Highly accurate two-gene classifier for differentiating gastrointestinal stromal tumors and leiomyosarcomas

Abstract

Gastrointestinal stromal tumor (GIST) has emerged as a clinically distinct type of sarcoma with frequent overexpression and mutation of the c-Kit oncogene and a favorable response to imatinib mesylate [also known as STI571 (Gleevec)] therapy. However, a significant diagnostic challenge remains in the differentiation of GIST from leiomyosarcomas (LMSs). To improve on the diagnostic evaluation and to complement the immunohistochemical evaluation of these tumors, we performed a whole-genome gene expression study on 68 well characterized tumor samples. Using bioinformatic approaches, we devised a two-gene relative expression classifier that distinguishes between GIST and LMS with an accuracy of 99.3% on the microarray samples and an estimated accuracy of 97.8% on future cases. We validated this classifier by using RT-PCR on 20 samples in the microarray study and on an additional 19 independent samples, with 100% accuracy. Thus, our two-gene relative expression classifier is a highly accurate diagnostic method to distinguish between GIST and LMS and has the potential to be rapidly implemented in a clinical setting. The success of this classifier is likely due to two general traits, namely that the classifier is independent of data normalization and that it uses as simple an approach as possible to achieve this independence to avoid overfitting. We expect that the use of simple marker pairs that exhibit these traits will be of significant clinical use in a variety of contexts.

Fig. 1.

Kit protein expression in GIST and LMS. (A) The expression of Kit detected by Western blotting for 47 tumor samples (22 GIST and 25 LMS). Blue arrows indicate samples diagnosed as GIST, and red arrows indicate LMS. Each GIST sample for which no Kit protein was observed and each LMS sample that (weakly) expressed Kit are marked with a ∗. (B) Immunohistochemistry staining of two Kit-positive GIST samples. (C) Immunohistochemistry staining of two Kit-negative GIST samples.

Fig. 2.

Expression values of the two genes involved in the TSP classifier on the Agilent microarrays after quantile normalization. (Note: The classification is independent of normalization, because the decision is based only on which gene is higher, but the magnitude of the expression shown does vary somewhat with normalization technique.) The separating line (slope = 1) represents the cutoff for which gene is more highly expressed. It is not a fit to the data.

Fig. 3.

Relative expression (OBSCN/C9orf65) as measured using RT-PCR. (A) Results are shown from 20 samples also used in the microarray experiments and an additional independent set of 17 samples. Classifier prediction as LMS or GIST is determined by which gene (OBSCN or C9orf65) is more highly expressed. Clinical diagnosis is shown as “X” for GIST and “O” for LMS. (B) The raw data in terms of the distance of C_t average from the RT-PCR experiments (OBSCN C_t − C9orf65 C_t).

Fig. 4.

Comparison of TSP classifier with Kit gene expression for separating samples of GIST (X) and LMS (O). The values shown are the expression values from the Agilent micorarrays after quantile normalization. (Note: The OBSCN/C9orf65 classifier is independent of the normalization chosen, but any classification based on c-Kit expression alone would not be.)