Several studies have described chemotherapy response in ovarian cancer using gene expression profiles, as summarized by Helleman et al
. However, the number of ovarian cancer specimens used for the gene selection in those studies was relatively small, ranging from 6 to 119, and the corresponding gene sets discovered to be associated with platinum-based chemotherapy resistance exhibited a wide range of 14 to 1,727 genes where only seven genes were observed as an overlap and each between only two gene sets 
. Lack of overlap between the discovered gene sets is likely due to the limited sample size in most studies. However, ours is the first study performed on such a large scale, two genes in the 227-gene set, EPH receptor B3 (EPHB3) and nuclear factor I/B (NFIB), had been identified in one of the previous studies 
, and one gene, RNA binding protein 1 (RNABP1), had been identified in a different study 
. More prominently, a gene set discovered on a large data set undoubtedly has high statistical power and robustness in accurately predicting chemotherapy response. Recently, the TCGA research network identified 193 prognostic gene signatures predictive of OS, but the gene association with chemotherapy response remains unexplored 
. Here we used a large sample set (493 samples from TCGA and 244 samples from an external source) for identification of molecular and morphologic signatures that are associated with chemotherapy response. The predictive model on the basis of gene signature revealed an accuracy of 87.9% in correctly classifying refractory from responsive tumors in the TCGA training set and stratified patients in both the TCGA validation set and the Australian data set into groups that demonstrated significant discrepancy in tumor progression, suggesting the capacity of the gene signature to serve as a mechanism to stratify patients with respect to treatment.
The imaging approach stratifies the cells into 10 bins based on nuclear size and accounts for the heterogeneity of cells in a tumor population. Our stratification revealed that most significant morphologic features differed between the chemosensitive and chemoresistant groups in the larger nuclei (range, 300 to 500 pixel2
; Table S2
). However, nuclei within this size range account for a very small percentage (approximately 2.0%), and the majority of the nuclei (approximately 98.0%) do not show a significant difference in chemotherapy response. This observation not only is consistent with the Goldie-Coldman hypothesis 
that only a small cell population may contribute to differential response to chemotherapy, but also suggests the difficulty of a conventional approach of simply correlating the overall morphologic differences with chemotherapy response, owing to the “dilution” effect 
. Therefore, our imaging approach allows us to interrogate different cell populations separated on the basis of nuclear size in a high throughput and automated fashion.
The 15 morphologic features (Table S4
) most weighted in achieving the patient separation are highly instructive. The same nuclear parameter might exhibit different or even opposite patterns. The average roundness of nucleus in Bin 8 (Mean_Ro_Bin8) is significantly higher in the chemoresistant group (P
, Figure S2A
), on the contrary, the same nuclear parameter in Bin 9 (Mean_Ro_Bin9) shows significant decrease in the chemoresistant patients (P
0.0015, Figure S2B
). The average roundness of the entire nucleus per sample (Mean_Ro_Total) shows no significant difference (P
0.56, Figure S2C
). In addition, none of the image features calculated from the entire nucleus per sample, the way similar to those used in other studies 
, show significant difference between the chemoresistant and chemosensitive patients. This discrepancy from the previous studies 
likely results from the number of nuclei used in the feature calculation. We used approximately 4000 nuclei per sample for feature value calculation, almost 80 times more than the amount used in the other studies 
. Taken together, our approach of binning the nucleus size and then assessing the image feature in each individual bin improves the image feature resolution and enhances the discriminating power. Furthermore, our approach of calculating the morphologic features in separate bins (with smaller size variations) is capable of alleviating the size dependence of some of the features, such as circularity and roundness 
Aside from the potentially practical value, the morphologic features also provide insights into cancer morphogenesis. The chemosensitive patients exhibit a smaller value of nuclear roundness in Bin 8 (Mean_Ro_Bin8), but with a larger variability (Std_Ro_Bin8) and a larger aspect ratio (Mean_AR_Bin8). Such morphologic differences likely result from the active response of the cells to their environment and heightened cellular metabolism, that is contributable from different molecular regulations (, Table S7
). This is further corroborated by pathway analysis, which revealed the gene enrichment in the morphologic function at cellular, tissue, and tumor levels (). The gene content of this table offers potential insight into the structural and molecular mechanisms of the chemotherapy response. The importance of A2M gene expression is of particular interest, in view of past work suggesting a correlation between decreased A2M levels with sensitivity to drugs 
. A2M is an inhibitor of matrix metalloproteinase activity, which is reported to contribute to tissue remodeling and morphogenesis 
. PAX6, which is associated with drug response, is strongly activated by cotylenin A in retinoblastoma cell lines 
. Decreased expression of EPHB3 in the chemoresistant group may have promoted chemoresistance by impairing the apoptotic response to cell damage 
Morphologically related genes at cellular, tissue, and tumor levels.
In conclusion, a gene signature discovered on a large data set provides robustness in accurately predicting chemotherapy response in serous OvCa. Meanwhile, we propose a novel approach for tumor nuclear image profile generation by characterizing patients with nuclear features (such as size, aspect ratio, and roundness etc) in incremental bins, and we demonstrate that the tumor nuclear image profile exhibits a strong association with chemotherapy response. This imaging approach is capable of accounting for cell heterogeneity and improving the discriminating power. The integrated approach herein, using gene expression profile that predicts chemotherapy response coupled with the morphologic features to stratify patients to the most appropriate treatment regimen, represents an important step toward the goal of personalized cancer treatment by identifying the area where novel drugs can be developed. Although our observations suggest that the tumor image profile is capable of defining prognosis and yielding mechanistic insights into the process of chemoresistance, one limitation of this study is the lack of validation of the image analysis due to unavailability of the independent image sets especially in a large population. This issue should be addressed in the future in order to determine the ultimate value of this technique in clinical practice. Besides, the resolution dependence of the morphologic features in separate bins has not been systematically investigated yet in this study and deserves attention in the follow-up studies. Future work also consists of inclusion of more possible morphologic features and verification of the gene-feature relation identified in this study.