The extraordinarily rapid drug development necessitated by the COVID-19 pandemic may have pulled AI- and HPC-accelerated pharmacology into the spotlight, but similar efforts have been underway for a wide range of disease applications for years. One of these applications is a subfield called safety pharmacology, which is exactly what it sounds like: understanding the dangers posed by medications. At the University of California, Davis, researchers are hard at work operating a novel safety pharmacology pipeline through cloud HPC, allowing them to rapidly screen drugs for their cardiotoxicity by scaling molecular-level data to tissue-level conclusions.
Cardiotoxicity is a particularly relevant issue for safety pharmacology: over the last few decades, most of the drugs that were approved by the FDA and subsequently pulled from the market were pulled due to cardiotoxicity, and as a result, cardiotoxicity is responsible for the withdrawal of a sizable portion (nearly ten percent) of approved drugs. This is to say nothing of the majority of drug candidates that don’t make it to FDA approval due to potential cardiac effects.
The issue, in essence, is that these effects are difficult to accurately model and predict. In clinical testing, researchers tend to look for a lengthening in the “QT interval” – one of the measurements seen on an electrocardiogram (ECG), but this metric has proved less than perfect in predicting serious cardiac effects from a drug.
Linking distant scales
Meanwhile, researchers at UC Davis had been tackling a different question.
“It was many years ago that we started to think about how we could link together a variety of modeling and simulation approaches that were being done by different laboratories in isolation,” said Colleen Clancy, a professor of pharmacology at UC Davis. “There were the atomistic scale people that were doing the molecular dynamic-type simulations, there were people using packages like Rosetta to do protein modeling and then there were a lot of people like me doing more function-scale modeling, where we’re looking at how variables change in time over much longer time scales.”
“We started to think: maybe we can use molecular-scale or atomistic-scale simulation to generate parameters for those higher-order models,” she continued. “And that led us to the idea of thinking about drug interactions. Could we use molecular dynamics and really think about how a drug interacts with its receptor, and use simulation to predict the rates of that interaction?”
So the researchers began building that pipeline. First, Igor Vorobyov, an assistant professor of physiology and membrane biology at UC Davis and a chemist by training, helps to load a 3D all-atom model of the drug in question into nanoscale molecular dynamics (NAMD), where the code simulates hundreds of thousands of atoms to understand how the drug binds to specific cardiac proteins. This is done around a hundred times using different permutations of the drug, capturing 40 to 50 nanoseconds of motion with each simulation, with the researchers compiling the overall rates of interaction.
“The rates that are produced by those simulations are then plugged into higher-order function scale models to represent how a drug interacts with its receptor over much longer timescales,” Clancy explained, crediting much of this translation work to Parya Aghasafari, a postdoctoral researcher at UC Davis. “And then we can put that drug receptor model into a model of a cardiac cell, connect those cells together through specialized HPC techniques to create virtual tissues and even whole organs, and then compute the signal average in time and space over that electrically stimulate tissue to create a pseudo-electric cardiogram.”
The researchers say that, to their knowledge, this is the first time those vastly different scales of analysis have been meaningfully connected.
The computational cost
Running such a large ensemble of detailed simulations, of course, comes at a computational cost. The researchers say they’ve been using a medley of HPC resources, including systems like Anton 2 at the Pittsburgh Supercomputing Center (PSC). But, Vorobyov said, “we have pretty limited resources there.”
This has led to expanded use and benchmarking of cloud HPC systems, including Microsoft Azure and Oracle Cloud. The researchers say they’re able to run about a microsecond (1000 nanoseconds) of simulated time on an Oracle Cloud virtual cluster (to which the researchers were provided access for benchmarking) within a couple weeks.
Still, the process is laborious.
“Currently, the pipeline is not very efficient, so the area for improvement that we’re going to try to be addressing in the future is to really think about how we can optimize and make the simulation pipeline more efficient,” Vorobyov said. “Maybe we don’t need to have a full umbrella sampling simulation for every single drug – maybe we could do that for 12 drugs, and then we could have a library of drugs that we sample other drugs again or compare our other drugs to.”
Despite those limitations, the research has already produced meaningful results that may help to inform analysis of drugs’ cardiotoxicity. “Parya’s work in that study also showed which parameter we need to track in the simulation that is most indicative or most predict of proarrhythmia – and it’s not the one that people have historically tracked – long QT – which is kind of the hallmark of a drug-induced arrhythmia,” Clancy said.
Instead, she explained, it was beat-to-beat variability in the QT interval that proved most predictive – a result that aligns with recent research stemming from other approaches, giving these researchers confidence that they’re on the right track.
Next, Clancy said, they need to work with pharmaceutical companies to obtain raw chemical signatures of unapproved drugs without foreknowledge of the companies’ internal safety pharmacology studies, which would allow the team to test their approach in a real-world, blinded environment.