Radiating pattern revealed by a deep learning model traces the evolutionary dynamics of the Archaea domain

Archaea, one of the primary domains of life, have coevolved with Earth for billions of years and play critical roles in biogeochemical cycles across diverse ecosystems. However, resolving deep archaeal phylogeny is confounded by among-site heterogeneity and compositional biases, which together exacerbate long-branch attraction in lineages like DPANN. Here, using variational autoencoders and other deep learning models along with phylogenomics, we re-evaluated archaeal evolution and identified a radiating evolutionary landscape of three major archaeal superphyla, DPANN, TACK-Asgard and Euryarchaeota. Our integrated results also support the scenario in which DPANN evolved as an early monophyletic lineage, potentially relying on acetate-related metabolism under lower temperature conditions, whereas TACK-Asgard and Euryarchaeota possibly share a hydrogen-dependent methyl-reducing methanogen ancestor with the complete Wood–Ljungdahl pathway in higher temperature habitats. Environmental factors, including temperature, oxygen, and salinity, have kept influencing subsequent archaeal evolution in distinct patterns, highlighting the interplay between environmental factors and evolutionary trajectories of this ancient domain. By demonstrating that deep learning can be systematically integrated with phylogenomics to generate testable hypotheses on ancient metabolisms and environmental adaptation, this work provides a new analytical insight for archaeal evolutionary studies.

(Zhenbo Lv, Fengping Wang, Yinzhao Wang*)