A molecular map of mesenchymal tumors

Abstract

Background: Bone and soft tissue tumors represent a diverse group of neoplasms thought to derive from cells of the mesenchyme or neural crest. Histological diagnosis is challenging due to the poor or heterogenous differentiation of many tumors, resulting in uncertainty over prognosis and appropriate therapy.

Results: We have undertaken a broad and comprehensive study of the gene expression profile of 96 tumors with representatives of all mesenchymal tissues, including several problem diagnostic groups. Using machine learning methods adapted to this problem we identify molecular fingerprints for most tumors, which are pathognomonic (decisive) and biologically revealing.

Conclusion: We demonstrate the utility of gene expression profiles and machine learning for a complex clinical problem, and identify putative origins for certain mesenchymal tumors.

Figure 1

Multi-dimensional scaling of all 96 mesenchymal tumor samples. There are 19 types of tumor shown; the color coding of which is used consistently for all figures. All gene expression values were used to calculate the inter-sample Euclidean distance matrix. The distances are translated here onto a two-dimensional plane using the classical cmd-scale algorithm of R. The stress of the plot was 0.34; an index of the goodness of fit between the original distance matrix and the MDS distance (see Materials and methods). WLS, well-differentiated liposarcoma; LMA, lipoma; MLS, myxoid liposarcoma; EWS, Ewing's sarcoma; FMT, desmoid fibromatosis; CHS, chondrosarcoma; CHB, chondroblastoma; ARMS, alveolar rhabdomyosarcoma; ERMS, embryonal rhabdomyosarcoma; DCS, de-differentiated chondrosarcoma; MSS, monophasic synovial sarcoma; MPNST, malignant peripheral nerve sheath tumors; CMF, chondromyxoid fibroma; PMS, pleomorphic sarcoma; LMS, leiomyosarcoma; OS, osteosarcoma; CMA, chordoma; NFB, neurofibroma; SWN, schwannoma.

Figure 2

Schematic of two-step model. In order to successfully classify the sarcoma with the minimum of errors a two-step approach was used. A mixture of single sarcoma and composite classes were used for prediction in step one. Then, in step two, composite classes were separated into their constituent tumors. * The SPIN (spindle-like) group comprising PMS, LMS, and DCS could not be separated by our model. ** The WLS and LMA could not be separated by our model but were distinct from MLS. ADIP, adipocytic tumors; RHAB, rhabdomyosarcoma.

Figure 3

Pathognomonic fingerprints for many tumor types. In step one of our model, 61 genes were used in all folds of cross-validation. Average linkage clustering of this geneset reveals strong sets of distinct genes for many single mesenchymal tumors or composite groups. The sample types are color coded as before. A, adipocytic tumors; B, rhabdomyosarcoma; C, NFB/SWN; D, EWS; E, CMA; F, FMT; G, MSS/MPNST; H, CHB; I, CHS; J, CMF; K, OS; L, spindle-like tumors.

Figure 4

Pathognomonic fingerprints step two. Molecular fingerprints of genes for A and B: ARMS and ERMS; C and D: MSS and MPNST; E and F: NFB and SWN; G and H: WLS/LMA and MLS. We have selected genes based upon their inclusion in the majority of folds of cross-validation then clustered them by average linkage.