MUSC, HPE make innovative drug discovery software open source

In yet another innovative partnership between academia and industry, MUSC and Hewlett Packard Enterprise (HPE) are making available to researchers worldwide an innovative new drug discovery program they co–developed, PharML.Bind, in an open-source release. Through this release, MUSC and HPE aim to accelerate the search for effective therapies against COVID–19, the disease caused by the SARS–CoV–2 virus. 

The novel coronavirus pandemic has created many challenges to biomedical infrastructure, and chief among these is the need to identify effective therapies for treating COVID–19. The decision to share, rather than license and market, PharML, is in direct response to the world’s immediate and increasing need for effective therapies. 

“The smart move in the middle of a pandemic is to take a fast look at repurposing existing drugs, which we and many others are doing,” said MUSC researcher Yuri Peterson, Ph.D. “But what we really need, and where artificial intelligence like PharML can shine, is to get ahead of COVID–19 by finding the right drug — the perfect drug, if you will — that can limit this virus’ ability to survive, reproduce and continue to wreak havoc on our world.”

Troy Huth, J.D., Ph.D., associate director of the MUSC Foundation for Research Development, agreed, adding that with PharML’s massive scale and speed, researchers can rapidly search for drugs that target the pathogen itself, as well as processes critical to its survival and lifecycle, for vulnerabilities that could result in a treatment or cure. 

“The MUSC Foundation for Research Development strongly supports making this highly valuable software open source in order to drive global innovation and positively impact drug discovery efforts aimed at solving the COVID–19 crisis,” Huth said. “The open sourcing of PharML also demonstrates an important way in which universities and their technology commercialization offices can make significant progress toward solving the challenges of our times, including this pandemic.” 

While PharML is in its early phase and was meant to be the core in a number of long–term drug discovery and personalized medicine applications being developed by Peterson and his team, he, along with colleagues at HPE, came to the conclusion that making the code and training files open–source was the best course of action to facilitate, most rapidly, drug development in an effort to fight COVID–19. 

 

Jacob Balma, an artificial intelligence engineering researcher with HPE and part of the PharML team, said that by open–sourcing this work, they hope to change the types of problems people can solve at home on their laptops and open the door to a new class of problems for the scientific community to explore on supercomputers. 

“Making PharML widely available is an important step toward providing the world with true high–throughput, open-therapeutics technology,” said Balma. “The drug discovery pipeline, which this framework aims to accelerate, is currently the limiting factor in the time it takes to design and repurpose compounds to treat diseases. With PharML, we are making it easier for others to join the race to find treatments to address the COVID–19 pandemic.” 

PharML solves that. PharML can be used to predict which drugs will work if/ when a virus or other pathogen mutates so researchers can try to get ahead of future outbreaks and create a pipeline of future medicines.

PharML uses neural network architecture to estimate what drugs might be effective for specific disease targets, and it does so on a massive scale. PharML was trained on an extensive real-world data set containing millions of data points. It achieved a prodigious 98.3% accuracy rate, rapidly and accurately searching known drugs to find potential candidates for treating COVID–19 and other diseases. 

Peterson explained that AI in the biomedical field takes on many forms, much of it focused on improving human interactions with expansive data like medical records, scientific abstracts or medical diagnostic imaging. PharML, however, takes a chemistry-oriented approach. 

“We set out to test if we could use machine learning and neural nets to predict accurately all of the characterized drug–protein interactions that can happen in a human body,” Peterson said. Our testing indicates that it is very feasible to process with high accuracy millions of distinct and complex interactions while being light on computing resources,” he said. 

The team recognized that this provides the kernel of a much bigger project in which open therapeutics could serve to remove roadblocks in drug development and open the door to more complex problems like emerging pathogens and personalized precision medicine. 

The software, an innovative drug and mechanism–of–action (MOA) evaluation graph-based deep neural network (NN) architecture, began in early 2017 as a learning and discovery project. The team is now doing hands–on testing of the feasibility, utility and accuracy of artificial intelligence in the drug development process.

The team initially mapped out a strategy to consolidate more collaborators, test real–world predictions in a preclinical setting and publish in an academic journal. However, the advent of the COVID-19 pandemic accelerated this timeline, and the urgency of releasing the code and training files under an open– license became apparent. 

The developers expect the code to be a huge time saver for coders, developers and researchers wanting to use artificial intelligence to find new COVID–19 therapies. The code was designed on the front end by Balma and Aaron Vose at NanoSemi to be fast, efficient and reliable in parallel processing environments, making it easy both to modify and to use. PharML can be used as a stand–alone drug discovery or evaluation program that will predict the likely target (MOA) of any drug that has some similarity to a drug that has existed before. PharML can be integrated with more processes to do fundamental things like predict individual drug disposition, aiding in drug combination studies and in personalized medicine, both in terms of outcomes and genetic variations. 

PharML allows users to explore solutions to the long–standing problem in pharmacology of drug-induced side effects that emerge from the tendency for disease–implicated proteins to become the primary focus for further study, leaving behind the potential interaction drugs might have with all of the other proteins in the body. 

“Because PharML can rank–order full proteins in a compound–focused study, or just as quickly solve the inverse problem of rank-ordering compounds relative to a specific protein in a target-focused study, we can effectively see where further study is warranted from the short lists generated by PharML,” Balma explained. “This quickly narrows the search space across databases containing millions of compounds and tens of thousands of proteins down to the few which matter most … and we can do that in minutes.”

The collaboration between MUSC and HPE, Peterson said, has been a great showcase of teamwork. He stressed that both teams bring a lot to the table, and the skills of each team dovetail perfectly. 

“MUSC brings to the table a strong foundation in pharmacology and therapeutics and a practical understanding of the principles of coding, software and hardware. My friends at HPE have good fundamentals in chemistry and biology and are fantastic at coding, software and hardware. By drawing on each other’s strengths and staying open, we have been able to communicate effectively as a team, ask important questions, interrogate the results and make consistent and positive progress.” 

For more information on PharML or Peterson’s work, contact the MUSC Foundation for Research Development at frd@musc.edu or (843) 876-1900.