Scalable phylodynamic inference of mutational and epistatic fitness effects with multi-feature birth-death models
- 1North Carolina State University
Abstract
Fitness is a complex phenotype determined by many different genomic features, their interactions with each other and the environment. For rapidly evolving microbes like viruses and bacteria, phylodynamic multi-feature birth-death (MFBD) models provide a way to infer the fitness of lineages with different genotypes from reconstructed phylogenies. However, existing phylodynamic methods assume very simple additive or multiplicative genotype-to-fitness (G-to-F) mappings that typically ignore epistatic interactions among features. We developed a gradient-based Bayesian MFBD framework for estimating the fitness effects of multiple features while allowing for both additive and epistatic G-to-F maps. The framework uses Hamiltonian Monte Carlo for posterior inference and incorporates shrinkage priors to regularize weakly supported fitness effects without cross-validation. In simulations, Bayesian inference performs well at recovering additive fitness effects, with accuracy and precision improving consistently with increasing tree size. While epistatic effects are more difficult to estimate because they depend on the co-occurrence of mutations, shrinkage improves estimation when interaction effects are weak or sparsely supported. Simulations with misspecified additive models show that ignoring moderate or strong epistasis can bias estimates of individual mutational effects. We applied the framework to a time-calibrated \textit{Escherichia coli} ST131 phylogeny with 1763 samples and estimated fitness effects for several mutations in the quinolone resistance-determining region (QRDR). These results show that Bayesian phylodynamic inference can be used to learn the genomic determinants of fitness while quantifying uncertainty in mutational and epistatic fitness effects.
Related
- Mutational fitness landscapes of diverse human influenza H1N1 neuraminidases
- Phylodynamic inference with the bounded coalescent: a point process perspective
- Simulation based-inference of epidemiological and phylodynamic models via Neural Posterior Estimation
- SLAFIT: Viral fitness inference and structural variant detection from TGS data
- Scalable and semi-automated Boolean network inference from multi-condition single-cell transcriptomes with scBOLT
BibTeX
@misc{cen2026scalable,
title = {{Scalable phylodynamic inference of mutational and epistatic fitness effects with multi-feature birth-death models}},
author = {S. Cen and L. M. Kepler and C. Lanzas and B. J. Reich and D. A. Rasmussen},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.19.760109},
url = {https://www.biorxiv.org/content/10.64898/2026.10.19.760109v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Scalable phylodynamic inference of mutational and epistatic fitness effects with multi-feature birth-death models
AU - S. Cen
AU - L. M. Kepler
AU - C. Lanzas
AU - B. J. Reich
AU - D. A. Rasmussen
PY - 2026
DA - 2026/10/19
PB - bioRxiv
DO - 10.64898/2026.10.19.760109
UR - https://www.biorxiv.org/content/10.64898/2026.10.19.760109v1
AB - Fitness is a complex phenotype determined by many different genomic features, their interactions with each other and the environment. For rapidly evolving microbes like viruses and bacteria, phylodynamic multi-feature birth-death (MFBD) models provide a way to infer the fitness of lineages with different genotypes from reconstructed phylogenies. However, existing phylodynamic methods assume very simple additive or multiplicative genotype-to-fitness (G-to-F) mappings that typically ignore epistatic interactions among features. We developed a gradient-based Bayesian MFBD framework for estimating the fitness effects of multiple features while allowing for both additive and epistatic G-to-F maps. The framework uses Hamiltonian Monte Carlo for posterior inference and incorporates shrinkage priors to regularize weakly supported fitness effects without cross-validation. In simulations, Bayesian inference performs well at recovering additive fitness effects, with accuracy and precision improving consistently with increasing tree size. While epistatic effects are more difficult to estimate because they depend on the co-occurrence of mutations, shrinkage improves estimation when interaction effects are weak or sparsely supported. Simulations with misspecified additive models show that ignoring moderate or strong epistasis can bias estimates of individual mutational effects. We applied the framework to a time-calibrated \textit{Escherichia coli} ST131 phylogeny with 1763 samples and estimated fitness effects for several mutations in the quinolone resistance-determining region (QRDR). These results show that Bayesian phylodynamic inference can be used to learn the genomic determinants of fitness while quantifying uncertainty in mutational and epistatic fitness effects.
ER -