July | August 2006 | Identifying decisive proteomic biomarkers with mass spectrometry has proven difficult. Noisy data, differences between instruments, and the lack of standards (protocol and data reporting formats) have all complicated the effort. To help solve these problems, the Natonal Cancer Institute (NCI) has awarded roughly $1.4 million to statistics software specialist Insightful to develop analytic methods and a software suite — S+ Proteome — for processing and analyzing protein mass spectra data.
The award is for Phase II work on a Small Business Innovation Research (SBIR) contract following successful completion of the six-month, proof-of-concept Phase I.
“What’s become clear is that biological data are inherently noisy. It’s got a lot of junk in it that we’d like to properly filter,” says Keith Baggerly, a consultant to the project and professor in the department of biostatistics and applied mathematics, M.D. Anderson Cancer Center. “We’re working [with Insightful] on a software system that can take a set of spectra generated by profiling tumors from a group of cancer patients and from healthy controls that can quickly establish which peaks are associated with the development of the disease.”
M.D. Anderson played an early role in flagging shortcomings with efforts to define proteomic biomarkers: “We wrote one or two cautionary papers pointing out in some cases proper processing had not been applied and some of the results were, in our opinion, wrong. We wound up going around giving several talks on how you need to design these experiments. In the process, I and others in my group wound up chatting with Insightful.”
Insightful, known for its rigorous statistics software used in microarray analysis, had responded to a NCI request for proposal. “A typical microarray has about 50,000 genes, and on some of these SELDI and MALDI (mass spec) platforms for proteomics, you’ve got 20,000 or more markers as well. It’s very similar problem,” says Michael O’Connell, director of life sciences at Insightful.
Baggerly notes the NCI-sponsored effort seeks to do more than just produce software; it also seeks to identify standard analysis methods. “One of the things plaguing this field is that, in effect, people have been rolling their own with respect to experimental protocols and with respect to analysis methods. It can be extremely difficult, even given the raw spectra from another group, to reproduce the analysis,” he says.
Some support has been gathering around mzXML, though that standard isn’t yet set, and the Human Proteome Organization (HUPO) also has ongoing standards efforts. It turns out that S+, Insightful’s programming language, “is actually pretty good at loading in data in various formats,” according to Baggerly, and should help ensure that S+ Proteome has the needed flexibility to handle different data formats and protocols.
During Phase I, Insightful outlined the structure of the tools and developed some of the data import capacity, leveraging its expertise with time series data and wavlet analysis techniques, which are used to do much of the de-noising and peak selection on mass spectra data.
“We’ve already released what you might call a beta version,” says Jill Goldschneider, research director for Insightful, “and it’s being evaluated by our consultants, and we made it available to NCI’s Early Detection Research Network (EDRN). Our plan is to do a release every six months.”
Using mass spec for protein biomarker identification presents several challenges; for example, jagged spectra obtained at different times must be aligned and the inherent instrument “noise” — typically both background hiss and high frequency — must be accounted for.
“It sounds like a simple question, if I look at a graph, where’s a peak on that graph?” Baggerly says. “But there are some weirdnesses. Whether you’re on the left end or the right end, peaks will have different shapes. It’s very hard to design a filter for peak type A and have it universally applied. So we’ve been working with Insightful to develop adaptive methods for different shapes based upon wavlet techniques that will go through and identify these peaks more readily and extract them quickly.”
Also critical are so-called “bookkeeping” features being built into the software. They track where the samples came from, what preprocessing (e.g., filtering) has been done, etc. Baggerly cites an instance in which his lab received a set of 192 spectra to analyze, but the results were confounding.
“We went back and looked at all 192, one at a time. The problem was they had intended to calibrate them, match peaks to masses, all at once,” says Baggerly. “They got the appropriate equation to calibrate everything and went to the spectra and selected them all and said, OK, we’re going to apply this calibration to all of them. Then somebody’s finger slipped. Instead of having all the spectra selected, they had one selected. Only one of the 192 spectra we got was properly calibrated. All the others were off.”
More detailed tracking would have found the error quickly.
Presumably, Insightful will make S+ Proteome compliant with NCI’s caBIG (Cancer Biomedical Informatics Grid) initiative. Timing of the release and cost of a commercial product based on the SBIR work are still undetermined.