Amphibians exhibit diverse responses to environmental challenges, but their responses to infection risk remain poorly understood. This study investigates how the presence of ranavirus, a deadly viral pathogen, affects growth, development, and resource allocation in wood frog (Rana sylvatica [Lithobates sylvaticus]) tadpoles. Using three years of pond survey data from a wood frog metapopulation in northeastern Connecticut, USA, we compared tadpole physiological rates across three scenarios: ranavirus-free ponds, ponds with sustained ranavirus infection, and ponds experiencing ranavirus die-offs. In ranavirus-positive ponds, tadpoles exhibited increased growth and resource allocation early in their development. These differences waned following die-off events in some ponds but persisted where widescale infection did not lead to die-off. This study provides evidence that an important disease agent appears to induce growth and developmental responses in its host that may help tadpoles survive severe infection by providing a buffer against the associated energetic demands. Alternative hypotheses, such as size-biased mortality, should be evaluated in experiments aimed at evaluating underlying mechanisms.

Pond Surveys

Ranavirus surveys spanned ~mid-April to mid-July each year, during which we sampled each pond at least once every two weeks. At each visit, we collected up to 20 live wood frog tadpoles via timed dip net surveys with durations based on estimated pond area (m2) and estimated tadpole density per unit effort (Werner et al. 2007; "rasy_density" column in dataset). Density surveys were not conducted on 34 occasions (14% of sampling events) in our analysis dataset. When surveys were missed due to early season hatchling clustering near the oviposition site (~76% of cases), we used density values from the following survey ~two weeks later. For surveys missed due to other reasons (e.g., rain, equipment failures; ~14% of cases), we estimated tadpole densities as the average from the nearest surveys before and after the missing session. If both neighboring surveys were unavailable, we used the density from the most recent survey. We also recorded the presence of the predatory salamander Ambystoma opacum ("amop_detected" in dataset)

Following (Hall et al. 2018), we considered the detection of ≥ 5 amphibian carcasses to be the start of a die-off ("start_of_dieoff" in dataset). We estimated the relationship between ranavirus status at the pond level and tadpole growth and development by categorizing each pond-year by ranavirus status ("ranavirus_status" column in dataset): no/low ranavirus (L-RV/Uninfected), sustained ranavirus infection without a die-off (M-RV/Infected), and sustained ranavirus infection with a die-off event (H-RV/Dieoff). All tadpoles collected from pond-years with ranavirus detected in tissue at ≤ 1 sampling session were classified as Uninfected. All tadpoles collected from pond-years with ranavirus detected in tissue at ≥ 2 sampling sessions without a ranavirus die-off event were classified as Infected. All tadpoles collected from pond-years where a ranavirus die-off event occurred were classified as Dieoff.

Larval measurements

We used measurements of the wild-captured wood frog larvae to estimate growth and development rates. We used individual measurements of unique individuals in a pond cohort rather than repeated measurements of the same tadpoles because ranavirus testing requires destructive sampling, and mark-recapture with wild tadpoles at this scale is impractical. Nonetheless, this method provides a strong proxy for individual growth and development - wood frogs are explosive, synchronous breeders, and so most larvae in a cohort follow a very similar trajectory of growth and development. Each larva was assigned a developmental stage (Gosner 1960; "stage_gosner" in dataset) and measured (snout-vent length [SVL] and total length [TL]) using digital calipers ("svl_mm" and "tl_mm" in dataset).

Ranavirus sample processing

For pond-years (i.e., samples from a single pond in a single year) with adequate samples, we assessed ranavirus infection in ≥ 10 larvae from each of ≥ three time points. This sample size gives a ~72% probability of detecting at least one infection at ≥ two time points for prevalence ≥ 10% and a ~97% probability for prevalence ≥ 20%. Testing ten individuals across three time points provides a reasonably high likelihood of detecting sustained infection in a given pond-year, as ranavirus infection prevalence tends to remain stable/increase over time in larval wood frog populations (Hall et al. 2018).

We dissected liver tissue from each tadpole using forceps sterilized in a 50% bleach solution and extracted DNA using Qiagen DNeasy Blood & Tissue Kits (Qiagen, Hilden, Germany). We also included extraction negative controls to detect any potential contamination from equipment.

To determine the viral load of samples, we used quantitative polymerase chain reaction (qPCR) primers and a fluorescent probe from Leung et al. (2017), which target a 97-bp region of the ranavirus major capsid protein. Each 20 uL reaction contained 2 μL template DNA, 2.0 μL forward primer (10 μM), 2.0 μL reverse primer (10 μM), 0.05 μL FAM-labeled fluorescent probe (100 μM), 10.0 μL SsoAdvanced™ Universal Probes Supermix (Bio-Rad Laboratories, Hercules, CA, USA) and 5.95 μL nanopure water. Each qPCR plate included a standard dilution series (2.79 ✕ 106 - 2.79 ✕ 101) using a 350 bp synthetic gBlock (IDT, Coralville, IA USA) gene fragment of the major capsid protein gene and a negative control. Each control, standard curve, and unknown sample was run in duplicate on a 96-well qPCR plate on a CFX Connect™ (Bio-Rad Laboratories) with cycling conditions consisting of 98 °C for 3 min followed by 40 cycles of 98 °C for 15 s and 60 °C for 45 s, with a plate read at the end of each cycle. 

Viral quantities for positive samples were averaged across the duplicate wells ("rasy_SQ_avg" in dataset). All duplicate unknown samples that peaked before cycle 40 were considered positive ("infection_status" in dataset. In cases where there were discrepancies between duplicates, a third reaction was run and if the third reaction amplified, the sample was considered positive. We calculated viral load as viral copies per ng DNA in the sample ("virus_per_ngDNA" in dataset), measured using a NanoDrop 2000c (Thermo Fisher Scientific, Waltham, MA, USA). 

Pond temperature data

Using temperature logger data for each pond-year, we calculated the average pond temperature as the average daily temperature in a pond from 1 May to 30 June, which captures most of the wood frog larval period in our system ("avg.temp" column in dataset). However, water temperature data was not recorded for some pond-years due to logger loss or equipment failure; in total, ~30.6% of the total daily water temperature data was missing. Therefore, we implemented a random forest model to impute the missing data (randomForest; Liaw & Weiner 2002). Imputation used a combination of climate variables and a local variable characterizing pond canopy cover. Climate variables were extracted from the DayMet database v.3 (Thornton et al. 2022; tile: 11754, N 41.9198, W 72.1604) and included daily temperature (minimum, maximum, and average), precipitation, shortwave radiation, snow water equivalent, and water vapor pressure. We included 1- and 2-day time lags for all climate variables. The canopy variable was the weighted average global site factor (GSF; the ratio of below-canopy radiation to above-canopy radiation) for the wood frog larval period. Briefly, five hemispherical photographs were taken along the shore at each cardinal point and at the center of each pond during leaf-off and leaf-on seasons, and we used Gap Light Analyzer (Frazer et al. 1999) to estimate GSF. Canopy estimates were unique to each pond but the same across years, while climatological variables were unique to each day and year but the same across all ponds. We also included day-of-year, year, and pond as variables.

Our model was grown from 500 regression trees and used 10-fold cross-validation to evaluate the predictive accuracy of imputed points. The random forest model fit well (mean out-of-bag R2 = 0.95) and had high predictive accuracy (RMSE = 1.0). On average, the predicted daily pond temperatures were within 0.68 °C of the observed daily temperatures in the hold-out cross-validation dataset. Daily average air temperature was the most important variable, followed by daily average air temperature with a 1-day lag and daily minimum temperature.

We then used the model fit to all our observations to impute daily water temperature data. We imputed daily water temperature data both for pond years missing temperature data and for pond years with temperature data (i.e., the entire dataset) to ensure that every value in the dataset was generated using the same logic and assumptions and to avoid introducing potential biases. After averaging daily observed and predicted pond temperatures for each pond-year from 1 May to 30 June, the difference between predicted average pond temperatures and observed average pond temperatures was 0.21 °C, or ~2.8% of the range of average observed pond temperatures (Range: 11.27 - 18.65 °C).