Voucher Specimens in Taxonomy and Simpson’s Hypodigm

Abstract

The problem of specifying representative specimens to serve as vouchers or ground truth specimens for species is viewed from a perspective of object recognition based on training using exemplars recognized by personal perception. In taxonomy, an ‘exemplar’ mirrors the hypodigm concept of Simpson, which refers to certain specimens being unequivocal members of a species. His concept has been discarded in most taxonomies because he did not provide procedures that distinguished it from ‘material’ or ‘sample’. However, his underlying view of a morphospecies was of a group united by shared character resemblances, probabilistically related. On this basis, the hypodigm of a taxonomic species is here restricted to specimens demonstrated as likely belonging to one population. To raise objectivity in this task, personal perception should be supplemented by morphometric analyses, several of which are demonstrated using living and Holocene samples of the oceanic zooplankter Truncorotalia crassaformis (Galloway and Wissler) from the Atlantic and Caribbean Oceans.

1. Introduction

For Simpson ([1], p. 185) “The hypodigm of a given taxonomist at a given time and for a given taxon consists of all the specimens personally known to him at that time, considered by him to be unequivocal members of the taxon, and used collectively as the sample on which his inferences as to the population are based”. Although the concept is mentioned in few taxonomies and was designated a synonym of ‘material’ by Mayr ([2], p. 270), its largely unrecognized implications are significant because they imply that the hypodigm is a set of voucher or ground truth specimens. This status has come into focus with the development of machine learning (ML) algorithms that use training sets of named specimens to construct models which enable unnamed specimens to be identified. These algorithms mimic the human ventral visual pathway ([3], p. 2) and use human knowledge encapsulated in the training set to name specimens. The approach draws attention to a prior step: how did Simpson recognize his hypodigm? More generally, how do taxonomists categorize specimens and select training sets? This question moves attention to a strand in neuroscience.

Objects of biological origin are just one group that we categorize in our daily lives, and the process is best explained by the exemplar theory [4]. Our visual concept of a group is built from stored representations of already identified individuals (labelled exemplars), from summary representations of multiple such individuals (prototypes), or from a combination of exemplar and prototype data [5]. Labelled exemplars are particularly valuable [6]. In practise, an observer tasked with categorizing an object autonomously compares its visual representation with those in their stored memory. Generated neural responses, equivalent to statistical distances, vary according to the closeness of the match. If sufficiently strong, the name of the exemplar is recalled [7]. The acuity of these visual processes is very high, which reflects their importance as a primary source of information about our environment. The visual attributes (e.g., gross shape, constituent parts, colour, texture, …) of objects in the group are critical for categorization. If they are static, as in many artificially made objects, one exemplar is adequate and categorization errors among observers are minimized. This restriction would also apply to biological groups categorized by the presence of specific diagnostic attributes. However, that essentialist approach to taxonomy [8] has been replaced by the polytypic concept of species [9], in which attributes are variably shared among members and none are essential for their recognition. Although this concept fits well with population genetics, it obfuscates categorization.

Clearly, the selection of voucher specimens has several problems. Vision theory suggests that categorization is an autonomic process using training sets that are personal to the observer. The matching processes are hidden from conscious recall, and outcomes are subjective. Taxonomic theory considers species are groups whose members have shared attributes, but no one is sufficient for membership; this complicates categorization. Simpson made a significant contribution towards the resolution of this nexus of problems. Although he did not directly acknowledge it, the polytypic concept is almost explicit in these statements: “… a taxonomic species is an inference as to the most probable characters and limits of the morphological species from which a given series of specimens has been drawn ([10], p.148); “… a morphological species is a group of individuals that resemble each other in most of their visible characters, sex for sex and variety for variety, and such that adjacent local populations within the group differ only in variable characters that intergrade marginally” ([10], p.147). His essential advance was recognition of the role probability plays in categorization of polytypic species. “Taxonomic studies are always statistical in nature … Populations, not individuals, are the units of systematics. … They are real entities whose properties are inferred from samples” ([1], p. 65). Nevertheless, he did not show how hypodigm theory could be applied, and clouded his concept by asserting that membership was unequivocal. Here, it is advocated that the first step in finding members of a hypodigm (also known as voucher specimens) within a sample is to demonstrate their probability of belonging to a single population. Outlined here are two approaches to this problem.

2. Material

Studied here are data from Truncorotalia crassaformis (Galloway and Wissler [11]), a globally distributed mesoscopic testate planktonic foraminifer. This group lives in the upper ocean and has an excellent fossil record, which is a valuable source about past ocean climates [12]. The use of specimen abundances to unlock the archive has promoted the application of ML techniques (e.g., [13]).

Gulf of Mexico sediment trap (GOM): This trap (Figure 1) is a time series (2008–2012) of foraminiferal and particulate flux at 700 m on the northern Gulf of Mexico continental shelf (27.5° N; 90.3° W). Dr. Caitlin Reynolds supplied 38 specimens (212 μm–425 μm fraction) from the GMT21 sediment trap [14]. Specimens were collected between 21 and 27 April 2008.

Cariaco Basin sediment trap (CAR): Dr. Robert Thunell supplied 29 specimens of Truncorotalia crassaformis collected in a sediment trap at 150 m in Cariaco Basin (Figure 1) in January 2007. Tedesco and Thunell [15] give details of its location, sampling procedures, and species fluxes. This anoxic basin is on the continental shelf, separated from the Caribbean by a sill at about 150 m.

DSDP Site 366A 1-1W-3-5 cm (SLR): This site (05 40.7° N; 19 51° W; 2853 mbsf, Figure 1) is on the Sierra Leone Rise in the eastern tropical Atlantic and lies under the Equatorial Counter Current. I extracted 35 specimen randomly from the >149 µm fraction. It is near Core 234 examined by [16]. From the model of Lazarus et al. [17], the age of the sampled horizons is <3 kyr.

ODP Site 925B (CER) is on the Ceara Rise in the western equatorial Atlantic Ocean (04 12.12.2° N; 43 29.3° W; 3053 mbsf, Figure 1). I extracted 29 specimens randomly from the >149 µm fraction of core 1H-1A-3–5. From the model of Chaisson and Pearson [17,18], Table 1) the age of the sampled horizons is <2 kyr.

Figure 1. (A) Localities sampled in this study plotted on global sea surface temperatures from [19]. Shown are records of Truncorotalia crassaformis in the ForCenS 43 database [20] with abundance ≥2%. (B–D) Features of shell architecture in axial orientation and source of outline data.

Figure 1. (A) Localities sampled in this study plotted on global sea surface temperatures from [19]. Shown are records of Truncorotalia crassaformis in the ForCenS 43 database [20] with abundance ≥2%. (B–D) Features of shell architecture in axial orientation and source of outline data.

3. Methods

Although taxonomists take a holistic view of a specimen, experimental neuroscience shows that subliminally, its outline is an important guide to its recognition [21,22]. Truncorotalia crassaformis builds a shell by incremental addition of ~15 chambers that expand isometrically and are arranged in a low trochospiral of about 3 whorls. A view of the outline in the plane of the coiling axis (Figure 1B–D) encapsulates much of the ontogeny including rate of whorl translation (height of early whorls), gross radial/axial dimensions, and the axial extension of late-formed chambers (conical form). While this orientation captures only one plane of a 3D object, it is a highly informative trait that is suitable for the procedures described here.

Specimen outlines were manually captured (tpsdig2, http://life.bio.sunysb.edu/morph/soft-dataacq.html, accessed on 5 June 2024) from SEM images as 180 equally spaced coordinates, and are available in Supplementary Data. Raw data were processed using generalized procrustes analysis (GPA, R package shapes https://CRAN.R-project.org/package=shapes, accessed on 5 June 2024), which aligns specimens on their centroids, removes size and positional differences, and enables visualization of most of the high-dimensional shape data when projected onto several principal coordinate (PC) axes [23].

The focus of the investigation is the integrity of the samples drawn from single populations. Density maps [24] function kde2d in R package MASS, https://CRAN.R-project.org/package=MASS, accessed on 5 June 2024 for PC 1:2) provide an overview of the distribution of specimens in shape space. A bivariate normal model is fitted to each individual, and contours map their distribution in the sample. The covMCD function ([25] 2017; https://CRAN.R-project.org/package=robustbase, accessed 5 June 2024) makes minimal initial assumptions about the distribution of the specimens by first finding the subgroup that occupies the smallest space in a sample, which is then used to compute revised estimates of mean and covariance. These provide an improved basis for judgement about the recognition of a group of shape-related individuals. Potential outliers are flagged by their distance from the revised mean, assuming that the latter now better conforms with the normal distribution.

4. Results

The analysis focused on objective estimates of the properties of the underlying populations. Mardia’s tests for normality (https://CRAN.R-project.org/package=MVN, accessed on 5 June 2024) show that for all samples, PC1 and PC2 are univariate normal (p > 0.05), but only GOM and CER are bivariate normal at that level. Nevertheless, the approximation to normality justifies its use for estimating population parameters. Density maps (Figure 2) of PC 1:2 data represent between 67 and 80% of information on outline shape and provide graphical overviews. This mapping of individual specimen tests for normality shows that the p ~< 0.05 level identifies outliers. While CAR has the best-defined mode and highest p-value contour, troughs between peaks in other samples are very shallow and reflect differences in shape over the restricted regions of the specimen outline.

Scatterplots (Figure 3) are projections of the procrustes analyses onto three principal component axes; this increases the representation of shape data to 79–84%. Specimens are plotted by their distance from the revised mean from the covMCD analysis; while this allows for an alternative assessment of their conformity with a normal model using additional shape information, outliers (p < 0.05) differ little from those found in the bivariate analysis.

Whether the analysis correctly identifies the statistical outliers as outside the taxonomist’s concept of the morphospecies is dependent on their knowledge of the functional morphology of specimens, for example, specimens #11, #14, and #34 in the GOM sample are recognized as statistical outliers due to the atrophied and misshapen final chamber (kummerform condition, Figure 1C), but I accept them as part of the taxonomic species. In contrast, I recognize specimens #9, #11, #25, and #29 in the CAR sample as likely to belong to Globorotalia crassula Cushman and R.E Stewart.E. Stewart 1930 [26].

The preceding data provide estimates of local populations, whereas Simpson’s hypodigm concept is not spatially limited. Following that approach, a procrustes analysis (Figure 4) of the pooled data with the addition of three named specimens of Truncorotalia oceanica shows that most specimens are within the 95% data ellipse of their respective samples; the partial intersection of data ellipses is a principal feature of the projection.

5. Discussion

My interpretation of Simpson’s hypodigm seeks to identify which specimens in a sample represent one population using these steps. 1. The taxonomist uses their visual perception to select specimens they provisionally identify as species A. 2. Morphometric data on a trait are gathered. 3. Statistical analyses provide probabilities of specimens belonging to a group. 4. The taxonomist evaluates this statistical grouping against their knowledge of the functional significance of the trait. 5. Voucher specimens are selected, informed by the probability that they belong to Species A. In this example, some might be rejected by the statistical model, but be recognized as aberrant individuals in a living population. In summary, the hypodigm is a filtered version of the material, aimed at providing data that facilitate its assessment as one from a single population, and which aids the selection of voucher specimens. Following Simpson [1] the hypodigm is statistically based, but contrary to him, assigned specimens remain contestable, except for the holotype, which is a de facto member. Importantly, this analysis is of local populations in space/time, whereas Simpson’s hypodigm represents a compendium of hypodigms.

A principal result of the analysis is that in each sample the shape data, as represented on PC axes, tend to be normally distributed. This supports an interpretation that they are sourced from a single population; it also suggests that the trait is under selection for an optimal form, and that the density maps might be viewed as metaphors of Wright’s [28] adaptive landscapes. That axial shell shape in planktonic foraminifera is a powerful trait, as these data suggest, was recognized by Cifelli [29], who documented its iterative evolution over the past 65 myr, but the relation between axial shell shape and ecology remains poorly understood. Caromel et al. [30] related shell shape to position in the water column, but it might also be related to the passive ambush trophic strategy of planktonic foraminifera [31] and the positioning of their adhesive rhizopodal nets to sinking particles, as in other non-motile plankton [32].

Particularly for formally named groups, identification by personal perception is highly dependent on the taxonomist’s training history. For taxonomic species, this leads to variations in the identification and selection of voucher specimens. Its significance in the planktonic foraminifera was demonstrated by Fenton et al. [33] in an operator error study of expert and novice taxonomists, which found a median identification accuracy of 57% of specimens viewed. The potential for such errors was recognized by Hsiang et al. [13], who used a panel of 24 expert taxonomists to identify images of planktonic foraminifers from 35 Atlantic coretops. Those images, for which there was 75% agreement amongst experts, were retained as labelled training sets. While such consensus sampling solutions are widely used [34], they average potentially different training histories and neglect the properties of source populations.

Voucher specimens are widely cited in taxonomy. A search (Google, 25 March 2024) in the literature for the combination ‘voucher’ ‘taxonomy’ reported ~2.7 × 10⁶ items, likely heightened by the use of ‘voucher’ in genomic studies. A cursory survey suggests that the usage of ‘voucher’ examined here is consistent with Culley’s ([35], p. 1) broad definition, being that it is “… a representative sample of an expertly identified organism that is deposited and stored in a facility from which researchers may later obtain the specimen for examination and further study”. There are many modifications, often concerning curation, but the essential theme concerns named specimens whose attributes are representative of a group of organisms. While revisions might lead to a demotion of vouchers, judgements about their recognition have remained firmly based on personal perception and are fully exposed to the problem of training histories and the autonomous nature of that process.

In contrast, ‘hypodigm’ is seldom mentioned beyond vertebrate taxonomy, where it is widely used, particularly in paleoanthropology (e.g., Anthreya and Hopkins [36]), which Simpson mentored [37]. Although the definition from Simpson [1] is used in this study, it is one of several that he offered. Simpson [38] introduced ‘hypodigm’ as “All of the specimens used by the author of a species as his basis for inference, and this should mean all the specimens that he referred to the species, constitute his hypodigm of that species. In a subsequent comparison or identification, the basis of comparison is not correctly a ‘type’ in any restricted sense, but a hypodigm”. Simpson ([39], Table 2) provided an example “Hypodigm: American Museum specimens of Ectocion from the Sand Coulee of Granger”. This aligns with Mayr ([2], p. 405), who defined hypodigm as “The entire material of a species that is available to a taxonomist”. However, Simpson ([1], Figure 1; adapted here as Figure 5) showed a schema for recognition of a hypodigm, in which the first step is from ‘sample’ to ‘hypodigm’, and stated the following: “The process begins with observation of the specimens in hand, the objective materials. The specimens studied and believed to be related in some biologically relevant way are a sample. If they are believed to represent a definite taxon (as determined at another level of inference), they constitute a hypodigm”. “Unequivocal membership” from Simpson [1] was clarified to refer to catalogued specimens that have repository data. This usage is followed by Mayo ([40], p. 8). This grants voucher status to specimens identified by personal perception because they are labelled; it overrides considerations of their population status.

6. Conclusions

As a supporter of the New Systematics (Huxley [41] and inspired by contemporary advances in genetics and statistics, particularly by R.A. Fisher [42] Simpson sought to place the recognition and attributes of morphospecies in a population context. Species were real, but their populations could only be sampled. This was a major contribution to the understanding of taxonomic species. While he outlined the theory underlying the concept of a hypodigm, he did not show how the population might be inferred from the sample. Currently, ‘hypodigm’ is often used as a synonym for material and might be discarded. But rather than concerns about nomenclature, the focus should be on methods that raise objectivity in the selection of vouchers for morphospecies that have been recognized by personal perception. The morphometric approach trialled here requires discretionary judgement, but it is reduced relative to consensus judgements about vouchers by expert panels. The algorithms used here are several of many that may aid the identification of members of polytypic taxa, a major problem in taxonomy since Linnaeus.