{
  "abstract": "Background Melanoma is a highly aggressive skin cancer with unpredictable metastatic behavior. While staging and histopathology remain standard in clinical assessment, they lack sufficient resolution to anticipate metastatic risk. Recent advances in transcriptomics and interpretable machine learning have enabled the identification of gene expression signatures that may serve as robust predictors of tumor progression. This study aimed to develop an RNA-seq-based predictive model for metastasis in melanoma, integrating data from The Cancer Genome Atlas (TCGA) and interpretable AI techniques.Methods A total of 110 melanoma RNA-seq samples were analyzed, comprising 56 primary tumors and 54 metastatic lesions. Raw counts were normalized as log2(TPM + 1), and annotated with clinical labels. A gradient-boosted decision tree model (XGBoost) was trained using stratified 5-fold cross-validation, with hyperparameter optimization via randomized search. Class imbalance was addressed using SMOTE during training. Feature importance was quantified with SHAP (SHapley Additive exPlanations), and stability of gene importance was evaluated using 300 bootstrap iterations.Results The final model achieved an AUC of 0.89 and an accuracy of 83%. SHAP interpretation revealed a gene signature inversely associated with metastasis, led by NOTCH2, whose high expression was more frequent in non-metastatic samples and contributed negatively to the probability of metastatic classification ( figure 1, figure 2). This suggests a potential suppressor role for NOTCH2 in the progression of melanoma, consistent with recent findings proposing that NOTCH2 downregulation may facilitate metastasis via the TRAF6/Akt/mTOR pathway. Other top genes included RNF169, KLK6, CRABP2, and MMP27, associated with extracellular matrix remodeling and DNA damage response, and showing stronger contributions to metastatic prediction. SHAP beeswarm visualization confirmed the inverse association of NOTCH2 expression with metastatic output.Conclusions This work supports the hypothesis that NOTCH2 may function as a metastasis suppressor in melanoma. The model reveals a compact and biologically meaningful gene expression signature associated with metastatic status in this TCGA cohort. However, due to limited sample size, the model’s generalizability is constrained. As such, it should be interpreted as an explanatory framework that highlights promising biomarker candidates rather than a deployable clinical tool. Further validation in larger, independent cohorts and functional studies is required to determine the translational potential of the identified gene set.Abstract 141 Figure 1ROC curve of the XGBoost classifier for predicting metastatic melanoma receiver operating characteristic (ROC) curveAbstract 141 Figure 2Interpretability of gene contributions to metastatic melanoma classification using SHAP values",
  "authors": [
    {
      "affiliations": [
        "SMED Clinical Research, Alicante, Spain"
      ],
      "name": "Sonia Macia Escalante"
    },
    {
      "affiliations": [
        "SMED Clinical Research, Barcelona, Spain"
      ],
      "name": "Ruben Lopez Aladid"
    },
    {
      "affiliations": [
        "SMED Clinical Research, Sevilla, Spain"
      ],
      "name": "Rebeca Muñoz Tovar"
    },
    {
      "affiliations": [
        "SMED Clinical Research, Alicante, Spain"
      ],
      "name": "Ana Navarro Sellés"
    },
    {
      "affiliations": [
        "SMED Clinical Research, Alicante, Spain"
      ],
      "name": "Mariana Fossati"
    },
    {
      "affiliations": [
        "SMED Clinical Research, Alicante, Spain"
      ],
      "name": "Roberto Ledezma"
    }
  ],
  "title": "141 Gene expression-based machine learning model identifies a NOTCH2-associated signature inversely linked to melanoma metastasis",
  "uid": "dd0d7640-c45c-522e-b088-24a57a6802a4"
}