A comparison of non-surgical methods for sexing young gopher tortoises (Gopherus polyphemus)

Study animals

We tested several sexing methods on three age classes of immature gopher tortoises. “Neonates” were assessed within 1 week of hatching, “hatchlings” were sampled at ∼4–8 months of age (mean carapace length: 66.2 mm, range: 58–76 mm; for definitions of age/size classes of gopher tortoises, see Ashton & Ashton, 2008), and “juveniles” were sampled at approximately 45 months of age (mean carapace length: 145 mm, range: 101–170 mm). To obtain young tortoises, eggs were incubated in natural nests in the ground at Fort Stewart Army Reserve and George L. Smith State Park in southeastern Georgia, USA (Rostal & Jones, 2002; Hunter & Rostal, 2021). After the critical sex determination period (Rostal & Wibbels, 2014) but prior to hatching, we moved eggs to incubators where they were maintained at 29.9 °C until hatching. Eggs of assayed neonates and hatchlings were collected in 2018–2020, and eggs of juveniles were collected in 2015. Captive tortoises were reared in groups of 1–10 in plastic bins with plastic or ceramic hides. They were fed daily with a diet of mixed greens and provided with water twice weekly. All procedures using these animals were registered with and approved by the Institutional Animal Care and Use Committee of Georgia Southern University (#I18023) and the Georgia Department of Natural Resources (#1000545889 and #1000838720). The protocol entailed euthanasia by a registered veterinarian in the case of damage to internal organs during surgery. No euthanasia was required. Surviving animals were released after at least 1 month of recovery in captivity.

Quantitative testosterone hormone assay for Gopherus polyphemus

We collected approximately 100 µl of blood from the subcarapacial sinus (neonates and hatchlings) or the brachial vein (juveniles). Blood was quickly transferred into lithium heparin-treated glass capillaries (70 mm; Fisher Biosciences¹) and immediately centrifuged for 3 min at 6,000 rpm. Plasma was stored at −20 °C until extraction. We extracted testosterone from 30–50 ul of plasma by combining with one mL of anhydrous diethyl ether (Acros Organics) in a glass test tube, then vortexing for 20 seconds. The tube was then held against a block of dry ice until the aqueous layer was frozen (20 seconds), and the organic layer was then decanted into a clean test tube. Ether was evaporated in a fume hood overnight, and samples were then ready for enzyme-linked immunosorbent assay (ELISA). Samples were re-eluted the next morning in ELISA buffer (Item No. 582701, Cayman Chemical, Ann Arbor, MI) and vortexed for 20 seconds. We then performed ELISAs using the Cayman Chemical Testosterone ELISA kit following manufacturer instructions, running all samples in duplicate.

We validated that the ELISA detected the same target as the standard by checking for parallelism between an 8-point, twofold serial dilution of the manufacturer-provided testosterone standard and a 10-point, twofold serial dilution of an extracted pool of tortoise plasma (20 adult females from the same population). We calculated extraction efficiency and checked for matrix interference by spiking 25–30 µl aliquots of juvenile female plasma (n = 6) with 10 pg of testosterone (5 µl of a 1:100 dilution of the Cayman supplied standard). Samples were then extracted and run in duplicate as described above. Extraction efficiency was calculated for each sample using the formula E = (M_spiked − M_unspiked)/S where M is the measured amount of testosterone in the spiked or unspiked sample, and S is the known amount of testosterone added in the spike (in pg).

Samples were run in duplicate with up to 32 samples per plate, as well as reference standards. The reference standard was a dilution of pooled plasma from 20 adult female G. polyphemus collected in 2007 and stored at −20 °C. We used these standards to calculate between- and within-plate percent coefficients of variation (% CV).

Laparoscopy to determine sex

Evaluating non-invasive sexing methods requires knowing the sex of subjects to determine the accuracy of each approach. We used laparoscopy to determine sex for tortoises when they were 6–12 months old, using the method described by Rostal et al. (1994). However, we modified this protocol by administering lidocaine as a local anesthetic (Hernandez-Divers, Stahl & Farrell, 2009; Rakotonanahary, Kuchling & Routh, 2015; Emmel et al., 2021), and sealing the incision with Surgi-Lock 2oc (Meridian) glue following the procedure. Tortoises were monitored individually for 24–48 hours after surgery until they were observed feeding and behaving normally, then returned to their housing. Surgeries were performed at least 1 week after each individual’s last blood draw to ensure they had fully recovered before the operation. Laparoscopies were performed when tortoises were at least 50 g body mass, and after at least 6 months post-hatching to ensure the yolk had been completely absorbed. For juveniles hatched in 2015, laparoscopies occurred at approximately 1 year of age. The same person (D. Rostal) conducted laparoscopies in all cases. The surgeon was blind to individual sex, and all associated data collection (testosterone, morphology) was done blind to the sex determined by laparoscopy.

Sexing approach 1: baseline plasma testosterone

Using the validated ELISA method described above, we quantified plasma testosterone in neonates, hatchlings and juveniles (see sample sizes in Table 1). Samples were run in haphazard order by individual ID (not determined by sex). Sample sizes were determined by the number of available tortoises large enough to laparoscopy (>50 g) in each year. No animals with sex determined by laparoscopy were excluded.

Sexing approach 2: testosterone following follicle stimulating hormone challenge

We injected tortoises with FSH to see if FSH differentially elevates male testosterone, allowing the separation of males and females based on a subsequent plasma testosterone assay. Juveniles were not challenged with FSH, as naive testosterone levels clearly distinguished the two sexes (see results). In 2018, we tested three doses of FSH: young juveniles were injected with either 0.1 (n = 12), 0.05 (n = 12), or 0.01 (n = 15) units of porcine FSH (MP Biomedicals 0210172750) in 100 ul sterile saline. Injections were performed intracoelomically (Lance, Valenzuela & Von Hildebrand, 1992). If the individual had been sampled for baseline plasma testosterone, the injection was performed at least 6 days after the earlier blood draw. Four to six hours after injection, ∼100 µl of blood was drawn from the subcarapacial sinus. Samples were processed, stored and testosterone was quantified using the methods described above. For a subset of individuals (n = 15), we repeated this sampling six days after injection to determine the longevity of the effects of FSH treatment. A subset (n = 11 males and n = 14 females) were later laparoscopied to determine sex.

To verify that observed effects were not due to the stress of handling and injection rather than FSH itself, four hatchlings (later determined to be two males and two females) were injected with 100 ul of sterile saline and sampled 4 hours later. These individuals were later laparoscopied, and 1 month following surgery treated with 0.01 units of FSH and sampled (included in the sample size reported above).

After establishing the lowest effective dose, we similarly administered 0.01 units of FSH by injection (followed by blood draws and testosterone quantification) for neonates in 2019 and hatchlings in 2020 (Table 1). These animals were then maintained in the lab and sexed via laparoscopy. For subsequent analysis of post-FSH hatchlings, we only used data from 2020 individuals, when all individuals were administered the same dose of FSH.

Testosterone analysis

All analyses were performed in R v4.1.0 (R Core Team, 2021). We estimated several thresholds to use in distinguishing males and females for each testosterone dataset (sexing approaches 1 and 2). First, we estimated the threshold that minimized the total percentage of erroneous sex assignments (herafter T_m). This total percentage is calculated by estimating the percentage of erroneously assigned males and the percentage of erroneously assigned females and summing them; this avoids bias that would occur if we used total numbers of erroneous assignments for datasets with unequal sex ratios. Then, since T_m could be biased toward one type of error (i.e, mistaking males for females or vice versa), resulting in a bias in the overall estimate of sex ratio, we estimated the threshold associated with an equal rate of misidentifying true females and true males. This was calculated by generating a sequence of possible testosterone values from the minimum to the maximum observed values by 0.1 units, then estimating the error rate (the fraction of observed females with testosterone values above the threshold, or the fraction of males with values below the threshold) for each sex for each possible testosterone value. We then selected the testosterone value that resulted in the minimum difference in male and female error rates. Finally, given the substantial overlap in testosterone concentration distributions between sexes for some datasets, we sought a method to filter ambiguous individuals and only assign sexes in individuals with testosterone values outside of the ambiguous range. To do this, we calculated the threshold below which the probability of an individual being female was 90% (T_F90), and the threshold above which the probability of an individual being male was 90% (T_M90). To calculate the former, we used Bayes’ Theorem: (1)${\displaystyle P\left(\text{sex}=\text{“female”}|\mathrm{T}<x_0\right)=\frac{\mathrm{P}\left(\mathrm{T}<x_0|\text{sex}=\text{“female”}\right)\mathrm{\ast \; P}\left(\text{sex}=\text{“female”}\right)}{\mathrm{P}\left(\mathrm{T}<x_0\right)}.}$ When we assume a given individual has equal probability of being either sex, this reduces to: (2)${\displaystyle \frac{\mathrm{P}\left(\mathrm{T}<x_0|\text{sex}=\text{“female”}\right)}{\mathrm{P}\left(\mathrm{T}<x_0|\text{sex}=\text{“female”}\right)+\mathrm{P}\left(\mathrm{T}<x_0|\text{sex}=\text{“male”}\right)},}$ which can be estimated from our datasets for each sex. To calculate T_F90, we generated a sequence of testosterone values from the minimum to the maximum observed testosterone by 0.1 units and used the above formula to calculate corresponding probability of being female given observed concentrations of testosterone less than a threshold value. We then fit a smoothed spline to the probabilities and from this curve estimated the testosterone value below which an individual is 90% likely to be female. We calculated T_90M in the same way, as well as analogous thresholds for 80% and 95% likelihood for each sex. This approach allowed us to avoid fitting distributions to our testosterone datasets, as the male testosterone values were in some cases not well fit by lognormal or gamma distributions. To validate error rates from the application of T_E and T_M thresholds, we estimated error rates using leave-one-out cross-validation of each sample, generating errors that are comparable to out-of-bag (OOB) error rates from random forest models (see below). To quantify sampling variation, we computed 95% quantiles from bootstrapped sampling distributions (n = 1, 000 replicates) for all thresholds and error rates.

To test for a differential effect of FSH challenge on testosterone levels for males and females, we used permutation t-tests (perm.t.test(), Kohl, 2020) on the absolute and percentage change in testosterone for 23 male and 65 female hatchlings for which we had both pre- and post-FSH testosterone levels.

Sexing approach 3: Morphological measurements

For each neonate used in the testosterone experiments, we took five morphological measurements (midline carapace length (MCL), shell height (SH), carapace width (CW), minimum and maximum plastron length (PMIN and PMAX)) in the first 1–2 weeks after hatching in 2018, 2019 and 2020. Hatchlings were measured at 7.5–8.5 months post-hatching (n = 78 in 2020), and for these individuals we also measured anal width (AW), anal notch depth (AN), gular width (GW) and tail length (TL) (Fig. 1; Burke et al., 1994). Juveniles were not measured. Sex was determined by laparoscopy. We calculated several derived variables based on these measurements that were suggested to be sex-informative in an earlier study (VOL = MCL * SH * CW, MCL.V = MCL/VOL, SH.V = SH/VOL, CW.V = CW/VOL, PMIN.V = PMIN/VOL, PMAX.V = PMAX/VOL, CW.MCL = CW/MCL, CW.SH = CW/SH; Burke et al., 1994). All raw measurements, as well as the composite measurements, were included in a random forest model (Liaw & Wiener, 2002) to see if they could be used to classify neonate sex, using mtry = 5 and ntrees = 10⁵. Because sexes were not equally represented in some datasets, we specified sampsize = c(x,x) to downsample the larger class such that class frequencies were equal, where x is the number of samples in the rarer class. We created models using only morphological data for the full dataset, and using morphological data and testosterone data for two subsets of neonates, those with an FSH challenge (2019) and those without (2020), and compared these to models with only testosterone included (in these cases, mtry =1), to see if morphological data improved sex identification by testosterone level alone. We did not use linear discriminant function analysis (as done by Burke et al., 1994) due to high multicollinearity among predictors.

Sexing approach 4: Geometric morphometrics

To test for subtle differences in shape between sexes, we used a geometric morphometric approach to quantify plastron and carapace morphology in neonates (n = 21 females; n = 15 males). We followed the procedures detailed by Valenzuela et al. (2004) to quantify morphometric differences among individuals. First, we took standardized photographs of the top of the carapace (top) and plastron (bottom) with one measurement (suture intersection measurement, SIM) per photograph to provide a scaling factor that was most in-line with the plane of important shell landmarks (Fig. S1). We marked the x,y coordinates of standardized landmarks (scute intersection points) and aligned them using TpsDig software (Rohlf, 2018). We delineated 28 top and 14 bottom landmarks (Fig. S1). We extracted partial warp scores from the thin-plate spline using TPSRelw software (Rohlf, 2018) which capture the variation in shape among individuals (n = 51 top and 24 bottom partial warp scores). We created a random forest model following the process for morphological measurements described above using all partial warp scores.

Testosterone ELISA validation

The Cayman ELISA assay demonstrated good parallelism between serially diluted tortoise plasma and serially diluted standards in the range of 10–90% bound/maximum bound tracer (B/B₀), the range in which male and female testosterone concentration distributions overlapped and thus the range where accurate testosterone concentrations were most important (Fig. S2). This parallelism demonstrated that the kit was likely detecting testosterone in the tortoise plasma, and that the plasma lacked matrix interference that could bias quantification. Estimated extraction efficiency was high, averaging 94.1% (+/-16% SD) for six spiked samples. Between-plate % CV for plates with 2018 samples was 7.6%, with a within-plate % CV of 9.2%. Between- and within-plate % CVs for the remaining plates were 14.0% and 6.0%, respectively. Within-replicate error (for the technical duplicates) was 3.3%.

Sexing approach 1: Naive plasma testosterone

For naive neonate tortoises, males had higher average testosterone than females (female mean ± SE in pg/mL: 17.9 ± 1.1, n = 44; male: 51.7 ± 13.5, n = 18; Fig. 2A). The ranges of testosterone levels for male and for female neonates overlapped extensively, but a subset of males exhibited testosterone values well above the female range (Fig. 2A). Adopting an equal-error threshold (T_E) to separate males from females would result in a substantial error (approximately 22% for each sex; Table 1), with wide confidence intervals on threshold and error rates (Table 2). Adopting the 90% likelihood thresholds (T_90%) to exclude ambiguous individuals would result in excluding over half of all tortoises (Table 2; Fig. 2).

Male and female naive hatchling (4–8 months old) tortoises were better separated by testosterone levels (Fig. 2B; Table 2), and male levels were much higher than females (female mean ± SE in pg/mL: 27.9 ± 1.7, n = 66; male: 206.0 ± 49.6, n = 25). Adopting the T_E would result in ∼10–12% misclassification for males and females in our dataset (Table 2), with 95% CIs in error rates of 3–18% and 5–13% for females and males (Table 2). Using the T_90% thresholds would exclude ∼13% of females and ∼24% of males in our dataset (Table 2). The lower “female” T_80% was greater than the higher “male” threshold, indicating that the T_E had a likelihood of >80% accuracy for each sex.

Testosterone levels completely separated male and female juveniles (∼3.5 years old; female mean ± SE in pg/mL : 79.8 ± 7.8, n = 11; male: 3,341 ± 432, n = 12; Fig. 2C). The maximum observed female testosterone concentration was 121 pg/mL, while the minimum male concentration was 1,230 pg/mL. A threshold of 700 pg/mL approximates the midpoint between the two distributions (Fig. 2C).

Sexing approach 2: Testosterone following FSH challenge

In our preliminary trials of three different doses of FSH administered to hatchlings from 2018, visual inspection of the results revealed that all three dose levels increased testosterone in males and females (Fig. 3). We did not perform statistics, as there was an inadvertent association of dose with sex (sexes were not known at time of challenge trials), and the sample size within each sex and dose group was small. Nonetheless, it was clear that the lowest dose (0.01 units) effectively elevated testosterone, and that testosterone levels generally returned to their pre-challenge levels after 6 days (Fig. 3), making an FSH challenge using the lowest dose suitable for temporarily stimulating testosterone production. Injections of saline did not elevate testosterone, indicating that FSH itself is responsible for the change in testosterone levels (Fig. S3).

Neonates challenged with FSH in 2019 had higher testosterone levels than naive testosterone levels in 2020 neonates (data were not paired, and were collected in different years; Fig. 4A), suggesting that FSH similarly elevates testosterone in neonates. In hatchlings in 2020, samples were taken from the same individuals before and after FSH challenge, and the challenge consistently elevated testosterone levels within individual males and females (Figs. 4B–4C). The average difference between naive and post-FSH levels for males was 392 ± 84 pg/mL (n = 23), representing an average increase of 261%, and for females the difference was 25.1 ± 3.5 pg/mL (n = 65), with an average increase of 104%. Permutation t-tests confirmed that there was a significant difference in both the numeric increase and percent increase between sexes (difference in numeric increase: 95% CI [262–502] pg/mL; p < 0.001; difference in percent increase: 95% CI [85.2–236.5]%; p < 0.001). Thus, FSH differentially elevated testosterone in male hatchlings, relative to females.

Comparing the distributions of male and female neonates post-FSH challenge, we observed substantial overlap in male and female testosterone concentrations (Fig. 5A), such that the T_E of 29.5 pg/mL resulted in ∼30% errors for each sex in our dataset, with 95% CIs up to 43.8% for both sexes (Table 2). Adopting a T_90% would exclude 94% of females and 66% of males, making FSH-challenge in neonates a poor method for determining sex.

In contrast, male and female hatchlings (from 2020 when all were given the same dose of 0.01 units) were better separated after FSH challenge (Fig. 5B), such that adopting a T_E of 128 pg/mL would result in <5% error for each sex in our dataset (95% CI errors: 0–11% for females, 1–11% for males; Table 2), allowing high confidence in sex assignment (Table 2). The lower “female” T_80% and T_90% were greater than the higher “male” threshold, indicating that the T_E had a likelihood of >90% accuracy for each sex. The T_95% thresholds could be adopted with ∼6% loss of males and females, allowing a greater percentage of individuals to be highly accurately sexed when compared to the naive hatchlings (Table 2).

Sexing approach 3: Morphological measurements

Neonates—The random forest model trained on morphological measurements from 174 neonates weakly predicted sex (out-of-bag (OOB) error = 33.3%, with 29.6% error for females and 38.2% error for males; Table 3). Analyzing the subset of 18 male and 44 female neonates from 2020 for which we also had naive testosterone concentrations, OOB error rate was 24.2%, which was not substantially different from a random forest model trained on the testosterone data alone (OOB error = 27.4%). For post-FSH neonates (n = 16 females + 32 males from 2019), the OOB error for the model including morphological and testosterone data was 8%, substantially better than the OOB error for the testosterone-only model (25%). For this model, testosterone, CW/SH, Pmax and SH were the four most important predictors (Table S1).

Hatchlings—The random forest model for 55 female and 23 male hatchlings including only morphological data had an OOB error rate of 19.3%, with 16.3% error for females and 26% error for males (Table 3). Removing the AW, AN, TL and GW measurements (and their volume-corrected versions) resulted in the same model performance, indicating these variables contributed little to sex discrimination. Combining morphological measurements with naive testosterone levels (measured at the hatchling age) gave an OOB error of 10.2%, improving on the 19.2% error rate for testosterone alone. Combining the morphological measurements with post-FSH testosterone levels actually reduced model performance from 4% to 12% OOB error.

Sexing approach 4: Geometric morphometrics

The random forest model for 21 female and 15 male neonates including only geometric morphometrics data had an OOB error rate of 52.8%, with 23.8% error for females and 93.3% error for males.