To date, most of the high-throughput studies of cancer were focused on the gene signatures. Here we present the technique that uses the genome-wide expression pattern as a mean to diagnose tumor. Each human cells is a dynamic system occupying a specific position in the multidimensional phase space spanned by expression levels for of its genes (an ‘attractor’, or a stable position). Our studies of the cancer cell dynamics showed that the malignant transformation could be reflected by a combination of attractor-like behavior of the dynamic cell transcriptome with an action of ‘master genes’. This observation led us to investigate possible middle ground between diagnostic use of the discriminating signatures and entire expression landscapes in cancer biomarker research.

We extracted: A) datasets comprised of paired normal and tumor tissue samples collected from the same individual (N = 8); B) datasets comprised of a group of normal and a group of tumor samples collected from the same tissue type across a number of subjects (N=16). A total of 7 datasets were comprised of three or more groups of samples with different degrees of malignancy. For each sample, the global and specific expression distances (Dglobal and Dspecific) were calculated both from the Normal Sample Space (DN) and Cancer Sample Space (DC) based on: 1) all transcripts present on the chip and 2) genes with significant expression changes according to Mann-Whitney test. For each of the datasets, linear graphs depicting the relative distance of every given sample to the Normal Sample Space as defined by DNGlobal and DNSpecific metrics were generated. Both DNGlobal and DNSpecific correctly placed the most malignant tumors farther from the normal tissue control than the least malignant tumors or relatively benign tumors precursor states. In cases when easy visual discrimination of the tumor and normal/benign samples was achieved, the performances of DNGlobal and DNSpecific were comparable. These results suggest that the genome-wide metrics may help to assess the ‘degree of malignancy’ of the tumor cells. Mutual correlation between DNGlobal and DNSpecific distances was estimated by the PCA on the four dimensional space spanned by these four indexes (DCglobal, DNglobal, DCspecific, DNspecific). The patterns of the component loading were remarkably consistent across all the 24 (including multi-stage) datasets analyzed. In analyzed datasets, the relative importance of cell-kind driven gene expression regulation (PC1) was ranged from 77% to 98%, while the distinction between normal and cancer poles (PC2) was ranged from 22% to 1%. PCA analysis supported the two-pole structure of the cell-kind attractor with normal and cancer poles. The remarkable similarity of the observations made using multiple independent datasets, including these comprised of multiple types of samples demonstrates robustness of the genome-wide expression signatures as a mean to diagnose tumors.

Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 101st Annual Meeting of the American Association for Cancer Research; 2010 Apr 17-21; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2010;70(8 Suppl):Abstract nr 111.