Machine learning–integrated transcriptomic clustering for prognostic stratification in epithelioid pleural mesothelioma.

Abstract

8051 Background: Survival outcomes in pleural mesothelioma (PM), particularly within epithelioid PM (ePM), are heterogeneous and inadequately captured by histological classification. We applied machine learning (ML) methods integrating transcriptomic and clinical features to improve prognostic stratification in ePM. Methods: RNA sequencing was performed on FFPE tumor samples from 125 ePM patients treated at two Italian centers (2006-2021). Overall survival (OS) was categorized as long (> 36 months), intermediate (12-36 months), or short (≤12 months). Unsupervised clustering of highly variable genes identified transcriptomic subgroups, and GSEA was performed. A previously published 20-gene sarcomatoid-like transcriptional score (S-score), derived from single-cell data was calculated. ML models were developed using age, sex, S-score, and a cluster-derived score related to cell proliferation as input features. Logistic Regression (LR) and Random Forest (RF) classifiers were trained to predict long and short survival, using class-balanced weighting and 5-fold stratified cross-validation. Model performance was assessed using AUC-ROC and balanced accuracy. Model interpretability was evaluated with SHapley Additive exPlanations (SHAP). External evaluation was performed in the TCGA PM cohort with cohort-specific standardization. Results: Unsupervised transcriptomic analysis identified three clusters with distinct survival patterns. Cluster 1 was enriched for immune-related pathways and associated with longer survival; Cluster 2 exhibited proliferative and cell-cycle-related programs and was enriched in short-survival patients. Cluster 3 showed immune-related transcripts, with variable expression patterns across histology and survival groups. Given the convergent GSEA results, Cluster 2 was selected as a ML feature. Both the S-score and Cluster 2 score were associated with OS, with stronger survival discrimination for Cluster 2. ML models showed good discriminative performance for long survival prediction (cross-validation AUC-ROC: LR 0.83 ± 0.10; RF 0.83 ± 0.04), with similar results in the TCGA cohort (LR AUC-ROC = 0.84; RF = 0.77). Predictive performance for short survival was moderate. SHAP analysis identified Cluster 2 score as the main contributor to ML predictions, with higher values associated with increased risk of short survival. Conclusions: Machine learning models integrating transcriptomic clustering with clinical variables improve prognostic stratification in ePM. ML-based risk prediction builds upon transcriptomic features, capturing survival-relevant heterogeneity beyond histology. These findings support the use of ML as a complementary tool to transcriptomic profiling for prognostic assessment in ePM.