PD10-09: A Novel Model Integrating Clinical, mp-MRI, and Epigenomic Features to Predict Lymph Node Invasion in Prostate Cancer Patients Undergoing Radical Prostatectomy and Pelvic Lymph Node Dissection
Introduction: Multivariable models should be used to identify prostate cancer (PCa) patients candidates for extended pelvic lymph node dissection (ePLND) during radical prostatectomy (RP) to spare unnecessary ePLNDs without missing lymph node invasion (LNI). Improving LNI detection in PCa is key in reducing ePLND-related morbidity. We hypothesized that LNI can be better predicted by integrating clinical, radiologic and epigenomic information. Methods: We recruited 172 PCa patients with a risk of LNI >5% diagnosed by target + systematic biopsy undergoing RP + ePLND between 2014-2021. Epigenetic profiles of tumor DNA biopsy cores were sequenced via reduced representation bisulfite conversion. MethylKit R package assessed the percentage methylation differences among CpG sites of patients with and without LNI. A 50% cut-off methylation difference (50-MD) identified significant (False Discovery Rate <0.001) CpGs. Enrichment analysis tested for gene pathways that were expressed among patients with and without LNI by using differentially methylated (50-MD) CpGs. Analyses were performed for target and systematic biopsy samples independently. Two signatures were created from hypermethylated CpGs from target + systematic samples and integrated with PSA, mpMRI stage, and grade group at target biopsy to develop two models predicting LNI which underwent 500 internal train-test validations and were compared with existing tools. Results: Overall, 37 patients (21.5%) had LNI. We identified 508 and 511 CpGs sites within target and systematic samples that were differentially methylated among patients with and without LNI. Gene pathways involved in the transcription of potassium channels were associated with LNI in target samples. The epigenetic signatures including only hypermethylated CpGs were correlated with LNI on univariable regression (target samples LogOdds 0.12, p<0.001; systematic samples LogOdds 0.08, p<0.001). Clinical and mpMRI variables (PSA, mpMRI stage, and ISUP grade group at target biopsy) were associated with LNI (all p=0.01). An AUC of 86% and 83% was achieved for the target model and systematic model. Both models outperformed the previous versions of the Briganti nomogram at any LNI threshold risk. Conclusions: We developed two LNI prediction models that integrated clinical, mpMRI and epigenetic features which outperformed available tools. Epigenetic features from target tumor samples appeared to better predict LNI compared to their systematic counterparts. SOURCE OF Funding: None
.jpg "Marco Bandini, MD photo")