In this study, we developed a signature-based method to index cell cycle phase distribution from microarray profiles under consideration of cycling and non-cycling cells, providing two sources of valuable information on cancers.
One source of information is the proportion of cycling cells in the sample. The rationale of most current cell cycle phase estimation methods, including mitotic index, S phase fraction and IHC against cell cycle markers, is that the high proliferative tumors leading to poor prognosis contain more cycling cells. In the analysis of the human breast cancer datasets, higher CCScycling
scores for the total gene dataset, indicative of a larger number of cycling cells in the sample, did associate with poor prognosis. Naturally, it can be thought that an increase in the number of cycling cells leads to a uniform increase in the number of cells at all cell cycle phases. However, some patients showed non-uniform changes in CCSphase
scores for the total gene dataset (Fig. , upper panel), suggesting that each cell cycle phase was not evenly changed. Similarly, Whitfield et al
. observed that some cell cycle-regulated genes did not express in correlation with proliferation status in some breast cancers [11
]. Furthermore, although the G1 phase is a part of the cell cycle, G1 phase marker cyclin D1
often negatively correlates with poor prognosis of breast cancers [2
]. Therefore, considering only the proportion of cycling cells seems insufficient.
The other source of information is cell cycle phase distribution. A number of oncogenic events are known to perturb the duration of cell cycle phases. For example, activation of oncogenes such as v-H-ras
, cyclin D1
, cyclin E
, and c-myc
shortens the G1 phase [24
]. Loss of tumor suppressor Pten
shortens the G1 phase [27
] and loss of Lzts1
shortens the M phase [28
]. Viral infections such as SV40-Tag and HTLV-1 Tax also shorten the G1 phase [30
]. Such perturbations in the cell cycle phase duration subsequently alter the cell cycle phase distribution. Thus, the cell cycle phase distribution per cycling cells would reflect the biology of cancers. Actually, in the analysis of mouse tumor models, oncogenic-event specific cell cycle phase distributions were observed. This suggests that the cell cycle phase distribution under consideration of both cycling and non-cycling cells has a potential for cancer characterization.
A model of tumors with different cell cycle phase distributions is proposed in Fig. . Oncogenic events perturb the cell cycle each in a unique way which in turn alters the cell cycle phase distribution as well as the proliferation rate. High proliferative tumors grow rapidly and thereby produce a large number of cycling cells. The opposite is the true for low proliferative tumors. However, high proliferative tumors with a small number of cycling cells or low proliferative tumors with a large number of cycling cells would exist at a low probability. This model would account for non-uniform changes in CCSphase
scores for the total gene dataset found in some breast cancer patients, the Whitfield et al
.'s observation, and the adverse prognostic value of cyclin D1
. Current cell cycle phase estimation methods are insufficient for detecting such cancers. Mitotic index and S-phase fraction do not recognize non-cycling cells. Combinatorial IHC [32
] still needs improvement and validation. Shetty et al
. reported a relationship between breast cancer grade and G1 phase length estimated from the ratio of geminin
IHC measurements; however, it was not significant [33
]. The CCS method, on the other hand, indexed the cell cycle phase distribution under consideration of cycling and non-cycling cells, and showed a potential for characterizing cancers.
Figure 5 A model of tumors with different cell cycle phase distributions. Oncogenic events perturb the cell cycle each in their unique way, which alters cell cycle phase distribution as well as proliferation rate. High proliferative tumors grow rapidly and produce (more ...)
Previously, as an alternative microarray-based cell cycle analysis technique, Lu et al
. introduced the "expression deconvolution" method [34
]. To predict the cell cycle phase distribution of yeast, they prepared about 700 equations with 5 variables representing 5 cell cycle phases and searched for the optimal solution. The method has comparable or even better potential to improve cancer characterization than the CCS method. However, it requires a tremendous amount of computational resources to find the optimal solution and avoid the local minimum, especially as the number of variables increases (18 + 1 phases were analyzed in our study). There are some hurdles that need to be overcome before high resolution cell cycle phase analysis is practical and we are currently tackling some of them.