Genomic landscape of metastatic hormone sensitive prostate cancer (mHSPC) vs. metastatic castration-refractory prostate cancer (mCRPC) by circulating tumor DNA (ctDNA).

5043 Background: mCRPC carries a poor prognosis, and targeted therapies have had minimal success in mCRPC. Novel genomic targets could improve drug development. To date, large ctDNA studies in metastatic prostate cancer have been descriptive with limited or no clinical annotation. Herein, we hypothesize that profiles of genomic alterations (GAs) in ctDNA not only differ significantly between, but can also be used to predict mCRPC vs. mHSPC. These findings could help identify new drug targets for mCRPC treatment. Methods: Men with mHSPC or mCRPC who underwent NGS of ctDNA using G360 (Guardant Health Inc.) at the Huntsman Cancer Institute were included. Men were classified as mCRPC or mHSPC (patients with current or no prior ADT). G360 detects somatic mutations in selected exons of 73 genes, amplifications in 18 genes, and selected fusions in 6 genes. Two-sided students t-test was used to compare the %cfDNA and total GAs. The Chi squared test was used to compare the frequency of each GA. Machine learning (ML) algorithms were trained on GAs and benchmarked by cross-validated performance. GAs contributing to mCRPC vs. mHSPC classification were measured by ML feature importance (e.g. odds ratios, regression coefficients). Results: Of the 259 men included, 119 men had mHSPC and 140 had mCRPC. Men with mCRPC had more GAs (4.5 vs. 1.86, p<0.0001) and higher %cfDNA (9.56% vs. 5.02%, p=0.02). In mHSPC, there was no significant difference in the number of GAs or %cfDNA between men on ADT and those who hadn’t yet started ADT. ML algorithms used GAs to predict mCRPC with 78.1% sensitivity, 64.0% specificity, 76.7% PPV, 65.1% NPV, and 70.3% overall accuracy. mCRPC was enriched with GAs in AR, ARID1A, BRAF, BRCA2, CCNE1, CTNNB1, EGFR, FGFR1, KIT, MET, MYC, PDGFRB, PIK3CA, and TP53. Of note, many of these genes are involved in MAP/ERK signaling. Conclusions: Men with mCRPC have more GAs, higher %cfDNA, and enrichment of GAs in the MAP/ERK pathway compared to men with mHSPC. The distinct GAs seen in mCRPC represent novel therapeutic targets, especially in the MAP/ERK pathway. We also show that machine learning can differentiate mHSPC and mCRPC based on GAs detected in ctDNA.