1194 Prediction of lymph node metastasis in melanoma using machine learning models

<h3>Background</h3> The presence of lymph node (LN) metastases has a strong prognostic factor for survival in melanoma. Sentinel lymph node biopsy (SLNB) is used to assess LN involvement but carries potential morbidities in a proportion of patients. Thus, there is a great need to predict patients with a higher risk of LN involvement before performing SLNB. Here, we built different machine-learning models to predict LN involvement in melanoma using clinical multi-omics data from The Cancer Genome Atlas (TCGA) skin cutaneous melanoma patients (SKCM). <h3>Methods</h3> We performed differential expression analysis (DEA) for mRNA and miRNA sequences between SKCM patients with nodal metastasis (N) and patients without nodal metastasis (N0) using the DESeq2 package with a log fold-change threshold of 1.5 (|logFC|&gt;1.5) and corrected p-value&lt;0.01. Differentially methylated genes (between N and N0 were analyzed using a linear model on beta-values from the limma package with a p-value&lt;0.01. Single-sample Gene Set Enrichment Analysis (ssGSEA) and Gene ontology enrichment analysis were performed to investigate significant enrichment in biological processes (BP). Three random forest (RF) machine-learning ensembles were fitted to predict nodal metastasis. The first was using mRNA gene expression, while the second model was using the genomics model while the third trained on the highest contributors from the previous models. The models were evaluated on the testing set using a mean bootstrap estimate with 95% confidence interval,10-fold cross-validation, and area under the receiver operating characteristic curve (AUC/ROC). <h3>Results</h3> N0 had a higher tumor mutational burden and mutation counts compared with N. We found 37 up-regulated genes and 46 down-regulated genes in N compared to N0. No genes were differentially methylated after p-value correction. We also did not find any miRNA that is differentially expressed. Using ssGSEA for DEGs, we did not find any significant enrichment in hallmark pathways. Different cutaneous-related functions were among the BP enriched in the N cohort. We found that M2 and plasma cells are more enriched in the N cohort compared to N0. Model performances and transcriptomics significant features are shown in table 1 and figure 1, respectively. <h3>Conclusions</h3> Our preliminary work showed low variability in between LN involvement. Our results should be confirmed in different and larger cohorts. Future are also needed to determine the omics drivers of LN metastasis in melanoma. <h3>Ethics Approval</h3> All analysis was done retrospectively based on the publicly available TCGAdata. Due to the nature of the study, no institutional review board approval.