Samples were washed with 70% ethanol, then kept at −80°C for 12 h (to facilitate cell wall breakage) and milled in an MM400 (Retsch Inc.) grinder. Genomic DNA was extracted using a modified-CTAB protocol (Doyle and Doyle, 1987), further adjusted to improve yield from herbarium samples (Supplementary Data Sheet S2A). Key changes include incubating samples in CTAB at 65°C for 12 h (to optimize DNA isolation) and in isopropanol at −20°C for 48 h (to improve precipitation of fragmented DNA). Precipitated DNA pellets were washed twice with 70% ethanol and resuspended in Milli-Q ultrapure (Type 1) water (Merck KGaA). We measured relative DNA concentration (ng/μL) with the QuantiFluor^® dsDNA System (Promega Corp.). Samples were purified using Agencourt AMPure XP beads (Beckman Coulter Life Sciences) (Supplementary Data Sheet S2B). Extractions from pre-1970 collections or with concentrations <10 ng/μL were treated as having highly fragmented DNA and were cleaned using AMPure beads and isopropanol following Lee (2014) to reduce loss of small (<300 bp) DNA fragments (Särkinen et al., 2012). Where material was available, we repeated extractions to obtain at least 200 ng of DNA from each sample. DNA fragment size distribution was determined by gel electrophoresis (Supplementary Data Sheet S3A). Extractions with fragments predominantly above the target insert size (≥ 500 bp) were sonicated with an ME220 Focused-ultrasonicator^TM (Covaris Inc.).

We prepared genomic libraries using the NEBNext^® Ultra^TM II DNA Library Prep Kit and Multiplex Oligos (Dual Index Primers, sets 1 and 2) for Illumina^® (New England Biolabs) at half-volumes to reduce per sample costs. Target insert size was 350 bp and size selection was not required where DNA template was highly degraded (<300 bp). Size-selected libraries were amplified with 13 PCR cycles and all others with 14 cycles. Libraries were re-amplified, where required, using KAPA HiFi HotStart ReadyMix (Roche) with i5 and i7 forward and reverse primers (as described in Meyer and Kircher, 2010) to obtain at least 75 ng/library.

Libraries were multiplexed (11–12 per pool) for hybridization. Equimolar library pools were made homogenizing phylogenetic distances (if known) and avoiding combinations of libraries originating from different quality DNA to reduce uneven bait competition within hybridization pools. We considered the following criteria: (i) whether re-amplification was required, (ii) whether sonication was required, (iii) whether size selection was required, and (iv) whether there was sufficient library template (Supplementary Table S3B). Outgroup taxa were pooled separately from Papuasian taxa where possible to even out occupation of bait binding sites. Each pool contained 500–1,000 ng DNA in total.

Pools were enriched using the Angiosperms-353 myBaits^® Expert Panel target capture kit (Arbor Biosciences). They were hybridized at 65°C for 24 h, then amplified with i5 and i7 forward and reverse primers (Meyer and Kircher, 2010) for 14–18 PCR cycles to obtain pools at least 1 nM. Each pool was quality-controlled with a TapeStation 4200 (Agilent Technologies) (Supplementary Data Sheet S3C). Due to funding constraints, we only sequenced the eight library pools with the highest quality, which covered the broadest diversity range and consisted of 90 accessions in total (though only 74 passed quality filters). These were denatured and diluted following manufacturer’s specifications (Illumina^® protocol # 15039740) and loaded at 16 pM for sequencing in an Illumina^® MiSeq using two v2 (300-cycles) reagent kits (Illumina^®, Inc.) at the Jodrell Laboratory (Royal Botanic Gardens, Kew, Richmond, United Kingdom).

Sequences were trimmed using Trimmomatic 0.38 (Bolger et al., 2014), employing “palindrome mode” adapter removal and Maximum Information quality filter settings to favor longer reads (ILLUMINACLIP:TruSeq3-PE-2.fa:2:30:10:2:TRUE MAXINFO:40:0.2 LEADING:3 TRAILING:3 MINLEN:36). We examined sequence quality with FastQC 0.11.7 (Andrews, 2018) before and after trimming to ensure complete adapter removal and to identify surviving artifacts that could affect downstream analyses.

HybPiper 1.3.1 (Johnson et al., 2016) was used to retrieve target sequences of nuclear genes (exonerate script) and flanking off-target regions (intronerate.py script, which reruns exonerate but, instead of removing flanking regions, it keeps them), with the BWA mapper (Li, 2013) and the SPAdes assembler (Bankevich et al., 2012) (–bwa -cov-cutoff = 3). We investigated polymorphic sequences for paralogy using exploratory trees built from MAFFT 7.215 (Katoh and Standley, 2013) (–auto) alignments in FastTree 2 (Price et al., 2010) (-nt -gtr). We discarded any sequence that may have resulted from gene duplication and retained the most common allele when sequences were homologous (Kates et al., 2018). We used our own custom script, max_overlap³ (Supplementary Data Sheet S4A), to calculate a coverage score for each sequence that is proportional to representativeness (proportion of accessions/loci with sequences), data matrix completeness (percent sequence recovered against overall length of target), and evenness of distribution (adapted from Pielou, 1966) across accessions per locus and so, to reduce noise in our data matrixes by filtering out underrepresented, incomplete, and unevenly distributed sequences.

Internal Transcribed Spacer (ITS1-5.8S-ITS2) nuclear ribosomal DNA sequences (hereafter ITS) were also retrieved with HybPiper using a target file we made from aligned ITS sequences – from Li and Wen (2014) and Plunkett et al. (2004, 2005) – deposited in GenBank (Benson et al., 2018). Off-target sequences corresponding to 142 plastid loci (genes and intergenic spacers) were similarly rescued with HybPiper using a plastid-target file we generated from the complete plastid genomes of S. heptaphylla (L.) Frodin (Zong et al., 2016: KT748629), Aralia elata (Miq.) Seem. (Kim and Yang, 2016: KT153023), S. delavayi (Franch.) Harms, and Metapanax delavayi (Franch.) J.Wen & Frodin (Li et al., 2013: KC456166, KC456165).

All sequences were aligned with UPP (Nguyen N.-P.D. et al., 2015) to produce accurate alignments from fragmentary datasets (i.e., historical herbarium DNA template), using only those with >95% of the longest available sequence length for the backbone dataset. UPP uses hidden Markov models (HMM) for multiple sequence alignment (MSA) and it relies on PASTA 1.8.4 (Mirarab et al., 2015) (a divide-and-conquer MSA method) to generate initial backbone alignments. In turn, PASTA relies on FastTree 2 (Price et al., 2010) for tree estimation, MAFFT 7.215 (Katoh and Standley, 2013) for alignment, and OPAL 2.1.3 (Wheeler and Kececioglu, 2007) for merging. We trimmed alignments using our own custom script, optrimAl⁴ (Supplementary Data Sheet S4B), which optimizes the gap threshold value in trimAl 1.2 (Capella-Gutiérrez et al., 2009), to obtain the highest proportion of parsimony-informative characters (P_PIC) while retaining adequate sequence length (Shen et al., 2016). Alignments where trimming resulted in data loss exceeding 30% were interpreted to contain low ratios of phylogenetic signal to noise and were discarded. Alignment statistics were calculated in AMAS 0.98 (Borowiec, 2016) using the summary function. Sequence capture statistics were calculated using R 3.5.3 (R Core Team, 2019).

Gene trees were inferred with IQ-TREE 1.6.10 (Nguyen L.-T. et al., 2015) after selecting substitution models with ModelFinder (-m TEST) (Kalyaanamoorthy et al., 2017). Outlier branches that increased the diameter of each gene tree by more than 20% were identified using TreeShrink 1.3.1 (Mai and Mirarab, 2018) with centroid re-rooting (-b 20 -c) and removed. Each locus was then realigned without the outlier sequences. We estimated bipartition support with 1000 UFBoot2 (Hoang et al., 2018) bootstrap replicates using IQ-TREE (-bb 1000) and contracted branches with support values below 10% (Mirarab, 2019) using Newick Utilities 1.6 (nw_ed “i & b ≤ 10”) (Junier and Zdobnov, 2010).

Pseudo-coalescent species trees were inferred using ASTRAL III v5.6.3 (Zhang et al., 2018). Species trees were inferred separately for the nuclear and chloroplast genomes as they represent different evolutionary pathways (Ravi et al., 2008). We used the local posterior probabilities (PP_local) calculated in ASTRAL to estimate quartet support for the recovered topology at each node. Conflict, concordance, and phylogenetic signal were assessed with phyparts (Smith et al., 2015) and displayed with the PhyPartsPieCharts script⁵, which depicts the number of gene trees that support, oppose or provide no information with respect to the dominant species tree topology. Unresolved polytomies in the final species tree were tested in ASTRAL (-t 10) to determine if they are due to insufficient data or could possibly reflect a true polytomy (Sayyari and Mirarab, 2018).

To verify placement of our sampled Papuasian and outgroup accessions within family Araliaceae, we inferred the ITS gene tree for these together with sequences of other Araliaceae accessions from Li and Wen (2014) and Plunkett et al. (2004, 2005). We used two accessions from Myodocarpaceae (Delarbrea paradoxa Vieill. and Myodocarpus fraxinifolius Brongn. & Gris.) as the outgroup. The gene tree was estimated in MrBayes 3.2.6 (Ronquist et al., 2012) under a GTR + Γ substitution model (with settings: nchains = 4, nruns = 4, and MCMC ngen = 100M).

To limit variation in substitution rates and minimize overestimation of recent divergence times (Ho et al., 2005), we selected nuclear genes that (i) were at least 10% concordant with the species tree (bipartition support > 0.1), (ii) likely evolved according to a strict molecular clock (we set root-tip variation to <0.024), (iii) contained the most information (tree length > 0.1), and (iv) represented the most accessions (at least 25) with SortaDate (Smith et al., 2018). To reconstruct the biogeographic history of Papuasian Schefflera (Crisp et al., 2011), we included in the data matrix ITS sequences from other Asian Schefflera clades sampled by Li and Wen (2014). All other loci for these Asian accessions were coded as missing data. This resulted in a 51-taxa data matrix, partitioned by locus (we used independent substitution models for each locus, as selected by ModelFinder), consisting of only Asian and Papuasian Schefflera sequences. This data matrix comprised 31,496 bp across 10 nuclear exonic regions (6,469 bp), 15 nuclear flanking regions (24,177 bp), and nuclear rDNA ITS. The data matrix had 11.6% parsimony-informative sites and, as 16 taxa (32% of the total sampling) are represented only by ITS sequences, 55.5% missing data.

Divergence times were estimated in BEAST 1.10.4 (Suchard et al., 2018) on the CIPRES Science Gateway v3.3 online platform (Miller et al., 2011). Since known Araliaceae fossils lay outside our focus group, three secondary time constraints – drawn from Li and Wen (2014) – were imposed on: (i) the root node (normal prior distribution with mean = 42.0 and st.dev. = 8.0); (ii) the Heptapleurum crown node (normal, mean = 36.0, st.dev. = 7.0); and (iii) the Papuasian Schefflera crown node (normal, mean = 22.0, st.dev. = 5.0). To reduce search-space and avoid miss-rooting problems with BEAST analyses, we enforced monophyly on the Agalma, Brassaia, S. elliptica alliance, Heptapleurum, Heptaphylla + Hypoleuca, Ischyrocephalae, and Papuoschefflera s.s. clades (fully supported in our species tree), as well as the root node, to prevent inverted ingroup-outgroup topologies (for further details see Springer et al., 2018).

Concurrently, ancestral areas were reconstructed using a fully probabilistic approach – first described by Sanmartín et al. (2008) and implemented in BEAST by Lemey et al. (2009) – that infers diffusion processes among discrete locations in timed evolutionary histories under Bayesian stochastic search variable selection (BSSVS). Continuous-time Markov chains (CTMC) are used to model instantaneous geographic locations of any given sequence, together with the transition and migration rates between these locations. Since this Bayesian CTMC phylogeographic model assumes ancestral ranges are limited to single regions, it requires discretizing the entire distribution of any given taxon (i.e., a given taxon can be represented by multiple accessions), while tolerating incomplete sampling (Drummond et al., 2012). Moreover, unlike other approaches (e.g., dispersal-vicariance parsimony or dispersal-extinction cladogenesis), it can be implemented in scenarios where the number of areas is large (>10 areas), allowing for fine-scale area explorations (e.g., Mairal et al., 2015, 2017). Thus, we included a partition with collection localities – coded according to tectonic plate boundaries (Bird, 2003) – in our BEAST input file (generated in BEAUTi 1.10.4; Suchard et al., 2018) and we tested six models, comprising all possible combinations of three clock priors – strict, random local, and uncorrelated relaxed lognormal – and two species-tree priors robust to incomplete sampling (our case) – Yule process (Yule, 1925) and birth-death incomplete sampling (Stadler, 2009). We selected the best-supported model by estimating marginal likelihoods (MLEs, path steps = 100, chain length = 1M), using path sampling (PS) and stepping-stone sampling (SS) (Baele et al., 2013), from runs that converged after 100M iterations. BEAST log files were loaded into Tracer 1.7.1 (Rambaut et al., 2018) and visually inspected to check that the chains had converged, and that mixing and Effective Sample Sizes (ESS > 200) were adequate for all parameters (after 100M iterations). After discarding burn-in iterations, trees were annotated and posterior probabilities (PP) summarized in TreeAnnotator 1.10.4 (Suchard et al., 2018) on the tree in the posterior sample with the maximum sum of the posterior clade probabilities (MCC tree), rescaling to reflect median node heights for clades contained in said tree. The resulting MCC tree was visualized in FigTree 1.4.4 (Rambaut, 2018).

Though our estimated crown age of ∼26.3 Ma for the Papuasian Schefflera clade is older than the crown age of ∼21.9 Ma estimated by Li and Wen (2014) – which is the source of our secondary time constraints –, as expected, the highest posterior density intervals for both these estimates overlap. These divergence dates may seem to be at odds with the assertion by Lohman et al. (2011) that most of New Guinea was submerged prior to 20 Ma. Yet, several studies have also inferred the origin of Papuasian taxa back to the Oligocene (Cibois et al., 2014; Kodandaramaiah et al., 2018) or even as early as the late Eocene (Jønsson et al., 2011; Cozzarolo et al., 2019). During the Oligocene, there may have been an island archipelago, located where present-day New Guinea is, formed by the collision of both the Philippine-Sea and the Pacific plates with the Australia plate (Hill and Hall, 2003). The largest landmass would probably have been where the present-day Papuan Peninsula is and have resulted from the docking of the East Papua Composite Terrane (EPCT; part of the Woodlark plate) onto the Australia plate (Davies, 2012). Stratigraphic examination of sediments deposited in the Aure Trough provides evidence for mountain building on the Woodlark plate during this period (van Ufford and Cloos, 2005), indicating that the terranes forming the Papuan Peninsula were already emergent. Our reconstruction of the Woodlark plate as the best-supported ancestral area for Papuasian Schefflera (Figure 5) is consistent with the above scenario.

Hall (2012) estimated that the Sahul and Sunda shelves first collided ∼25 Ma (see Figure 32 in Hall, 2012). Phylogenies reconstructed by Li and Wen (2014) and Plunkett et al. (2005) and this study support Papuasian Schefflera as nested within Sundan Heptapleurum. It is thus within reason that our ancestral area reconstruction suggests that the common ancestor of Papuasian Schefflera dispersed from Sunda to colonize the Woodlark plate right before this contact took place (Figure 5). In this scenario, we hypothesize that a light-loving and pioneer ancestral Papuasian Schefflera would have rapidly colonized areas of the New Guinea land mass as they gradually emerged above sea level. The Woodlark plate could, therefore, have functioned as a source area for the colonization of New Guinea along the predominantly west-to-east axis of the Sunda-Sahul floristic exchange (Crayn et al., 2015).

Our reconstruction of the Sahul shelf as the ancestral area for Brassaia points to a vicariance scenario for the early evolutionary history of this clade (Figure 5). The Brassaia crown is dated at ∼18.1 Ma, which is earlier than the ∼5 Ma date estimated for the emergence of the Sahul shelf to form the southern half of New Guinea (Hall, 2009). As Brassaia also occurs on the southern fall of the Owen Stanley Range (near present-day Port Moresby, in the “tail” of Papua New Guinea) and is partly formed by the Sahul shelf, it is possible that ancestral Brassaia evolved in isolation from ancestral Papuoschefflera s.l. on either side of the proto-Owen Stanley Range during the early Miocene.

The placement of S. kraemeri in Brassaia supports the morphological circumscription of the group, despite this species’ disjunct distribution with regard to the rest of the group. Schefflera kraemeri is found only in the Truk Islands, more than 800 km away from Papuasia. It is most likely to have arrived via long distance dispersal as there are no intervening landmasses in the Pacific Ocean to serve as stepping-stones. Human-mediated dispersal is unlikely, not only because of the inferred timing (which predates humanity), but also because Schefflera has limited uses in Papuasia and S. kraemeri has not been recorded among the region’s indigenous people as a useful species (Cámara-Leret and Dennehy, 2019a,b). The estimated divergence time of ∼6.8 Ma is consistent with the geological age of these islands, which were determined to have been the result of volcanism ∼11 Ma (Keating et al., 1984).

The divergence of the major clades of Papuoschefflera s.l. between ∼24.6 and 16.7 Ma overlaps with the period in the early Miocene when most of New Guinea was submerged (Figure 5). While Ischyrocephalae and the Eastern clade are inferred to have remained on the Woodlark plate at this time, Papuoschefflera s.s. may have arisen from an early dispersal to then-emerged islands in the Maoke plate, corresponding to the present-day Maoke Mountains. The splitting of the Eastern clade into Cephaloschefflera and the Oreopolae + Barbatae clade probably resulted from another dispersal to Sunda shelf terranes between ∼17.3 and 12.6 Ma. Indeed, zoochorous dispersal across narrow water barriers has been found to play an important role in the intercontinental floristic exchange of the Malesian flora (Crayn et al., 2015). The Papuasian Schefflera, with their fleshy drupaceous fruits, could have been widely dispersed (for example by birds; Snow, 1981) across the proto-Papuasian archipelago.

Accessions from islands located in the Bird’s Head plate illustrate how Schefflera species may have colonized islands in geological times. The shallowest seafloor connecting the islands of Biak and Waigeo to mainland New Guinea is about 200 m deep, precluding an overland connection even during the Pleistocene, when sea level was up to 120 m shallower than nowadays (Voris, 2000). The divergence of S. “bougainvill.biak” sp. ined. ∼5.6 Ma coincides with the end of a 6-Myr upper Oligocene hiatus in deposition in the Biak Basin near Biak Island (Gold et al., 2014), indicating that sea level was shallower at that point in time. A larger area of land would have been exposed on both the island and the mainland, facilitating dispersal over a narrower water barrier. Similarly, the divergence of Schefflera sp. (Raja Ampat) ∼4.7 Ma coincided with an active period of Pliocene deformation, resulting in more land emerging above sea level (Charlton et al., 1991). The placement of this collection as sister to S. “KB22” (Sorong) sp. ined. suggests descent from a common ancestor that occupied the NW Bird’s Head peninsula.

Ischyrocephalae is restricted to upper montane forests and subalpine grasslands, which could be explained in terms of phylogenetic biome conservatism – phenomenon that has been observed in vascular plants from the Malesian island of Borneo (Merckx et al., 2015) and also worldwide (Crisp et al., 2009). Schefflera ischyrocephala and S. pilematophora are found in the Saruwaged Range on the NE coast of New Guinea, which is interpreted to be a terrane of the South Bismarck plate. These taxa are inferred to have arrived separately from the Woodlark plate sometime between ∼11.6 and 10.7 Ma to the present, respectively, concurrent with the Finisterre volcanic arc accretion to the EPCT (Davies, 2012). This pattern could suggest that adaptation to montane regions in ancestral Papuoschefflera s.l. may have a role in promoting the highly diverse “sky island” flora of New Guinea (Sklenáø et al., 2014). Given the prevalence of high- and mid-elevation taxa both within the clade and its most recently diverging sister lineages, as well as a putative temperate origin for Asian Schefflera (Valcárcel and Wen, 2019), Papuasian Schefflera may have been pre-adapted to these environments (Antonelli, 2015). Pre-adaptation has also been invoked in a broad range of high-elevation taxa on Gunung Kinabalu (Merckx et al., 2015), a Malesian mountain located in Borneo, with uplift timing similar to that of the New Guinea Highlands. In both cases, however, this hypothesis remains to be tested by ancestral trait reconstruction, while accounting for diversification rate shifts.

Based on morphological characters, Philipson (1978) mused that Schefflera may be “vigorously diversifying at the present time.” In Papuasia, the species richness of various plant and animal taxa has been shown to peak at mid-elevations (Colwell et al., 2016; Vollering et al., 2016). Papuasian Schefflera exhibits a similar distribution in species richness, which may have arisen from rapid adaptive radiation into new ecological niches created by mountain-building. This evolutionary mechanism has also been observed in the genus Pseuduvaria (Annonaceae) on New Guinea (Su and Saunders, 2009), various genera of Andean alpine plants (Nürk et al., 2018), and Ericaceae in montane regions worldwide (Schwery et al., 2015). Many biotic and environmental factors, including its presumed temperate ancestry and the ready availability of new uncolonized habitats resulting from the uplift of the New Guinea Highlands, would have predisposed Papuasian Schefflera to speciation via biome shifts (Donoghue and Edwards, 2014). Our chronogram hints to a speciation burst in the last 12–5 Myr and we hypothesize recent adaptive radiation, facilitated by mountain building, could be the primary driver for the present-day diversity of Schefflera. Further insights into this phenomenon could be achieved by examining an expanded sample of taxa for trait shifts and diversification rates along an elevational gradient.