ERICA-trio: an outgroup-free deep learning method for topology inference and introgression detection.

Genetic admixture is a widespread phenomenon across diverse organisms. Conventional algorithms for inferring species relationship and detecting introgression often rely on outgroup data, which can be challenging to obtain and may introduce bias. Given the demonstrated efficacy and flexibility of neural networks in topology inference and introgression detection, we developed a deep learning-based method, ERICA-trio, to reconstruct evolutionary relationships among three taxa without requiring outgroup data. We trained and evaluated this network model using extensive simulated data that cover a broad range of evolutionary scenarios and parameter spaces. Our results demonstrate that ERICA-trio achieves accuracy and robustness comparable to the outgroup-dependent model. Leveraging the predicted topological proportions, we employed two strategies to identify genomic regions with potential introgression: one based on topological symmetry, and the other on the proportions of topology corresponding to gene flow. Both approaches were highly effective, particularly for detecting signatures of adaptive introgression. We further applied ERICA-trio to real genomic data from the Heliconius butterflies and successfully identified adaptive introgressed loci associated with mimicry wing patterns. In summary, our work extends the application of deep learning frameworks in evolutionary genomics, and presents a new tool for outgroup-free phylogenetic inference and introgression detection.