16/04/2026, 15:00 — 16:00 —
Room P3.31, Mathematics Building
Francisco Berkemeier, University of Cambridge, UK
Inverse problems and machine learning in DNA replication
Before a cell divides, it must first duplicate its genome accurately and in full. DNA replication is therefore essential for genome integrity, and its disruption is a major source of replication stress, a hallmark of cancer and a key target of modern therapies. A central signature of this process is the replication timing programme, whereby different genomic regions are copied at different times during S phase. Although genome-wide experimental assays have made this programme accessible, such measurements are often costly, time intensive, and indirect, motivating the need for mathematical models that can infer the underlying replication dynamics from observable data. We show that replication timing can be understood within a mathematical framework based on nucleation and crystal growth ideas related to the Kolmogorov-Johnson-Mehl-Avrami model, casting genomic data as an inverse problem for inferring initiation landscapes across the genome. This yields a quantitative view of replication kinetics and identifies regions where standard models break down, revealing links with transcription and genomic fragility. We then explore complementary machine learning approaches, from genome language models that predict replication initiation directly from sequence, to physics-informed machine learning for modelling replication under stress. Together, these approaches point towards predictive models of replication with potential relevance for targeted therapies.
