Identification of New Isolates of Classical Swine Fever Virus Based on Untranslated Regions with Traditional and Machine Learning Classification

Classical swine fever (CSF) is a communicable infectious disease of pigs, including both domestic and wild types, with a devastating impact on the swine industry, as it causes large economic losses (Postel et al., 2012). China provides more than half the world’s pigs, therefore being the largest producer in the world (Zhou, 2019). CSF is caused by the classical swine fever virus (CSFV), which has a single-stranded, positive-sense RNA genome of 12.3 kb. It is a member of the Pestivirus genus and belongs to the Flaviviridae family (Zhou, 2019). On either side of the nucleic acid, there are 5’ and 3’ untranslated regions (UTRs) and a single open reading frame (ORF). It encodes eight nonstructural Npro, p7, NS2, NS3, NS4A, NS4B, NS5A, and NS5B and four structural proteins i-e. C, Erns, E1, and E2 (Rümenapf et al., 1993; Paton et al., 2000). The structural E2 protein aids in the generation of neutralizing antibodies through virus attachment and entrance into the cell, hence protecting pigs from a virulent challenge (Sánchez et al., 2008). The 5’- and 3’-UTRs are important regulatory components of the genome. 5’-UTR is involved in polyprotein translation and replication (Hsu et al., 2014). It has been shown that the 5’-terminal nucleotides (nt) 1–29 had an inhibitory effect on the translation of CSFV that is facilitated by IRES (the internal ribosome entry site) (Xiao et al., 2009). The 3’-UTR is involved in negative-strand RNA synthesis by generating many structural elements (Deng and Brock, 1993). Furthermore, the 3’-UTR affects translation by interacting with the 5’-UTR or in certain other ways (Ito et al., 1998). The conserved sequences at the UTRs suggest the signals of viral RNA replication (Fletcher and Jackson, 2002; Xiao et al., 2004).

CSFV genotypes 1, 2, and 3 have a total of 11 sub-genotypes i.e., including 1.1, 1.2, 1.3, and 1.4; 2.1, 2.1a, 2.1b, 2.1c, 2.1d, 2.2, 2.3; and 3.1, 3.2, 3.3, and 3.4 (Beer et al., 2015; Hu et al., 2016). Since the 1980s, genotype 2.1 has been circulating, particularly the 2.1b strain was dominant and endemic in most regions of China (Zhang et al., 2015).

All RNA viruses that include CSFV have higher mutation rates than DNA viruses, the reason behind this is that RNA-dependent RNA polymerase (RdRp) lacks proof-reading ability, resulting in poor fidelity replication (Drake and Holland, 1999; Hughes and Hughes, 2007). Nonetheless, the high mutation frequency might be advantageous for the viruses, allowing them to protect from the host immune responses (Sanjuán et al., 2010).

Recently, sequencing based on untranslated and coding regions of CSFV has been a potent method to trace changes in the virus population over time (Postel et al., 2012) or to characterize strains responsible for outbreaks in combination with epidemiological surveys (Greiser-Wilke et al., 2000; Shen et al., 2013). Moreover, several authors hypothesized that mutations (nucleotide substitutions or insertions/deletions) may accumulate in CSFV strains due to vaccination pressure and thus, in the near future, C-strain-based vaccines may lose the efficacy of protecting pigs against all strains of CSFV (Sanjuán et al., 2010; Ji et al., 2015; Yoo et al., 2018).

Although extensive vaccination strategies have been implemented in China, CSF outbreaks are still common, as demonstrated by recent cases in pig herds vaccinated with C-strain (Luo et al., 2014; Zhang et al., 2018). Therefore, the present study aims to identify all CSFV genotypes and sub-genotypes circulating in different regions of China based on analysis of the untranslated regions.

The present study includes eight strains of CSFV Heilongjiang (HL2018-0494, HL2018-016, HL2016-0205, HL2018-0490, HL2018-0462), Inner Mongolia (NM2016-0323, NM2016-0333) and Shengdong (SH2018-0461), which have previously been separated from tissue samples (lymph node, spleen, kidney, and tonsil). These samples were taken from deceased or ill pigs suspected of having CSF at various swine farms in different locations in eastern China. Furthermore, these strains were also identified and characterized earlier by PCR.

Genomic material (RNA) of CSFV strains extracted from cell cultures and synthesis of cDNA performed and used to amplify the 5’-UTR and 3’-UTR regions by RT-PCR using two pairs of designed primers CSFV-FP-UTR (1-21)-5’-GTATACGAGGTTAGCTCATCC-3’, CSFV-RP-UTR (988-1008)-3’-CTCTACCACAATCGTAGCATC-5’; CSFV-FP-UTR (11054-11075)-5’-CTATGCACATGTCAGAAGTACC-3’, CSFV-RP-UTR (12259-12279)-3’-ACCTTAGTCCAACTATGGACG-5’. The resultant product was stained with gene SafeViewTM (Applied Biological Material, Richmond, BC, Canada) and electrophoresed on 1.5% gel of agarose before being seen using the Gel Doc XR+ system. The amplified product was cloned into the pMD18-T vector before it was incorporated into E. coli (DH5 α) cells (TaKaRa, Dalian, China). The colonies were grown before being delivered to Comate Bioscience in Jiangsu, China, for sequencing.

Multiple sequence alignments of these eight strains of CSFV were generated using the Laser gene (Version 7.1., DNASTAR Inc., Madison, WI, USA) and MEGA6 Software (Center for Evolutionary Functional Genomics, the Biodesign Institute, Tempe, AZ, USA). The neighborhood-joining (NJ) based phylogenetic trees were constructed with MEGA6 (Tamura et al., 2011).

Virus confirmation and titration in the cell line were determined by immunofluorescence assay (IFA) based on the EU Diagnostic Manual (EU Reference Laboratory for CSF (EURL), 2020).

To further validate our results, we employed a Support Vector Machine (SVM) model (Shapshak et al., 2019) to perform class predictions on 8 new chains. SVM is a supervised learning algorithm, which realizes classification by finding an optimal hyperplane. SVM is particularly suitable for complex datasets and high-dimensional feature spaces, making it an ideal choice for our research. Its ability to find the best decision boundary ensures accurate classification of the new chains. At first, we used k-mers as features to convert the CSFV gene sequence data into feature vector representations. This method, as described by Rahman et al. (2018), can transform sequence data (such as DNA sequences or protein sequences) into k consecutive character Subsequence, called k-mers. These k-mers can be used as feature representations for sequence data. By calculating the frequency of each k-mers in the sequence, a feature vector can be constructed, where each element represents a different k-mers and represents the number or frequency of occurrences of that k-mers in the sequence. Then, we trained on previous data sets using the SVM model and classified the new 8 CSFV chains. In the process of classifying CSFV, a linear kernel function is used for classification and a one-to-many strategy is used to handle multi-class classification problems. Furthermore, the performance of the model is evaluated through indicators such as Accuracy, Recall, and F1-score. Finally, to gain insights into the classification results, we employed PCA dimensionality reduction (Groth et al., 2013) to visualize the distribution of the sequences additionally. PCA is a dimensionality reduction technique that allows us to represent the data in a lower-dimensional space while preserving its essential characteristics. In our study, we applied PCA to reduce the dimensionality of the feature vectors extracted from the CSFV gene sequences. By reducing the dimensionality to three components, we could visualize the classification results in a three-dimensional scatter plot.

The samples were detected CSFV with IFA and processed next (Fig. 1). All selected monoclonal antibodies were specifically generated against the epitopes of the E2 protein. Epitopes of structural E2 proteins are important not only for the early stage of CSFV infection but also for serological differentiation of CSFV from other ruminant pestiviruses, given strong cross-reactivity with sheep and goat sera in ELISA and neutralization assay (Huang et al., 2021). The obtained sequence data sets of UTRs regions were subjected to phylogenetic analysis and investigated systematically to gain the discriminatory ability of the individual genomic regions. In total, 25 and 30 genome sequences including the 8 obtained strains together with 21 and 22 reference sequences were used in the construction of the phylogenetic trees based on the 5’- and 3’-UTRs. The isolates (HL2018-0494, NM2016-0323, HL2018-016 and HL2016-0205) were clustered into sub-genotype 2.1d, while the rest isolates (HL2018-0490, SD2018-0461, HL2018-0462 and NM2016-0333) were clustered into sub-genotype 2.1b (Fig. 2).

Table II. Results of SVM model predictions. The SVM model successfully predicted the category of each chain, regardless of whether the 5’-UTR or the 3’-UTR data was used for training.

The classification results of the SVM model on our dataset achieved perfect Accuracy, Recall, and F1-score, as indicated in Table II. Whether using only the 5’-UTR or the 3’-UTR data for training the SVM model, we were able to successfully predict the categories of the new eight chains, and the prediction results were consistent with our phylogenetic tree analysis. That is, chains HL2018-0462, HL2018-0490, NM2016-0333, and SD2018-0461 were predicted as subgenotype 2.1b, while chains NM2016-0323, HL2018-0494, HL2018-0416, and HL2016-0205 were predicted as subgenotype 2.1d. Figure 3 illustrates the distinct separation of different categories in the PCA-reduced and visually represented classification results, indicating that the SVM model effectively captures

the underlying patterns and discriminative features within the data. The integration of SVM in our study adds a layer of validation and enhances the reliability of our results. The present result agrees with several studies indicating sub-genotype 2.1 strains have been dominant in China since 2000, although other sub-genotypes (1.1, 2.2 and 2.3) were also found in the field (Luo et al., 2017; Zhang et al., 2018; Gong et al., 2019; Hao et al., 2020). We examined the mutations and deletions on domain Ia, which spans nucleotide (nt), 1 to 29 at the 5’-terminal sequence of the CSFV genome and formed a stem-loop structure. By comparison of other vaccine strains and sub-genotypes (used as reference strains), the 5’-UTRs of the newly isolated strains had a nucleotide T deletion at positions 15 and 21 (T15/T21) and HL18-462, HL18-490, and NM16-333 strains had a nucleotide G insertion at position 18 (Fig. 4A). The mutations in domain la were responsible for high IRES-mediated translation but were less important for CSFV replication (Xiao et al., 2011). In addition, CSFV isolated strains showed nucleotide C and A deletions at 44 (C44) and 45 (A45) positions, while HL2016-0205, HL2018-0416, HL2018-0494 and NM2016-0323 strains showed mutations at 67(G67), 122(A122), 125(C125), 163(T163), 250(G250) and 293(C293), clearly distinguishing isolated strains as two sub-genotypes (2.1d and 2.1b) (Fig. 4A). Deletions in the 5’-UTR may change structural characteristics and integrity, some studies showed that CSFV virulence varied due to the secondary structure of the 5’-UTR, with differing numbers and shapes of the pseudoknot loop (Fletcher and Jackson, 2002;Li et al., 2006). Compared with the new isolates, the 3’-UTR of all vaccine strains have a continuous 12-nucleotide (CTTTTTTCTTTT) insertion at positions 45 to 69. The HL2016-0205, HL2018-0416, HL2018-0494 and NM2016-0323 strains showed mutations at 48(C48), and HL2018-0462, HL2018-0, NM2016-0333 and SD2016-0461 strains showed mutations at positions 99 (T99), 120 (A120) and 124 (A124) (Fig. 4B). The 3’-UTR was the most variable region in the genome and poly (T) deletion was an important virulence factor in that region (Wang et al., 2008). Previously, several isolates, including RUCSFPLUM, Brescia, Kolovos, Margarita, Shimen, and Rovac, with different virulence, were found with deletions in the same region (Li et al., 2006). Furthermore, it has been reported that the CSFV Pinar del Rio strain had a unique poly(T) tract in the 3’-UTR, while other CSFV strains had no deletions (Pérez et al., 2012; Coronado et al., 2017). The reason that unique poly(T) insertion affects virulence is still unknown. Some studies mentioned that 3’-UTR, together with the NS2-3, NS5A, and NS5B genes, can control RNA expression and synthesis (Sheng et al., 2012; Chen et al., 2012). The 3’-UTR of the new isolates and sub-genotype 2.1 strains exhibited some discontinuous nucleotide deletions compared with 1.1 isolates. Moreover, the interaction between 3’-UTR nucleotide deletions and other genes needs further investigation.

In conclusion, the present findings provide that genetic typing based on targeting UTRs of CSFV may be useful for evaluating methods to detect mutations in the virus population and characterize the strains involved in outbreaks. This could help to establish a national program on prevention and control strategies to protect pigs against CSFV infection.

Prof. John Morris (KMITL, Bangkok, Thailand) is thanked for his help in translating the manuscript and the authors extend their appreciation to Dalia Fouad for supporting this research through Researchers Supporting Project number (RSPD2024R965), King Saud University, Riyadh, Saudi Arabia.

This work was supported by the National Natural Science Foundation of China (Nos. 3157240 and 31402194), Natural Science Foundation of Heilongjiang Province of China (No. ZD201410), and the State Key Laboratory of Veterinary Biotechnology, Harbin Veterinary Research Institute, CAAS, China (no. SKLVBP2015013).

The genomic sequences of the 5’- and 3’-UTR regions of 8 isolates have been deposited in the NCBI GenBank with accession numbers OQ150757-OQ150772.

The authors have declared no conflict of interest.