The study sites that were used for the estimation of dietary profiles of domesticated cattle were located within the eastern semi-arid savanna in the Limpopo Province of South Africa (SA) across two bioregions, namely the Central Bushveld- and the Lowveld Bioregion (Mucina and Rutherford 2006). Welverdiend (24°34'15"S, 31°20'42"E) is a rural village located close to Bushbuckridge in the SA Lowveld which supports a communal rangeland along the periphery of the village. This area is predominantly populated by broad-leaved woody species, like Senegalia nigrescens (Oliv.) PJH Hurter, Sclerocarya birrea subsp. caffra (Sond.) Kokwaro, Albizia harveyi E Fourn., and various species from the Combretaceae family (Mucina and Rutherford 2006). The grass layer is mostly composed of species such as Digitaria eriantha subsp. eriantha (Steud.) Maire, Eragrostis rigidior Pilg., Melinis repens (Willd.) Zizka, Schmidtia pappophoroides Steud. ex JA Schmidt, Urochloa mosambicensis (Hack.) Dandy, and Aristida congesta Roem. & Schult. Additionally, the area boasts the forb species from genera such as Commelina, Hibiscus, Sida, and various others from the Malvaceae, Fabaceae, Convolvulaceae, and Acanthaceae families (Mucina and Rutherford 2006). Syferkuil (23°50'36.86"S, 29°40'54.99"E) is the experimental farm of the University of Limpopo where rotational grazing is being practised (Scheiter et al. 2024). The vegetation of the Syferkuil study area has been described as an open thorn bush savanna, specifically belonging to the Pietersburg Plateau False Grassveld type (Acocks 1988). In this ecosystem, the herbaceous layer is mainly composed of perennial grass species such as Themeda triandra Forssk., Digitaria eriantha Steud., Schmidtia pappophoroides Steud. ex J.A. Schmidt, and various species of Eragrostis. Additionally, the woody component is predominantly represented by Vachellia tortilis (Forssk.) Galasso & Banfi (Low and Rebelo 1998). The climate of both sites is semi-arid with a mean annual rainfall of 350–500 mm, most of which occurs in the summer months (Shackleton 2000; Matsika et al. 2013; Phefadu and Kutu 2016; Munjonji et al. 2020). Comprehensive plant species lists were compiled as described by Botha et al. (2023). The rbc L and a trnL DNA barcode sequence reference libraries of these plants were developed and evaluated by Botha et al. (2023).

DNA was extracted from the faeces of 24 individual cattle from Welverdiend and 24 individual cattle from the Syferkuil Experimental farm during the end of the dry season, specifically directly before the first rains of the new season when forage supply from the grass layer is particularly scarce (Linstädter et al. 2013). DNA was also extracted in the same manner from a blank control made up of molecular water, labelled as sample C2. The faecal samples from Welverdiend were collected in October 2018 and those from Syferkuil in October 2019. Great care was taken to ensure that these faecal samples were representative of individuals either by visual observation or by allowing a sufficient gap in time and space in the collection efforts. Approximately 10 g of fresh faecal material was collected after defecation per individual and ~200 mg of this was used for subsequent DNA isolation using the Qiagen DNeasy mini–Plant DNA Extraction kit according to the manufacturer’s protocol. Faecal samples were collected without interacting with or disturbing animals.

The rbcL and trnL(UAA) intron chloroplast gene regions from 48 faecal samples as well as the extraction blank control, sample C2, were amplified using the barcode’s metagenomic sequencing library protocol (Illumina). Two PCR reactions were completed for rbcLa and trnL(UAA) library preparations and incorporated the Illumina overhang adapters onto the primers: rbcLa-F (Levin et al. 2003), rbcLa-R (Kress and Erickson 2007), trnLc and trnLd (Taberlet et al. 1991), which are further described in Suppl. material 1. Each PCR reaction had a final volume of ~25 µl which consisted of 12.5 ng DNA template, 5 µl (1 µM) of the forward and reverse primer, respectively, 12.5 µl 2X Kapa HiFi HotStart ReadyMix (KAPA Biosystems Woburn, MA), and 2.5 µl PCR-grade water.

The PCR amplification of these reactions, as well as negative control reaction without any template DNA, were carried out simultaneously as follows: heated lid 110 °C, 95 °C for 30 seconds, 25 cycles of 95 °C for 30 seconds, 55 °C for 30 seconds, 72 °C for 30 seconds, then 72 °C for 5 minutes and held at 4 °C. Gel electrophoresis (1X TAE buffer, 1.5% agarose, 100 V) was used to visualise the PCR results and successful PCR products were purified using AMPure XP magnetic bead-based purification (Beckman Coulter, IN). Multiplexing was made possible by the Illumina Nextera XT Index kit (Illumina, San Diego, CA) using two 8-base indices. Two indexing primers per sample were utilised and PCR reactions with the purified DNA template (5 µl DNA template, 5 µl × 2 index primer, 25 µl 2x KAPA HiFi HotStart ReadyMix and 10 µl PCR-grade water) were performed on a Bio-Rad C1000 Thermal Cycler (Bio-Rad, Hercules, CA). The cycling conditions consisted of an initial denaturation cycle of 95 °C for 3 minutes, followed by eight cycles of denaturation (95 °C for 30 seconds), annealing (55 °C for 30 seconds), and extension (72 °C for 30 seconds). The cycle was completed with a final extension cycle of 72 °C for 5 minutes. The indexed libraries were purified with Ampure XP beads before library pooling, and their quantities were determined using the Qubit dsDNA HS Assay Kit from Thermo Fisher Scientific (Waltham, MA). Prior to sequencing, the quality of the purified amplicons was evaluated using the Agilent Bioanalyzer (Agilent Technologies, Santa Clara, CA). Following instructions from Illumina, the sample pool (2 nM) was denatured with 0.2 N NaOH, diluted to 7 pM, and mixed with 10% (v/v) denatured 20 pM PhiX. Samples were sequenced on the Illumina MiSeq platform (2 × 300 bp) by the NICD Sequencing Core Facility service, in Pretoria, South Africa.

Cutadapt (v. 4.1: Martin 2011) was used to remove PCR primers from the rbcL and trnL amplicons and additional quality filtering, trimming, sample inference, dereplication and the removal of chimeric sequences with the "dada2" R package (v.1.24: Callahan et al. 2016) on forward reads. The default parameters of DADA2 included the removal of sequences containing ambiguous bases, sequences that were truncated at base pair positions at the first instance of quality scores less than 20, and sequences that were shorter than 50 bp. Default parameters were therefore used in most cases, except that reads were not merged. The reason for this is that a minimum of 20 base pairs are required to overlap to create a consensus sequence and it would be unsuccessful for larger amplicons expected for the trnL intron. This choice is not unusual, especially when large amplicons or amplicons with highly variable lengths or low-quality reverse reads are present as seen for both the rbcL and trnL reads (Pauvert et al. 2019; Tan et al. 2019). Data handling hereafter will therefore only be applied to unmerged, forward reads. The amplicon sequence variants (ASVs) left after chimaera removal and resulting from sample inference were subject to additional filtering and curation.

Frequency-based filtering was applied to both the rbcL and trnL datasets, which aided in removing ASVs only present in a single sample and an additional abundance filter, identified as necessary when tested with a simulated mock community (Botha 2023 and Suppl. material 2), was applied to the rbcL datasets which removes ASVs representing 0.005% of the total dataset (Bokulich et al. 2013). The removal of these low-frequency features, which may represent contamination, PCR, or sequencing errors, results in decreased false positives and a truer reflection of animal foraging behaviour. The filtered ASV tables for rbcL and trnL were then subject to post-clustering curation, or post-denoising in the case of DADA2, by the LULU algorithm. The R package “lulu” (v.0.1: Frøslev et al. 2017) identifies and merges often erroneous ‘daughter’ OTUs with consistently abundant co-occurring sequences, or ‘parent’ OTUs (Frøslev et al. 2017). Curated ASVs remaining in the feature tables for rbcL and trnL were assigned to the taxonomy contained in a custom, study-area-specific reference libraries (Botha et al. 2023) using the “blastn” command under the “BLAST+” tool to only keep one identification per ASV that has similarity threshold of 80% and an e-value of 0.00001. The rbcL reference library consisted of 1 238 sequences representing 318 genera and 562 species. The trnL reference library consisted of 921 sequences representing 270 genera and 461 species.

As part of an additional quality control step, the unassigned ASVs resulting from taxonomic assignment by the compiled reference libraries by Botha et al. (2023) were “BLAST searched” against the contents of NCBI’s non-redundant nucleotide (nt) database on the 13^th of September 2022. This was done to establish (i) whether the unassigned reads had non-plant origins, and (ii) whether the unassigned ASVs are attributed to poor taxonomic coverage of the reference libraries. Reads that were assigned to species other than plants, with a percentage identity of 80%, query coverage of 80% and e-values smaller than 0.001 were removed from the feature tables for each dataset since they were not a reflection of the inability of the reference libraries to identify the composition of herbivores foraging in an eastern semi-arid South African savanna and cannot truly be classified as “unassigned”.

To determine the sum of branch lengths (SBL) of the trnL and rbcL datasets, the ASVs were aligned in the Multiple alignment program for amino acid or nucleotide sequences (MAFFT, v.7: Katoh and Standley 2013) and then used to reconstruct neighbour-joining (NJ) phylogenetic trees for the trnL and rbcL datasets, as implemented by the package “ape” (v.5.6.2: Paradis and Schliep 2019) in R (v.4.3.1: R Core Team 2023) with bootstrap testing of 1 000 replicates using the Jukes-Cantor model. SBL was calculated for both the assigned and the unassigned reads after the removal of non-plant sequences, frequency and abundance-based filtering and the application of the LULU algorithm for the rbcL and trnL amplicon per bioregion.

Alpha-diversity was determined using the Shannon index and was calculated based on the abundance matrices obtained for each primer dataset for both assigned and unassigned reads sampled from each bioregion using the estimate-richness function of the "phyloseq" (v.1.48: McMurdie and Holmes 2013) package. All analyses were performed using ASVs that were taxonomically identified by the reference libraries created by Botha et al. (2023). The Shannon index provided insights into the species richness detected by each marker within each treatment group. The Kruskal-Wallis test was used to compare multiple independent groups within the filtered datasets. It was performed using R software to evaluate the statistical difference in plant species richness and diversity between the different markers and bioregions. Plots were created using the package “ggplot2” (v.3.3.6: Wickham 2016). The p-values less than 0.05 were considered significant.

Taxonomic assignments of the rbcL and trnL ASVs retrieved from the Syferkuil and Welverdiend cattle faecal samples were visualised via read abundance, frequency of occurrence (FOO), also known as absolute data, as well as the relative read abundance (RRA) of ASVs assigned to family and genera per sampled individual. Each taxon was grouped according to plant life forms, i.e., woody taxa (trees/shrubs), woody taxa/forbs (shrubs/forbs), forbs, and graminoids (grasses). All statistical analyses were carried out in R (v.4.3.1: R Core Team 2023). Absolute read abundances were calculated and visualised as bar plots with the “phyloseq” package to show the number of detected ASVs for each marker per study site. These bar plots were employed to summarise and visualise the number and proportion of assigned and unassigned ASVs (NA reads). Distinct approaches regarding treating unassigned reads were used for FOO and RRA calculations and visualisations to ensure a robust analysis and accurate portrayal of taxonomic proportions across samples. NA reads were omitted for FOO calculations to focus on the occurrence of taxa that could be assigned to the respective rbcL and trnL datasets. The FOO percentages of families and genera were calculated by summing all present values per specified family or genus (1 = present; 0 = absent) for samples in the dataset, dividing it by the number of samples and multiplying this total by 100. FOO used bar graphs to visualise the data. In contrast, RRA calculations and visualisations retained NA reads to accommodate for accurate interpretations of taxa proportions. NA reads are prevalent in amplicon sequencing datasets due to the absence of taxa in reference databases or inadequate and differential amplification of certain taxa by barcoding primers. The omission of NA reads from RRA visualisations (100% stacked bar plots) would lead to discrepancies in the interpretation of taxa proportions, particularly when comparing samples with different degrees of NA read removal. Furthermore, each sample was assigned to a dominant life form category (Graminoids, Forbs, Forbs/Woody taxa and Woody taxa) to contextualise RRA within a dietary analysis framework. In cases where NA reads were dominant, the second most abundant life form assigned to families or genera was used for categorisation. Samples were then grouped into facets and arranged according to the descending order of abundance of these dominant life forms. RRA percentages were reported based on taxa proportions of assigned reads constituting each sample and were calculated by dividing the count of each unique ASV belonging to a specific genus or family in a sample by the sum of all assigned ASVs in that sample. RRA and FOO bar graphs were created with the “ggplot2” package.

The Welverdiend rbcL dataset comprised of 10 209 509 sequences and 4 989 550 ASVs, of which 494 were left after filtering and post-clustering curation. For the trnL dataset, 3 894 691 sequences resulted in 1 852 442 ASVs of which 1 217 were left after filtering and post-clustering curation. The Syferkuil rbcL dataset consisted of 1 858 096 sequences and resulted in the creation of 1 444 146 ASVs of which 1 403 were left after filtering and curation. Of the 1 985 791 sequences that were input for the Syferkuil trnL dataset, 1 234 380 ASVs were created and 3 320 were left after filtering and trimming. ASVs that were assigned to non-plant references and were subsequently removed from further analysis included 17 from the Welverdiend trnL dataset, 51 from the Syferkuil rbcL dataset and 18 from the Syferkuil trnL dataset. Furthermore, the blast results indicated that ASVs classified as unassigned by the compiled reference libraries could also not be matched with any plant species in NCBI’s non-redundant database within the specified parameters.

To compare the phylogenetic distances of the species detected with the trnL and rbcL at each of the foraged communities, Welverdiend and Syferkuil, NJ trees from each treatment for the trnL and the rbcL ASVs were compared. The phylogenetic distances or SBL for the trnL ASVs were higher than the rbcL ASVs at both sites. Specifically, at Welverdiend, the SBL for trnL (295.51) exceeded that for rbcL (137.58), while at Syferkuil, trnL (934.78) also exceeded that of rbcL (358.77). The SBL for both markers was higher at Syferkuil than at Welverdiend. The alpha diversity metrics (Shannon index) revealed that the highest diversity and richness of consumed plant species (both assigned and unassigned) was detected by the trnL marker at the Syferkuil sampling site (Fig. 1). This site also showed the highest diversity for the rbcL marker. Plant species richness differed significantly between bioregions and markers (Kruskal-Wallis test, p < 0.05).

Figure 1.

Boxplots representing Shannon Diversity Index for unassigned (a), and assigned ASVs (b), calculated per barcode (rbcL and trnL) for Welverdiend (WV) and Syferkuil (SK).

The FOO analysis of both the rbcL and trnL datasets from Welverdiend (Fig. 2) showcased the diverse utilisation of various life forms within this semi-arid ecosystem at the end of the dry season, revealing discernible patterns of variations in the FOO among these life forms. The FOO% calculations for the Welverdiend rbcL dataset showed a consistent presence of woody taxa across the majority of samples. Albizia and Combretum were predominant (FOO = 96%), whereas forbs, Malvastrum and Thunbergia, and graminoids, Trachypogon, had lower FOO values of 8%. Despite the low FOO% values displayed by forb taxa identified by the rbcL marker, the trnL marker analysis showed the significance of forbs in the dietary composition, with Malvaceae emerging as the most frequently ingested plant family with a FOO of 92%. Ebenaceae and Poaceae, representing woody taxa and graminoids respectively, followed closely with 84% each (Fig. 2). Conversely, the forb family, Asteraceae was the least frequently occurring family in the dataset (FOO = 8%). Among the genera identified by the trnL marker, the woody taxa Euclea and Senegalia as well as the forb, Ipomoea, showed high frequencies, with FOO values of 84%, 72%, and 48%, respectively. Other genera belonging to the woody taxa were the genera Ficus and Ehretia, representing the least frequent dietary items with each recording a FOO of 12%.

Figure 2.

Bar chart showing the percentage of frequency of occurrence (FOO%) of the plant (a) families and (b) genera, identified as part of the diets of Welverdiend cattle, faceted according to their life form. White bars represent taxa identified with rbcL and grey bars represent those identified with trnL.

In terms of forb abundance, the Welverdiend rbcL dataset showed lower abundance for genera like Malvastrum and Thunbergia compared to the dominance of woody taxa, but trnL identifications highlight the importance of forbs with Malvaceaea being the dominant family. This marked difference in datasets underscores the need for a dual-locus approach to fully capture the richness and abundance of forbs in cattle diets. The RRA calculations for the Welverdiend rbcL dataset (Fig. 3), showed that forbs/woody taxa were most abundant in 18 samples, followed by six samples where woody taxa were most abundant. Fabaceae, Combretaceae, and Ebenaceae were identified as the main food items across most samples. These families represent a variety of life forms, especially including woody taxa like Albizia, Combretum, Euclea, and Diospyros, indicating the significant contribution of shrubs and trees to the diets of Welverdiend cattle. However, variations from this overall pattern were observed for specific samples. Sample 16 exhibited a dominance of Malvaceae (forbs/non-woody shrubs), with greater inclusion of Poaceae (graminoids) compared to other samples. Samples 22 and 23 stood out with Anacardiaceae (woody taxa) as the dominant food source. The genus Rhus (shrub with synonym Searsia) belonging to this family, was identified as an important food item in samples 17, 18, 19, 4, and specifically samples 22 and 23, where Rhus represented more than 75%.

Figure 3.

Relative read abundance (RRA) of families (a, b) and genera (c, d) identified with the rbcL (a, c) and trnL (b, d) barcoding marker from the Welverdiend dataset. The stacked bar plots the distribution of plant taxa in individual samples arranged according to their dominant life forms relative to the number of unassigned reads, with codes indicating graminoids (G), forbs (F), forbs/woody taxa (F/W) and woody taxa (W). For the trnL charts, the sample with no families identified (C2) and samples with no genera identified (4 and 16) were excluded from the visualisations.

Forbs/woody taxa were again identified as the dominant life form in most samples in the Welverdiend trnL dataset followed by woody taxa (Fig. 3). Malvaceae, composed of forbs and non-woody shrubs, was the most abundant family in many samples. Ebenaceae, composed of woody taxa such as Euclea, also played a significant role in the contribution of identified families in samples 6, 2, 3, and 21. While less abundant than the latter families, Convolvulaceae was also important across many samples and included the forb Ipomoea. Grasses (Poaceae), including Eragrostis, Setaria, and Enteropogon, served as a vital food item across samples, however, it did not dominate the dietary compositions; its relative abundance was lower compared to other families like Malvaceae and Ebenaceae in most samples (Fig. 3). Other forb taxa such as Asparagaceae, Rubiaceae, and Euphorbiaceae also contributed to cattle diets in smaller relative amounts. Furthermore, the analysis of genera identified for the Welverdiend trnL dataset revealed that Euclea (woody taxa) was the most abundant, followed by Ipomoea (forb), Vachellia (shrub/tree), Senegalia (shrub/tree), Asparagus (forb) and Eragrostis (grass) as well as Ficus (shrub/tree) in samples 18 and 23.

At Syferkuil, a different pattern emerged, with graminoids and woody taxa dominating both marker datasets, but particularly so in rbcL (Fig. 4). Forb and woody taxa frequency was lower overall compared to Welverdiend, with graminoids dominating the cattle diets. This is evident from the rbcL dataset where Poaceae was the most frequently occurring family (88%). Fabaceae emerged as the second most frequently occurring family (75%), which comprises both forbs and woody taxa. This was followed by Rhamnaceae, containing woody taxa, with a 67% FOO. Conversely, the families representing forbs, Geraniaceae, and Solanaceae were the least frequently occurring families, with 8% each. Tricholaena, a graminoid, was the most frequently occurring genus (79%). This was followed by Cenchrus (graminoid) with 71%, and Ziziphus (woody taxa) with a 71% FOO. The least frequent genera were Berchemia (woody taxa), and the forbs Monsonia and Solanum with FOO values of 8%, each. In the trnL dataset (Fig. 4), the dominance of grasses was confirmed with Poaceae having 96% FOO. Like in the rbcL dataset, this was followed by Fabaceae (woody taxa/forbs) with 79% FOO. In the trnL dataset, Caryophyllaceae (forb) emerged as one of the most frequently utilised families with a FOO of 25%. Other families representing forbs such as Solanaceae, Lamiaceae, and Euphorbiaceae were the least frequent (8%). The genus Setaria, representative of graminoids, was the most common (88%), followed by Vachellia (woody taxa) with a 58% FOO. Pollichia and Indigofera, consisting of forbs each had a 25% FOO while other forbs such as Leucas, Senna, and Solanum, and a woody taxon, Senegalia, only had an 8% FOO.

Figure 4.

Bar chart showing the percentage of frequency of occurrence (FOO%) of the plant (a) families and (b) genera, identified as part of the diets of Syferkuil cattle, faceted according to their life form. White bars represent taxa identified with rbcL and grey bars represent those identified with trnL.

Reinforcing the dominance of grasses in the diets of cattle at Syferkuil, the rbcL marker (Fig. 5) primarily identified graminoids across most samples, with notable genera such as Setaria, Cenchrus, and Tricholaena. Woody taxa such as Vachellia (Fabaceae), Flueggea (Ebenaceae), and Ziziphus (Rhamnaceae) were also significant components. Forbs such as Pollichia showed some diversity across the samples, although in smaller abundances compared to other life forms. Forbs were identified as the dominant life form in 5 samples. The forb family, Caryophyllaceae, was especially abundant in sample 25 where it was the only family detected, and Solanaceae which appeared in small amounts in samples 33 and 46. Another important forb family was Apocynaceae which represented more than 90% of the identified plant families in sample 41. Following the patterns of grass-dominated diets at Sygerkuil, trnL RRA calculations showed that Poaceae was the dominant life form across 20 samples. Woody taxa (Fabaceae) also contributed a relatively large proportion (Fig. 5). Additionally, Setaria (grass) and Vachellia (shrub/tree) were identified as dominant genera in Syferkuil cattle diets with Euclea (shrub/tree), Pollichia (forb), Cleome (forb) and Indigofera (forb/shrub) in lower relative abundances. Leucas (forb) was also identified as an important food item in samples 28 and 47 and Senna (forb) in smaller proportions in samples 28 and 37.

Figure 5.

Relative read abundance (RRA) of families (a, b) and genera (c, d) identified with the rbcL (a, c) and trnL (b, d) barcoding marker from the Syferkuil dataset. The stacked bar plots the distribution of plant taxa in individual samples arranged according to their dominant life forms relative to the number of unassigned reads, with codes indicating graminoids (G), forbs (F), forbs/woody taxa (F/W) and woody taxa (W). For the trnL charts, sample C2 had no family or genus identifications and was excluded from the visualisations.

In summary, the RRA analysis across both sites showed a predominant reliance on graminoids and woody taxa, with forbs contributing to the diets, but in smaller relative amounts. The cattle diets at Welverdiend showed a higher overall dominance of forb and woody taxa compared to those of Syferkuil, which relied more heavily on grasses. The use of both the rbcL and trnL markers further highlighted the differences in the identification of plant families and genera, with trnL providing more detailed insights into forb consumption. Despite these differences in resolution between loci, there was consistency across both datasets in the identification of key families such as Fabaceae, Ebenaceaea, and Poaceae. These families were prominent in both Welverdiend and Syferkuil. However, the broader diversity captured by trnL, especially forbs, emphasises the importance of using multiple markers to obtain a comprehensive view of the dietary composition of the cattle foraging in east semi-arid African savannas.

In this study, we determined the dietary composition of cattle foraging in two bioregions of the eastern semi-arid South African savanna and demonstrated that forbs and woody taxa make a substantial contribution to the overall diet of cattle during the dry periods. The taxonomies described using the different barcoding regions highlight the need for the use of a dual-locus identification system for plants (CBOL Plant Working Group 2009; Hollingsworth et al. 2011). The tendency of barcoding loci to fail or weakly amplify specific taxonomic groups (CBOL Plant Working Group 2009; Zhu et al. 2022) allows for the systematic underrepresentation of inefficiently amplified species and the omission of food items. Using multiple loci to recreate diet profiles allows for mutual verification, which helps reveal such disparities (Camp et al. 2020). The use of the rbcL and trnL chloroplast genes resulted in complementing taxonomic identifications, facilitating the reconstruction of herbivore diets from an eastern semi-arid South African savanna (Figs 2, 4, Table 1). Frequently consumed food items in the diets of Welverdiend cattle, according to the rbcL and trnL marker, were from the Fabaceae, Combretaceae, Malvaceae, Ebenaceae and Anacardiaceae. The family Fabaceae, Poaceae and Ebenaceae also occurred in relatively greater proportions in the diets of these cattle according to the RRA graphs except for Malvaceae in the rbcL dataset, where it made up ~63% of the diet of one individual and Combretaceae, which the trnL barcode could not identify. A well-established reason for the differing success of plant barcoding loci is due to the completeness of barcode reference libraries, especially those containing representative barcode records from South Africa (Rattray et al. 2024). Dissimilarity in the recorded plant diversity identified by barcoding loci could be due to added information due to differing taxonomic breadths (Fahner et al. 2016), which may explain why Combretaceae was missing from identifications with the trnL dataset. This family was represented by 12 unique species in the combined reference libraries, of which only three were barcoded with trnL and only consisted of one reference sequence each (Botha et al. 2023). The rbcL barcode, however, was represented by all 12 species, of which nine were unique to this dataset, with two or more reference sequences each. Interestingly, although Malvaceae was one of the most frequently occurring families in the Welverdiend trnL dataset, it could not be identified to genus level, but it was in the rbcL dataset in the form of Malvastrum where it accounted for ~8% of the FOO and ~63% RRA in sample 16.

The high number of trnL ASVs left after filtering, combined with its greater phylogenetic distances and alpha diversity metrics, indicates that trnL is more robust than rbcL for capturing the diversity and phylogenetic structure of food items contained in herbivorous diets. Syferkuil exhibited the highest alpha diversity (Shannon index) and species richness for both markers, especially for trnL. This suggests that plant communities contained in cattle diets are more diverse compared to those at Welverdiend (Kruskal-Wallis test: p < 0.05). A substantial number of ASVs remained unassigned after comparison with reference libraries and could also not be assigned to plant species in the NCBI nt database. This indicates potential gaps in these reference sequence libraries or the presence of unknown or poorly characterised plant species in eastern semi-arid South African savannas, further emphasising the need for expanded reference databases for accurate metabarcoding analyses that will ideally facilitate multi-barcode approaches for plant identifications (Botha et al. 2023).

Except for the differences in families identified, the rbcL and trnL barcoding markers have different affinities for amplifying certain genera from a common family identified in the diets of cattle foraging in each study site. The taxa under consideration, as well as the marker used, have a significant impact on the success of species identification using DNA metabarcoding (De Vere et al. 2012; Mallott et al. 2018; Amandita et al. 2019) and dissimilarities in the identified taxa may increase at lower taxonomic levels to result in marker disagreement at genus level plant diversity (Yoccoz et al. 2012; Fahner et al. 2016). RbcL and trnL has modest discriminatory power (Chase et al. 2005; Newmaster et al. 2006; Taberlet et al. 2007; Hollingsworth et al. 2009; Valentini et al. 2009; Quéméré et al. 2013), but the observation of an overlapping identification between these markers can provide increased support for the findings (Kress and Erickson 2007; Fahner et al. 2016). In the Welverdiend study site, the family Fabaceae was represented by Albizia, Dichrostachys and Philenoptera as identified by the rbcL barcoding marker and Senegalia and Vachellia as identified by the trnL barcoding marker. For Syferkuil this same family was represented by Dichrostachys and Vachellia (shrubs and trees) by the rbcL marker and Senegalia (shrub/tree), Vachellia (shrub/tree), Indigofera (forb) and Senna (shrub/forb) by the trnL marker. The rbcL barcoding marker therefore exclusively identified the genus Dichrostachys across both datasets and the trnL marker exclusively identified Senegalia.

Identified food items across the studied sites were representatives from 25 different plant families, of which 16 included forbs as a potentially foraged life form (Table 1). Forb genera were frequently identified within the sampled datasets as foraged items at both Welverdiend and Syferkuil. Genera representing forb species that were identified for the Welverdiend cattle are Malvastrum and Thunbergia as identified by the rbcL marker and Asparagus, and Ipomoea identified with trnL. These genera (Malvastrum and Thunbergia) are not present in the trnL reference dataset and did not appear frequently across the sampled dataset with values of 8%. Forb genera were identified at a much higher frequency in the trnL dataset with values of 40% for Asparagus and 48% for Ipomoea. The trnL barcoding marker therefore could allow for the identification of the genera Asparagus and Ipomoea even though there were reference sequences available in the rbcL barcoding library. For the Syferkuil cattle the genera Pollichia (forb; FOO = 25%), Adenium (succulent shrub; FOO = 50%), Solanum (forb/shrub; FOO = 8.33%), and Monsonia (forb; FOO 8.33%) were included in their diets as identified by the rbcL marker, as well as Solanum (forb/shrub; FOO = 8.33%), Cleome (forb; FOO = 16.7%), Indigofera (forb; FOO = 25%), Leucas (forb; FOO = 8.33%) and Pollichia (forb; FOO = 25%) identified by the trnL marker. Only the genera of Pollichia and Solanum could be identified by both the barcoding markers, again indicating the complementary nature of this dual-locus identification system that aids in the determination of dietary profiles when reference sequences are not available for all barcodes (Taberlet et al. 2007) or when a barcode shows preferential amplification of certain plant taxa (Kartzinel et al. 2015; Mallott et al. 2018; Zhu et al. 2022).

According to the results, forbs are foraged by a noticeable percentage of individuals in the sampled dataset, especially species from the genera Asparagus and Ipomoea at Welverdiend and Adenium, Pollichia and Indigofera at Syferkuil. Different forb community compositions and diversity across these two study sites (Mucina and Rutherford 2006) were confirmed by the RRA and FOO visualisations. Welverdiend had the highest frequency of occurrence of forb families (families including forbs/woody taxa) across the sampled datasets (i.e., higher forb species richness in the diet), while Syferkuil results illustrated a greater relative abundance of the forbs identified in individual diets (i.e., lower richness, but higher relative abundance per species). This suggests that although the forbs occurring at Welverdiend were foraged upon by many members of the sampled dataset, they have not foraged heavily, and their diets still included a high abundance of other life forms. In contrast a higher preference for forbs was observed at Syferkuil. The overall dietary composition of cattle foraging in a semi-arid African savanna at the end of the dry season shows a transition from usual grass-based forage to a stronger reliance on forbs and leaves from woody taxa such as trees and shrubs.

The interaction between availability and selection determines the actual diet consumed by animals (Julian et al. 2024). When availability is high, animals can be more selective, potentially improving their nutritional intake. When availability is low, selection becomes constrained, and animals may be forced to consume less preferred or lower-quality foods. In Welverdiend, forbs and woody taxa from families such as Malvaceae and Ebenaceae are the main parts of cattle diets. This indicates how crucial these life forms become when there is less grass available, or as shown by Craine (2021) and Craine et al. (2015), when graminoids have lower nutrient concentrations with warming climate and large mammalian herbivores rely on shrubs and forbs for their protein intake. This was also observed at Syferkuil, though grasses still dominate the cattle diets, consumption of woody plants and forbs from families such as Ebenaceae and Caryophyllaceae also occurs. This reliance on forbs and woody life forms in the dry season suggests that they have a flexible foraging strategy that allows them to survive through periods of drought while maintaining dietary diversity and a nutritional balance.

Recognising and understanding these dietary shifts in cattle diets can guide adaptive management practices and will ensure the sustainable use of forage in semi-arid landscapes while sustaining livestock productivity. To mitigate the impacts of nutrient stress imposed by increasing temperature and decreasing precipitation, managers should consider implementing rotational grazing, providing supplemental feeding when necessary, and conserving the diversity of not only grasses but of shrubs and forbs in grazing areas to support flexible foraging strategies during periods of scarcity in dryland landscapes (Samuels et al. 2016; Scasta et al. 2020; Julian et al. 2024).

While the analysis provides valuable insights into the foraging behaviour of these herbivores, it’s important to consider potential biases inherent to the methodology. PCR bias may result in selective amplification of certain taxa, leading to an over- or underrepresentation of specific plant taxa. Furthermore, not all taxa that are present in the samples may have been successfully amplified and identified, potentially skewing the perceived diversity and richness of the cattle diets. Moreover, the faecal samples that were collected represent a “snapshot” of what an individual consumed immediately before sampling, rather than providing an overview of their long-term dietary habits. The authors recommend exploring methods to mitigate PCR bias and improve the representativeness of sampling by incorporating multiple sampling points over time. Additionally, integrating dietary observations or tracking methods alongside genetic analyses could offer a more holistic understanding of cattle diets and foraging behaviour.

The authors have declared that no competing interests exist.

The North-West University Animal Production Sciences Research Ethics Committee approved the study as zero-risk and granted written permission for its completion. The ethics number is NWU-00177-18-A5.

We thank the North-West University for its support of this research. The research was also supported by the German-South African research project “South African Limpopo Landscapes Network” (SALLnet) funded by the German Federal Ministry of Education and Research (BMBF; Grant number: 01LL1802C).

Conceptualization: SB, AL, FS. Data curation: FS, SB, DB. Formal analysis: DB. Funding acquisition: AL, FS. Investigation: AL, FS, DB, KB, MVM, SB. Methodology: DB, MP, SB. Project administration: FS, KB, SB, AL. Resources: ZTHM, SB, FS, MA, AI, AL. Software: DB, MP. Supervision: SB, FS. Validation: SB, DB, FS, AL. Visualization: DB. Writing - original draft: DB, FS, SB. Writing - review and editing: AI, MVM, KB, AL, MA, MP, ZTHM.

Danielle Botha https://orcid.org/0000-0002-7321-5382

Sandra Barnard https://orcid.org/0000-0002-9299-2769

Morné du Plessis https://orcid.org/0000-0003-4154-7830

Mushal Allam https://orcid.org/0000-0002-9875-6716

Kai Behn https://orcid.org/0000-0003-2598-9575

Arshad Ismail https://orcid.org/0000-0003-4672-5915

Anja Linstädter https://orcid.org/0000-0003-0038-9557

Malesela Vincent Mokoka https://orcid.org/0000-0002-1026-2982

Zamantungwa Thobeka Happiness Mnisi https://orcid.org/0000-0002-7923-5077

Frances Siebert https://orcid.org/0000-0001-5549-8211

All of the data that support the findings of this study are available in the main text or Supplementary Information.