Machine-learning to predict utility of circulating tumor DNA (ctDNA) for somatic genotyping.

Abstract

232 Background: ctDNA genotyping is increasingly used to evaluate patient eligibility for genomics-driven treatments ( e.g., PARP inhibitors) in metastatic castration-resistant prostate cancer (mCRPC), but low ctDNA-fraction (ctDNA%) causes false negatives and is unpredictable. We investigated whether a machine-learning model exploiting routine clinical prognostic markers can predict if a mCRPC patient will have sufficient ctDNA% for informative ctDNA testing. Methods: We analysed plasma cell-free DNA (cfDNA) collected at baseline prior to first-line therapy from 463 consecutive mCRPC patients. ctDNA% was quantified using somatic allele frequencies and genome-wide copy number profiles via deep targeted sequencing. We built an XGBoost model (70:30 train-test split) using 18 clinical factors to predict synchronously measured ctDNA% (binary classification >2% or ≤2%, i.e., the conventional ctDNA% limit of detection for mutations employed by commercial tests), with performance evaluated using area under the curve (AUC) of receiver operating characteristics. Results: Median age was 73 (range: 45-98) and 86% of patients had ECOG 0-1. Median ctDNA% was 5% (range: 0-89%) and correlated with serum and radiographic metrics of disease burden, including total cfDNA concentration (ng/mL of plasma, reflecting cfDNA released by both normal and tumor cells; Spearman ρ=0.55), alkaline phosphatase (ALP) per upper limit of normal (ULN) ( ρ=0.46), lactate dehydrogenase (LDH) per ULN ( ρ=0.41), PSA ( ρ=0.3), presence of liver metastases, and ≥10 bone lesions. We trained an XGBoost model incorporating these and 13 additional clinical factors achieving an AUC for ctDNA >2% of 0.83 (F1 score: 0.79). SHAP interpretability scores indicated that cfDNA concentration most strongly informed prediction of ctDNA >2%, followed by ALP/ULN and PSA—consistent with our prior bivariate rank correlations—whereas features associated with initial prostate cancer diagnosis (Gleason Grade Group, de novo versus metachronous M1 disease) were less informative. Recognizing that comprehensive and standardised clinical annotation is not always available in real-world settings, we built a secondary parsimonious ctDNA%-prediction tool restricted to 8 highly informative and clinically practical factors (cfDNA concentration, ALP/ULN, LDH/ULN, PSA, albumin, ECOG, liver metastases, lung metastases) and flexible to incomplete input data, achieving a comparable AUC for ctDNA>2% of 0.76. Conclusions: Our results demonstrate the feasibility of a machine learning framework to estimate ctDNA% in patients with mCRPC. This point-of-care tool would enable prioritisation of mCRPC patients for ctDNA somatic genotyping with a predicted ctDNA >2%, versus tissue or germline-only testing in patients with a predicted ctDNA ≤2%.