Unsupervised Analysis Uncovers Changes in Histopathologic Diagnosis in Supervised Genomic Studies

Transcription

1 Technology in Cancer Research and Treatment ISSN Volume, Number 2, April (2006) Adenine Press (2006) Unsupervised Analysis Uncovers Changes in Histopathologic Diagnosis in Supervised Genomic Studies Human gastrointestinal stromal tumors () have recently emerged as a distinct mesenchymal tumor type that has a unique phenotype characterized by a gain of function mutations in c-kit. In contrast, leiomyosarcomas (LMS) of the gastrointestinal tract or retroperitoneum, which were previously classified together with s as gastrointestinal sarcomas, have much less frequent mutations of c-kit. We performed microarray analyses to gain a comprehensive understanding of the difference between the two types of soft-tissue sarcomas at the level of gene expression. Microarray experiments were performed on 0 s and 0 LMSs that were collected at the time of surgical resection. These tumors were categorized based on the histopathologic diagnosis recorded in our institutional database. Prior to our search for genes that are differentially expressed between these two types of cancers, we first carried out an unsupervised analysis using multidimensional scaling (MDS) to determine whether the two groups have marked overall differences in gene expression. Initially, the MDS did not reveal a good separation between the two groups. We then re-reviewed the histopathology of these tumors and realized that some of the cases included in our study were acquired 10 years ago when the diagnosis of gastrointestinal sarcoma was made according to histopathologic criteria alone without immunohistochemistry for c-kit. An experienced pathologist reviewed all of the specimens and this revealed that a number of the cases were classified as LMS in the clinical database. Correction of the histopathologic diagnosis and relabeling of the samples resulted in a much more pronounced separation of and LMS in the MDS analysis. This study underscores the need to re-review histopathology as reclassification occurs. While updating the clinical database may be desired, this is usually impractical. For molecular studies that use archival samples, it is critical to have the archival samples re-reviewed by a pathologist. Further, unsupervised analysis often proves to be a critical quality control step in identifying structural problems that may exist. Finally, MDS analysis further supports that is a distinct type of sarcoma. Matti Nykter, M.Sc. Kelly K. Hunt, M.D. 2 Raphael E. Pollock, M.D., Ph.D. 2 Adel K. El-Naggar, M.D. 1 Ellen Taylor, B.S. 1 Ilya Shmulevich, Ph.D. 4 Olli Yli-Harja, Ph.D. Wei Zhang, Ph.D. 1 1 Departments of Pathology and 2 Surgical Oncology The University of Texas M. D. Anderson Cancer Center Houston, Texas Institute of Signal Processing Tampere University of Technology 720 Tampere, Finland 4 Institute for Systems Biology Seattle, Washington Key words: Microarray; Database; Quality control; ; ; and MDS. Introduction Cancers are classified in different categories based on their site of origin and histopathologic features. While classification for hematologic, epithelial, and mesenchymal types of cancer is usually straightforward, subclassification within each of these categories can be challenging for the pathologist. For example, lymphomas can be categorized as Hodgkin s and non-hodgkin s lymphomas; non-hodgkin s lymphomas can further be classified as follicular, B-cell, et cetera (1). * Corresponding Author: Wei Zhang, Ph.D. wzhang@mdanderson.org 177

2 178 Nykter et al. As new discoveries are made that allow for improved classification systems, new subgroups of tumors can emerge or the existing groups can be redefined (1). While revised classifications increase the possibilities for additional treatment options, this also creates challenges for clinicians and pathologists who store the data based on currently available classification schemes. Clinical and pathologic information for each patient is generally stored in a clinical database that uses broad classifications for clinical information. At the time of initial surgical resection, the tumors are analyzed in detail and the information regarding the tumor is stored in the database. As additional information becomes available, tumors may be studied and reclassified based on the new information. However, it is impractical to recode all of this information in clinical databases. The correct classification of tumor types is important for clinical management and therapy because even when the cancer types appear to be very similar based on standard histopathology, unique molecular features may render the tumor sensitive or resistant to different therapies. While in some cases the classification of tumor types is obvious based on their morphology, this is not always the case. Recent studies have shown that in the independent reevaluation by different pathologists, the diagnoses are not always in consensus (2, ). This is mainly because the classification of the cancer types is now frequently based on both morphologic and immunohistochemical differences. Microarray experiments are being increasingly utilized in cancer research. In order to conduct a successful microarray experiment yielding conclusive results, a large number of samples need to be obtained. This is a major challenge for investigators who study rare tumor types, such as sarcomas of the gastrointestinal tract (4-6). Thus, conducting an experiment requires the utilization of all the samples that are available. This may include samples obtained over several decades that have been extracted and classified using different clinical and pathologic features. If new knowledge about the cancer types has changed the classification system, then the data obtained from clinical databases may be inaccurate and may yield incorrect conclusions. can be used as a quality control tool for the clinical and biological aspects of datasets. Materials and Methods We used Agilent human whole-genome microarray chips in our experiments. Tumor specimens from 60 patients were obtained, of which 0 were labeled as LMS and 0 were labeled as, in our institutional database (Table I). These samples were collected and hybridized during a -month period. RNA was isolated using the RNeasy Mini Kit (Qiagen Inc., Valencia, CA) after the tissues were ground to powder under frozen conditions and lysed in the lysis buffer TRI Reagent (Molecular Research Center, Inc., Cincinnati, OH). After preparation, samples were assigned randomly to the slides. For each slide, a sample from a and from a LMS was hybridized using different dyes. To avoid the possible effect of dye bias, dye colors were selected randomly for each slide (8). While the standard approach is to use one channel of the array for the experiment and another for the reference sample (8), we chose to hybridize two experiment samples to the same slide. As the quality of microarrays has improved, it has become more common to use intensity measurements directly in two-channel microarrays (e.g., Agilent microarray chips) ( 9, 10). Data from microarray slides were extracted using Agilent Feature Extraction software version 7..1 with default settings. The preprocessing done in Feature Extraction software which included Lowess normalization, was determined to be adequate (11, 12). After normalization, there were no observable biases between the samples due to the different slides or dyes (Figure 1). At the quality control phase, all control data was removed and all saturated values were replaced by the largest non-saturated expression value mea- Our case study exemplifies the need for pathologic re-review of banked tumor specimens prior to inclusion in microarray studies. Other investigators have also reported that a large number of gastrointestinal stromal tumors () have been misclassified as leiomyosarcomas (LMS) (7). We demonstrate here the importance of pathologic re-review when tumor samples are studied for molecular and genetic analysis. We show that false conclusions can be drawn from molecular studies when clinical information is inaccurate. Further, we propose that unsupervised learning Figure 1: Scatter plot from microarray data with dye swap. The samples labeled as R and G are taken from different slides. There is no visible bias due to the dye or slide. Technology in Cancer Research & Treatment, Volume, Number 2, April 2006

3 Unsupervised Analysis in Genomic Studies 179 Table I Overview of samples used in this study. Sample classification from our institutional database and the result of the pathological rereview are shown. Barcode Completed Number Surgery date Tissue type Initial diagnosis (institutional database) Diagnosis on re-review Small intestine, colon tumor normal mucosa Small intestine-tumor Stomach Stomach tumor Stomach, omentum, spleen, left diaphragm T-retroperitoneum, normal mucosa colon Abdomen-tumor Stomach-greater curve (esophagogastric junction) Retroperitoneum tumor , Metastasis Liver-, Metastasis Small bowel NORMAL, small bowel-4 nodules Stomach-tumor Tumor-stomach, Normal mucosa Chest wall/soft tissue of, chest Metastatic Ovary, uterus, Metastatic Small bowel mesenteric tumor Abdominal, colon Abdominal mass metastatic Segment of transverse Recurrent colon and intra abdominal tumor Retroperitoneum Colon Retroperitoneum-tumor, Kidney-normal Lt colon & stomach N-gallbladder, T- retroperitoneal Pelvic Retroperitoneum Bowel, normal mucosa Abdominal wall Abdomen-tumor Normal stomach, tumor T intra abdominal & prox sm bowel T-rectum, uterus, met N-sm Possible intestine high grade Normal-periteneum, adipose, Tumor Tumor-sm bowel, uterine, normal-sm bowel, mucosa Pelvic mass Favor leiomyosarcoma +++SMA Cy Dye Technology in Cancer Research & Treatment, Volume, Number 2, April 2006

4 180 Nykter et al. Table I Continued Overview of samples used in this study. Sample classification from our institutional database and the result of the pathological rereview are shown. Barcode Completed Number Surgery date Tissue type Initial diagnosis (institutional database) Diagnosis on re-review Pelvic mass omentum tumor Rectum Intraabdominal tumor, Mesentery tumor recurrent with peritoneal metastasis Liver met, liver normal, met Small bowel mesentery, recurrent tumor & normal Tumor, periadrenal retroperitoneal Retroperitoneum, metastasis Stomach-tumor Small bowel, tumorretropertoneum, normalcolon mucosa, colon wall Cy Dye sured for that specific gene. Finally, we had expression profiles of 41,67 genes available for the analysis. We used unsupervised analysis to analyze the obtained microarray data. It should be noted that instead of processing the data as ratios, we performed the analysis on the gene expression values directly (9, 10). We wanted to determine whether different cancer types could be separated based on gene expression data. Before unsupervised analysis can successfully be applied, genes with nearly constant expression profiles should be removed from the analysis as they are not likely to be informative. We computed the th and 9th percentiles for each gene expression on the log scale. If the ratio of these two statistics was less than two, the gene was deemed to be non-informative and removed. This filtering was done without any knowledge of the class labels using all the samples in the computation of the percentiles. After the filtering,,12 informative genes remained. Using these genes, classical multidimensional scaling (1) was applied with correlation as the distance measure. All the analyses were done using Matlab version 6.. Two-dimensional MDS representation of the microarray data with the class labels from our institutional database is shown in Figure 2. The spread of the and LMS populations is shown with the ellipses. The size of the ellipse is relative to the variance. Axes correspond to the direction that maximizes the variance within the populations. As the ellipses are significantly overlapping, all the samples seem to belong to one large cluster. We note that the ellipses should only be considered as a heuristic illustrative tool and do not reflect class separation obtained with real classification algorithms. Based on this result we might conclude that there is no visible separation between these two cancer types. Instead, the Results and Discussion Using multidimensional scaling, we can visualize the structure of the high-dimensional data in low-dimensional Euclidean space. Eigenvalue analysis (scree plot) of the MDS representation indicated that 2- dimensions are sufficient for data visualization (14). The first two and three dimensions included 2 and 4 percents of the total variance of all 60 dimensions, respectively. Since we have used MDS only as a data visualization tool, including the third dimension did not improve the visually observable class separation. Figure 2: Two dimensional MDS representation with labels from our institutional clinical database. Samples from patients with are denoted by squares and those from patients with LMS by circles. The spread of the populations is demonstrated by ellipses. Dashed and dotted ellipses correspond to LMS and, respectively. Ellipses are overlapping, thus there is no visible separation between LMS and. Technology in Cancer Research & Treatment, Volume, Number 2, April 2006

5 Unsupervised Analysis in Genomic Studies 181 majority of the samples are mixed together in one large cluster of points. This was surprising since there is usually some type of separation between different cancer types observed with an unsupervised analysis (1, 16). In addition, the distinction of and LMS based on microarray data has been reported earlier (6). This raised significant concerns regarding the validity of our analysis. After studying the data in more detail, the error source led back to the clinical and pathologic information from our institutional database. The earliest samples used for this experiment were over ten years old and thus, the clinical classification of these samples corresponded to the knowledge at that time. New knowledge about the classification of these mesenchymal tumors had not been incorporated into the institutional database since there had been no general pathologic re-review of samples in this histologic classification (7). In order to assign the clinical labels in a more accurate fashion, the samples were re-reviewed by a single pathologist. This pathologic re-review revealed that many of the samples had the wrong clinical label since they were classified according to the old histologic classification scheme (Table I). Of 0 samples that were originally classified as LMS, 11 were reclassified as. Thus, we actually had 41 samples that were and 19 that were LMS. The reason for such a significant reclassification was the new information revealing that have expression of CD117 whereas LMS tumors are positive for smooth muscle actin (7). The pathological features and c-kit and smooth muscle actin staining patterns for and LMS are shown in Figure. The same MDS representation with new corrected class labels is shown in Figure 4. While the class separation still is A C Figure : Composite photomicrograph of and LMS. (A) Profileration of mesenchymal cells with mild polymorphism (). (B) c-kit positive immunostaining in the cytoplasm of neoplastic cells. (C) A hematoxylin and eosin stain of LMS revealing interlacing bundeles of neoplastic spindle cells. (D) Positive smooth muscle actin in the cytoplasm of tumor cells of LMS. B D not perfect, we observed that the majority of the samples of same type are clustered together. If we compare this to the result in Figure 2, we observe that now the ellipse showing the spread of LMS population is smaller and the cluster is partially separate from the cluster. Thus, even though the separation is not perfect, LMS samples appear to be clustered in a cluster of their own. This shows that unsupervised analysis can potentially uncover underlying similarities between the samples from the same cancer type. Figure 4: Two dimensional MDS representation with labels after pathological re-review and validation. Samples from patients with are denoted by squares and those from patients with LMS by circles. The spread of the populations is demonstrated by ellipses. Dashed and dotted ellipses correspond to LMS and, respectively. There is a clear class separation as the samples from LMS are clustered in a small group that is separate from the cluster. Multidimensional scaling applied in this work is just one of the many unsupervised learning approaches proposed for gene expression studies. For example, different clustering methods (1, 16), such as hierarchical and k-means clustering, have been proposed for the same purpose. To get a more comprehensive understanding of the data, several different methods could be applied. If all unsupervised learning methods produce results that are inconsistent with the class labels, this would further strengthen the assumption that the labels are wrong. To demonstrate the effect of false class labels on supervised analysis, we performed a two-tailed t-test. As a result p- value for the significance of each gene was obtained. The number of differentially expressed genes is shown as a function of p-value in Figure for original, updated, and random class labels (17). As there are more differentially expressed genes with updated class labels, this indicates that the results obtained with supervised learning are consistent with the unsupervised approach. If the original class labels are used, the number of observed differentially expressed genes is closer to the number of genes observed using randomly selected class labels. With updated class labels, the num- Technology in Cancer Research & Treatment, Volume, Number 2, April 2006

6 182 Nykter et al. ber of observed differentially expressed genes is higher [17 genes after Bonferroni correction (14) for multiple testing] than with the original labels (8 genes). Figure : Number of differentially expressed genes shown as the function of p-value, obtained using a t-test. Data are shown for original and updated class labels. In addition, 1% and % significance levels are shown for p-values, obtained permuting class labels randomly 100 times. Shown p-values are raw p-values from the t-test (i.e., they have not been corrected to compensate for multiple testing). It is important to note that had we ignored the structural problems observed with unsupervised learning approaches and proceeded with the analysis, the results and conclusions would have been completely different and incorrect. For example, if we would have used supervised learning to construct classifiers with false class labels, we would have obtained a list of genes that may have no role in the true class separation. However, this error would not have been noticed prior to incurring time and expenses for laboratory experiments designed to validate such classifiers. Therefore, it is vital that the clinical labels are verified before any additional analyses are performed. This is especially true when older samples are used that have been classified using earlier classification schemes. Re-review of samples by a single pathologist or a team of pathologists using the same criteria for classification can ensure that a more homogeneous population is being examined. While unsupervised analysis can uncover new information from the data, it can also be used as a quality control tool for clinical information. If the findings of the unsupervised analysis are not consistent with the expected behavior, the investigator may find it prudent to question the initial clinical information. Acknowledgments This work was supported by a grant from NCI/NIH (CA09870 to WZ and REP). Support from National Technology Agency of Finland is gratefully acknowledged (MN and OY). References 1. Harris, N. L., Jaffe, E. S., Stein, H., Banks, P. M., Chan, J. K., Cleary, M. L., Delsol, G., De Wolf-Peeters, C., Falini, B., Gatter, K. C., et al. A Revised European-American Classification of Lymphoid Neoplasms: A Proposal from the International Lymphoma Study Group. Blood 84, (1994). 2. Trotter, M. J. and Bruecks, A. K. Interpretation of Skin Biopsies by General Pathologists: Diagnostic Discrepancy Rate Measured by Blinded Review. Arch. Pathol. Lab. Med. 127, (200).. Xavier, A. C. G., Siqueira, S. A. C., Costa, L. J. M., Mauad, T., Saldiva, P. H. N. Missed Diagnosis in Hematological Patients an Autopsy Study. Virchows Archiv 446, (200). 4. Koh, J. S., Trent, J., Chen, L., El-Naggar, A., Hunt, K., Pollock, R., Zhang, W. Gastrointestinal Stromal Tumors: Overview of Pathologic Features, Molecular Biology, and Therapy With Imatinib Mesylate. Histol Histopathol. 19, 6-74 (2004).. Shmulevich, I., Hunt, K., El-Naggar, A., Taylor, E., Ramdas, L., Laborde, P., Hess, K. R., Pollock, R., Zhang, W. Tumor Specific Gene Expression Profiles in Human : An Evaluation of Intratumor Heterogeneity. Cancer 94, (2002). 6. Antonescu, C. R., Viale, A., Sarran, L., Tschernyavsky, S. J., Gonen, M., Segal, N. H., Maki, R. G., Socci, N. D., Dematteo, R. P., Besmer, P. Gene Expression in Gastrointestinal Stromal Tumors is Distinguished by KIT Genotype and Anatomic Site. Clin. Cancer. Res. 10, (2004). 7. De Schipper, J. P., Liem, R. S. L., Van den Ingh, H. F. G. M., Van der Harst, E. Revision of Gastrointestinal Mesenchymal Tumours with Cd117. Eur. J. Surg. 0, (2004). 8. Zhang, W., Shmulevich, I., Astola, J. Microarray Quality Control. John Wiley And Sons (2004). 9. Yang, Y. H., Thorne, N. Normalization for Two-color cdna Microarray Data, Science and Statistics: A Festschrift For Terry Speed, 40, pp Ed., D. Goldstein. Ims Lecture Notes, Monograph Series (200). 10. t Hoen, P. A. C., Turk, R., Boer, J. M., Sterrenburg, E., De Menezes, R. X., Van Ommen, G. B., Den Dunnen, J. T. Intensity-based Analysis of Two-colour Microarrays Enables Efficient and Flexible Hybridization Designs. Nucleic Acids Res. 2, E41 (2004). 11. Quackenbush, J. Microarray Data Normalization and Transformation. Nat. Genet. 2, (2002). 12. Shmulevich, I and Zhang, W. Binary Analysis and Optimizationbased Normalization of Gene Expression Data. Bioinformatics 18, -6 (2002). 1. Borg, I. and Groenen, P. J. F. Modern Multidimensional Scaling, Springer Series. In Statistics, 2nd Ed. (200). 14. Johnson, R. A. and Wichern, D. W. Applied Multivariate Statistical Analysis. Prentice Hall, 4th Ed. (1998). 1. Alon, U., Barkai, N., Notterman, D. A., Gish, K.,ybarra, S., Mack, D., Levine, A. J. Broad Patterns of Gene Expression Revealed by Clustering Analysis of Tumor and Normal Colon Tissues Probed by Oligonucleotide Arrays. Proc. Natl. Acad. Sci. USA 96, (1999). 16. Eisen, M. B., Spellman, P. T., Brown, P. O., Botstein, D. Cluster Analysis and Display of Genome-wide Expression Patterns. Proc. Natl. Acad. Sci. USA 9, (1998). 17. Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J. P., Coller, H., Loh, M. L., Downing, J. R., Caligiuri, M. A., Bloomfield, C. D., Lander, E. S. Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring. Science 286, 1-7 (1999). Received: October 1, 200; Revised: January, 2006; Accepted: January 12, 2006 Technology in Cancer Research & Treatment, Volume, Number 2, April 2006