We study how population history and natural selection shape genomes, and how to reliably infer that history from genetic data. To do so, we build and benchmark methods for inferring demographic history the distribution of mutation fitness effects, applying them to systems from humans to polyploid plants to tumors.
Below we highlight a few recent projects; see our Publications page for a complete list.
The Distribution of Fitness Effects
The distribution of fitness effects (DFE) of new mutations governs how much genetic variation populations harbor, how fast they adapt, and how they respond to changes in environment and population size. Our group builds on the diffusion approach of dadi (Diffusion Approximations for Demographic Inference) to infer the DFE from population genomic data while accounting for demographic history.
A key question is whether a mutation that is harmful in one population is equally harmful in another. We developed a method to infer the joint DFE across pairs of populations, which estimates how strongly fitness effects are correlated between them (Huang et al. Mol Biol Evol 2021). We have also shown that analyzing the full DFE, including beneficial mutations, can reveal signatures of domestication (Castellano et al. Peer Community J 2025).
The DFE also shapes the evolution of the mutation process itself. With Lindi Wahl’s group, we showed theoretically that shifts in the spectrum of mutation types can favor mutator alleles by altering the DFE (Tuffaha et al. Am Nat 2023). We are now connecting empirical DFE estimates to theory for the evolution of mutation rates.
Demographic History
Demographic history is both interesting in itself and a confounder that must be modeled to infer selection accurately. We continue to extend dadi to new questions and new kinds of data. We developed methods for species with nonrandom mating and for polyploids, which inherit chromosomes in ways that span from disomic to polysomic (Blischak et al. Mol Biol Evol 2020; Blischak et al. Genetics 2023), and we are now studying how polyploidy alters the dynamics of selection. We have also developed methods to correct for the biases introduced by low-pass genome sequencing, enabling accurate inference from cheaper data (Fonseca et al. Mol Biol Evol 2025).
We are also exploring modern machine learning methods for population genomic inference. We developed a neural network approach for demographic inference from allele frequency spectra that is far faster than likelihood-based optimization (Tran et al. Mol Biol Evol 2024). Because machine learning methods can be opaque, we also developed an approach to interpret what information networks trained on haplotype data actually use, by systematically permuting their inputs (Tran et al. Mol Biol Evol 2025).
Community Resources and Benchmarking
We strive to make population genomic inference reliable and broadly useful. We maintain dadi along with tools that make it easier to use at scale, including GPU acceleration (Gutenkunst Mol Biol Evol 2021) and a command-line interface (dadi-cli). We contribute to the stdpopsim project, which provides standardized, realistic simulations of genomes, including with natural selection (Lauterbur et al. eLife 2023; Gower et al. Mol Biol Evol 2025).
To evaluate how well inference methods work in practice, we organize the Genomic History Inference Strategies Tournament (GHIST). In GHIST, participants infer hidden histories from simulated data, and their results are compared against the truth (Struck et al. Mol Biol Evol 2025).
Cancer Evolution
Cancer is fundamentally an evolutionary process, as tumor cells acquire mutations that let them grow, outcompete their neighbors, and evade the immune system. We apply concepts and tools from population genomics to understand tumor development.
A key goal of cancer evolution research, with applications to cancer prevention, is to understand what hastens the acquisition of driver mutations. Tumors can do so by elevating their overall mutation rate, but many do not. In collaboration with Lindi Wahl’s group, we used data from thousands of tumors to show that cancers without elevated mutation rates can instead reach driver mutations by reversing the germline mutational bias (Tuffaha et al. Mol Biol Evol 2025). Our analysis also suggested that hypoxia drives particularly potent changes in mutation spectra, a prediction we are following up on.
Tumor mutations can also generate novel antigens that provoke an immune response. By comparing tumors from immunocompetent and immunocompromised patients, we found that immunoediting restricts clonal neoantigens in primary, untreated tumors (Borden et al. Immunity 2026). We are using these insights to better identify tumor vaccine targets, and we are developing models of how the immune system shapes early tumor evolution.