Performance evaluation of adaptive introgression classification methods
This article has been Reviewed by the following groups
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
- Evaluated articles (Peer Community in Evolutionary Biology)
Abstract
Introgression, the incorporation of foreign variants through hybridization and repeated backcross, is increasingly being studied for its potential evolutionary consequences, one of which is adaptive introgression (AI). In recent years, several statistical methods have been proposed for the detection of loci that have undergone adaptive introgression. Most of these methods have been tested and developed to infer the presence of Neanderthal or Denisovan AI in humans. Currently, the behaviour of these methods when faced with genomic datasets from evolutionary scenarios other than the human lineage remains unknown. This study therefore focuses on testing the performance of the methods using test data sets simulated under various evolutionary scenarios inspired by the evolutionary history of human, wall lizard ( Podarcis ) and bear ( Ursus ) lineages. These lineages were chosen to represent different combinations of divergence and migration times. We study the impact of these parameters, as well as migration rate, population size, selection coefficient and presence of recombination hotspots, on the performance of three methods (VolcanoFinder, Genomatnn and MaLAdapt) and a standalone summary statistic (Q95( w , y )). Furthermore, the hitchhiking effect of an adaptively introgressed mutation can have a strong impact on the flanking regions, and therefore on the discrimination between the genomic windows classes ( i.e. AI/non-AI). For this reason, three different types of non-AI windows are taken into account in our analyses: independently simulated neutral introgression windows, windows adjacent to the window under AI, and windows coming from a second neutral chromosome unlinked to the chromosome under AI. Our results highlight the importance of taking into account adjacent windows in the training data in order to correctly identify the window with the mutation under AI. Finally, our tests show that methods based on Q95 seem to be the most efficient for an exploratory study of AI.
Article activity feed
-
-
Detecting adaptive introgression (AI) is a key challenge in evolutionary genomics. Although several statistical and machine learning tools have been developed to tackle this, most were designed and trained on human genomic data, particularly to study admixture between Homo sapiens, Neanderthals, and Denisovans. Whether these methods perform well in other biological systems has remained unclear. In this study, Romieu et al. (2025) address this gap by systematically comparing four AI detection approaches—Q95, VolcanoFinder, MaLAdapt, and Genomatnn—across simulated scenarios inspired by humans, Iberian wall lizards (Podarcis), and bears (Ursus).
Using a set of simulations that vary in divergence time, selection strength, timing of gene flow, effective population size, and recombination, the authors evaluate how these factors influence …Detecting adaptive introgression (AI) is a key challenge in evolutionary genomics. Although several statistical and machine learning tools have been developed to tackle this, most were designed and trained on human genomic data, particularly to study admixture between Homo sapiens, Neanderthals, and Denisovans. Whether these methods perform well in other biological systems has remained unclear. In this study, Romieu et al. (2025) address this gap by systematically comparing four AI detection approaches—Q95, VolcanoFinder, MaLAdapt, and Genomatnn—across simulated scenarios inspired by humans, Iberian wall lizards (Podarcis), and bears (Ursus).
Using a set of simulations that vary in divergence time, selection strength, timing of gene flow, effective population size, and recombination, the authors evaluate how these factors influence method performance. They also test different types of genomic regions, including those near selected sites and on separate chromosomes, to assess how background signals can interfere with AI detection. One of the most notable outcomes of the study is that Q95, a straightforward computing summary statistic, performs remarkably well across most scenarios. It often outperforms more complex machine learning methods, especially when applied to species or demographic histories different from those used in the training data.Beyond comparing method performance, Romieu et al. provide valuable guidance for researchers studying AI in non-human species. They emphasize the importance of tailoring detection approaches to the evolutionary history of the study system and address practical challenges, such as retraining machine learning models and establishing suitable score thresholds. Last but not least, they share all code and data utilized in the study, enhancing transparency and reproducibility.
This work sets a new benchmark for evaluating AI detection tools. It makes clear that no single method is universally best, and that choices should be informed by biological context. For researchers working in systems beyond humans—whether in speciation, invasion, conservation, or domestication—this study provides a critical roadmap for selecting and applying the right AI-detection tools.
References
Jules Romieu, Ghislain Camarata, Pierre-André Crochet, Miguel de Navascués, Raphaël Leblois, François Rousset (2025) Performance evaluation of adaptive introgression classification methods. bioRxiv, ver.3 peer-reviewed and recommended by PCI Evolutionary Biology https://doi.org/10.1101/2024.06.12.598278
-
