Predicting the presence and titer of rabies virus neutralizing antibodies from low-volume serum samples in low-containment facilities

Serology is a core component of the surveillance and management of viral zoonoses. Virus neutralization tests are a gold standard serological diagnostic, but requirements for large volumes of serum and high biosafety containment can limit widespread use. Here, focusing on Rabies lyssavirus, a globally important zoonosis, we developed a pseudotype micro-neutralization rapid fluorescent focus inhibition test (pmRFFIT) that overcomes these limitations. Specifically, we adapted an existing micro-neutralization test to use a green fluorescent protein–tagged murine leukemia virus pseudotype in lieu of pathogenic rabies virus, reducing the need for specialized reagents for antigen detection and enabling use in low-containment laboratories. We further used statistical analysis to generate rapid, quantitative predictions of the probability and titer of rabies virus neutralizing antibodies from microscopic imaging of neutralization outcomes. Using 47 serum samples from domestic dogs with neutralizing antibody titers estimated using the fluorescent antibody virus neutralization test (FAVN), pmRFFIT showed moderate sensitivity (78.79%) and high specificity (84.62%). Despite small conflicts, titer predictions were correlated across tests repeated on different dates both for dog samples (r = 0.93), and for a second dataset of sera from wild common vampire bats (r = 0.72, N = 41), indicating repeatability. Our test uses a starting volume of 3.5 μL of serum, estimates titers from a single dilution of serum rather than requiring multiple dilutions and end point titration, and may be adapted to target neutralizing antibodies against alternative lyssavirus species. The pmRFFIT enables high-throughput detection of rabies virus neutralizing antibodies in low-biocontainment settings and is suited to studies in wild or captive animals where large serum volumes cannot be obtained.


Summary 11
Serology is a core component of the surveillance and management of viral zoonoses. Virus 12 neutralization tests are a gold standard serological diagnostic, but requirements for large volumes of 13 serum and high biosafety containment can limit widespread use. Here, focusing on Rabies lyssavirus, 14 a globally important zoonosis, we developed a pseudotype micro-neutralization rapid fluorescent 15 focus inhibition test (pmRFFIT) that overcomes these limitations. Specifically, we adapted an existing 16 micro-neutralization test to use a green fluorescent protein-tagged murine leukemia virus 17 pseudotype in lieu of pathogenic rabies virus, reducing the need for specialized reagents for antigen 18 detection and enabling use in low-containment laboratories. We further used statistical analysis to 19 generate rapid, quantitative predictions of the probability and titer of rabies virus neutralizing 20 antibodies from microscopic imaging of neutralization outcomes. Using 47 serum samples from 21 domestic dogs with neutralizing antibody titers estimated using the fluorescent antibody virus 22 neutralization test (FAVN), pmRFFIT showed moderate sensitivity (78.79%) and high specificity 23 (84.62%). Despite small conflicts, titer predictions were correlated across tests repeated on different 24 dates both for dog samples (r = 0.93), and for a second dataset of sera from wild common vampire 25 bats (r = 0.72, N = 41), indicating repeatability. Our test uses a starting volume of 3.5 µL of serum, 26 estimates titers from a single dilution of serum rather than requiring multiple dilutions and end point 27 titration, and may be adapted to target neutralizing antibodies against alternative lyssavirus species. 28 The pmRFFIT enables high-throughput detection of rabies virus neutralizing antibodies in low-29 biocontainment settings and is suited to studies in wild or captive animals where large serum volumes

Introduction 33
The last few decades have seen a surge in newly emerging human viruses that originate from wildlife 34 ( of GFP fluorescence. Next, the command "Analyze Particles" was used to count the total number of 169 fluorescent cells per field (i.e. infected cells). This command grouped and counted the white 170 neighboring pixels with a predetermined size area and circularity to be a single cell (size area: 5-50 171 circularity: 0.80-1.0), so counts corresponded to the number of infected cells. Cell count outputs were 172 converted into a standardized spreadsheet using a Python version 3.7.2 script (Python Core Team,  173 2019) (script available in supplementary materials). At the end of the image processing step, each 174 serum sample was described by 10 data points consisting of the number of the fluorescent cells in 175 each of 5 fields (photographs) in the 1:10 and 1:25 dilutions (Figure 1 B). 176

Statistical analysis 177
All statistical analyses were executed in R (R Core Team, 2018 (scaled to improve model convergence) and 2) the serum dilution level (two factors: 1:10 and 1:25 187 dilution). Random slope and intercept terms were also considered for the date the test was run ("test 188 date") to account for observed variation in the relationships between SRIG titers and infected cell 189 counts across dates (Figure 2) and for the field number (1 to 5) within each microscope well ("field") 190 to account for variation in cell counts between fields (the middle field, field 5, had more agglomerated 191 cells in particular). To evaluate whether a simpler, single dilution test produced comparable results, 192 the full dataset was then subset to the 1:10 dilution only. The binomial and log-normal models fit to 193 this data subset included only the fixed effect of the virus-infected N2A cell counts, but the random 194 effects were identical to those explained above (i.e. test date and field). Models were fit using the 195 'lme4' package (Bates et al., 2015). The 'predict' function was used to generate the predicted 196 probability that a SRIG concentration or serum sample was seropositive (binomial model) and its 197 corresponding RVNA titer (log-normal model). Predictions per field were averaged to obtain results 198 per sample (Figure 1 C) with a threshold of > 0.1 IU/mL (positives) were used as the benchmark reference. 234

235
To understand the variability of the pmRFFIT, replicate SRIG titer concentration curves were produced 236 on 6 different dates between 30/05/2019 and 27/06/2019 (hereafter "Test 1" through "Test 6"). As 237 expected, the number of infected cells declined at higher SRIG titers in all replicates; however, the 238 shape of the antibody decay curve varied across test dates (Figure 2). At the 0.1 IU/mL SRIG 239 concentration, infected cell counts were more dispersed in the 1:25 dilution than in the 1:10 dilution, 240 as indicated by higher interquartile range (IQR) within each of the 6 test dates. Across all the SRIG 241 concentrations in all test dates (N = 36), 77.78% of the count comparisons were less dispersed in the 242 1:10 dilution suggesting this dilution could be more precise for downstream statistical analysis (SI 2). 243

Prediction of seropositivity using binomial GLMM 244
Binomial GLMMs accurately predicted seropositive and seronegative SRIG concentrations (Figure 3). 245 The best random effects for the binomial model included a random slope and intercept for test date 246 ( Table 1). The models built with the 1:10 dilution data only ("one-dilution model") and from both the 247 1:10 and 1:25 dilution data ("two-dilution model") had equivalent specificity (100%), but the one-248 dilution model was more sensitive (100% versus 58.33%, Figure 3 A, B). Furthermore, the two-dilution 249 binomial model failed to correctly predict the seropositive controls on 4 out of the 6 test dates, 250 confirming improved performance of the one-dilution model (Figure 3 B). 251

Prediction of titer values using log-normal GLMM 252
The log-normal GLMMs gave repeatable predictions of RVNA titers from the datasets generated 253 through our protocol across test dates (Figure 3 C). The best log-normal model included a random 254 intercept and slope for test date. Although the most complex model had the lowest AICc, the simpler 255 model (without the random intercept of field) had a ΔAICc < 2 ( Table 1). Observed and predicted SRIG 256 titers were highly correlated for both the one and two-dilution models (r = 0.95, Figure 3 C). When 257 comparing test dates (i.e. one-to-one comparison between correlations of the one-dilution and the 258 two-dilution model from the same test dates), the correlation coefficients were similar, suggesting the 259 simpler one-dilution model is sufficient for titer prediction (Figure 3 D). The most commonly applied serological tests to detect RVNA titers challenge a range of serial dilutions 286 of serum with infectious RV. This process is labor-intensive and requires laboratory capacity to grow 287 large quantities of pathogenic RV. Here, we provide an alternative serological framework that uses a 288 combination of digital image analysis and statistical analysis to estimate the presence and titer of 289 RVNA from a single dilution using only 3.5 µL of serum. 290 The pmRFFIT differs from other lyssavirus neutralization tests in several key aspects. It uses an 291 MLV(RG) pseudotype rather than pathogenic RV, allowing the pmRFFIT to be performed in any low-292 containment laboratory with appropriate cell culture and microscopy facilities. The addition of GFP 293 expression is significant, since it removes the need for FITC-conjugated antibody (reducing reagent 294 costs) and the fixation and staining steps used in traditional RFFIT or FAVN. One potential drawback 295 of using GFP expression to measure infectivity is the prolonged neutralization period (66 h versus 24 296 h RFFIT and 48 h FAVN) required to gain sufficient fluorescence for image processing (Aubert, 1992;297 Smith et al., 1973). Longer neutralization requirements (60 h) were also required in a FAVN 298 modification using a GFP expressing recombinant CVS-11-eGFP but did not alter results relative to the 299 test run with CVS-11 (Xue et al., 2014). Fortunately, extended incubations are unlikely to alter 300 neutralization outcomes since MLV(RG) pseudotype is replication incompetent, preventing infection 301 of additional cells during the incubation (Temperton et al., 2015). The pmRFFIT also uses an imaging 302 pipeline that combines systematic photography of microscope fields with automated digital image 303 processing to count infected cells. Microscopy in neutralization tests is time consuming and presents 304 challenges for interlaboratory comparisons due to multiple sources of variation, especially those that 305 affect the manual readout (e.g. laboratory user, manual pipetting, uneven cell monolayer) ( The pmRFFIT standardized approach minimizes these sources of error while potentially reducing 308 microscope operator time. Moreover, the imaging process generates traceable and permanent 309 electronic records of the raw data, eliminating the need to manually digitize records of field counts. 310 Several investigators have previously incorporated image processing into RV neutralization tests. to count pixels using a microRFFIT but did not make full use of the quantitative nature of imaging data 315 to obtain RVNA titers and used pathogenic RV rather than a viral pseudotype. A final distinction is that 316 instead of scoring microscope field or wells as virus positive or negative, the pmRFFIT predicts 317 serological status and RVNA titer from infected cell counts in a single serum dilution using statistical 318 modelling. The efficacy of this approach highlights the value of historically underutilized quantitative 319 data on cellular infectivity for lyssavirus serology. 320 Model selection indicated substantial day-to-day variation in the SRIG dilution series. This was 321 unsurprising, since virus neutralization tests are biologically dynamic systems that can be influenced 322 by many factors (e.g. variability in the humidity of the incubator, technical manipulation, light 323 condition of the microscope, variability in GFP expression in the cells) (Briggs et al., 1998;Hammami 324 et al., 1999;Kostense et al., 2012). Since our statistical approach handles this variability through the 325 random effect of test date, the pmRFFIT is best suited for large numbers of serum samples that require 326 testing to be carried out across multiple batches. However, performance is only marginally reduced 327 when running single models for each test date, implying the pmRFFIT may still be useful when fewer 328 samples are available for testing (see SI 5, 6). Surprisingly, fitting the GLMMs to data from a single 1:10 329 dilution of SRIG predicted both seropositivity and RVNA titer more accurately than models fit to both 330 the 1:10 and 1:25 dilutions. The reduced performance of the two-dilution model reflected higher 331 variability in the 1:25 dilution compared to the 1:10 dilution, as evidenced by greater IQR values (SI 332 2). Ultimately, this variability likely reflects both higher stochasticity in infected cell counts at lower 333 serum concentrations and pipetting error. Regardless, the ability to detect low titers (< 0.

Conflicts of Interest Statement 410
The authors declare no conflict of interest. 411 perform the cell count of the fluorescent cells to construct a database to fit the statistical models. C) 668 Construction of the statistical models with two different types of prediction. 669 SRIG data to fit models~6 6 hrs 5 fields/well