LIMITATIONS, EXTENSIONS AND FUTURE RESEARCH DIRECTIONS

Perhaps the biggest challenges the current computational approaches are facing still originate from the limited coverage and accuracy of the interaction data available. Protein-protein interaction maps for yeast and human were estimated to contain only roughly 50% and 10% of all of the possible interactions in 2006 (Hart et al., 2006). There are various experimental techniques for identifying protein interactions, the major differences arising from their scale and ability to detect binary or complex interactions (Shoemaker & Panchenko, 2007a). However, all the techniques currently used in large-scale interactions screens are poor at detecting very transient interactions and those that depend on posttranslational modifications. Another technical complication comes from the fact that these techniques detect interactions that are biophysically possible but may not actually occur in the living cells. Moreover, the measured interactions are very much dependent upon the cell types studied and on the developmental stages or external conditions under which the measurements were performed. In particular, the completion of the human interactome will pose many experimental problems, not least because of its large estimated scale and relative sparseness (Stumpf et al., 2008). The context-dependency together with the large number of false positive and false negative interactions, that are bound to occur in the current human protein interaction datasets, also explains their small overlap (Ramírez et al., 2007; Futschik et al., 2007). These experimental limitations should be understood when developing and applying data mining methods. Computational strategies can also help to increase the coverage and accuracy of the interaction networks by predicting and prioritizing those components and their interactions, either observed or missing, that warrant further experimental testing (Qi et al., 2006; Qiu & Noble, 2008; Skrabanek et al., 2008). It can be expected that the development of data mining approaches to protein interaction networks will benefit greatly also from the ongoing community efforts to standardize the representation and storage of the interaction datasets (Orchard et al., 2007; Kerrien et al., 2007; Klipp et al., 2007), as well as from the integrated tools for accessing and querying the public interaction databases (Cerami et al., 2006; Strömbäck et al., 2006; Dhanapalan & Chen. 2007; Newman et al., 2008; Splendiani, 2008).

Albeit being error-prone and incomplete, the protein interaction networks mapped so far have already proved useful in understanding biological processes and have also triggered many biomedical applications with relevance to human health. The large-scale experimental datasets for human cell biology, in particular, that are constantly improving both in quality and coverage, are enabling computational methods to systematically address exciting biological questions, such as identifying the key players and their roles responsible for the multi-factorial behavior in complex diseases. An approach with a particular biomedical potential involves using protein interaction networks as priors when analyzing genome-wide gene or protein expression datasets; network modules can be used for distinguishing those genes and proteins directly involved in a particular disease process from the background variation originating from other biological factors or experimental noise. Besides such false positives detections, the use of connectivity patterns can help to decrease the number of false negatives as well, caused e.g. by those members of the module that are expressed at low or constant levels. The need of such integrated analysis is emphasized when the sample sizes are small and the phenotypes of interest involve also components with subtle expression changes. Accordingly, such network-based analysis of complex phenotypes is likely to be extremely useful in diseases like diabetes (Liu et al., 2007). Possible future extensions of the network-based identification of potential biomarkers include accounting for inter-individual variability in the disease responses: by aggregating the module-level expression changes between different subjects, it may be possible to provide multiple sets of molecular markers that correspond to different subsets of patients. Such modular approach to biomarker discovery should be robust to high measurement noise and intrinsic biological variation, thus enabling in the future tests to distinguish between different disease subtypes, progression stages, individual risk factors, or even individual response to different treatments.