By Ben Butkus
DUBLIN – Bioinformatics startup Azure PCR is developing an automated, machine-learning-based data-analysis method for qPCR that it says will remove the ambiguity of data obtained using different testing methods and potentially improve the development and performance of molecular diagnostic assays, PCR Insider has learned.
The company is currently collaborating with various academic institutions to validate the technology, and hopes to have its first commercial product – web-based software for analyzing large PCR datasets from any type of experiment – ready for market by the end of next year, company executives said.
In addition, the firm is eyeing the development of a sub-$10 microchip that could be integrated into benchtop or point-of-care qPCR testing platforms to automate and improve data analysis and increase the accuracy of tests run on the platforms, the executives said.
Azure, co-founded late last year by software engineers Aron Cohen and Ze'ev Russak, debuted at Select Biosciences' qPCR Europe conference, held here this week.
In a presentation at the conference, Cohen, currently CEO of Azure; CTO Russak; and President David Kennard outlined some of the inherent problems in current qPCR data analysis — in particular, the fact that data from experiments can be limited in its accuracy and reproducibility due to difficulties in human interpretation, sample and test variability, and other mathematical complications.
One key problem, outlined by Russak, is that manual data visualization often becomes difficult when attempting to observe subtle but meaningful differences in graphed PCR curves.
"Curve fitting tells you nothing about the PCR data behind it," Russak said. A result of this is "a mistrust of PCR results because of ambiguity between different data-analysis methods," he added.
Azure's solution is designed to fully automate interpretation of real-time PCR data accurately and rapidly, using machine learning techniques that do not require the input of parameters by the user. As described on the company's website, this "eliminates the need for manual interpretation and training altogether and provides objective, standardized results."
During their presentation, Azure's executives, citing proprietary issues, were guarded about sharing the exact algorithms it has developed to automate and improve data analysis, leading some audience members to question exactly how the company's technology works.
In an interview following the presentation, Russak and Cohen further explained that the need for improved data analysis has arisen because many researchers just accept the multiple "black-box" analyses of data provided by different PCR platforms without questioning the methodology or results.
"The problem is that people have been doing the same kind of analysis … and using the same kinds of tool sets for years," Russak told PCR Insider. He added that while he expected some pushback from conference attendees regarding the new approach, it was not as strong as he anticipated "because I think most of the people here know of the problems that I spoke of. I did not invent these problems. I've heard this from many scientists in the field about the ambiguity of interpretation."
Russak cited an idea, born from multiple research studies of machine and human learning, that "the worst machine is better than even the best people when it comes to [interpreting] a large data set. This has been proven statistically over and over again."
Cohen added that "once people become reliant on certain tool sets, and especially if these are being used routinely and commercially, it's very difficult … to admit the genuine flaws and errors that can occur. Mathematically we've demonstrated how and why these problems can occur."
Around the time Azure was founded in December, the company filed for PCT patents surrounding its methods. The company then turned its strategy to developing interfaces for its algorithms "to make it easy and demonstrate its effectiveness," Cohen said. Azure also began validation studies with various academic institutions.
Thus far, Azure, which is headquartered in London but maintains an R&D facility in Israel, has disclosed initial data from studies carried out in collaboration with London-based teaching hospital St. George's Hospital; and the West of Scotland Specialist Virology Centre.
Specifically, the company used its software to analyze 644 patient samples from H1N1 flu testing, and it correctly identified 2,554 of 2,576 target markers for 99.1 percent accuracy. In addition, the Azure software correctly identified all 2,442 target markers from 1,221 samples from hepatitis B molecular tests.
Both studies were conducted using a Qiagen RotorGene PCR platform, Cohen said. According to the company's website, the software works with the RotorGene 3000/6000 cyclers and Life Technologies' ABI 7500 platform.
Cohen said that the company is also collaborating with other institutions in the UK and overseas, but declined to identify them, citing the early nature of the research.
Azure will continue to conduct these validation studies with various collaborators, and hopes to begin testing its software in a clinical setting early next year.
In the nearer term the company will offer its software via the cloud, primarily for high-throughput use such as assay development and validation at larger companies or academic institutions conducting large-scale molecular diagnostic studies.
"The initial offering will likely be a fully scalable, automated cloud solution, which will deliver results almost instantaneously for as many samples as you can throw at us from wherever," Cohen said.
Longer term, however, Azure is focused primarily on the molecular diagnostic market, because "that's where we see this having the most immediate impact.
"We know that within the diagnostics [market] … there is either difficulty or resistance to using the web for routine services," Cohen said. "In developing nations, it's because web access isn't permanent in field hospitals; and in Western countries, people are scared that if the computer goes down, they can't do any more diagnostics."
As such, Azure hopes to offer to this market "an extremely low-computation-intensity solution … on a sub-$10 microchip that can be incorporated with existing cyclers … or handheld machines," Cohen added.
Eventually, the two commercial paths may cross, "so the cloud would be used for test development and validation, and when the test is developed, the chip approach would be used."
The diagnostic application of Azure's technology may depend on the successful development of POC molecular testing platforms, of which there are few on the market but many in development.
"We're waiting," Cohen said. "I would imagine that the major problem that portable device manufacturers are having has to do with their data analysis, and the way in which data is handled. And we have other advantages to our technology that we can't go into right now that we believe will be useful in portable devices."
Azure has obtained an undisclosed amount of early-stage funding, Cohen said, but declined to provide details, adding that the company is "doing OK" in that aspect.