Protein tertiary structure is no better than poorly modelled amino acid sequences at resolving bilaterian relationships

The exponential increase in the availability of protein tertiary structures has re-ignited interest in their suitability for phylogenetic inference. Despite claims that this data type vastly outperforms primary sequences in resolving deep-time relationships, they have not held outside of paralog-rich protein family tree inference. Subsequent work on a standard, though taxon-poor, single-copy ortholog data set found that sequence-only data recovered considerably more uncontested metazoan relationships than the structure-only data. Here, we have extended these experiments to a taxon-rich data set, compared sequence-only, structure-only and combined data sets under model-based maximum likelihood tree inference, using supermatrix and supertree approaches. We also compared the performance of the model and distance-based structure-only tree inferences. We found that, even with minimal mitigation of sources of error, species trees inferred from sequence-only data were more similar to the canonical metazoan tree than all structure-aware trees, with conflicts between the canonical and inferred sequence trees explained by well-known cases of systematic error due to long-branch attraction (LBA). Within the structure-only analyses, which were just as prone to LBA, we found the neighbour-joining trees inferred from the 3Di-based Fident distance matrices to outperform the model-based analyses of 3Di “sequences”. Thus, while promising and exciting, we do not yet have the methodological tool set that will enable us to routinely use tertiary protein structures in the context of single-copy ortholog phylogenomics.