(P 112) MACHINE LEARNING ANALYSIS TO DEFINE CELL LINEAGE IN LEIOMYOSARCOMA

Objective: Cellular differentiation and lineage commitment are associated with differences in DNA methylation. Leiomyosarcoma (LMS) is a tumor believed to originate from smooth muscle cells in vessel walls or the uterine myometrium. Extra-uterine LMS and uterine LMS are treated using different therapies. In a novel approach we decided to study the lineage of smooth muscle in LMS samples. We first detected significant differences in the methylation patterns of smooth muscle cells from normal vessel wall vs. myometrium, and a classifier was trained to identify the signature of smooth muscle lineage. Application of this classifier to 67 uterine and extra-uterine LMS cases matched the majority of these tumors to their site of origin. However, a subset of samples showed an unexpected association with the classifier. This research provides insights into the pathogenesis and heterogeneity of LMS, and suggests that recognition of, for example, a uterine LMS with a signature indicating vessel lineage may have important consequences for treatment.

Methods: EPIC methylation profiling was performed on DNA from 49 formalin-fixed paraffin-embedded (FFPE) smooth muscle samples from normal vessels and myometrium. DNA and RNA were isolated from FFPE samples of 67 LMS, followed by transcriptome sequencing using RNA-seq and identification of methylation profiles using EPIC arrays. A Random Forest classifier, with 20,000 trees trained on the top 10% most variable CpGs, was generated on the normal smooth muscle training set. In parallel, a Support Vector Machine (SVM) was also trained on this data. Both classifiers had an accuracy score of 1.0 on the test smooth muscle dataset. Principal Component Analysis (PCA) was used on the methylation EPIC array dataset to identify clusters within the methylation and RNA-seq data.

Results: Employing machine learning classifiers, we could show a strong correlation between the site of origin of a tumor (uterine vs. extra-uterine) and the classifier developed by machine learning for myometrium vs. vessel wall smooth muscle lineage (FIG 1A and 1B). In a subset of samples, the site of origin did not match with the classification; from 45 uterine LMS samples, 7 unexpectedly classified as originating from vessel (15%), while among the 22 extra-uterine LMS samples - 7 had a myometrial signature (32%). When total methylation data was used in a principal component analysis (FIG 1C), it could be seen that the differential methylation patterns in tumors, derived from vessel wall or from myometrium, significantly contributed to the major principal components (PC1 and PC2). This indicates that the lineage-induced methylation patterns are a major contributor to the total methylation signal.

Conclusion: We investigated the smooth muscle subtype origin for LMS and found that the samples could be classified as more similar to either myometrium or vessel wall. The majority of cases showed a correlation between the cell of origin indicated by the classifier and the actual known site of origin of the tumor. However, a subset of cases did not show this correlation and future studies will need to show whether these cases might be better treated based on the cell lineage indicated by the classifier than on the site of origin for the tumor.