Compandia, a Nottingham, UK-based startup, is taking a bioinformatics-focused approach to the competitive field of biomarker discovery.
The company recently launched its third software tool based on artificial neural network technology developed by co-founder Graham Ball, a reader in bioinformatics at Nottingham Trent University. Ball, along with Robert Rees, a mass spectrometry specialist and director of NTU's John van Geest Cancer Research Centre, founded the company around 18 months ago to identify biomarkers in high-dimensional data sets generated by mass spec, microarrays, next-generation sequencing, and other high-throughput experimental platforms.
The new software, called Pathway Distiller, uses ANNs to predict the interactions between biomarkers associated with a particular biomedical question. It joins the company's flagship technology, called Biomarker Distiller, and a follow-on to that called Risk Distiller. All of the company's software is offered through a services model.
The key to Compandia's technology is an iterative, stepwise approach to ANN modeling that first assesses the prognostic potential of each marker individually and then adds further markers in a sequential, multivariate approach in order to improve the accuracy of the classifier.
According to the company, the approach is able to identify much smaller biomarker panels than methods based on support vector machines, genetic algorithms, or decision trees. Furthermore, the ANN-based approach can identify sets of biomarkers that have higher predictive accuracy than other methods.
In a paper published in the February issue of Breast Cancer Research and Treatment, Ball, Rees, and colleagues describe how they applied the ANN approach to a breast cancer microarray data set generated by researchers at the Netherlands Cancer Institute. That study, published in Nature in 2002, identified a 70-gene signature that ultimately served as the foundation for Agendia's MammaPrint microarray-based breast cancer recurrence test, which the US Food and Drug Administration cleared in 2007.
Ball et al. noted that when they applied the ANN approach to the same data set, they found that "just nine genes were necessary to predict metastatic spread" with sensitivity and accuracy of 98 percent — a finding that "compares favorably with the computational approach used in the original manuscript," which resulted in 70 genes, sensitivity of 90 percent, and accuracy of 83 percent.
Andy Sutton, Compandia's CEO, told BioInform that Biomarker Distiller is the company's "workhorse tool," and the others are built upon the same underlying technology.
Pathways Distiller starts with a rank-ordered list of markers in a signature as determined by Biomarker Distiller, and then "turns the algorithm in on itself to start to look at interactions between biomarkers within the context of a given question," Sutton said.
Once the biomarkers are ranked, "then we ask, 'Which of those 100 or 200 biomarkers, within the context of the question that we're asking — so responder from non-responder, or patient versus control — is predictive of the others within that set?'"
The end result is a pathway map that shows directionality between individual markers, as well as the strength of the interaction, and whether the genes are up- or down-regulated — a data representation that Sutton said differs from pathway maps based on literature curation, which tend to present "single-line notation."
Risk Distiller, meantime, allows the company to predict the individual risk for each patient in a population. While Biomarker Distiller "forces a sort of binomial output — biological state A or B — Risk Distiller takes a continuous output, and then allocates a risk factor to each member of the population in transit from state A to state B," Sutton said.
That analysis allows the company to produce a "prospective Kaplan-Meier curve" for any given data set, he said. When the company applied the approach to the Netherlands Cancer Institute breast cancer data set, the prospective Kaplan-Meier curve "almost overlapped with the data that was published."
That capability is "getting a lot of interest from industry," Sutton said.
The company has not disclosed any of its current clients, but Sutton said that it "varies from very large pharma companies through to diagnostics companies and a large number of smaller companies who don't have the bioinformatics capacity and maybe the mass spec capacity that we have."
Compandia is still small, however, with only three full-time employees in addition to Ball and Rees, who still hold their appointments at NTU.
Last month, the firm appointed Dave Tapolczay, CEO of the technology transfer arm of the UK's Medical Research Council, as non-executive chairman. Tapolczay, a 23-year veteran of GlaxoSmithKline who led the pharma's technology-evaluation efforts, will advise the firm on its strategic direction.
In December, Compandia received £340,000 ($500,000) in seed funding from two regional economic development funds in the UK. Sutton said the firm will seek additional funding to support its current commercialization efforts, as well as to expand its model into diagnostic development over the next three to four years.
"We started life primarily as a service company … but what's become apparent quite quickly is that we're generating really interesting IP related to target discovery and biomarker discovery, and we're starting to explore grant funding to start to validate some of these signatures en route to products," Sutton said.
In the shorter term, he said, "we're now looking for grant funding to test out that nine-gene [breast cancer] signature for potential use as a product in its own right."
Sutton said that the company's ability to identify a handful of markers rather than a complex signature appeals to its clients, "who are looking for smaller biomarker panels on more robust platforms."
Currently the biomarker field is "falling into two camps," he said. "There are those who think that biomarker signatures with hundreds or thousands of genes, based on a chip-based platform, are going to be the way to go. And then there are other people in the market who feel that smaller biomarker panels on a qPCR platform is very much going to be the way to go."
While "it's yet to play out which is going to be the better model," Sutton noted that for most developers, "a small panel of six markers on a qPCR platform is potentially cheaper and perhaps a little bit more robust."
Likewise, while the firm has its roots in protein markers, "what became apparent very quickly is that it's probably of more value for us to look at nucleic acid-based data sets. I think in terms of where the market is going in terms of biomarkers, that's clearly where there's a bigger market opportunity for us."