MetaPhOrs statistics. The orthology assignments for 829 complete genomes were mapped onto the tree of life (NCBI taxonomy tree). Bar charts around the tree represent the fraction of each genome for which orthologs have been identified (green) and with no orthologs identified (grey). The total length of each bar (grey + green fractions) is proportional to the logarithm of the number of genes in the genome. A higher resolution, interactive figure, showing the coverage of each independent dataset (PhylomeDB, Ensembl, EggNOG, Fungal Orthogroups, COG and TreeFAM) is available online (MetaPhOrs Overview at: http://orthology.phylomedb.org/?q=stats). The figure was constructed using iTOL MetaPhOrs statistics. Detailed statistics of MetaPhOrs and all of its subsequent databases are provided in Supplementary Table S2.

Parameter evaluation. The accuracy of predictions were investigated applying various cut-offs for the likelihood filter (A), orthology consistency score (B) and EL (C). The harmonic mean (F_1.0, precision and recall equally weighted, see ‘Materials and Methods’ section) was calculated based on a subset of TreeFam-A reference set for human–mouse, human–zebra fish and human–fruit fly [100 orthogroups as in (24)] and YGOB reference set for S. cerevisiae-C. glabrata, and S. cerevisiae-A. gossypii. For the sets evaluated on TreeFam-A benchmark we did not use trees coming from this database.

Accuracy of the MetaPhOrs approach using different datasets. Recall and precision scores of our pipeline applied to individual datasets (blue rhombus), combined datasets (red squares) and the full MetaPhOrs approach (black double triangles) were calculated based on TreeFam-A reference set (see ‘Materials and Methods’ section). Note that results on accuracy do not correspond to predictions as given by a given repository (e.g. OrthoMCL), but to our phylogeny-based approach based on trees derived for data contained in such repository (e.g. species-overlap algorithm applied on trees derived from OrthoMCL families). In order to avoid circularity in our benchmark, trees coming from TreeFam-A were not considered in any dataset. For the combined methods, predictions from two or more sources were summed together: orthology was assigned if confirmed by at least one repository, paralogy was assumed only if there were more paralogy signals than orthology. For the full MetaPhOrs approach, we used several level (EL) thresholds; for instance, for EL = 2 only predictions confirmed by any combination of 2 independent sources (phylomes or databases) are taken into account. A consistency threshold (CS_o) of 0.5 is applied. Plotted curves represent combinations of recall and precision providing identical Fβ scores as the best performing method. Two scenarios are considered: recall and precision are equally weighted (blue thin line, F1.0 = 0.817); or precision is two times more important than recall (grey thick line, F0.5 = 0.837). The ranking of the best methods can be defined based on relative distance of each method to the curve representing F score of the best scoring method. MetaPhOrs with EL cut-off of 2 (MO el = 2; F1.0 = 0.817) and MetaPhOrs with EL cut-off of 3 (MO el = 3; F1.0 = 0.797) are the best performing approaches in the first scenario. In the second scenario, MetaPhOrs with EL cut-off of 3 (MO el = 3; F0.5 = 0.837), MetaPhOrs with EL cut-off of 4 (MO el = 4; F0.5 = 0.824) and MetaPhOrs with Evidence level cut-off of 2 (MO el = 2; F0.5 = 0.807) perform the best.