Personalized disease signatures through information-theoretic compaction of big cancer data.

Background: Every individual tumor develops and grows in its own specific way, giving rise to a recognized need for the development of personalized cancer diagnostics. Tumors often grouped by the source of the cancer and a handful of potentially informative biomarkers. And yet cancers from the same source and with similar biomarker expression can have radically different responses to treatment. We suggest that this is due, at least in part, to different unbalanced processes within in the tumor, and by identifying the unbalanced processes suitable diagnosis and treatment can be selected for each individual tumor.

Methods: We have recently pioneered the application of surprisal analysis, a thermodynamic based information-theoretic approach to biological systems. This technique allows to analyze large datasets and to discern the altered, unbalanced molecular processes made of specific transcriptional networks that govern tumor heterogeneity. We study a number of datasets spanning five different cancer types, aiming to capture the extensive inter-patient heterogeneity that exists within a specific cancer type as well as between cancers of different origins.

Results: Using 506 tumors, from 5 different cancer types, we show that this diverse collection of tumors can be characterized by only 12 unbalanced processes, demonstrating extreme compaction of the data. Every patient is characterized by a small specific subset of unbalanced processes, leading to 144 different patients-specific signatures, which represent 144 different cancer diseases instead of 5.

We show that different patients may display similar oncogene expression levels, albeit carrying biologically distinct tumors that harbor different sets of unbalanced molecular processes. Thus, tumors may be inaccurately classified, and addressed as similar.

Conclusions: The approach enables us to combine datasets from different experiments and enable us to extract significant signals from the large dataset, gaining in depth, unbiased, patient-specific information. Surprisal analysis efficiently uncovers the altered transcriptional networks in every individual patient, potentially allowing improved classification of cancer patients. Our finding that similar oncogene expression levels in different patients may stem from distinct sets of unbalanced processes underscores the need to extend the initial analysis of tumors based on tumor specific biomarkers to patient-specific, comprehensive transcriptional networks for increasing the resolution of cancer patient diagnosis.