DNA extraction, amplification and sequencing

For Salvia, new sequence data are presented for the internal transcribed spacer (ITS; 39 species), the external transcribed spacer (ETS; 38 species) and the plastid marker rpl32-trnL^UGA (57 species). The latter was selected based on the results of previous primer screening (trnL-F and rpl32-ndhF). Total genomic DNA was obtained from silica-dried or herbarium leaf material. DNA was extracted according to the manufacturer's protocol for the NucleoSpin^® plant DNA extraction kit (Macherey-Nagel, Düren, Germany). The standard 25 μL PCR mix consisted of 2 mm MgCl₂, 200 μm dNTPs, 1 pm primer, 0·025 U μL^–1 Taq polymerase and 0·5–1·0 μL of DNA extract in the reaction buffer provided by the manufacturer of the polymerase.

The PCRs were carried out in a Biometra T3 or a PTC 100 MJ Research thermocycler using the following program: 60 s at 94 °C; followed by 35 cycles of 20 s at 94 °C, 30 s at 55 °C and 60 s at 72 °C; and a post-treatment of 80 s at 55 °C and 8 min at 72 °C for each marker. The whole ITS region was sequenced as a single piece using the ITS-A (Noyes and Rieseberg, 1999) and ITS-4 primers (White et al., 1990). The ETS region was sequenced using 18S-E (Baldwin and Markos, 1998) and ETS-B (Beardsley and Olmstead, 2002). For plastid sequences, we used the rpl32 and trnL^UGA primers (Shaw et al., 2007). PCR products were purified according to the manufacturers' protocols using ExoSAP-IT PCR Product Clean-up (Affymetrix UK Ltd, Wooburn Green, UK) or NucleoSpin^®Extract II-kit (Macherey-Nagel).

Cycle sequencing was performed using ABI Prism Big DyeReady Reaction Mix (Perkin Elmer/Applied Biosystems, Foster City, CA, USA) using the primers listed above and following the manufacturer's protocol. Products were purified with Sephadex™ G50 (VWR International GmbH, Darmstadt, Germany) and sequenced on a 16-capillary ABI 3130 xl automated sequencer (Life Technologies GmbH, Darmstadt, Germany).

DNA sequence alignment and phylogenetic analyses

Sequencing was straightforward for each marker. Forward and reverse sequences were edited manually, merged into consensus sequences using Sequencer™ 4.1.2. (GeneCodeCorp., Ann Arbor, MI, USA) and aligned manually in McClade4.1 (Maddison and Maddison, 2000). Ambiguously alignable regions (identified manually) were excluded from analyses. The three data sets were analysed separately. In order to increase resolution, we combined nuclear and plastid data (combined data set). Partitions were defined for the combined data set before the best-fit models of nucleotide substitution were selected with jModeltest 2.1.1 (Darriba et al., 2012). Under the Akaike information criterion (AIC), the GTR+I+G model was selected for the ITS data set and TVM+G for ETS and the plastid marker. Two tree searches, one under maximum likelihood (ML) with bootstrapping (BS; RAxML-HPC BlackBox v.7.4.4; Stamatakis, 2006; Stamatakis et al., 2008) and one under Bayesian inference (BI; MrBayes v.3.1.2 on XSEDE; Ronquist and Huelsenbeck, 2003), were performed on the CIPRES Science Gateway v.3.3 server (Miller et al., 2010). Since MrBayes does not allow nst = 5, required for TVM+G, we chose the more complex model (nst = 6). For BI, we ran four Markov chains simultaneously for 10 million generations analysing the plastid and ETS data sets. Two independent runs of 40 million generations were performed for the ITS and combined data sets. Every thousandth generation was sampled. The burn-in was determined with Tracer v.1.5 implemented in BEAST. We generated 50 % majority rule consensus trees with posterior probabilities (PPs) using MrBayes v.3.1.2.

The ITS, ETS and plastid data were analysed separately to identify incongruences. To combine data without conflict, strongly supported (PP = 1·00 or ≥92 % ML BS) incongruences were dealt with by the duplication of the corresponding individuals, with one duplicate having only ITS and ETS sequences, and one only having plastid sequences (Pirie et al., 2008). The absent sequences were coded as missing data (‘?’). Sequences from the same or different species that were completely identical were reduced to one haplotype. In the text or figures, sequence identity is indicated by a slash separating the corresponding accessions. The existing concept of clades sensu Walker and Sytsma (2007) was adopted, except for ‘Clade III’. The latter was split into two independent clades, i.e. the S. trichocalycina group (Clade III) and the S. miltiorrhiza group (Clade IV) (Fig. 1).

ETS (Supplementary Data Fig. S1)

The alignment contains 69 accessions (53 species), 60 (44 species) representing the genus Salvia. The aligned length of the data set is 466 bp, 230 (40·4 %) of which are potentially parsimony informative. There are few major conflicts with the ITS topology. Differences concern a clade which is moderately supported by BI (PP = 0·98) including: (1) Clade II plus Meriandra and Dorystaechas; (2) Clade III without S. aristata (here called Clade III-A); (3) Zhumeria plus S. aristata; and (4) S. przewalskii (Clade IV represented by only one species in this data set). This clade is not supported (but also not contradicted) in the ITS analyses. The incongruence between the ITS and ETS data sets is in the position of Zhumeria. However, this difference might be based on the slightly different sampling in Clade III. Furthermore, support for a monophyletic Clade II is lacking. Instead, the three lineages of NW Salvia spp. form a polytomy with Meriandra and Dorystaechas (Supplementary Data Fig. S1].

Compared with the ITS data set, support for Clade I (Salvia s.s.) is low (PP = 0·96). The four sub-clades (sub-clades I-A through I-D) form a polytomy. Differences in the topology of sub-clade I-C are mainly based on additional accessions in the ETS data set, e.g. S. canariensis 464. The latter renders S. canariensis paraphyletic with respect to S. broussonetii, but this relationship is not strongly supported. Similarly, adding S. disermas 454 causes S. disermas to be paraphyletic with respect to S. radula. Sub-clade I-A is better resolved in the ETS than in the ITS data set. The former supports sister relationships for (1) S. cf. repens 437 and S. stenophylla and (2) S. aurita and S. dolomitica. Furthermore, S. africana-caerulea, S. albicaulis, S. chamelaeagnea and S. lanceolata × africana-caerulea appear in a moderately supported clade not supported in ITS data (PP = 0·98).

rpl32-trnL (Fig. 6)

Open in a separate window

Open in a separate window

Fig. 6.

Analyses of the rpl32-trnL data set. Non-Salvia genera are highlighted (bold); only support values ≥75 % (BS) and ≥0·95 (PP) are illustrated. Species distribution is indicated by different colours.

The aligned length of the plastid data set is 929 bp, with 246 (26·5 %) potentially informative nucleotide positions. Hyptis is again found in a trichotomy with Collinsonia and a strongly supported clade that includes all other accessions. The latter splits into two major lineages one only including non-Salvia samples (Horminum, Glechoma, Thymus, Clinopodium and Mentha). The second clade contains all Salvia samples and six additional genera. Melissa and Lepechinia are moderately supported (ML) as sister genera. They are found in a polytomy with two strongly supported clades.

The first includes Clade IV, sub-clades III-A, III-B, Zhumeria, Clade II and Dorystaechas, as found in the ETS data set. The position of Zhumeria differs from the ETS data; the genus is placed in a trichotomy with S. aristata (III-B) and the S. aegyptiaca group (III-A) based on plastid data. As in the ETS data set, the monophyly of Clade II is not supported. Instead, three sub-clades (II-A, II-B and II-C) are found in a polytomy with Dorystaechas. The most striking incongruence between nuclear and plastid data is the position of S. deserti (Fig. 6; III-A). It is sister to S. aegyptiaca based on ETS data (Supplementary Data Fig. S1) but strongly supported in a sister relationship to all species nesting in sub-clade III-A based on the plastid data set (Fig. 6). In the ITS data set, S. deserti is not represented.

The second major clade consists of a trichotomy composed of Rosmarinus, Perovskia and a strongly supported Clade I (Salvia s.s.). Compared with nuclear data, the latter is better resolved, splitting into three major lineages (sub-clades): (1) I-D (topology corresponding to nuclear data); (2) I-C; and (3) I-B plus I-A. Within sub-clade I-C, neither S. sclarea (244 and JQ669373) nor S. palaestina (400 and 200) is supported as monophyletic. Sister grouping of sub-clades I-A (Africa) and I-B (America) is in conflict with ITS data (Fig. 5). Sub-clade I-A was strongly supported by the nuclear data, but the relationships among its species were largely unresolved. However, its topology slightly differs. Instead of being part of a basal polytomy, S. chamelaeagnea 52 and S. africana-caerulea 230 are sister species in the plastid data set, closely related to S. aurita and S. scabra. The ITS data do not suggest any relationships for S. aurita, whereas the ETS data weakly support a sister relationship to S. dolomitica. Furthermore, in contrast to the ITS topology, different accessions of S. namaensis (78 and 435) and S. lanceolata (264 and 58) are not supported to be monophyletic based on plastid data.

Combined analyses of the nuclear and plastid data sets (Fig. 7)

Open in a separate window

Fig. 7.

Analyses of the combined data set. Non-Salvia genera are highlighted (bold); GenBank accessions are marked with an asterisk (*); for taxa with only one or two markers, the corresponding marker is given after the taxon name and extraction number. Uncapitalized letters (m, f) indicate taxa described after the revision of African Salvia (Hedge, 1974); classification in the corresponding species group is based on morphology and the corresponding species description (Santos and Fernández, 1986; Thulin, 1993, 2009; Van Jaarsveld, 1999). For detailed information about ancestral character reconstruction see also Supplementary Data Table S1 and Figs S2–S6.

The aligned length of the combined data set is 1982bp, of which 582 (29·4 %) are potentially parsimony informative. The sequence duplication approach was not suitable to resolve the conflicting placement of sub-clade I-B within Salvia s.s., which was placed either at the base of Clade I (ITS; Fig. 5) or as sister to sub-clade I-A (plastid; Fig. 6). Thus, we used the tree with the best topology for illustration, being aware of the unresolved conflict for the two clades.

The combined tree largely reflects the topology of the plastid data set (e.g. sister relationship of sub-clade I-A and I-B), but shows better resolution and higher support within sub-clade I-A. Melissa is in a trichotomy with two clades containing Salvia spp. The first covers (1) Clade II, Dorystaechas and Meriandra, and the latter two moderately supported as sister; (2) sub-clades III-A, III-B and Zhumeria; and (3) Clade IV. The second includes Perovskia, Rosmarinus, and Salvia s.s. (Clade I). The latter is strongly supported and falls into the same three major lineages as in the plastid data set. As to sub-clade I-A, S. taraxacifolia is sister to all remaining taxa. Salvia nilotica splits next, followed by S. somalensis, which is sister to a large clade including only taxa from southern Africa and Madagascar. Monophyly of S. namaensis, S. repens, S. sessilifolia and S. leucodermis, each of which was represented by more than one accession, is not confirmed.

Interspecific relationships in African Salvia: species groups sensu Hedge

Subgeneric classification of the genus is based on morphology and distribution (Bentham, 1832–1836, 1848, 1876; Briquet, 1897). Hedge (1974) established 23 ‘species groups’ to address relationships among African Salvia spp. and their affinities beyond the continent. For four of these groups, molecular data can be used to discuss their monophyly.

The two taxa placed in species group V (S. disermas and S. radula) form a clade within sub-clade I-C (Fig. 7). Both species occur in southern Africa but do not overlap in their distribution (Fig. 9E: 26, 34). They have similar flower morphology but differ in flowering time and indumentum (Hedge, 1974). We confirm monophyly of this group and, based on ETS and combined data sets, find some support for Hedge's (1974) idea that S. radula could be a subspecies of S. disermas (Supplementary Data Table S1).

Open in a separate window

Fig. 9.

Distribution of African Salvia spp. (except North Africa and circum-Mediterranean area). (A) Madagascar: (1) S. cryptoclada; (2) S. leucodermis; (3) S. parvifolia; (4) S. perrieri; (5) S. porphyrocalyx; (6) S. sessilifolia. (B) Canary Islands: (7) S. broussonetii; (8) S. canariensis; (9) S. herbanica; (10) S. aegyptiaca; (11) S. verbenaca. (C, D) East Africa and Arabian Peninsula: (12) S. areysiana; (13, black asterisk) S. bariensis; (14) S. deserti; (15) S. geminata; (16) S. merjamie; (17) S. nilotica; (18) S. schimperi; (19) S. somalensis. (E–G) southern Africa: (20) S. africana-caerulea; (21) S. africana-lutea; (22) S. albicaulis; (23) S. aurita; (24) S. chamelaeagnea; (25) S. dentata; (26) S. disermas; (27) S. dolomitica; (28) S. garipensis; (29, black asterisk) S. granitica; (30) S. lanceolata; (31) S. muirii; (32) S. namaensis; (33) S. obtusata; (34) S. radula; (35) S. repens; (36) S. runcinata; (37) S. scabra; (38) S. schlechteri; (39) S. stenophylla; (40, white asterisk) S. thermarum; (41) S. triangularis; (42) S. tysonii. Based on Codd (1985), Hedge (1974), Santos and Fernández (1986), Thulin (1993, 2009) and Van Jaarsveld (1999). Note the overlapping distributions of species in southern Africa (Fig. 7E–G) and the disjunct area of S. stenophylla (39) and S. disermas (26).

Species group F originally included three African species (Hedge, 1974). The two species included in this study (S. aegyptiaca and S. deserti) form a clade with four more recently described species (Fig. 7; III-A), all of which are adapted to arid or semi-arid habitats. Except for the widespread S. aegyptiaca (Fig. 10C: 10), they are all local endemics, e.g. in East Africa and the Arabian Peninsula (S. areysiana, S. bariensis and S. geminata) or Fuerteventura (S. herbanica) (Fig. 9B: 9; C: 12, 13, 15, D: 14). As the clade is well supported by synapomorphies, e.g. growth as dwarf shrubs with simple, revolute leaves, straight upper corolla lips with exposed stamens and minute flowers (Hedge, 1974; Santos and Fernández, 1986; Scholz, 1993; Thulin, 1993, 2009), we not only confirm monophyly of species group F but extend it to include at least the six sampled species. Based on the unique character syndrome, more species are likely to be included in this species group (Bokhari and Hedge, 1977; M. Will and R. Claßen-Bockhoff, unpubl. data). Furthermore, our data confirm the close relationship proposed for African and SW Asian Salvia (Davis and Hedge, 1971; Hedge, 1974).

Open in a separate window

Fig. 10.

Distribution of Salvia spp. from North Africa and the circum-Mediterranean area. (A–C) (10) S. aegyptiaca; (11) S. verbenaca; (43) S. argentea; (44) S. barrelieri; (45) S. fruticosa; (46) S. lavandulifolia; (47) S. phlomoides; (48) S. sclarea; (49) S. spinosa; (50) S. viridis; (51) S. algeriensis; (52) S. balansae; (53) S. chudaei; (54) S. dominica; (55) S. gattefossei; (56) S. interrupta; (57) S. jaminiana; (58) S. lanigera; (59) S. mouretii; (60) S. palaestina; (61) S. pseudojaminiana; (62) S. taraxacifolia. Based on Hedge (1974); the distribution of S. sclarea on the Iberian Peninsula is based on Rosúa and Blanca (1986). Note the overlapping distribution of S. phlomoides (47) and S. candelabrum (endemic to the Iberian Peninsula, indicated by ‘x’; based on Rosúa and Blanca (1986).

In contrast, relationships for the monospecific species groups H and N were not predicted before. While Hedge (1974) supposed that the allies of S. canariensis (H) occur in southern Africa, he described S. broussonetii (N) as a relict species without any close ally. However, our data point to a close relationship between these two endemics from the Canary Islands (Fig. 7).

Morphological characters used for classification: stamen types

Stamen morphology was used by Walker and Sytsma (2007) to distinguish two major lineages within Clade I. However, our analysis clearly shows that stamen morphology is much more variable. Character state reconstruction revealed that stamen type A (Supplementary Data Fig. S5) is the ancestral state for Clade III-A and Salvia s.s. (Clade I). Consequently, the reduction of the lower lever arm evolved several times in parallel. In Salvia s.s. (Fig. 8), sub-clade I-A covers the whole range of stamen modifications described in African Salvia (Fig. 8; Hedge, 1974), including the rare stamen type C (S. namaensis). In the OW, this stamen modification is only known from the Eurasian S. verticillata group (four species), which is also part of Salvia s.s. (Clade I; Will, 2013), but not closely related to S. namaensis from SSA. A third species with stamen type C and the same ontogeny as S. verticillata is Rosmarinus officinalis (unpubl. res.). Parallel evolution is thus evident (Supplementary Data Fig. S5), restricting the use of stamen types to lower taxonomic levels.

Open in a separate window

Fig. 8.

Trends in the evolution of stamen types in African representatives of Salvia s.s. (Clade I). Proceeding from the ancestral stamen type A, the hypothetical stamen evolution is illustrated. Stamen of S. whitehousei modified after Whitehouse (1949), S. interrupta modified after Rosúa and Blanca (1986), schemata of all other stamens modified after Hedge (1974); filament (medium grey); connective (light grey); theca (dark grey); stamen types (in circles; ± A = reduced type A); and clades (I-A to I-D) represented by the species are given above the branches; scale bar = 5 mm.

Floristic links of Salvia spp. in North Africa

Salvia spp. distributed in North Africa are clearly members of two different clades (Fig. 7; I-A and III-A). Based on the molecular data, they show many floristic links to southern Europe, SW Asia, East and southern Africa (Fig. 5). One example is S. taraxacifolia, a relict species endemic to the High Atlas, Morocco (Hedge, 1974). It is most closely related to East (S. nilotica and S. somalensis) and southern African species (Fig. 7; I-A). Our data suggest dispersal from North to East Africa followed by a second dispersal to southern Africa. Salvia taraxacifolia and the two East African species are adapted to mesic habitats, whereas their southern African relatives prefer arid habitats (Fig. 9E–G). This indicates that the common ancestor of sub-clade I-A might have been adapted to mesic conditions and that within the large SSA radiation (sub-clade I-A), adaptation to arid localities evolved in southern Africa.

The floristic links between North Africa and southern Europe already proposed by Hedge (1974) (Fig. 7; sub-clade I-C) were confirmed by the close relationship of S. interrupta (SW Morocco; Fig. 10B: 56) and S. candelabrum (southern Spain; Fig. 10B). Both are thermophilic and partly overlap in their distribution (Fig. 10B; Rosúa and Blanca, 1986, 1990). They have a similar habit (divided leaves, most of them at the base of the stem), conspicuous, elongated inflorescences and the same chromosome number (2n = 14) (Hedge, 1974; Rosúa and Blanca, 1985, 1990). Salvia interrupta is considered as a Tertiary relict which was more widely distributed when the climate was more mesic (Rosúa and Blanca, 1990). Both species are obviously derived from a common, probably mesic-adapted, ancestor. Their relationship might reflect allopatric speciation probably triggered by different edaphic factors in the corresponding habitats.

The strongly supported sister relationship of S. daghestanica (Caucasus) and S. phlomoides subsp. phlomoides (North Africa and southern Europe; Fig. 10B: 47) reflects floristic links between the Mediterranean and SW Asia. Since contact between the African and Eurasia floras should have increased during the Messinian Salinity Crisis in the late Miocene [5·96–5·33 million years ago (Mya)], plant colonization across the Mediterranean is expected to have occurred often during this time frame (Caujapé-Castells and Jansen, 2003). We assume that this ‘route’ was also used repeatedly by Salvia. We thus support the hypothesis of Davis and Hedge (1971) that the SW Asian origin of some NW African species was triggered by a westward shift of Irano-Turanian elements.

Repeated colonization of the Canary Islands and long-distance dispersal in Salvia

The Macaronesian flora is composed of endemics derived from an ancient Tertiary relict flora and more recently introduced species (e.g. Helfgott et al., 2000; Manen et al., 2002; Carine et al., 2004). This general pattern also appears to hold for the Macaronesian S. canariensis, S. broussonetii and S. herbanica (Fig. 9B: 7–9). Since the three species clearly differ in their morphology (Fig. 4F–H) and habitat preference (Fig. 3D, G), they were never expected to be closely related (Hedge, 1974; Carine et al., 2004). However, S. canariensis and S. broussonetii are sister species forming one clade within sub-clade I-C (Figs 5 and 7). The proposed allies of S. canariensis (Hedge, 1974) are not closely related to these two species, thus contradicting Hedge's (1974) hypothesis of a link between the Canary Island and southern African Salvia.

In contrast, S. herbanica is found in a different clade (Fig. 7; III-A) and has clear links to species from East Africa and the Arabian Peninsula. Our findings indicate the non-monophyly of the Canary Island endemics and support the hypothesis of repeated dispersals to the archipelago from different mainland sources (Emerson et al., 2000; Arnedo et al., 2001; Fuertes-Aguilar et al., 2002; Carine et al., 2004, and references therein; Vargas, 2007).

The colonization of the Canary Islands raises the question of how dispersal might have taken place. The question of how (if at all) Salvia might be adapted to long-distance dispersal (LDD) is not yet answered. For the Canary Islands, the proximity to the African continent might have eased dispersal. The oldest islands, Lanzarote and Fuerteventura, are presently 100 km from the coast of North Africa (Francisco-Ortega et al., 2000; Acosta et al., 2005), but at some periods during the last 20 million years they were probably much closer (García-Talavera, 1997). García-Talavera (1997) suggested that the volcanic sea mounts served as ‘stepping stones’ when the sea level dropped during glacial periods. In addition, recent studies have assumed convective updraft to be the key mechanism for LDD of even heavy diaspores (Nathan et al., 2002; Tackenberg et al., 2003). Long-distance dispersal to the archipelago mediated by wind is conceivable for S. herbanica, S. aegyptiaca and their potential common ancestor. Salvia aegyptiaca is distributed within the area of the Saharan Air Layer, a westward-directed wind of comparably high velocity (Carlson and Prospero, 1972; Tackenberg et al., 2003)

East Africa and the Arabian Peninsula as ‘melting pots’ for Salvia

East African Salvia is found in three independent clades (Fig. 7; III-A, I-C and I-A) and consequently has various floristic links beyond the continent.

Salvia merjamie (sub-clade I-C; Fig. 9D: 16) is a frequent and extremely variable species in the montane forest belts from Ethiopia to Zimbabwe (Hedge, 1974). It is part of a strongly supported (ITS) clade with two widely distributed species from SW Asia. Salvia merjamie is moderately supported (ML) as the sister to S. verbenaca 141 from Turkey. Both are polymorphic, occasionally have cleistogamous flowers and share the chromosome number of 2n = 42, which is uncommon in Salvia (Reese, 1957; Gadella et al., 1966; Hedge, 1974; Hedberg and Hedberg, 1977; Haque and Ghoshal, 1980; Codd, 1985; Vogt and Aparicio, 1999; Foley et al., 2008).

Salvia nilotica (sub-clade I-A; Fig. 9D: 17) has a broader distribution range than S. merjamie. Hedge (1974) assumed that S. nilotica was a distinct, taxonomically isolated species but also discussed its similarities with species restricted to the eastern Cape. With S. taraxacifolia and S. somalensis, S. nilotica is found in a basal position of sub-clade I-A, suggesting a dispersal from North to southern Africa via East Africa. The relationship of S. nilotica and the African members of section Heterosphace Benth. proposed by Hedge (1974) is confirmed, since all of these species are placed in the same sub-clade (I-A).

Salvia deserti (sub-clade III-A; Fig. 9D: 14) is an endemic of the Egyptian and Arabian deserts (Boulos, 2008; Hedge, 1974; Migahid, 1978). It is morphologically and genetically distinct from Salvia s.s. Incongruences detected in nuclear and plastid data support a hybrid origin of this species.

Pollinator diversity and evolution of bird pollination in Sub-Saharan Africa

Based on flower morphology, bees are seen as the most important pollinators in Salvia overall (Wester and Claßen-Bockhoff, 2006, 2011). This was also assumed for the African and in particular for the southern African species (Hedge, 1974). However, only a limited number of field observations confirm this view, e.g. the first report of small bees on S. africana-caerulea (Marloth, 1908), data on Anthophora diversipes (Goldblatt et al., 2000a, b), and observations of Xylocopa caffra, Amegilla spp., further solitary bees and honey-bees by P. Wester, R. Claßen-Bockhoff and H. Technau (pers. comm.).

Flowers with long tubes and freely accessible pollen were expected to be pollinated by long-tongued insects, e.g. flies. This pollinator guild is characteristic for South Africa, especially in the western Cape Region (Goldblatt and Manning, 2000). Potgieter and Edwards (2001, 2005) assumed that S. scabra and S. repens are pollinated by long-tongued flies (Stenobasipteron wiedemanni and Prosoeca spp.), but field observations are still lacking.

Although bird pollination is quite frequent in NW Salvia (Wester and Claßen-Bockhoff, 2007), the only known ornithophilous Salvia spp. in the OW appear in southern Africa (Wester and Claßen-Bockhoff, 2006). Three, S. lanceolata, S. thermarum and S. africana-lutea, are restricted to southern Africa (Fig. 4I, R, S). Probably two more (S. leucodermis and S. sessilifolia; Fig. 4P, Q) occur in Madagascar. Our data confirm that ornithophily evolved repeatedly in the NW (Fig. 7), but also indicate at least two pollinator shifts from bee to bird pollination in Africa (Supplementary Data Fig. S6). Most probably, this pollination system evolved three times in parallel within sub-clade I-A: (1) in the Madagascan sub-clade; (2) in the S. africana-lutea and S. lanceolata clade; and (3) in S. thermarum. Including species adapted to bee, bird and, most probably, long-tongued fly pollination, sub-clade I-A represents a further example of a monophyletic lineage having undergone pollinator-driven diversification in southern African (e.g. Van der Niet and Johnson, 2012; Sun et al., 2014; Van der Niet et al., 2014).

In bird-pollinated species, bees are largely excluded from nectar access but not from collecting pollen. They might therefore trigger hybridization between bird- and bee-pollinated species (Van Jaarsveld, 2002; P. Wester, University of Düsseldorf, Germany, pers. comm.), e.g. in S. africana-caerulea. The species is morphologically well adapted to bee pollination but is also occasionally pollinated by birds (Wester, 2013). It might be a species in which the exploitation of a food plant by pollinators (birds) can be observed even though both species are not yet perfectly adapted to each other (Thomson and Wilson, 2008). Thus, S. africana-caerulea might represent an example of a pollinator shift in progress (Rodríguez-Gironés and Santamaría, 2004).

We thank Dirk Albach (Oldenburg), Safi Bagherpour (Ankara), Benny Bytebier (KwaZulu-Natal), Ferhat Celep (Nevşehir), Ahmet El-Banhawy (Ismailia/Redding), Ahmed Kahraman (Ankara), Alexander P. Sukhorukov (Moscow), Mats Thulin (Uppsala) and Petra Wester (Düsseldorf) for sampling, and the following herbaria for offering plant material: ACECR (Iran), B, E, EA, GOET, HUH, M, MJG, MO, MPU and MW. Photographs were kindly provided by Ferhat Celep, Rafael N. B. Groneberg (Mainz), Dylan Hannon (San Marino, CA), Peter B. Phillipson (Paris), Hen Technau, Mats Thulin and Petra Wester. We thank Berit Gehrke and Michael D. Pirie (both Mainz) for assistance with data analyses, and Natalie Schmalz, Abigail J. Moore (Providence, RI) and two anonymous reviewers for helpful comments to improve the manuscript. This work was supported by the DFG (Deutsche Forschungsgemeinschaft; Cl 81/10–1) and the Fachbereich Biologie (Universität Mainz).