Recent advances in genome sequencing have produced phylogenomic datasets containing tens to hundreds of thousands of species. However, methods for reconstructing species trees from gene trees have not scaled to match these data volumes. The authors introduce ASTRAL-X, a reengineered version of the widely used ASTRAL coalescent-based estimator, designed to enable statistically consistent species tree inference at an unprecedented scale while using modest computational resources.
Coalescent-based estimators such as ASTRAL offer statistical consistency under the multispecies coalescent model, making them attractive for species tree reconstruction from gene trees. Despite this, existing implementations have faced substantial computational and memory bottlenecks that limit their applicability to ultra-large datasets. These limitations have prevented routine, statistically consistent inference on datasets demanded by Tree of Life–scale initiatives.
ASTRAL-X implements a complete algorithmic redesign of the original ASTRAL framework. The redesign focuses on three broad areas: underlying data representations, core algorithms, and the overall computational framework. By fundamentally changing how input and intermediate data are represented and processed, ASTRAL-X reduces both running time and memory consumption.
According to the authors, memory requirements in ASTRAL-X are lowered to nearly the size of the input, which they describe as the asymptotically optimal bound. The preservation of ASTRAL’s formal statistical guarantees is explicitly reported: ASTRAL-X maintains the same statistical consistency as its predecessor while providing large practical improvements in scalability.
The authors report substantial performance gains for ASTRAL-X. On modest computational resources, ASTRAL-X reconstructed species trees containing 200,000 taxa in 5 hours and 300,000 taxa in 12 hours. The memory footprint is reported as being reduced to near the input size, although precise memory numbers and the exact hardware configuration are not reported in the source abstract.
ASTRAL-X is also claimed to achieve accuracy comparable to other state-of-the-art methods on both simulated and empirical datasets while retaining statistical consistency. Specific benchmark datasets, comparative methods, and quantitative accuracy metrics are not detailed in the abstract; those details were not reported in the source provided here.
As an empirical demonstration, the authors report that ASTRAL-X reconstructed the evolutionary history of 9,524 angiosperm species in 16 minutes. This example highlights the method’s practical applicability to large biological datasets commonly encountered in plant phylogenomics and other fields.
Beyond this example, the abstract states that ASTRAL-X performs well across simulated and empirical datasets, but further information on dataset composition, gene tree inputs, and evaluation criteria is not provided in the abstract and therefore cannot be summarized here.
ASTRAL-X is publicly available from the authors’ GitHub repository: https://github.com/aaniksahaa/ASTRAL-X-releases. The source material notes that ASTRAL-X preserves the statistical guarantees of ASTRAL and reports improved practical performance, but the article is a preprint and has not been peer reviewed. Where the abstract omits implementation details, hardware specifications, full benchmark protocols, or quantitative comparisons, those details were not reported in the provided source and should be consulted in the full preprint or supplementary materials for reproducibility and evaluation.
ASTRAL-X represents a full redesign of a leading coalescent-based species tree estimator to meet the computational demands of very large phylogenomic datasets. It aims to combine statistical consistency with practical scalability by reducing running time and bringing memory use close to an optimal bound. Reported reconstructions include trees for 9,524 angiosperms in 16 minutes and trees with 200,000 and 300,000 taxa in 5 and 12 hours, respectively. The software is available on GitHub, and the work is presented as a preprint pending peer review.