This paper reports a more accurate—or sensitive—version of the program PSORTb, the most precise computational method for the prediction of bacterial protein subcellular localization. Such predictions are a key component of genome-based studies of bacteria, and an important step in the prioritization of possible new drug and vaccine targets against bacterial pathogens. Our paper also describes the first version of PSORTb that can make predictions for all bacteria.

“Perhaps most interesting to the biomedical research community is that fact that by using our tool, a researcher can identify surface-exposed proteins encoded in a genome very rapidly – roughly 100 times faster than using high-throughput laboratory methods.”

As bioinformaticians, we are writing for two audiences—the computer science community and the biology community. In the case of this particular paper, the method we report represents not only an important computational advance in the use of machine-learning methods for classification problems, but it also describes a tool that the microbiologist can implement to great effect in their everyday research.

Scientists from both of these fields cite our paper in their work—computer scientists refer to it when discussing approaches to biological classification problems, while microbiologists cite it in papers describing the characterization of particular proteins or even whole genomes. I think this speaks to the success of our interdisciplinary approach to research. By building a team combining the expertise of both computer scientists and biologists, we not only created a final product whose performance exceeded that of any tool we would have developed, but we also created something that has been well-received by a very diverse audience.

Our paper describes PSORTb v.2.0—an updated method for the prediction of protein subcellular localization in bacteria. In 2003, we released the first version of PSORTb, which implements a multi-component approach to prediction. It uses several different computational techniques to identify amino acid sequence features that are known to influence protein localization—signal peptides, transmembrane secondary structures, or sequence motifs, for example. In PSORTb v.2.0, however, we expanded the program’s predictive capability by introducing a new type of analysis, which combines data mining and support vector machine (SVM).

Previously, methods implementing SVM for biological classification had looked at proteins only in terms of their overall amino acid composition. While this does work, it doesn’t attain very high precision—amino acid composition can be thought of as a "low-resolution" characteristic of a protein, and one ends up with a fairly significant number of false positives. What our group did, however, was to look at proteins in terms of their structural and functional motifs.

By using data mining techniques to extract "frequent subsequences"—amino acid motifs that occur frequently in proteins found in a particular cellular compartment—we were able to use these frequent subsequences as the basis for classification. These motifs can be thought of as a "high-resolution" characteristic, and they result in much higher precision.

The software we’re describing in our paper takes a protein sequence—a string of text—as input, and returns a prediction of where in a bacterial cell that protein might be located—on the surface, inside, or in a membrane, for example. Our method is the most precise tool available for doing this, with a false positive rate of only 4%.

Microbiologists can now use this tool to generate highly confident predictions of protein localization—information that has a range of potential applications. Perhaps most interesting to the biomedical research community is the fact that by using our tool, a researcher can identify surface-exposed proteins encoded in a genome quite rapidly—roughly 100 times faster than using high-throughput laboratory methods.

These surface-exposed proteins are key players in the development of new drugs and vaccines. Thus our software is potentially able to speed up the discovery of new therapeutic targets significantly, a process which normally takes years.

The Brinkman Lab (Simon Fraser University, BC, Canada), which led development of PSORTb, has always been interested in combining bioinformatics with pathogenomics—the study of pathogenic organisms at the genome level. A secondary avenue of research that stems from Principal Investigator Fiona Brinkman’s earlier work is the analysis of bacterial outer membrane proteins. These proteins not only have a beautiful barrel shape, but they are also of significant medical importance—they are members of the group of surface-exposed proteins that represent drug and vaccine targets mentioned earlier.

The PSORTb project came about in 2001, when Fiona realized that the sole existing method for bacterial protein localization prediction, PSORT I, performed poorly at identifying this group of proteins. After a bit of investigating, she realized that the method had a number of other significant limitations, and so, with the permission of PSORT I developer Kenta Nakai of the University of Tokyo, she set out to develop a new localization prediction method.

I undertook the project as my PhD research in the Brinkman Lab, and over the next few years, we built up a team of collaborators from different fields, different institutions, and different countries. Although I’ve now moved on to a position as a postdoc at a new lab, Fiona and her lab are continuing development of PSORTb, and are also working on a version that makes predictions for yet more organisms.

By recognizing that we could not take on a project of PSORTb’s magnitude on our own and seeking out the best possible collaborators, we managed to avoid obstacles and roadblocks, and the development of PSORTb v.1.0 and 2.0 proceeded quite quickly and smoothly. Perhaps one of the biggest challenges in an interdisciplinary collaboration is effective communication between researchers from very different fields, but this is something we appreciated and we made an early and successful effort to overcome this.

I think the aspect of our work that matters most to the general public is its utility for drug and vaccine target discovery. According to WHO data, infectious diseases are the leading cause of productivity loss and are responsible for roughly a third of annual deaths worldwide.

Antimicrobial resistance is increasing at an alarming rate and researchers haven’t been able to generate new antibiotics as quickly as the bacteria have gained resistance to old ones. Newly emerging diseases are also causing considerable concern—including worries about the potentially significant economic and social impacts that new national or global disease outbreaks could have.

The discovery of new vaccine and drug targets has in the past been an expensive and slow process, but PSORTb has the potential to speed this process up significantly and really make an impact on the drug/vaccine development pipeline.