Unravelling single-cell DNA replication timing dynamics using machine learning reveals heterogeneity in cancer progression

Genomic heterogeneity has largely been overlooked in single-cell replication timing (scRT) studies. Here, we develop MnM, an efficient machine learning-based tool that allows disentangling scRT profiles from heterogenous samples. We use single-cell copy number data to accurately perform missing value imputation, identify cell replication states, and detect genomic heterogeneity. This allows us to separate somatic copy number alterations from copy number changes resulting from DNA replication. Our methodology brings critical insights into chromosomal aberrations and highlights the ubiquitous aneuploidy process during tumorigenesis. The copy number and scRT profiles obtained by analysing >119,000 high-quality human single cells from different cell lines, patient tumours and patient-derived xenograft samples leads to a multi-sample heterogeneity-resolved scRT atlas. This atlas is an important resource for cancer research and demonstrates that scRT profiles can be used to study replication timing heterogeneity in cancer. Our findings also highlight the importance of studying cancer tissue samples to comprehensively grasp the complexities of DNA replication because cell lines, although convenient, lack dynamic environmental factors. These results facilitate future research at the interface of genomic instability and replication stress during cancer progression.