Scientific Papers

Infectious bursal disease virus: predicting viral pathotype using machine learning models focused on early changes in total blood cell counts | Veterinary Research

Description of Image

Viral strains

IBDV strains listed in Table 1 were propagated ex vivo on B cell primary cultures as described below.

Table 1 Viral strains used in this study, their pathotype and genotype classification.

Propagation of IBDV on primary chicken bursal cells and preparation of viral stocks

Bursa of Fabricius (BF) were aseptically collected from four-to-ten-week-old specific-pathogen-free (SPF) White Leghorns chickens (ANSES, Ploufragan, France) and were processed as previously described [18]. Bursal cells were maintained in lymphocyte culture medium at 40 °C in a humidified 5% CO2 incubator. This medium was prepared using Iscove’s modified Dulbecco’s Medium (IMDM, Fisher) with L-glutamine and HEPES (reference 21980-032, Gibco, Thermo Fisher) supplemented with 8% FBS (Fetal Bovine Serum), 2% SPF chicken serum (ANSES, Ploufragan, France), 1X insulin transferrin selenium (reference 41400–045, Gibco, Thermo Fisher), 50 µM beta-mercaptoethanol, 1 µg/mL Phorbol 12-myristate 13-acetate (PMA, reference tlrl-pma, Invivogen), penicillin (200 IU/mL), streptomycin (0.2 mg/mL) and fungizone (2 µg/mL). PMA was reconstituted as previously described [18]. Bursal cells were diluted into phosphate buffered saline (PBS) 1X, pH = 7.2 (Gibco reference 20012027) containing 0.1% (m/v) erythrosin B (reference 200964, Sigma-Aldrich) and counted in a Thoma’s chamber to estimate cell viability and concentration after cell isolation.

Ten million chicken B cells per mL were seeded in 400 mL of the lymphocyte culture medium in 150 cm2 flasks. Chicken B cells flasks were individually infected at a multiplicity of infection (MOI) of 0.001 with each viral strain. The infected cells were maintained at 40 °C in a humidified 5% CO2 incubator for 48 h. Flasks content were then centrifuged at 1500 g for 4 min (4 °C) to pellet cell debris and supernatants were recovered. The latter were transferred to Amicon Ultra-15, PLHK, membrane Ultracel-PL, 100 kD (reference UFC910024, Merck-Millipore) and centrifuged 45 min at 3500 g (4 °C) in order to eliminate as much as possible proteins and to obtain high-titer viral stocks (80-fold concentration capability). Concentrated supernatants were recovered and stored at –80 °C.

Viral titration revealed by immunochemistry (ICC)

Ten-fold serial dilutions of viral stocks in IMDM were distributed into 96-well U bottom plates (50 µL/well, eight replicates per viral dilution). Freshly prepared bursal cells in lymphocyte culture medium were added in each well (106 cells in 150 µL/well) and incubated at 40 °C for 48 h in a humidified 5% CO2 incubator. Forty-eight hours post-infection, the cells were washed with PBS and pelleted twice by mild centrifugation, then were fixed with ethanol and acetone solution (1:1 volumetric ratio) at −20 °C for at least 30 min. After removal of the fixation solution, the plates were air-dried under a chemical hood and processed immediately or stored at −20 °C until further processing. The plates were subjected to ICC as previously described [18].

Viral titers expressed as Log10(TCID50)/mL (Tissue Culture Infectious Dose) or Log10(TCID50)/g (depending on the nature of the tissue from which the viral particles were extracted) were determined using the Reed and Muench formula [19].

Next-generation sequencing (NGS) of viral stocks

In order to assess the purity of the viral stocks used for this study, the viral stocks were sequenced using NGS. RNA from 150 µL of each viral stock were extracted using the QIAamp viral RNA mini kit (reference 52904, Qiagen) following the manufacturer’s instructions. However, linear acrylamide (reference AM9520, Thermo Fisher) at 0.025 mg/mL was used instead of carrier RNA. RNA concentration was determined by using Qubit RNA HS assay kit (Invitrogen, Q32852) on the Qubit®2.0 Fluorometer.

For RNA sequencing, NGS was performed on the RNA extract after ribosomal RNA (rRNA) depletion with NEBNext rRNA Depletion Kit (NEB), as described by the manufacturer. A RNA library was obtained using Ion total-Seq Kit v2 (Life Technologies) according to the manufacturer’s recommendations and was then sequenced using Ion Torrent Proton technology. Reads were cleaned with the Trimmomatic (0.36) software (ILLUMINACLIP:oligos.fasta:2:30:5:1:true LEADING:3 TRAILING:3 MAXINFO:40:0.2 MINLEN:36). Then a Bowtie 2 (version 2.2.5) alignment was performed (–very-fast –score-min L, −0.5, −0.2 –non-deterministic -N 1) with down-sampled reads on a local nucleotide (nt) database to identify virus references. Another Bowtie 2 alignment was performed (–very-fast –non-deterministic -N 1) using Gallus gallus genome against cleaned reads; unmapped reads were extracted with samtools (1.8). The IBDV references with the highest number of matching reads were used for an alignment with bwa mem (0.7.8) against unmapped reads. The reads from this third alignment were collected then down-sampled to fit a global coverage depth estimation of 80 × and were submitted to the SPAdes (3.10.0) de novo assembler and Mira (4.0.2) de novo assembler (related raw reads for mira). The de novo contigs were then submitted to Megablast on the local nt database.

The best matching sequences (selected using their accession number) were used as references for a bwa mem alignment. Finally, the de novo assemblies and the alignment on the references were compared and the strict identities of the de novo and aligned sequences were assessed for validation of the final sequences using the Integrated Genomic Viewer 2.8.10 program.

Kraken [20] was used in a metagenomics approach to assign taxonomic labels to all the sequences found in the samples in order to highlight the presence of adventitious agents.

Ethical statement

All animal trials were conducted in animal facilities approved for animal experiments (n° C-22–745–1); chickens were raised and humanely euthanized in agreement with EU directive number 2010/63/UE. Pathogenicity assessment in SPF chickens was approved by ANSES ethical committee, registered at the national level under number C2EA-016/ComEth ANSES/ENVA/UPEC and authorized by French Ministry for higher education and research under permit number APAFiS#4945-20 16041316546318 v6.

Animal experiments: experimental design

A first experiment was designed to characterize, under standardized experimental conditions, the pathogenicity of the five pathogenic viruses for 21 days post-inoculation (dpi) (Experiment 1). This experiment was complemented by two additional experiments designed to study the early responses to infection by attenuated (Experiment 2) or pathogenic (Experiment 3) IBDV strains. Previous experiments performed in our laboratory revealed a peak in virus cloacal shedding at 2 dpi and blood B cell depletion at 4 days post-infection by a very virulent strain [21]. Therefore, 2 dpi (time of clinical signs onset), and 4 dpi (peak of clinical signs), were chosen as experiments 2 and 3 time points.

Three-week-old SPF White Leghorns chickens (ANSES, Ploufragan, France) were distributed into groups of similar weight and sex, housed in separate negative-pressure filtered-air isolators (except for the mock-inoculated groups which were housed in separate positive-pressure isolators), as presented in Table 2. Three days before inoculation, blood samples were collected from one third of the flock in order to confirm seronegativity against IBDV using a viral neutralization assay, as previously described [22]. Viral inocula were prepared by diluting viral stocks in PBS supplemented with penicillin (200 IU/mL), streptomycin (0.2 mg/mL) and fungizone (2 mg/mL). Chickens in the infected groups were inoculated by the intranasal route with 0.1 mL of virus (106 TCID50/mL, equivalent to 105 EID50/chicken). Mock chickens were mock-inoculated with diluent.

Table 2 Experimental design.

Clinical and pathological follow-up

Mortality rates were followed throughout the animal experiments. Clinical monitoring was performed from day 0 to day 10 (Experiment 1) or until termination of the experiment at day 4 (Experiments 2 and 3). Clinical signs were measured daily based on a symptomatic index (Additional file 2) previously developed, which ranges from 0 to 3 with increasing severity, an index of 3 representing the ethical endpoint of the experiment [23]. At the end of the experiments, all remaining chickens were weighed, humanely euthanatized, necropsied and their spleens and BFs were collected and weighed for calculating the spleen-to-body-weight ratio (SBR) and the bursa-to-body-weight ratio (BBR), respectively.

Histopathological analysis and bursal lesions scoring (Experiments 2 and 3)

Two to three tissue samples of bursa per day and per group stored in a 75% ethanol solution were analyzed by a pathologist (Labocea, Ploufragan, France) to score the histopathological lesions according to Skeeles et al. [24].

Determination of the viral load in the bursa at 2 and 4 dpi (Experiments 2 and 3)

At 2 dpi, for the groups infected with Vv1, Vv2 and Cla viruses (Experiment 3), ten chickens of each group were selected according to their high clinical score. In the mock, im1 and im2 viruses infected groups (Experiment 2), 10 chickens of each groups were randomly chosen, weighed and humanely euthanatized. During necropsy (see above), a piece of bursa was collected and processed for viral titration. At 4 dpi, the same procedure was repeated on all the remaining chickens. Viral particles from each 2 and 4 dpi collected BF were extracted as follows. Bursal tissue was homogenized using Tissue Lyser (Qiagen) homogenizer to process individual BFs. All steps were carried out on ice. Briefly, bursae were weighed and cut into small fragments. PBS was added (1 mL of PBS per gram of bursa) to the chopped bursa before homogenizing with a stainless steel bead for 3 min at 30 Hz with a Qiagen Tissue Lyser. 1, 1, 1, 2, 3, 4, 4, 5, 5, 5-Decafluoropentane (reference 94884, Sigma) was added to the tissue suspension (at an approximate 1:1 volumetric ratio) and an additional homogenization step was carried out followed by centrifugation at 4000 g for 30 min. Supernatants were collected and stored at −80 °C for a later titration using the above-described ICC.

Blood cell counts using flow cytometry (Experiments 2 and 3)

At 2 and 4 dpi, chickens were blood sampled at the venous occipital sinus before euthanasia, using commercial ethylenediaminetetraacetic acid (EDTA) coated blood collection devices (S-Monovette EDTA K2 2.7 mL, Sarstedt, reference 04.1915.100). Previous observations from the authors’ laboratory determined that jugular sampling during euthanasia, even with EDTA coated collection devices, led to significant coagulation of the samples. All blood samples were kept at room temperature and processed within 4 h after blood collection according to Seliger et al. [25]. Briefly, blood samples were diluted in staining buffer (PBS with 1% FBS), mixed with the labeled antibody mixture and incubated for 30 min with agitation (500 revolutions per minute or rpm) at room temperature and in the dark. The antibodies and fluorochrome conjugates used in this study and their dilution before use are indicated in Additional file 1. After incubation, Precision Count Beads (reference BLE424902, BioLegend) prepared in staining buffer were added to each sample to determine the absolute counts of cells. To inactivate the virus, formaldehyde (1% final concentration) was added and the samples were incubated for 15 min with frequent agitation (500 rpm) at room temperature and in the dark to inactivate the virus [26]. Samples were analyzed on a FC500 MPL flow cytometer (Beckman Coulter), with a previously used gating strategy [21]. The results were transformed by the logarithmic function for the rest of the analyses. The application of this transformation ensures the normal distribution of the variable studied (especially if there are outlier values among the results). Cell concentrations below the level of detection of the flow cytometer leading to a result of “0 cells/µL” were manually replaced with “1 cell/µL” to allow analysis of the logarithmic results since the value log10(0) is not defined.

Uric acid dosage (Experiments 2 and 3)

Dosage of uric acid was performed individually on mock- and virus-inoculated animals whose sera had been collected on serum tube. The Cayman uric acid titration kit (Cayman Chemical, 700320) was used according to manufacturer’s instructions. Briefly, 15 µL of each animal serum was mixed with 105 µL of diluted Assay Buffer, 15 µL of Fluorometric Detector and 15 µL of Enzyme Mixture. The mix was incubated for 15 min at room temperature. The fluorescence was measured using a Tecan Infinite M200 Pro.

Statistical analyses

Our first aim was to classify the very virulent, virulent (comprising Cla, im1 and im2 strains) and attenuated (i + vaccine and i vaccine) pathotypes of the strains with respect to 9 explanatory variables: the 7 blood concentrations (i.e., logarithms of concentrations of B cells, T cells, monocytes, granulocytes, erythrocytes, thrombocytes, uricemia), the bursal viral load and the clinical score. For this purpose, 182 observations were available, corresponding to 90 chickens at 2 dpi and 92 chickens at 4 dpi.

All datasets from experiments 2 and 3 at 2 dpi were arranged in a multivariate manner with animals in rows and, as columns, variables measured during the experiments: logarithmic blood concentrations of B cells, T cells, monocytes, granulocytes, erythrocytes and thrombocytes, logarithmic uricemia, logarithmic bursal viral load, clinical score at 2 dpi and group of infection. The same approach was applied to the data at 4 dpi in a different dataset.

A factorial discriminant analysis was applied to illustrate the differences between pathotypes and associate them with the explanatory variables under study [27]. All statistical analyses were performed using R (version 4.0.3) [28]. This analysis required the R packages data.table (version 1.14.2) [29], FactoMineR (version 2.4) [30], factoextra (version 1.0.7) [31], gridExtra (version 2.3.) [32], missMDA (Version 1.18) [33] (missing values imputation was performed), corrplot (version 0.92) [34] and the function discrimin of the ade4 package [35].

Then, a machine learning procedure was used to evaluate the predictive performance of ten classification models, the discriminant analysis being one of these models, and to select the model with the best predictive performances. The caret R package was used [36]. The ten selected models were: naïve Bayes classifier [37] (Version 0.9.7), Weighted k-Nearest Neighbors Classification (Version 1.3.1) [38], Random Forest (Version 4.7-1) [39], Kernel method (Version 0.9-29) [40], Neural networks (Version 0.4-14) [41], Bagged tree (Version 6.0-90) [36], C5.0 (Version 6.0-90) [36], Multi-Layer Perceptron (Version 6.0-90) [36] and L2 Regularized Support Vector Machine (dual) with Linear Kernel (Version 6.0-90) [36]. A repeated (100 times) two-fold cross-validation procedure (i.e., 70% of the observations = training dataset; 30% of the remaining observations = test dataset) allowed to assess the predictive performance of each of the ten models under study, which corresponded to the percentage of the well-classified observations. The required parameters of each model were tuned by means of a repeated (10 times) tenfold cross validation procedure applied to the training dataset. The global performance, or accuracy, of the models was evaluated throughout the percentage of animals correctly assigned to an infected group (for example, among all infected chicken, how many of them were correctly assigned to their infected group). A pathotype-specific performance of each model was also evaluated throughout the percentage of animals infected with a strain of a defined pathotype accurately classified (for example, among all the Cla infected chickens, how many of them were correctly assigned to the Cla infected group).

Our second aim was to describe the individual variations of the valuable parameters selected in the final model. Differences in percentages of mortality were analyzed using Fisher’s exact test followed by pairwise comparisons using the fisher.multcomp function from RVAideMemoire package version 0.9-79 [42]. All other quantitative parameters were analysed using Kruskal–Wallis test followed by Fisher’s least significant difference test with Holm adjustment method for multiple comparisons using the kruskal function from the Agricolae package version 1.3-3 [43].

Description of Image

Source link