The ability to reliably predict pathway activity of onco genic and cancer signalling pathways in individual tumour samples is an important goal in cancer geno mics. Given that any single tumour is characterised by a large number of genomic and epigenomic aberrations, the ability mGluR to predict pathway activity may allow for a more principled approach of identifying driver aberra tions as those whose transcriptional fingerprint is pre sent in the mRNA profile of the given tumour. This is critical for assigning patients the appropriate treatments that specifically target those molecular pathways which are functionally disrupted in the patients tumour. Another important future area of application is in the identification of molecular pathway correlates of cancer imaging traits.

Imaging traits, such as mammographic density, may provide important additional Everolimus RAD001 information, which is complementary to molecular profiles, but which combined with molecular data may provide criti cal and novel biological insights. A large number of algorithms for predicting pathway activity exist and most use prior pathway models obtained through highly curated databases or through in vitro perturbation experiments. A common feature of these methods is the direct application of this prior information in the molecular profiles of the study in question. While this direct approach has been successful in many instances, we have also found many exam ples where it fails to uncover known biological associa tions. For example, a synthetic perturbation signature of ERBB2 activation may not predict the natu rally occuring ERBB2 perturbation in primary breast cancers.

Similarly, a synthetic perturbation signature for TP53 activation was not significantly lower in lung cancer compared to normal lung tissue, despite the fact that TP53 inactivation is a frequent event in lung cancer. We argue that this problem is caused by the implicit assumption that all prior information associated with a given pathway is of equal importance Plastid or rele vance in the biological context of the given study, a con text which may be quite different to the biological context in which the prior information was obtained. To overcome this problem, we propose that the prior information ought to be tested first for its consistency in the data set under study and that pathway activity should be estimated a posteriori using only the prior information that is consistent with the actual data.

We point out that this denoising/learning step does not make use of any phenotypic information regarding the samples, and therefore is totally unsupervised. Thus, our approach can be described selective FAAH inhibitor as unsupervised Bayesian, and Bayesian algorithms using explicit posterior prob ability models could be implemented. Here, we used a relevance network topology approach to perform the denoising, as implemented in the DART algorithm. Using multiple different in vitro derived perturbation signatures as well as curated transcriptional modules from the Netpath resource on real mRNA expression data, we have shown that DART clearly outperforms a popular model which does not denoise the prior infor mation.