patient outcomes after treatment.
Scientists reported on several case studies of their counterfactual algorithm in an article recently published in a special issue of IEEE Transactions on Big Data on COVID19 and Artificial Intelligence. and it is applied and evaluated in a DARPA Modeling Adversarial Activity project that uses scaled causal knowledge graphs to combat COVID19.
Molecular modeling to assist drug repurposing
Although vaccines against the new coronavirus are increasingly available worldwide, it will take time to stop the spread of the virus and its variants. Therefore, drugs to treat COVID19 will continue to be needed, and existing licensed drugs originally developed for other diseases could help.
A team of scientists from PNNL and the University of Washington (UW) School of Medicine examined more than 13,000 compounds from existing drug libraries to determine the ability to inhibit a vital protein made by genetic information in the novel SARSCoV2 coronavirus . Using high-throughput biochemical measurements in combination with artificial intelligence-based detection, their work identified a molecule from this collection with promising antiviral activity against SARSCoV2.
Wesley Van Voorhis and his team at the University of Washington used a cascade of biochemical tests to reduce to three hits the thousands of molecules that were potent inhibitors in experiments with purified proteins.
At PNNL, data scientist Neeraj Kumar and colleagues used artificial intelligence-based molecular models to predict where each hit would bind to the viral protein called nsp15. Chemist Mowei Zhou performed mass spectrometry measurements of every hit associated with nsp15 in its natural folded form, using resources from the Environmental Molecular Sciences Laboratory (EMSL), a user facility of the USD Department of Energy Office of Science based in PNNL.These measurements provided information about how strongly each compound bound to nsp15 and confirmed that one of the three compounds, a molecule called Exebryl1, bound to the protein.
In the results published in the journal PLoS ONE, the team showed that Exebryl1 had modest antiviral activity against SARSCoV2.
Exebryl1 was originally developed to treat Alzheimer’s disease. In screening tests, it did not have enough antiviral activity to be considered an immediate candidate for COVID19 treatment. However, artificial intelligence can help scientists modify the structure of Exebryl1 to improve its antiviral activity against the new coronavirus.
This work was supported by the National Virtual Biotechnology Laboratory, a consortium of 17 national US Department of Energy laboratories focused on responding to COVID19, with funding provided by the Coronavirus Aid, Relief, and Economic Security, or CARES, Act.
Developing an approach to accelerate drug discovery during this pandemic could reveal new design steps that could help in the next outbreak.
“Drug discovery and development is a complex, expensive, and time-consuming process, especially when you consider that most advanced molecules fail at the design stage in clinical trials,” said Kumar. “Computer screening incorporates chemical information during the development process to increase the potential for success of a drug candidate in clinical trials.”
Graph neural networks could generate tailor-made therapeutics
Another way to use artificial intelligence for drug design could be to create libraries of potential drug candidates that have never been seen before.
Chemists who develop drugs can identify the key features of a molecular structure that enable them to function. You can also dissect a structure to gauge how difficult it might be to make a molecule.
NLP computer scientist Sutanay Choudhury, data scientist Neeraj Kumar, data scientist Jenna Pope and their colleagues at the Argonne National Laboratory can recreate the same thinking process with artificial intelligence. The team uses graphical neural networks to generate structures for molecules that could be candidates for drug development.
Graphics provide a mathematical representation of the connections between the elements of a network; for example, how the atoms of a molecule connect to create a potential drug candidate. Neural networks based on such molecular graphs can learn to find patterns in data that might otherwise not be obvious.
To test their drug design methods, Choudhury, Kumar, Pope and their colleagues mapped ways to connect chemical components to create a drug-like molecule and identified which components contribute to the behavior of a molecule as a medication. Finally, they tested two methods of using graphical neural networks to assemble molecules using chemically relevant equations.
One method learned structural patterns of more than 7,000 molecules that are known to inhibit various viral proteases. The team found that this analysis tended to produce molecules similar to those in the known database.
In the other method, the algorithm built a molecule atom by atom and bond by bond, optimizing the desired drug and synthesis properties throughout the virtual construction. The team found that this method tends to produce molecules that were previously unknown.
Each approach has different drug development benefits. Reusing approved drugs could be a quick route to the clinic, and creating entirely new molecular structures injecting variations in the beginning of the notoriously difficult search for antiviral drugs, Kumar said.
This graphical neural network research was part of PNNL’s contribution to the Department of Energy’s (DOE) ExaLearn CoDesign Center, a group of eight national laboratories focused on machine learning technologies.
This hub is a product of the DOE’s Exascale Computing Project, launched in 2016 to explore the most persistent supercomputing. Data scientist Draguna Vrabie leads PNNL’s involvement in the ExaLearn center.
