Reusable pangenome model reveals how to watch pneumococcal vaccine escape

A team of genomic epidemiologists has built a reusable mathematical model that captures how Streptococcus pneumoniae populations evolve after pneumococcal conjugate vaccines (PCVs) are introduced, and has used it to work out how countries with limited budgets should design their genomic surveillance programmes. The study, published in Genome Medicine, draws on bacterial genome data from Nepal, the United States and the United Kingdom, and offers practical guidance for public health agencies trying to track the phenomenon known as serotype replacement, in which non-vaccine strains gradually fill the ecological space vacated by vaccine-targeted ones.

S. pneumoniae remains one of the world’s most consequential bacterial pathogens, causing pneumonia and meningitis with the highest disease burden falling on young children and the elderly. When PCVs were first rolled out in the United States in 2000, they delivered striking reductions in disease and in carriage of the serotypes they target, which are defined by the bacterium’s capsular polysaccharide. But because the vaccines cover only a subset of the more than ninety known serotypes, they reshape competition within the species. Strains carrying non-targeted serotypes experience relaxed competition and can expand, replacing the vaccine types and eroding some of the public health gains.

Traditionally, pneumococcal epidemiology has been organised around serotypes, yet the capsular locus that defines them represents only a small slice of the species’ genetic diversity. Whole genome sequencing has revealed a far richer picture, including lineages called global pneumococcal sequence clusters (GPSCs), which are defined by variation across the entire genome. Because many GPSCs carry multiple serotypes, and individual serotypes appear in multiple GPSCs, the two classifications provide complementary information. Serotype replacement can therefore arise either from closely related strains switching capsule within a lineage or from genetically distant lineages expanding into the niche left open by vaccination.

A leading explanation for these dynamics is negative frequency-dependent selection, or NFDS, a form of selection in which a trait confers greater benefit when it is rare than when it is common. In bacteria, NFDS is thought to act on accessory genes such as antimicrobial resistance genes and bacteriocins, helping to maintain a diverse pangenome in which no single gene combination sweeps to fixation. Earlier modelling work, notably by Corander and colleagues in 2017, showed that NFDS could explain much of the post-vaccine reshuffling of pneumococcal populations, but those models were tightly coupled to specific datasets and difficult for non-specialists to redeploy.

The new study, led by Leonie Lorenz and John Lees of the European Molecular Biology Laboratory’s European Bioinformatics Institute together with collaborators across Nepal, the United Kingdom and Canada, addresses three gaps. First, the team rebuilt the population dynamics model in the odin modelling framework, separating model code, genomic inputs and fitting procedures so that public health bodies can adapt it to their own settings. Second, they tested whether cheaper data types, such as serotyping alone or targeted sequencing of a fixed gene set, could substitute for whole genome sequencing. Third, they used simulation to ask how sample size and sampling frequency affect the reliability of parameter estimates and forecasts.

The model itself is a compartmental extension of the Wright-Fisher framework, arranged on a two-dimensional grid in which one axis represents GPSC lineages and the other represents serotypes. Each generation, which corresponds roughly to a month of transmission, the population is replenished by offspring from the current generation plus immigrants drawn from an external reservoir of observed strain-serotype combinations. Four parameters are fitted to data by Markov chain Monte Carlo: vaccination effectiveness, the strength of NFDS, the proportion of intermediate-frequency genes subject to NFDS, and the immigration rate. Genes are summarised at the lineage level, and those changing least in frequency over time, ranked by a delta statistic, are flagged as candidates under balancing selection.

When fitted to carriage data from Kathmandu, where 1,881 samples were collected between 2009 and 2019 around the introduction of PCV10, and to previously published datasets from Massachusetts and Southampton, the model reproduced the observed serotype frequency changes closely, with model confidence intervals overlapping the data in nearly every case. Vaccination effectiveness estimates were consistent across all three locations, ranging from roughly 0.08 to 0.12 per generation, while immigration rates were similarly stable. Notably, the genes inferred to be under NFDS differed substantially between countries, and model comparison using likelihood-ratio tests and the Bayesian Information Criterion showed that only a subset of intermediate-frequency genes, not all of them, appear to be under NFDS. This suggests that other forces, such as balanced rates of gene gain and loss or ecological niche partitioning, also help maintain the accessory genome.

The search for a universal set of NFDS genes proved disappointing in an instructive way. A genetic algorithm applied to the same data found far more overlap between countries than the delta statistic did, yet even that shared set of 155 genes was no larger than expected by chance. Serotype-only model versions fit the data poorly. Together, these results indicate that neither serotyping nor targeted sequencing of a fixed gene panel can substitute for whole genome surveillance, underscoring that each country needs its own genomic monitoring programme to understand local replacement dynamics rather than importing conclusions from elsewhere.

The simulation experiments delivered the study’s most actionable finding. By generating a synthetic twenty-year dataset and then subsampling it under a fixed budget, the researchers showed that when resources are scarce, it is better to sample less often with larger samples per round. At the lowest budgets, annual and quadrennial sampling produced biased estimates, whereas biennial and triennial sampling performed best, balancing statistical power against the risk of missing critical change points. With larger budgets, sampling frequency mattered little. The team has released the model as an open-source R package called Stubentiger, giving surveillance agencies a practical tool for anticipating how pneumococcal populations will respond as vaccine formulations evolve.

Subject of Research: Mathematical modelling of negative frequency-dependent selection in the Streptococcus pneumoniae pangenome to inform genomic surveillance strategies during pneumococcal conjugate vaccine introduction

Article Title: A reusable model of pangenome selection informs optimal surveillance strategies over vaccine introductions

Article References: Lorenz, L. J., Hellewell, J., Horsfield, S. T., Russell, M. J., Shrestha, S., Pollard, A. J., Bentley, S. D., Lo, S. W., Colijn, C., Croucher, N. J., & Lees, J. A. (2026). A reusable model of pangenome selection informs optimal surveillance strategies over vaccine introductions. Genome Medicine, 18(1), Article 131. https://doi.org/10.1186/s13073-026-01672-4

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01672-4

Keywords: Streptococcus pneumoniae, pangenome, negative frequency-dependent selection, pneumococcal conjugate vaccine, serotype replacement, genomic surveillance, Wright-Fisher model, mathematical modelling, GPSC lineages, vaccine effectiveness, public health, Nepal