Background: Melanoma is the third most common and deadliest cancer in Australia. While surgery alone often cures early-stage cases, some patients relapse with advanced-stage disease requiring systemic treatment. Discovering new biomarkers is crucial to identify patients who only need surgery versus others requiring more intensive therapy to prevent relapse. Here we aimed to (i) develop a multi-modal, low-coverage, low-cost sequencing liquid biopsy assay using a machine learning (ML) framework that sensitively detects circulating tumour DNA (ctDNA) and (ii) investigate the assay’s utility in early-stage melanoma.
Methods: For assay development, we performed low-coverage whole-genome sequencing and methylated DNA immunoprecipitation sequencing of plasma DNA from 38 healthy individuals and 36 advanced melanoma patients (estimated plasma tumour fraction range: 0-42%). Our methylation-based classifier utilised differentially methylated regions (DMRs) found between our melanoma and healthy plasma samples as well as literature-sourced melanoma-specific DMRs. We optimised tumour specific copy-number profiles using a reference panel-of-normals and analysed mutational signature patterns using an in-house algorithm to estimate single/double-base substitutions and indel signatures in each sample. We subjected all ML classifiers to 100-fold cross-validation.
Results: Individual AUCs for mutational signature, copy-number, and methylation classifiers were 0.82, 0.90, and 0.96, respectively. We took the mean of their predicted values for an ensemble classifier (AUC=0.99, specificity=0.87, sensitivity=1). We validated all classifiers on an exploratory cohort of 15 healthy controls and 79 early-stage melanoma samples (33 baseline, 46 post-surgery). The ensemble classifier showed an AUC of 0.76 at the baseline timepoint (prior to surgery) for the detection of ctDNA. Copy-number and methylation ctDNA detection at the post-surgical timepoint both showed a strong association with inferior progression free survival (p<0.02, Cox regression).
Conclusion: Methylation, copy-number and mutational signature profiles can inform ML models to achieve a minimally invasive, tumour-agnostic approach for ctDNA detection with potential for prediction of relapse in early-stage melanoma.