This study is largely based on Appelhans et al. (2020). In addition to that dataset, we added two samples of the new species collected from spatially separated individuals in the same (Wai‘ahi) population, as well as one sample each of M. fosbergii, M. helleri and M. linearifolia Hosaka (Suppl. material 5). With the exception of the O‘ahu endemic M. degeneri Hosaka, all currently accepted species of Hawaiian Myrsine have now been included in a phylogenomic study based on RADseq. Appelhans et al. (2020) used two species of Ardisia Sw. as outgroups. In order to have more closely-related taxa as outgroups, we omitted the Ardisia samples and rooted the phylogenetic trees with the earliest branching clade of Hawaiian Myrsine that consists of M. lanaiensis Hillebr. and two specimens of the polyphyletic M. lessertiana (Clade C in Appelhans et al. 2020). The final dataset included 31 samples (Suppl. material 5).

Genomic DNA was extracted from silica-dried leaf material using the Qiagen DNeasy Plant Mini Kit (Qiagen, Hilden, Germany) following the manufacturer’s instructions. A Qubit 3.0 fluorometer (Thermo Fisher Scientific, Waltham, MA, USA) in combination with the Qubit dsDNA Broad Range assay kit was used to measure the quality and quantity of the DNA extractions. The normalised samples (30 ng/µl) were sent to Floragenex (Eugene, OR, USA) for library preparation using the restriction enzyme SfbI and for sequencing on an Illumina HiSeq machine, which produced 2 × 150 bp paired-end reads. De-multiplexed raw reads for all samples have been deposited at the Sequence Read Archive (SRA; https://www.ncbi.nlm.nih.gov/sra; Suppl. material 5) under the BioProject number PRJNA614459 together with the sequence reads from Appelhans et al. (2020).

All steps, from demultiplexing to the calculation of the alignments of RAD loci, were carried out using ipyrad 0.9.52 (Eaton and Overcast 2020). One mismatch in the barcode sequence was allowed for demultiplexing. Adapters were removed and read filtering was done by deleting reads with more than five low-quality bases (< 20), a phred Q score offset of 33 and removing trimmed reads shorter than 35 bp. RAD loci were assembled de novo using clustering thresholds of 85% for within and amongst sample clustering. Appelhans et al. (2020) tested two different clustering thresholds (85% and 90%) and did not record any significant differences, so that only the 85% threshold – which is the default setting – is used here. As a trade-off between number of RAD loci included and missing data, five datasets were assembled that differed in the minimum number of samples recovered per RAD locus. This minimum number was set to either > 25% (min8; a RAD locus was only included in the alignment in case a sequence had been recovered for at least eight of the 31 samples), > 33% (min11), > 50% (min16), > 66% (min21) or > 75% (min24). With increasing “min” numbers, the numbers of RAD loci, but also the amount of missing data, were expected to drop. The resulting alignments were used for phylogenetic reconstruction using RAxML 8.2.4. (Stamatakis 2014), applying the GTR + Γ model and calculating 100 bootstrap replicates.

All ipyrad and RAxML analyses were computed on the high-performance computing cluster of the “Gesellschaft für wissenschaftliche Datenverarbeitung Göttingen” (GWDG), Germany (https://gwdg.de/en/hpc/services/).

Herbarium specimens of the new species have been deposited at PTBG and other herbaria listed under specimens examined and in Suppl. material 5 (acronyms according to Thiers (2018)). All measurements were taken from dried herbarium specimens and field notes and are presented in the descriptions as follows: length × width, followed by units of measurement (m, mm or cm). The authors have examined all specimens cited. We assessed the extinction risk for the new species following the IUCN Red List Categories and Criteria (IUCN 2012) and guidelines of the IUCN Standards and Petitions Committee (IUCN 2022). The extent of occurrence and area of occupancy were calculated by using ArcMap 10.6.1 (ESRI 2011) in relation to coordinates recorded while collecting herbarium specimens and making field observations. The coordinates latitude and longitude have been truncated to protect the exact holotype location from unauthorised access.

The sequencing runs yielded an average of 4,425,834 raw reads per sample of which an average of 4,422,490 reads remained after trimming. This is less compared to the sequence reads from the Appelhans et al. (2020) study (5,505,232 raw reads, 5,281,476 reads after filtering; Suppl. material 5), but the newly-generated sequences had more bp overall because they were sequenced with the 2 × 150 bp paired-end chemistry instead of single end 100 bp in the previous study. Despite the higher number of bp in the newly-sequenced samples, the numbers of retained RAD loci was much lower for the new samples (18× lower in the min8 dataset; 19× lower in min11; 16× lower in min16; 5× lower in min21; 2× in min24 dataset), which might be due to the different sequencing strategies (2 × 150 bp vs. 1 × 100 bp) and the different fragment size selection during library preparation in particular.

As expected, the assembled datasets varied greatly in numbers of RAD loci, alignment length and amount of missing data. The min8 dataset contained 55,048 loci with a concatenated alignment length of 4,832,837 bp and 49.87% missing data, while the min24 dataset contained only 317 loci with an alignment length of 35,821 bp and 35.78% missing data (Suppl. material 5).

Despite the large differences in the datasets, the consensus trees showed a congruent backbone, in which the same three main clades (Clades A, B and C) and the division of Clade A in two subclades (Subclades A1 and A2) are inferred as in Appelhans et al. (2020) (Fig. 1, Suppl. materials 1–4). However, average branch support differs amongst the datasets. The min16 and min8 consensus trees had the highest average branch support of 89.2% bs and 88.9% bs, respectively, followed by the min11 consensus tree with 84.2% bp, the min21 consensus tree with 79.9% bs and the min24 consensus tree with only 60.0% bs. Due to the higher support values, we discuss the consensus tree, based on the min16 dataset (Fig. 1, Suppl. materials 1–4) and the other consensus trees are mentioned in case of supported differences.

The two samples of the new species M. cirrhosa and the specimens of M. helleri and M. linearifolia are resolved within Subclade A2 (Fig. 1). This subclade consists of species endemic to Kaua‘i. With the exception of M. denticulata (Wawra) Hosaka, which has small leaves with a dentate margin, all species in this subclade are characterised by linear, narrowly elliptic to narrowly lanceolate leaves (Wagner et al. 1999; Appelhans et al. 2020). The new species M. cirrhosa fits well into this clade regarding its distribution and morphology.

Myrsine cirrhosa is resolved as the closest relative of M. helleri. In the min21 and min24 phylogenies, the two species are resolved as monophyletic sister species (Suppl. materials 3, 4). The other phylogenies did not resolve M. helleri as monophyletic. In the min11 and min16 phylogenies, M. helleri forms a grade at the base of M. cirrhosa (Suppl. material 2, Fig. 1). In the min8 phylogeny, one sample of M. helleri is sister to a clade that consists of the second M. helleri sample as well as M. cirrhosa, M. fosbergii and M. linearifolia (Suppl. material 1). A denser taxon sampling is needed to address the correct placement and potential polyphyly of M. helleri.

Myrsine fosbergii (Kaua‘i, O‘ahu) is part of Clade B in all analyses, except the phylogeny of the min8 dataset (Suppl. material 1), where it also belongs to subclade A2. In all, but the min8 phylogeny, the backbone support of Clade B is rather low, which is potentially caused by the M. fosbergii specimen. In the min8 phylogeny, bootstrap support values of Clade B are generally higher and also the placement of M. fosbergii in Subclade A2 as sister to M. linearifolia is highly supported. The placement of M. fosbergii in Subclade A2 is highly supported by morphology and the species is characterised by narrowly elliptic leaves. The different phylogenetic placements of this species might be due to the low number of recovered RAD loci instead of a biological reason such as introgression or hybridisation. The number of loci is the second lowest after M. lanaiensis 3 for the min8, min11, min16 and min21 datasets and is the third lowest after M. lanaiensis 3 and M. knudsenii (Rock) Hosaka for the min24 dataset (Suppl. material 5). On the other hand, the estimated heterozygosity, which is indicative of introgression and hybridisation, in the M. fosbergii sample is only slightly above the average (Suppl. material 5).

To accommodate Myrsine cirrhosa, the following couplets can be inserted into the beginning of the existing key to Hawaiian Myrsine (in Wagner et al. (1999, p. 935)). Note: K = Kaua‘i; O = O‘ahu.

According to the guidelines set by the World Conservation Union (IUCN 2012, 2022), Myrsine cirrhosa is classified as Critically Endangered (CR), indicating a very high risk of extinction in its natural habitat. This assessment, summarised by the IUCN alphanumeric criteria (CR B1ab(i,ii,iii,v); B2ab(i,ii,iii,v); C2a(i); D), is based on the fact that the species has a severely limited Extent of Occurrence (EOO) of only 2 km², an Area of Occupancy (AOO) of approximately 1 km² and population size of fewer than 50 individuals. Threats to the habitat of Myrsine cirrhosa include introduced non-native animals that destroy native vegetation such as pigs (Sus scrofa), rats (Rattus spp.), slugs (Meghimatium striatum) and occasional goats (Capra hircus) and black-tailed deer (Odocoileus hemionus). Remote island ecosystems have low resistance to non-native competitors, especially introduced animals and plants which can be devastating to native species that have evolved in their absence (Carlquist 1974; Weller et al. 2011, 2018). Invasive non-native plant species that compound habitat degradation around M. cirrhosa include Andropogon virginicus L., Axonopus fissifolius (Raddi) Kuhlm., Miconia crenata (Vahl) Michelang. (syn. Clidemia hirta (L.) D.Don), Cyperus meyenianus Kunth, Erechtites valerianifolius (Link ex Spreng.) DC., Juncus planifolius R.Br., Paspalum conjugatum P.J.Bergius, P. urvillei Steud., Pterolepis glomerata (Rottb.) Miq., Rhodomyrtus tomentosa (Aiton) Hassk., Rubus argutus Link, R. rosifolius Sm., Sacciolepis indica (L.) Chase, Setaria parviflora (Poir.) Kerguélen, and Sphaeropteris cooperi (Hook. ex F.Muell.) R.M.Tryon. Landslides after heavy rains also can be a very serious threat, especially along the windswept ridge colonies of M. cirrhosa where a single landslide could destroy large sections of native habitat.

Myrsine cirrhosa most closely resembles M. helleri. RADseq resolved this as its closest relative (Fig. 1, Suppl. materials 1–4), but the new species can be consistently distinguished from the latter species by its strongly undulated leaf margins (Table 1, Figs 2, 3A, B). One collection was initially identified as M. helleri and distributed under that name (i.e. Perlman 14606), but it clearly represented M. cirrhosa on critical examination. Leaves of M. cirrhosa tend to be comparatively larger than those of M. helleri and have a combination of undulate margins and a cirrhose apex, in addition to the inflorescences having more numerous (4–7) flowers per fascicle and longer petals. Populations of M. helleri from Wahiawa and Namolokama may have leaves with a slightly hooked apex, but the margins are not undulate and inflorescences have fewer (1–3) flowers per fascicle. The type of Myrsine helleri is from the headwaters of the Wahiawa Stream area on Kaua‘i, where it may grow sympatrically with M. cirrhosa (see Wagner et al. (1999) for synonymy and Wagner and Shannon (1999) for typification). Myrsine cirrhosa also was observed to grow sympatrically with M. helleri at Wai‘ale‘ale summit (Wood & Nishek 11683, BISH, PTBG). However, the two species remain distinct morphologically. One collection from Wai‘ale‘ale summit (Wood 3894, PTBG, US) is intermediate, having larger leaves with undulate margins, but lacking a cirrhose apex. It likely represents a hybrid between M. cirrhosa and either M. helleri or M. fosbergii, which is also sympatric here, but needs further investigation.