Whole-genome sequencing reveals hidden antimicrobial resistance genes in phenotypically susceptible probiotic candidate lactic acid bacteria Authors: Yuda Disastra, Thidathip Wongsurawat, Piroon Jenjaroenpun, David John Hampson, Ratchnida Kamwa, Nuvee Prapasarakul Published: 2026 ABSTRACT Phenotypic assays commonly used to evaluate probiotic safety may fail to detect clinically relevant antimicrobial resistance (AMR), potentially allowing genetically concerning strains to appear acceptable based on MIC testing alone. To explore this issue, we applied whole-genome sequencing (WGS) to three lactic acid bacteria (LAB) isolates previously identified as probiotic candidates based on acid and bile tolerance, antagonism against enteric pathogens, and biofilm formation in vitro: Lactiplantibacillus plantarum L25F and L22F (from pigs) and Ligilactobacillus salivarius AF2319 (from a chicken). Genome annotation identified extensive repertoires of probiotic-associated genes (46–47 per strain) linked to stress tolerance, adhesion, immunomodulation, and quorum sensing, supporting functional potential. The two L. plantarum strains exhibited broader predicted metabolic capacities than L. salivarius AF2319. However, genomic analysis revealed acquired AMR genes with complex genotype–phenotype relationships not fully apparent from phenotypic testing. The L. plantarum strains harbored lnu(A) (99.79% identity) on extrachromosomal DNA, conferring the L-phenotype (lincomycin resistance, clindamycin susceptibility); clindamycin MICs (1 mg/L) were concordant with this genotype, though lincomycin MICs were not determined. L. salivarius AF2319 carried tet(M), tet(L), and erm(C) (99.48%, 99.49%, and 99.45% identity by ResFinder, respectively) on extrachromosomal DNA; notably, the erythromycin MIC (1 mg/L) was precisely at the EFSA breakpoint (≤ 1 mg/L), representing borderline genotype–phenotype discordance potentially due to silent gene expression. Under current EFSA QPS criteria, these acquired ARGs would preclude all three strains from approval as probiotic feed additives despite favorable functional profiles, underscoring the indispensable role of WGS-based AMR gene detection in modern probiotic safety evaluation.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57518-x. Introduction The use of probiotics as oral supplements in livestock production has increased greatly in recent years. These microbes are used to improve animal production and to help control enteric bacterial pathogens, especially those exhibiting antimicrobial resistance (AMR), and represent a critical strategy for reducing antibiotic dependence while maintaining animal health and productivity. Among probiotic candidates, lactic acid bacteria (LAB), particularly Lactiplantibacillus plantarum and Ligilactobacillus salivarius are increasingly recognized for their ability to colonize the gut, inhibit pathogens, and survive under diverse stress conditions1. In the European Union, however, the term “probiotic” is generally considered to constitute a non-authorized health claim under Regulation (EC) No 1924/2006 when used on food labels, as it implies a health benefit that has not been formally approved. Consequently, strains intended for food and feed applications are often more appropriately referred to as “probiotic candidates” or “live microbial cultures” in regulatory contexts until their safety and efficacy have been sufficiently demonstrated2. For instance, L. salivarius CGMCC20700, isolated from a chicken, demonstrated high bile and acid tolerance, broad-spectrum antimicrobial activity, and a complete lack of virulence or resistance genes, supporting its candidacy for livestock applications3. Similarly, a swine-derived strain of L. salivarius enhanced nutrient absorption and immunity through the production of bioactive metabolites and digestive enzymes, while exhibiting strong antimicrobial effects without harboring virulence genes4. L. plantarum, meanwhile, not only supports gut health but also improves the quality and stability of silage used in livestock feed. For example, co-fermentation with L. plantarum in corn-foxtail millet silage has been shown to enhance lactic acid production and reduce fungal contamination5. Several strains have also exhibited strong colonization capacity and pathogen inhibition, reinforcing their potential as alternatives to antibiotics in animal production systems. Despite their widespread use and general recognition of safety, the evaluation of these probiotic strains has historically relied on phenotypic assays, such as antimicrobial minimum inhibitory concentration (MIC) profiling and hemolysis screening, that are inherently limited in detecting latent genetic risks6. These methods often overlook silent carriage of AMR genes, mobile genetic elements, or virulence factors that may pose safety concerns under specific conditions, particularly when they are genetically present but are not expressed phenotypically. This phenomenon of genotypic-phenotypic discordance is of particular concern for plasmid-borne resistance determinants, which may be silently transferred to other bacteria even when not phenotypically detectable, potentially contributing to horizontal gene transfer in the gut microbiome. To address these limitations, whole genome sequencing (WGS) has become a central tool for the safety assessment of microorganisms used in the food chain. WGS enables the detection of antimicrobial resistance (AMR) genes, plasmids, insertion sequences, prophages, virulence factors, and markers of horizontal gene transfer, many of which remain undetectable through conventional assays, and is now explicitly recommended in international guidance, including that of the European Food Safety Authority (EFSA), for the characterization of bacterial strains intended as feed additives or production organisms7. In line with this, numerous WGS- and metagenomics-based studies have explored the antimicrobial resistance gene profiles of LAB from probiotic products, starter cultures, human milk and gut isolates, and fermented foods, and several have systematically examined the concordance between phenotypic susceptibility testing and genomic AMR predictions. These investigations generally report high overall agreement between genotype and phenotype but also describe notable exceptions in lactobacilli, where unexplained phenotypic resistance or silent genomic determinants can occasionally be observed8. While this growing body of work has substantially advanced our understanding of LAB antimicrobial resistance gene profiles, there remains a need for focused case studies that integrate harmonized phenotypic and genomic safety evaluation in concrete probiotic candidate strains from defined ecological niches, particularly livestock. In particular, studies that explicitly contrast strain-level safety determinants within and between commonly used probiotic LAB species isolated from farm animals, while considering the genomic context and potential mobility of acquired AMR genes, are still limited9. Distinguishing between intrinsic characteristics that are broadly conserved within a species and those that vary at the strain level is essential for evidence-based probiotic selection and regulatory decision-making, especially under current QPS and EFSA frameworks. Moreover, the safety implications of AMR genes located on plasmids or other mobile genetic elements, including cases where they are not clearly reflected in MIC-based phenotypes, remain underexplored in practical probiotic risk assessment despite their potential contribution to AMR dissemination in complex microbial ecosystems. To address these gaps, the present study performed an integrative genomic and phenotypic characterization of three lactic acid bacteria probiotic candidate strains, two L. plantarum (L25F and L22F) isolated from pigs and one L. salivarius (AF2319) isolated from a chicken. These strains were selected from a larger internal collection of livestock-derived LAB isolates based on their demonstrated probiotic-relevant phenotypic characteristics, including acid and bile tolerance, antimicrobial activity against enteric pathogens, and suitability for further safety evaluation as candidate probiotic strains. By combining MIC profiling, hemolysis screening, and comprehensive WGS analysis, we aimed to characterize strain-level safety traits and AMR determinants, evaluate the concordance between genotypic and phenotypic resistance profiles, and illustrate how WGS-based antimicrobial resistance gene analysis enhances strain-level safety assessment within the EFSA/QPS framework, rather than to derive definitive species-wide conclusions. The inclusion of strains representing two commonly used livestock-associated LAB species from a shared ecological context enables controlled exploratory comparisons of genomic safety features within this defined set of candidate strains, while recognizing that broader conclusions regarding species-level characteristics require larger and more diverse strain collections. These findings provide practical insight into the application of integrated genomic safety assessment for probiotic candidate evaluation in livestock production systems. Materials and methods Bacterial strains and growth conditions This study involved three lactic acid bacteria (LAB) strains that were isolated in Thailand: two L. plantarum isolates (L25F and L22F) from pigs and one L. salivarius isolate (AF2319) from a chicken. These strains, obtained from the culture collection of the Faculty of Veterinary Science, Chulalongkorn University, Thailand, were selected from our in-house culture collection based on the following criteria: (i) prior demonstration of probiotic-associated phenotypic traits, including acid and bile tolerance, antagonistic activity against key enteric and opportunistic pathogens, including Escherichia coli, Salmonella enterica and Staphylococcus aureus, and biofilm formation in vitro, (ii) isolation from livestock species of direct relevance to veterinary probiotic applications (pigs and chickens), and (iii) representation of two taxonomically distinct but widely used probiotic LAB species, L. plantarum and L. salivarius, thereby enabling a preliminary exploration of intra- and inter-species safety variation within a controlled ecological context10,11. Each strain was preserved in de Man, Rogosa, and Sharpe (MRS) broth containing 20% (v/v) glycerol and stored at − 20 °C. Prior to analysis, the strains were cultured in MRS broth at 37 °C for 24–28 h in a standard incubator. A loopful of the culture was subsequently streaked onto MRS agar plates and incubated for another 48 h to obtain isolated colonies. All procedures involving these strains were conducted in compliance with institutional biosafety protocols and approved by the Chulalongkorn University Institutional Biosafety Committee (IBC No. 2331008). Species confirmation of the LAB strains Initial species identification of the three LAB strains was performed using routine phenotypic methods, namely Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry (MALDI-TOF MS) and the VITEK 2 Compact automated biochemical system. These methods are part of the standard diagnostic workflow in our laboratory and were used to provisionally assign species prior to genome sequencing, whereas whole-genome sequence–based analyses (ANI and SpeciesFinder) were subsequently applied and considered definitive for taxonomic confirmation in this study, consistent with current standards recognizing ANI ≥ 95–96% as the gold standard for bacterial species delineation12. Single colonies grown on MRS agar were first analyzed by MALDI-TOF MS. Colonies were transferred to a steel target plate and overlaid with a matrix solution containing formic acid and α-cyano-4-hydroxycinnamic acid (HCCA). Mass spectra were acquired using MALDI Biotyper software (Bruker Daltonics), and identification confidence was interpreted according to the manufacturer’s database and published criteria: scores above 2.3 indicate secure species-level identification, scores between 2.0 and 2.3 signify probable species identification, and scores between 1.7 and 2.0 correspond to reliable genus-level assignment13. To validate results, parallel identification was performed with the VITEK 2 Compact system (bioMérieux). Colonies were suspended in 0.45% sodium chloride and adjusted to a McFarland standard of 0.50–0.63. Prepared suspensions were loaded onto VITEK GP ID and CBC identification cards, and incubated at 35 °C for approximately six hours, in accordance with the manufacturer’s guidelines14. Antimicrobial susceptibility testing The antimicrobial susceptibility of the three LAB strains was evaluated using the broth microdilution technique according to the latest Clinical and Laboratory Standards Institute (CLSI) guidelines. Although CLSI protocols were originally developed for clinical pathogens rather than lactic acid bacteria, broth microdilution is widely used for MIC determination in LAB, including probiotic and food-associated strains, and is compatible with EFSA’s MIC-based cut-offs for safety assessment15. The antibiotics tested were ampicillin, vancomycin, gentamicin, kanamycin, streptomycin, erythromycin, clindamycin, tetracycline, and chloramphenicol. Susceptibility testing was performed on THAPH and THA1CLF Sensititre® plates (Thermo Fisher Scientific), and MICs were determined using the Sensititre™ Vizion™ Digital MIC Viewing System. Each assay was conducted in triplicate, and MIC endpoints were reported as the mode of three independent results. MIC results were interpreted using microbiological cut-off values recommended by the European Food Safety Authority16, classifying strains as susceptible (S) or resistant (R). Two reference strains, Enterococcus faecalis ATCC 29,212 and Escherichia coli ATCC 25,922, were included in each assay as controls, and their MIC values were within the acceptable ranges specified by CLSI guidelines, confirming assay reliability. Incubations were performed at 37 °C for 24 h, following manufacturer protocols17. Hemolytic activity testing The three LAB strains were screened for hemolytic potential by streaking each onto tryptic soy agar (TSA) plates enriched with 5% (v/v) defibrinated sheep blood. The cultures were incubated anaerobically at 37 °C for 48 h. Following incubation, colony-associated hemolytic effects were evaluated by observing the nature of any discoloration or clearing in the medium surrounding the colonies18. Hemolysis was classified as follows: a clear, transparent area indicated complete (β) hemolysis; partial greenish coloration was interpreted as α-hemolysis; while no observable change around colonies signified γ-hemolysis, suggesting an absence of red blood cell lysis19. Whole-genome sequencing and assembly Cells from each of the three LAB strains were grown to late exponential phase and harvested for genomic DNA extraction using the ZymoBIOMICS™ DNA Miniprep Kit (Zymo Research, USA), following the manufacturer’s instructions20. DNA yield and purity were verified via NanoDrop™ spectrophotometry and fluorometric quantification with a Qubit® 4.0 fluorometer to ensure suitability for downstream applications21. Whole-genome sequencing was conducted using a hybrid approach that integrated the Oxford Nanopore GridION (long-read) platform with Illumina short-read sequencing. Raw Nanopore reads were assessed for quality and length distribution using NanoPlot (v1.33.0)22, followed by de novo genome assembly with Flye (v2.9.1-b1780)23. Assemblies were iteratively polished with NextPolish24, and plasmid structures were resolved utilizing Unicycler25. The completeness and accuracy of genome assemblies were evaluated using QUAST (v5.0.2) and CheckM (v1.1.9)26. For plasmid classification and mobility gene prediction, MOB-suite v3.0.327 was used. To confirm chromosomal origin, genome sequences were aligned against the GTDB-Tk reference database (release R207_v2)28. Genome annotation was conducted using Prokka v1.14.629, and genome maps were visualized using PyCircos and Proksee30. All experimental procedures were carried out at the Siriraj Long-Read Laboratory, Faculty of Medicine Siriraj Hospital, Mahidol University in Bangkok, Thailand. The sequencing data generated in this study have been deposited in the National Center for Biotechnology Information (NCBI) database under BioProject accession number PRJNA1369779, including whole-genome sequences of strains L. salivarius AF2319 (accession JBSYVJ000000000), L. plantarum L22F (accession JBSYVK000000000), and L. plantarum L25F (accession JBSYVL000000000) (see Supplementary Table S1 for details). Species-level resolution and phylogenomic comparisons The three LAB strains were taxonomically identified using orthologous average nucleotide identity (ANI) with the OrthoANI Tool (OAT) on the EzBioCloud platform31 and species assignments were cross-checked with SpeciesFinder v2.032. Evolutionary relationships were inferred from core single nucleotide polymorphisms (SNPs) using RAxML v8.2.12 with the GTRGAMMA substitution model to construct a maximum-likelihood phylogenetic tree including the three study genomes and 12 reference L. plantarum and L. salivarius strains from NCBI (Supplementary Table S1)33. The tree was visualized and annotated in Interactive Tree of Life (iTOL)34. Protein-coding genes from L. plantarum strains (L25F, L22F, and reference strain WCFS1) and L. salivarius strains (AF2319 and reference strain UCC118) were analyzed with Orthovenn3 (default settings) to identify shared and strain-specific orthologous gene clusters among the analyzed genomes35. Detection of antimicrobial resistance genes and virulence factors Antimicrobial resistance (AMR) genes were predicted by screening predicted protein sequences with the Resistance Gene Identifier (RGI, version 6.0.5) against the Comprehensive Antibiotic Resistance Database (CARD, version 4.0.1). RGI was run in “perfect and strict hits only” mode with high quality/coverage parameters, and the “nudge loose hits to strict” option was disabled; only perfect and strict hits, as defined in CARD, were considered confirmed ARGs36. Acquired ARGs were additionally screened with ResFinder (version 4.7.2) using minimum alignment coverage and identity thresholds of 70% and 80%, respectively, and genes meeting these criteria in either RGI (perfect/strict) or ResFinder were regarded as high-confidence acquired ARGs for safety assessment. Lower-identity or fragmented matches that did not meet these cut-offs were not interpreted as functional resistance determinants in the risk assessment37. To determine the genomic context of acquired ARGs, contigs carrying these genes were analyzed with MOB-suite v3.0.327 for plasmid prediction and typing. Contigs classified as plasmid-derived by MOB-suite were considered extrachromosomal (plasmid-like), whereas ARGs on contigs predicted to be chromosomal were interpreted as chromosomally encoded. KEGG annotation via BlastKOALA (BRITE category ko01504) was used to identify genes associated with antimicrobial resistance pathways at the functional level38. Virulence factors were predicted by querying genomic data against entries in the Virulence Factor Database (VFDB) (http://www.mgc.ac.cn/cgi-bin/VFs/v5/main.cgi39. Putative virulence factor genes (VFG) identified in all platforms were filtered using stringent thresholds: a minimum sequence coverage and identity of 80%, and an E-value cutoff of less than 1e− 10, to maintain confidence in gene assignment. This integrative computational framework enabled high-resolution mapping of possible resistance and virulence determinants in each bacterial genome. Secondary metabolite and bacteriocin gene cluster analysis Secondary metabolite biosynthetic gene clusters were predicted with AntiSMASH v6.040, and bacteriocin gene clusters were identified with BAGEL4, which detects putative core bacteriocin ORFs together with associated immunity and transport/export genes41. These analyses provided an overview of secondary metabolite biosynthesis and bacteriocin-related loci in the three LAB genomes. Genome stability and horizontal gene transfer potential To evaluate the genomic integrity and mobility potential of the three LAB strains, a combination of bioinformatics tools was employed. CRISPRCasFinder v1.1.242 was used to detect clustered regularly interspaced short palindromic repeats (CRISPR) and associated cas genes, providing insight into the presence of adaptive immune systems. Prophage regions were identified using PHASTER43, which differentiates intact, questionable, and incomplete prophage sequences within bacterial genomes. OriTfinder (https://bioinfo-mml.sjtu.edu.cn/oriTfinder/) was utilized to screen for known plasmid replicons and predict origin-of-transfer (oriT) regions, assessing plasmid mobility based on the co-occurrence of relaxase, type IV coupling protein (T4CP), and mating pair formation (MPF) genes44. Genomic islands were detected using IslandViewer 445, which integrates three prediction algorithms, SIGI-HMM, IslandPath-DIMOB and IslandPick for comprehensive island identification. Together, these analyses offered valuable insight into the structural genome architecture and the horizontal gene transfer potential of the three strains, which are key factors in assessing probiotic safety. Functional annotation of probiotic-related genes and metabolic pathways To evaluate the functional probiotic potential of the three LAB strains, a combination of annotation-based keyword mining and pathway reconstruction was employed. Functional gene categories associated with probiotic traits, such as acid and bile resistance, oxidative stress tolerance, adhesion, and immunomodulation, were initially identified through manual screening of annotated genome files. This process was complemented by subsystem-based functional categorization using the RAST server46, enabling genes to be grouped into biologically meaningful subsystems. Comprehensive functional annotations across multiple classification systems were achieved using eggNOG-mapper 5.0, which assigned proteins to Clusters of Orthologous Groups (COG) functional categories, KEGG ortholog identifiers (KO numbers), and Gene Ontology (GO) terms spanning Biological Process, Molecular Function, and Cellular Component domains47. Metabolic pathway reconstruction was performed using the KEGG Mapper “Reconstruct Pathway” tool, with KEGG ortholog identifiers (KO numbers) assigned via BlastKOALA, allowing integration of functionally annotated genes into canonical microbial metabolic pathways48. This integrated framework enabled a robust evaluation of both the functional repertoire and metabolic flexibility of each strain, providing key insights into their potential roles as effective probiotics. Results Species identification of LAB strains Species-level identification was performed on the three LAB isolates using two independent methods (Table 1). MALDI-TOF MS generated the following score values: L. plantarum L25F (2.13), L. plantarum L22F (2.31), and L. salivarius AF2319 (2.34). All three strains exceeded the score threshold of 2.0 for species-level identification. The VITEK 2 Compact automated identification system assigned L. plantarum to strain L25F with 90% probability and to strain L22F with 88% probability. For strain AF2319, the system returned a classification of “low discrimination” without a definitive species assignment. Table 1MALDI-TOF MS and VITEK 2 Compact identification scores for L. plantarum L25F and L22F, and L. salivarius AF2319.No.LAB StrainsMALDI-TOF MSVITEK 2 CompactOrganism (best match)Score ValueOrganism (best match)Score Value1L25F L. plantarum 2.13 L. plantarum 90%2L22F L. plantarum 2.31 L. plantarum 88%3AF2319 L. salivarius 2.34Low discrimination- Table 1MALDI-TOF MS and VITEK 2 Compact identification scores for L. plantarum L25F and L22F, and L. salivarius AF2319.No.LAB StrainsMALDI-TOF MSVITEK 2 CompactOrganism (best match)Score ValueOrganism (best match)Score Value1L25F L. plantarum 2.13 L. plantarum 90%2L22F L. plantarum 2.31 L. plantarum 88%3AF2319 L. salivarius 2.34Low discrimination- MALDI-TOF MS and VITEK 2 Compact identification scores for L. plantarum L25F and L22F, and L. salivarius AF2319. Genome assembly quality and structural features Whole-genome sequencing and assembly yielded draft genomes for all three LAB strains, with genome sizes around 3.2–3.4 Mb for the L. plantarum strains and 2.1 Mb for L. salivarius AF2319, GC contents typical of their respective species, and assemblies dominated by one large contig per genome, resulting in N50 values equal to the largest contig and L50 of 1 in all cases (Table 2; Fig. 1). CheckM analysis indicated high completeness (≥ 99.4%) and low contamination (≤ 2.8%) for all assemblies, and plasmid sequences were detected in each strain (Supplementary Table S2), together supporting that these represent high-quality draft genomes. Table 2Genome assembly quality metrics for L. plantarum L25F and L22F, and L. salivarius AF2319.StrainNo. of contigsTotal length (bp)Largest contigs (bp)Contig N50 (bp)Contig L50G + C content (%)CompletenessContaminationPlasmidL25F113,209,1013,036,5903,036,590145.199.42.2DetectedL22F153,399,3013,026,5643,026,564144.999.42.8DetectedAF231982,147,6901,861,9901,861,990132.7699.50Detected Table 2Genome assembly quality metrics for L. plantarum L25F and L22F, and L. salivarius AF2319.StrainNo. of contigsTotal length (bp)Largest contigs (bp)Contig N50 (bp)Contig L50G + C content (%)CompletenessContaminationPlasmidL25F113,209,1013,036,5903,036,590145.199.42.2DetectedL22F153,399,3013,026,5643,026,564144.999.42.8DetectedAF231982,147,6901,861,9901,861,990132.7699.50Detected Genome assembly quality metrics for L. plantarum L25F and L22F, and L. salivarius AF2319. Fig. 1Circular genome maps of L. plantarum L25F (A) and L22F (B), and L. salivarius AF2319 (C). Circular genome maps of L. plantarum L25F (A) and L22F (B), and L. salivarius AF2319 (C). Phylogenomic analysis and strain relationships Pairwise OrthoANI analysis confirmed clear species delineation and close intra-species relationships among the three LAB strains (Supplementary Figure S1). L. plantarum L25F and L22F shared 99.98% identity, indicating near-clonal relatedness. Both showed > 95% ANI with L. plantarum WCFS1 (95.66% and 95.62%, respectively), confirming their species identity. L. salivarius AF2319 exhibited 97.17% ANI with reference strain UCC118, well above the species threshold. Inter-species comparisons yielded < 68% ANI, supporting clear taxonomic separation. Phylogenomic analysis further resolved evolutionary relationships among the strains (Fig. 2). L. salivarius AF2319 clustered with L. salivarius reference strains UCC118 and ATCC 11,741, forming a subclade distinct from other L. salivarius references (CECT 5713, Ren, S01). Within the L. plantarum clade, L25F and L22F formed a subclade with L. plantarum ZJ316, separated from WCFS1 and other reference strains. Fig. 2Phylogenetic relationships of the three LAB strains and representative reference genomes inferred from core-genome alignment. The tree was constructed using a maximum likelihood approach (GTRGAMMA) with 1000 bootstrap replicates; node labels indicate bootstrap support values ≥ 70%, and the scale bar represents the number of substitutions per site. Phylogenetic relationships of the three LAB strains and representative reference genomes inferred from core-genome alignment. The tree was constructed using a maximum likelihood approach (GTRGAMMA) with 1000 bootstrap replicates; node labels indicate bootstrap support values ≥ 70%, and the scale bar represents the number of substitutions per site. Orthologous gene clustering (OrthoVenn) among L. plantarum L25F, L22F, and WCFS1, and L. salivarius AF2319 and UCC118 identified 1,093 shared core clusters (Supplementary Figure S2). ‘Unique’ gene clusters are operationally defined as ortholog groups present exclusively in a single strain and absent from all other compared genomes. L. plantarum L25F and L22F contained 2,935 and 2,983 clusters, respectively, with only L22F harboring seven strain-specific clusters. WCFS1 carried 2,595 clusters, including 24 unique ones. In contrast, L. salivarius AF2319 exhibited a smaller pangenome (1,638 clusters) but the highest unique content (38 clusters), while UCC118 contained 1,657 clusters without unique genes. These data confirm clear inter-species divergence and substantial intra-species genomic variation among the analyzed LAB strains. Antimicrobial susceptibility profiles Antimicrobial susceptibility testing using broth microdilution showed that both L. plantarum strains (L25F and L22F) had elevated ampicillin MICs of 16 mg/L and 8 mg/L, respectively, exceeding the EFSA cut-off of 2 mg/L, and tetracycline MICs of 16 mg/L, above the cut-off of 8 mg/L (Table 3). In contrast, L. salivarius AF2319 exhibited an ampicillin MIC within the susceptible range (1 mg/L) but a tetracycline MIC of 16 mg/L, also above the EFSA threshold. Across all three strains, vancomycin MICs (> 32 mg/L) and streptomycin MICs (≤ 1000 mg/L) were consistent with intrinsic, non-transferable vancomycin resistance described in L. plantarum and L. salivarius and other lactobacilli. Table 3Minimum inhibitory concentrations (MICs) of antimicrobial agents against L. plantarum L25F and L22F, and L. salivarius AF2319 relative to EFSA breakpoints.Antibiotics L. salivarius L. plantarum Cut-off (mg/L)AF2319Cut-off (mg/L)L25FL22FAmpicillin412 16 8 Clindamycin44411Gentamicin161616<=1<=1Erythromycin 1 1 1<=0.25<=0.25Kanamycin646464<=16<=16Tetracycline 8 16 321616Streptomycin 1000 ug/ml64<=1000N. R<=1000<=1000Chloramphenicol4<=84<=8<=8VancomycinN. R> 32N. R> 32> 32Cells highlighted in bold: MICs above EFSA cut-off. Cells highlighted in italic: MICs at EFSA cut-off. N.R.: not required by EFSA52. Table 3Minimum inhibitory concentrations (MICs) of antimicrobial agents against L. plantarum L25F and L22F, and L. salivarius AF2319 relative to EFSA breakpoints.Antibiotics L. salivarius L. plantarum Cut-off (mg/L)AF2319Cut-off (mg/L)L25FL22FAmpicillin412 16 8 Clindamycin44411Gentamicin161616<=1<=1Erythromycin 1 1 1<=0.25<=0.25Kanamycin646464<=16<=16Tetracycline 8 16 321616Streptomycin 1000 ug/ml64<=1000N. R<=1000<=1000Chloramphenicol4<=84<=8<=8VancomycinN. R> 32N. R> 32> 32Cells highlighted in bold: MICs above EFSA cut-off. Cells highlighted in italic: MICs at EFSA cut-off. N.R.: not required by EFSA52. Minimum inhibitory concentrations (MICs) of antimicrobial agents against L. plantarum L25F and L22F, and L. salivarius AF2319 relative to EFSA breakpoints. Cells highlighted in bold: MICs above EFSA cut-off. Cells highlighted in italic: MICs at EFSA cut-off. N.R.: not required by EFSA52. Antimicrobial resistance gene profiling Antimicrobial resistance gene analysis using the CARD and ResFinder revealed that acquired AMR genes were predominantly associated with extrachromosomal elements in all three strains, with additional low-identity homologues to vancomycin-associated genes detected on the chromosome (Table 4; Supplementary Table S3). Table 4Antimicrobial resistance genes identified by CARD in L. plantarum L25F and L22F, and L. salivarius AF2319.StrainGeneLocationAMR Gene FamilyDrug ClassResistance Mechanism% amino acid identity to reference ARGL25FvanY gene in vanB clusterChromosomevanY, glycopeptide resistance gene clusterglycopeptide antibioticantibiotic target alteration33.81 lnuA DNA elementlincosamide nucleotidyltransferase (LNU)lincosamide antibioticantibiotic inactivation99.38L22FvanY gene in vanB clusterChromosomevanY, glycopeptide resistance gene clusterglycopeptide antibioticantibiotic target alteration31.33 lnuA DNA elementlincosamide nucleotidyltransferase (LNU)lincosamide antibioticantibiotic inactivation99.38AF2319vanT in vanG clusterChromosomeglycopeptide resistance gene cluster, vanTglycopeptide antibioticantibiotic target alteration35.31 tet(M) DNA elementtetracycline-resistant ribosomal protection proteintetracycline antibioticantibiotic target protection98.75 tet(45) DNA elementmajor facilitator superfamily (MFS) antibiotic efflux pumptetracycline antibioticantibiotic efflux76.04 ErmC DNA elementErm 23 S ribosomal RNA methyltransferasemacrolide antibiotic, lincosamide antibiotic, streptogramin antibiotic, streptogramin A antibiotic, streptogramin B antibioticantibiotic target alteration95.02a Low-identity homology (< 40% identity); reported for completeness but not considered a confirmed ARG for safety assessment.b Intermediate-identity tet(45)-like hit; interpreted as a putative variant and not treated as a confirmed ARG. Table 4Antimicrobial resistance genes identified by CARD in L. plantarum L25F and L22F, and L. salivarius AF2319.StrainGeneLocationAMR Gene FamilyDrug ClassResistance Mechanism% amino acid identity to reference ARGL25FvanY gene in vanB clusterChromosomevanY, glycopeptide resistance gene clusterglycopeptide antibioticantibiotic target alteration33.81 lnuA DNA elementlincosamide nucleotidyltransferase (LNU)lincosamide antibioticantibiotic inactivation99.38L22FvanY gene in vanB clusterChromosomevanY, glycopeptide resistance gene clusterglycopeptide antibioticantibiotic target alteration31.33 lnuA DNA elementlincosamide nucleotidyltransferase (LNU)lincosamide antibioticantibiotic inactivation99.38AF2319vanT in vanG clusterChromosomeglycopeptide resistance gene cluster, vanTglycopeptide antibioticantibiotic target alteration35.31 tet(M) DNA elementtetracycline-resistant ribosomal protection proteintetracycline antibioticantibiotic target protection98.75 tet(45) DNA elementmajor facilitator superfamily (MFS) antibiotic efflux pumptetracycline antibioticantibiotic efflux76.04 ErmC DNA elementErm 23 S ribosomal RNA methyltransferasemacrolide antibiotic, lincosamide antibiotic, streptogramin antibiotic, streptogramin A antibiotic, streptogramin B antibioticantibiotic target alteration95.02a Low-identity homology (< 40% identity); reported for completeness but not considered a confirmed ARG for safety assessment.b Intermediate-identity tet(45)-like hit; interpreted as a putative variant and not treated as a confirmed ARG. Antimicrobial resistance genes identified by CARD in L. plantarum L25F and L22F, and L. salivarius AF2319. a Low-identity homology (< 40% identity); reported for completeness but not considered a confirmed ARG for safety assessment. b Intermediate-identity tet(45)-like hit; interpreted as a putative variant and not treated as a confirmed ARG. For L. plantarum L25F, CARD detected a low-identity hit to a vanY-like peptidase on the chromosome (33.81% amino acid identity to the vanB cluster reference), which likely represents a housekeeping peptidoglycan-related enzyme rather than an acquired vancomycin resistance gene. In line with current EFSA-aligned thresholds, this partial and low-identity match is not interpreted as a confirmed ARG and is reported only as a distant homologue for completeness (Table 4, footnote a). In contrast, a lnu(A) gene belonging to the lincosamide nucleotidyltransferase (LNU) family was detected on extrachromosomal DNA (contig 13) with high sequence identity (99.38% via CARD; 99.79% via ResFinder), consistent with a confirmed acquired resistance determinant. L. plantarum L22F harbored an analogous profile: a low-identity vanY-like hit on the chromosome (31.33% identity; not considered a confirmed ARG) and a highly conserved lnu(A) gene on extrachromosomal DNA (contig 17) with identical high identity values to those observed in L25F. In L. salivarius AF2319, CARD detected a vanT-like hit associated with the vanG cluster on the chromosome (35.31% identity), which we likewise report as a low-identity homologue rather than a confirmed vancomycin resistance gene. Three acquired AMR genes were located on extrachromosomal DNA, including erm(C) on contig 4 and two tetracycline resistance genes on contig 3: tet(M) and tet(L). ResFinder identified these genes with high sequence identity: erm(C) (99.45%), tet(M) (99.48%), and tet(L) (99.49%). CARD RGI provided concordant identification for tet(M) (98.75%) and erm(C) (95.02%), but annotated the tet(L) locus as a low-identity tet(45)-like variant (76.04%); we adopt the ResFinder annotation (tet(L)) based on superior sequence identity and database specificity. All three genes met our strict criteria for confirmed acquired ARGs based on ≥ 95% identity matches in ResFinder. The genetic organization of representative contigs carrying lnu(A), tet(M), tet(L), and erm(C), together with neighboring mobile element–associated genes, is shown in Supplementary Figure S3. Virulence factor screening In silico screening for virulence-associated genes using the VFDB (Table 5) did not identify any genes classified as classical virulence determinants (gelE, esp, hyl) or any gene associated with toxin production or cytolysin activity in any of the three strains (Table 5). Nevertheless, L. salivarius AF2319 harbored genes encoding elongation factor Tu (tuf), ATP-dependent Clp protease (clpB), and cps4I. L. plantarum L25F harbored genes encoding cell surface adhesins: clfA, sdrE, sdrC, and sdrD, and L. plantarum L22F harbored genes encoding cell surface adhesins: clfA, sdrE, sdrC, and sdrD. These genes were assigned to the VFG IDs shown in Table 5. Table 5Virulence-associated genes identified by VFDB in L. plantarum L25F and L22F, and L. salivarius AF2319.StrainGene(s) DetectedVFG IDsFunctionSafety AssessmentAF2319tuf, clpB, cps4IVFG046465, VFG000079, VFG001373Housekeeping, stress response, EPSNon-pathogenic; typical LAB profileL25FClfA, sdrE, sdrC, sdrDVFG001289, VFG001281, VFG001279, VFG001280Cell surface adhesinsProbiotic adhesion genes; non-pathogenicL22FClfA, sdrE, sdrC, sdrDVFG001289, VFG001281, VFG001279, VFG001280Cell surface adhesinsProbiotic adhesion genes; non-pathogenic Table 5Virulence-associated genes identified by VFDB in L. plantarum L25F and L22F, and L. salivarius AF2319.StrainGene(s) DetectedVFG IDsFunctionSafety AssessmentAF2319tuf, clpB, cps4IVFG046465, VFG000079, VFG001373Housekeeping, stress response, EPSNon-pathogenic; typical LAB profileL25FClfA, sdrE, sdrC, sdrDVFG001289, VFG001281, VFG001279, VFG001280Cell surface adhesinsProbiotic adhesion genes; non-pathogenicL22FClfA, sdrE, sdrC, sdrDVFG001289, VFG001281, VFG001279, VFG001280Cell surface adhesinsProbiotic adhesion genes; non-pathogenic Virulence-associated genes identified by VFDB in L. plantarum L25F and L22F, and L. salivarius AF2319. Hemolytic activity All three LAB strains displayed α-hemolysis on sheep blood agar plates, characterized by greenish discoloration of the agar surrounding bacterial colonies. No evidence of β-hemolysis (complete lysis and clearing of erythrocytes) was observed for any strain tested (Fig. 3). Fig. 3Hemolytic activity of L. plantarum L25F and L22F (A and B) and L. salivarius AF2319 (C) on sheep blood agar. Hemolytic activity of L. plantarum L25F and L22F (A and B) and L. salivarius AF2319 (C) on sheep blood agar. Bacteriocin-encoding gene identification Bacteriocin-encoding genes were identified using BAGEL4, revealing diverse antimicrobial production potential across the three LAB strains (Supplementary Table S3). L. salivarius AF2319 exhibited the most extensive bacteriocin repertoire, with three distinct bacteriocin loci distributed across multiple genetic elements. On chromosomal contig 1, BAGEL4 predicted an Enterolysin A-like bacteriocin cluster, and the core protein showed 99.69% amino acid identity to the reference Enterolysin A sequence, including a signal peptide domain-containing region. On plasmid contig 2, a Salivaricin P (ABP-118 family) cluster was detected, with the predicted core peptide displaying 100% amino acid identity to the reference Salivaricin P sequence in the BAGEL4 database. Additionally, on chromosomal contig 6, a Sakacin G (skgA1) cluster belonging to the leucocin A/sakacin P class II family was identified, and the core peptide showed 100% amino acid identity to the corresponding reference bacteriocin sequence. In contrast, both L. plantarum strains exhibited more limited but highly similar bacteriocin profiles centered on plantaricin production. For L25F, BAGEL4 identified a plantaricin J bacteriocin cluster on chromosomal contig 1, with the predicted two-peptide core component (Plantaricin JK subunit PlnJ) showing 100% amino acid identity to the reference PlnJ sequence. Similarly, L22F harbored a plantaricin J operon on chromosomal contig 1 with a more complex genetic organization, encoding multiple core peptides (PlnJ, PlnK, PlnE/F), each displaying 100% amino acid identity to their respective reference sequences; this more complete operon structure distinguished L22F from L25F, which carried only the PlnJ subunit. Functional genomic features related to probiotic properties Comprehensive genomic analysis identified 46–47 genes associated with probiotic-relevant functions across the three strains: L. salivarius AF2319 harbored 46 genes, while L. plantarum strains both contained 47 genes (Table 6). These genes encoded functions spanning gastrointestinal stress adaptation, intestinal colonization, and host-microbe interactions. Table 6Probiotic-related genes identified in L. plantarum L25F and L22F, and L. salivarius AF2319.Functional CategoryAF2319L25FL22FHeat stress tolerancerelA, ctsR, hrcA, clpC, clpE, clpL, clpP, dnaK, dnaJ, grpEdnaK, dnaJ, grpE, hrcA, ctsR, clpC, clpE, clpPdnaK, dnaJ, grpE, hrcA, ctsR, clpC, clpE, clpP, htrACold stress tolerance cspC cspC cspC Acid stress toleranceatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHBile tolerance cfa cfa cfa Adhesion fbpA fbpA fbpA Oxidative stress/antioxidantmsrA, msrB, trxA, trxB, nprtrxA, trxB, msrA, msrB, nprtrxA, trxB, msrA, msrB, nprImmunomodulationdltB, dltDdltA, dltB, dltC, dltDdltA, dltB, dltC, dltDOsmotic shock tolerance proV opuCB, opuCC, opuCDopuCB, opuCC, opuCDCell wall formationmurA, murB, murC, murD, murE, murF, murG, mraY, dltB, dltD, tagA, tagO, ltaS, pbpX, srtAmurA, murB, murC, murD, murE, murF, murG, mraY, pbpX, ltaS, srtA, tagA, tagOmurA, murB, murC, murD, murE, murF, murG, mraY, pbpX, ltaS, srtA, tagA, tagOBiofilm formation luxS luxS luxS GIT survival/stress relA relA relA Table 6Probiotic-related genes identified in L. plantarum L25F and L22F, and L. salivarius AF2319.Functional CategoryAF2319L25FL22FHeat stress tolerancerelA, ctsR, hrcA, clpC, clpE, clpL, clpP, dnaK, dnaJ, grpEdnaK, dnaJ, grpE, hrcA, ctsR, clpC, clpE, clpPdnaK, dnaJ, grpE, hrcA, ctsR, clpC, clpE, clpP, htrACold stress tolerance cspC cspC cspC Acid stress toleranceatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHatpA, atpB, atpC, atpD, atpE, atpF, atpG, atpHBile tolerance cfa cfa cfa Adhesion fbpA fbpA fbpA Oxidative stress/antioxidantmsrA, msrB, trxA, trxB, nprtrxA, trxB, msrA, msrB, nprtrxA, trxB, msrA, msrB, nprImmunomodulationdltB, dltDdltA, dltB, dltC, dltDdltA, dltB, dltC, dltDOsmotic shock tolerance proV opuCB, opuCC, opuCDopuCB, opuCC, opuCDCell wall formationmurA, murB, murC, murD, murE, murF, murG, mraY, dltB, dltD, tagA, tagO, ltaS, pbpX, srtAmurA, murB, murC, murD, murE, murF, murG, mraY, pbpX, ltaS, srtA, tagA, tagOmurA, murB, murC, murD, murE, murF, murG, mraY, pbpX, ltaS, srtA, tagA, tagOBiofilm formation luxS luxS luxS GIT survival/stress relA relA relA Probiotic-related genes identified in L. plantarum L25F and L22F, and L. salivarius AF2319. All three strains harbored comprehensive stress response gene repertoires encompassing heat, cold, acid, bile, and osmotic stress adaptation. For heat stress tolerance, a core set of genes was conserved across all strains, including molecular chaperones (dnaK, dnaJ, grpE), transcriptional regulators (hrcA, ctsR), and ATP-dependent proteases (clpB, clpC, clpE, clpP). Beyond this shared foundation, both L. plantarum strains additionally harbored supplementary heat stress genes (lwcA, hrcA) absent from AF2319, while cold stress adaptation was uniformly supported by the cspC gene across all strains. Acid stress tolerance was mediated by the F₀F₁-ATP synthase operon (atpA–atpH) which was conserved across all strains, as was bile tolerance encoded by the cfa gene. In contrast, osmotic stress adaptation mechanisms differed between species: AF2319 harbored the proV gene, whereas both L. plantarum strains harbored opuCB, opuCC, and opuCD genes. Protection against oxidative stress was mediated by genes encoding thioredoxin reductases (trxA, trxB), methionine sulfoxide reductases (msrA, msrB), and NADPH peroxidase (npr), all conserved across the three strains. Concurrently, cell wall biosynthesis genes varied in abundance, with AF2319 harboring 13 genes while the L. plantarum strains each possessed 17 genes. The core peptidoglycan biosynthesis pathway encoded by the mur gene family (murA–murF) was conserved across all strains, along with genes for cell wall modification and assembly (mraY, murG, tagO, ltaS, pbpX, tagA, and srtA). Genes supporting intestinal adherence and biofilm formation were universally present across all strains. The fibronectin-binding protein gene (fbpA) and the luxS gene, associated with quorum sensing and biofilm formation, were conserved across all strains. Additionally, immunomodulatory genes included the dlt operon (dltA, dltB, dltC, dltD), and the stringent response regulator gene (relA), were detected in all strains. Metabolic pathway analysis Functional annotation using the KEGG and RAST databases revealed distinct differences in metabolic gene distributions among the three strains (Supplementary Tables S6-S7). The two L. plantarum strains exhibited nearly identical profiles but contained higher gene counts than L. salivarius AF2319 in genetic information processing, cellular signaling, and carbohydrate metabolism (KEGG: 181–185 vs. 101; RAST: 188 vs. 105). Similar patterns were observed in amino acid metabolism (KEGG: 93–94 vs. 45; RAST: 159–160 vs. 87), cofactor/vitamin metabolism, and stress response. All strains encoded comparable numbers of genes associated with protein metabolism and core biosynthetic pathways. COG, KO, and GO analyses supported these patterns (Fig. 4; Supplementary Figures S4–S5). COG annotation identified 19 functional categories across all strains. L. plantarum strains showed higher proportions in COG-K (Transcription; 10.00% and 9.89% vs. 7.70%), COG-G (Carbohydrate transport and metabolism; 8.11–8.32% vs. 6.42%), and COG-E (Amino acid metabolism; 8.11–8.43% vs. 7.41%), whereas AF2319 exhibited higher representation in COG-L (Replication and repair; 11.26% vs. 9.20–8.54%). All strains possessed substantial COG-S (Function unknown; 18.90–20.03%). Fig. 4COG functional category distribution in L. plantarum L25F and L22F, and L. salivarius AF2319. “Functional categories are based on eggNOG-mapper COG classifications: C = Energy production and conversion; D = Cell cycle control and mitosis; E = Amino acid metabolism and transport; F = Nucleotide metabolism and transport; G = Carbohydrate metabolism and transport; H = Coenzyme metabolism; I = Lipid metabolism; J = Translation and ribosomal structure; K = Transcription; L = Replication and repair; M = Cell wall/membrane/envelope biogenesis; N = Cell motility; O = Post-translational modification, protein turnover, chaperones; P = Inorganic ion transport and metabolism; Q = Secondary metabolite biosynthesis; T = Signal transduction; U = Intracellular trafficking and secretion; V = Defense mechanisms; S = Function unknown. Categories with < 1% representation are grouped as ‘Other.’ COG functional category distribution in L. plantarum L25F and L22F, and L. salivarius AF2319. “Functional categories are based on eggNOG-mapper COG classifications: C = Energy production and conversion; D = Cell cycle control and mitosis; E = Amino acid metabolism and transport; F = Nucleotide metabolism and transport; G = Carbohydrate metabolism and transport; H = Coenzyme metabolism; I = Lipid metabolism; J = Translation and ribosomal structure; K = Transcription; L = Replication and repair; M = Cell wall/membrane/envelope biogenesis; N = Cell motility; O = Post-translational modification, protein turnover, chaperones; P = Inorganic ion transport and metabolism; Q = Secondary metabolite biosynthesis; T = Signal transduction; U = Intracellular trafficking and secretion; V = Defense mechanisms; S = Function unknown. Categories with < 1% representation are grouped as ‘Other.’ KEGG ortholog analysis identified 1,026–1,294 unique KOs. L. plantarum strains contained 26% more total KOs than AF2319. The transposase gene K07482 occurred 19 and 17 times in L22F and L25F, respectively, but only 10 times in AF2319. L. plantarum strains showed greater counts of genes encoding MFS transporters K15580 (13–14 vs. 0 copies in AF2319), peptide ABC transporters K02035 (12–13 vs. 0), and PTS systems K02755/K02761 (15–17 vs. 4 total copies). AF2319 uniquely possessed Type I restriction systems (K01154), recombination proteins (K03294), and chemotaxis genes (K01104). Core metabolic enzymes, including phosphoglycerate mutase (K01990; 10, 16, 16 copies) and enolase (K01992; 7, 11, 11 copies), were consistently detected in all strains. Gene Ontology (GO) analysis assigned 15,525–17,452 total GO terms, comprising 1,702–1,856 unique terms (Supplementary Table S4). Major categories included metabolic process (GO:0008152; 252–288 proteins), carbohydrate metabolism (GO:0071704; 235–267 proteins), catalytic activity (GO:0003824; 207–253 proteins), cellular organization (GO:0005623, GO:0044464, GO:0005737; 176–272 proteins each), and biosynthetic process (GO:0009058, GO:0044249; 135–161 proteins). AF2319 contained the highest average GO terms per protein (9.1 vs. 5.7–6.0 for L. plantarum). CRISPR-cas system analysis Comprehensive CRISPR-Cas analysis revealed complete Type I-E systems in both L. plantarum strains, while L. salivarius AF2319 contained only orphan CRISPR arrays lacking cas genes (Supplementary Table S5). Both L. plantarum L25F and L22F possessed identical Type I-E CRISPR-Cas modules (~ 24.5 kbp) comprising eight genes, cas6, cas5, cas7, cse2, cse1, cas3, cas2, and cas1, arranged in canonical order. Multiple CRISPR spacer arrays were detected across contigs 1_1–1_4, including two high-confidence (Evidence Level 1) repeat sequences (GGGGCCTTGGCTAGTGCGTCAGTCAGTCAGC; GGATCACCCCGGCATACATGGGAACAC) and two lower-confidence (Evidence Level 4) arrays spanning 7–14 bp. In contrast, L. salivarius AF2319 harbored two short CRISPR spacer regions (TTTGCTAATTGATTACACCACCTTGTAATGCTGAA; CCGGTTGGCGAGGCAACGGAAAAAC) but no adjacent cas genes, indicating the absence of a functional CRISPR-Cas system. Prophage identification Prophage prediction using PHASTER revealed distinct patterns in prophage distribution and completeness across the three strains (Supplementary Table S8). L. salivarius AF2319 harbored six prophage-like regions spanning 7.9–19.2 kb, with no intact prophages detected. Five regions were classified as incomplete, while one region (13.7 kb) was designated as questionable. The predicted prophages displayed associations with diverse bacterial taxa, including Lactobacillus, Clostridium, Clostridioides, and Bacillus species. In contrast, both L. plantarum strains harbored two intact prophage regions each, spanning 44.9–47.3 kb and 47.3–51.3 kb in length, respectively. These intact prophage regions demonstrated high sequence conservation, with the most prevalent prophages identified as PHAGE_Lactob_phijl1 and PHAGE_Lactob_Sha1, both detected consistently in L25F and L22F. Additionally, both L. plantarum strains harbored incomplete prophage sequences (42.8–47.3 kb) sharing homology with phages infecting Lactobacillus and Clostridioides species. Insertion sequence distribution Comprehensive insertion sequence (IS) profiling revealed marked inter-strain variation in IS diversity and distribution (Supplementary Table S9). L. plantarum L25F exhibited the most extensive IS repertoire, spanning at least six IS families across multiple contigs. These included IS1182 (ISP2), ISL3 (ISP1), and IS200/IS605 family elements (ISDha13, ISStin10, ISSsu4) on contig 1, along with IS3, IS30, and IS256 elements (ISLpI2, ISLpI1, ISPp1, IS153, ISBame1, IS1310) distributed on contig 2 and others, with closest homology to Enterococcus hirae, Leuconostoc mesenteroides, Lactiplantibacillus sakei, and Lactiplantibacillus sanfranciscensis. In comparison, L. plantarum L22F displayed a reduced IS complement, retaining five IS types, IS1182, ISL3, IS200/IS605, IS3, and IS30, primarily located on contigs 1 and 2. In contrast, L. salivarius AF2319 exhibited a sparse IS landscape dominated by IS30 family elements, including ISLpI1, ISPp1, ISCbt13, IS6770, and ISLhe30. Across all strains, IS30 family members were most prevalent, showing greatest expansion in L25F, where they were distributed across at least six contigs. Overall, L25F demonstrated the highest IS diversity, followed by L22F and AF2319, reflecting lineage-specific genome plasticity. Genomic islands and horizontal gene transfer Bioinformatic prediction of genomic islands (GIs) using IslandViewer revealed distinct differences in the quantity and distribution of predicted horizontally acquired regions among the three strains (Supplementary Table S10). L. salivarius AF2319 contained 10 genomic islands totaling 163 kb (8.2% of the genome), ranging from 4.9 kb to 37.7 kb, all identified by IslandPath-DIMOB only. In contrast, L. plantarum L25F harbored 18 genomic islands with a combined size of 404 kb (13.5% of the genome), including four regions larger than 30 kb and five supported by multiple prediction methods (IslandPath-DIMOB and SIGI-HMM). L. plantarum L22F exhibited 21 genomic islands totaling 427 kb (14.2% of the genome), with four islands exceeding 30 kb and eight predicted by multiple methods (38% consensus). Overall, both L. plantarum strains possessed 1.8–2.1 times more genomic islands than AF2319. Complementary analysis using OriTfinder identified conjugation-associated features across all three strains (Supplementary Table S11). All isolates carried detectable oriT regions and one relaxase gene. The number of type IV coupling protein (T4CP) genes varied, with L. salivarius AF2319 harboring five, and L. plantarum L25F and L22F harboring three each. Conjugation capability was predicted in all strains based on the presence of essential mobilization components. Importantly, several of these mobilisation-related genes (oriT, relaxase or T4CP) occurred on the same MOB-suite–predicted plasmid-derived contigs that carried tet(M), tet(L), tet(45)-like or erm(C), indicating that these ARGs are embedded within broader mobile genetic element backbones rather than residing on isolated insertion sequences alone (Supplementary Figure S4 and Supplementary Table S11). Discussion Whole-genome sequencing (WGS) of three Thai LAB strains that are being assessed as potential probiotics for use in animal production revealed critical genomic features of relevance to safety that were otherwise undetectable through conventional phenotypic testing. While antimicrobial susceptibility and hemolytic assays provided initial safety indicators consistent with traditional Generally Recognized As Safe (GRAS) and Qualified Presumption of Safety (QPS) criteria, WGS uncovered plasmid-borne AMR genes and mobile genetic elements warranting careful regulatory consideration. These findings underscore the indispensable role of genome-based analysis in modern probiotic safety assessment. An important observation in this study was genotype–phenotype discordance for specific antimicrobial resistance determinants, which illustrates the limitations of relying solely on MIC-based testing in probiotic safety assessment. Most notably, L. salivarius AF2319 harbored erm(C) (99.45% identity) yet exhibited an erythromycin MIC of 1 mg/L, precisely at the EFSA breakpoint (≤ 1 mg/L for L. salivarius susceptibility), a borderline result that technically qualifies as “susceptible” but indicates the gene may be present in a silent or incompletely expressed state, potentially due to regulatory control via the erm(C) leader peptide attenuation mechanism or unfavorable test conditions. L. plantarum strains displayed elevated ampicillin MICs (8–16 mg/L), exceeding the EFSA threshold of 2 mg/L, yet WGS detected no corresponding acquired β-lactam resistance genes. This phenotype–genotype pattern, which has been frequently reported in L.plantarum, is consistent with intrinsic β-lactam tolerance mediated by penicillin-binding protein (PBP) modifications rather than acquisition of classical β-lactam resistance genes49,50. However, we did not systematically investigate mutations in penicillin-binding proteins (PBPs) or their regulatory regions, so the precise genetic basis of the elevated ampicillin MICs in these strains remains unresolved. We therefore interpret the ampicillin phenotype as consistent with previously described strain- or lineage-associated intrinsic tolerance patterns in L. plantarum, rather than as evidence for a defined species-specific acquired resistance mechanism51. EFSA categorizes such chromosomally encoded intrinsic resistance as posing minimal horizontal transfer risk52. Moreover, the high vancomycin MICs observed for the L. plantarum and L. salivarius strains in this study are in line with previous reports describing intrinsic, non-transferable vancomycin resistance in many lactobacilli, including these species. Such intrinsic vancomycin resistance is generally attributed to species-specific features of the peptidoglycan biosynthesis pathway, including alterations in the D-alanine–D-alanine ligase (Ddl) and related enzymes, rather than to acquisition of van gene clusters, and is therefore not considered a disqualifying trait under current QPS and EFSA safety frameworks53. Thus, genomic evidence effectively reclassifies apparent phenotypic safety concerns into species- or lineage-associated intrinsic traits while simultaneously revealing cryptic acquired resistance genes that would remain undetected by phenotypic testing alone. Both L. plantarum strains carried lnu(A) (99.79% identity), a lincosamide nucleotidyltransferase that confers the L-phenotype: resistance to lincomycin with retained susceptibility to clindamycin due to differential substrate specificity54. The strains exhibited clindamycin MICs of 1 mg/L (well below the 4 mg/L EFSA cutoff for susceptibility), which is concordant with the known substrate preference of LnuA, this enzyme efficiently inactivates lincomycin through 3-O-nucleotidylation but exhibits minimal activity against clindamycin due to the 7-chloro-7-deoxy substitution that introduces steric hindrance in the active site55. Although located on extrachromosomal DNA (contig_17 in L22F, contig_13 in L25F) and not clearly associated with complete conjugative plasmids, its presence is still classified as acquired under EFSA guidelines. However, lincomycin MICs were not determined in the present study, precluding confirmation of the expected lincomycin-resistant phenotype associated with the L-phenotype. This represents a limitation in our phenotypic characterization, as comprehensive assessment of lincosamide resistance ideally requires testing both lincomycin and clindamycin to distinguish L-phenotype (lnu-mediated) from MLS phenotype (erm-mediated)56. Nevertheless, because lincosamides are considered clinically relevant antimicrobials, EFSA’s current rule that any acquired ARG affecting medically important antibiotics disqualifies a strain from QPS status applies irrespective of the clindamycin-susceptible phenotype or the absence of lincomycin testing57. This interpretation is consistent with EFSA-approved L. plantarum strains (e.g., DSM 34271), which have been reported to lack acquired lincosamide resistance genes58. Tetracycline resistance genes (tet(M) and tet(L)) were detected in L. salivarius AF2319, located on extrachromosomal DNA elements as confirmed by MOBsuite analysis. Tetracycline resistance in L. salivarius is strain-dependent rather than species-intrinsic, as evidenced by the well-characterized probiotic strain UCC118, which lacks tetracycline resistance genes entirely and demonstrates susceptibility to tetracycline59. In contrast, L. salivarius JCM1046 harbors tet(M) on the conjugative plasmid pCTN1046, indicating horizontal acquisition60. The presence of tet genes on extrachromosomal elements in AF2319, combined with the absence of flanking IS elements (ISfinder, E-value ≤1e-10), is consistent with the stable maintenance of previously acquired resistance determinants rather than with obvious signatures of very recent horizontal transfer; however, this remains an inference in the absence of targeted phylogenetic or experimental mobility assays. Under EFSA QPS guidelines, acquired antimicrobial resistance genes on mobile elements require case-by-case evaluation for potential transferability, particularly when located on conjugative plasmids or transposons. Metagenomic surveys of human and livestock gut microbiomes have consistently reported tetracycline resistance genes such as tet(M) among the most prevalent ARGs, underscoring their widespread dissemination in intestinal resistomes. However, linking the specific plasmid or contig backgrounds identified in our strains to concrete in vivo transmission events would require targeted comparative and metagenomic analyses, which were not undertaken in this study61. The absence of flanking mobilization machinery in AF2319 may reduce, but does not eliminate horizontal transfer risk, warranting continued genomic surveillance in probiotic safety assessments. Most notably, L. salivarius AF2319 harbored erm(C) (99.45% identity) on extrachromosomal DNA (contig_4, position 601–1332) yet exhibited an erythromycin MIC of 1 mg/L, precisely at the EFSA breakpoint (≤ 1 mg/L for L. salivarius susceptibility), representing a borderline genotype–phenotype relationship where the gene is genomically present but may be phenotypically silent or incompletely expressed. Unlike the tet genes, erm(C) was flanked by ISLhe30 (IS30 family, E-value ≤1e-10), a configuration that is consistent with an acquired origin and historical mobility of erm(C), as similar IS-flanked arrangements have been reported to mediate plasmid–chromosome transposition in L. salivarius and related LAB62. Despite the absence of conjugation genes or AMR-carrying mobilizable plasmids, which might otherwise suggest low transfer potential, EFSA remains unequivocal: the presence of acquired erm(C), conferring resistance to erythromycin, a WHO “highly important” macrolide, constitutes grounds for QPS disqualification, regardless of genomic context, borderline MIC, or apparent risk of transferability. EFSA and WHO prioritization frameworks treat tetracycline and macrolide resistance differently in the context of probiotic and feed additive strains63. In practice, the presence of an acquired erm(C) gene, which confers resistance to the clinically important macrolide erythromycin, is sufficient to exclude a strain from QPS status, whereas acquired tet genes may be evaluated more flexibly depending on their context. In AF2319, this means that detection of erm(C), together with the acquired lnu(A) genes in the L. plantarum strains, is enough to disqualify these otherwise promising candidates even though the ARGs are located on non-conjugative, plasmid-like contigs with apparently limited mobility. We did not, however, perform a detailed analysis of the erm(C) coding sequence, leader peptide region or regulatory elements, so we cannot determine whether the borderline erythromycin MIC reflects mutational inactivation, regulatory control, or limitations of the phenotypic test conditions. These interpretations of historical mobility based on insertion sequence context are necessarily indirect, and definitive assessment of transfer potential would require targeted phylogenetic analyses and experimental mobility assays. Beyond regulatory implications, the genomic context analysis provided further insight into the evolutionary dynamics underlying these resistance determinants. Comparative examination of genomic islands revealed strain-level differences in putative AMR acquisition patterns. The L. plantarum strains exhibited higher genomic island content (18–21 islands; 13.5–14.2% of the genome) than AF2319 (10 islands; 8.2%), consistent with a greater genome plasticity and DNA uptake capacity. Nearly one-third of L. plantarum islands were detected by both IslandPath-DIMOB and SIGI-HMM, a pattern that is compatible with segments that may have been acquired through horizontal gene transfer, although this inference remains indirect in the absence of phylogenetic or experimental validation45. AMR gene mapping revealed that lnu(A) in L. plantarum was embedded within large genomic islands (92–108 kb), a genomic context that is suggestive of horizontal acquisition. In AF2319, erm(C) was island-associated (37.65 kb), whereas tet(M) and tet(45) were not, which is compatible with older integrations or alternative acquisition mechanisms64,65. Despite its lower genomic plasticity, AF2319 bore the heaviest AMR burden, possibly reflecting selective retention of specific resistance traits during adaptation to host-associated niches, whereas the L. plantarum strains maintained a more dynamic genome flux typical of isolates from environmental or fermented-food sources66. Prophages are ubiquitous mobile genetic elements in lactic acid bacteria and important drivers of horizontal gene transfer (HGT) alongside conjugative plasmids and transposons67. Comprehensive genome-wide analyses have demonstrated that prophage fragments are detectable in nearly all (99.8%) Lactobacillus genomes, with intact prophages identified in 64.1% of surveyed strains across 16 species, and L. plantarum exhibiting the highest intact prophage occurrence rate (98.5%) among them68. Temperate phages mediate gene transfer through specialized and generalized transduction, both experimentally confirmed in lactobacilli through spontaneous plasmid transfer in L. gasseri and prophage-recombinase-driven chromosomal recombination in L. plantarum WCFS169. Importantly, prophages in L. plantarum have been shown to carry AMR-associated cargo including antibiotic efflux genes (efrA, lmrD), with prophage-borne lmrD linked to clindamycin resistance in multiple strains, demonstrating that prophages can function as AMR vectors in this species70,71. Similarly, L. salivarius prophages encode niche-adaptive cargo such as UDP-sugar biosynthesis and cysteine protease genes associated with host colonization, supporting prophage-mediated strain specialization72. Beyond gene transfer, prophages provide fitness benefits through superinfection immunity and lysogenic conversion, with prophage induction triggered by intestinal stressors including antibiotics, oxidative stress, and transit conditions relevant to probiotic use73. Collectively, these findings establish that prophages in L. plantarum and L. salivarius are active contributors to genome evolution and potential vectors for AMR dissemination. Prophage characterization using PHASTER revealed striking interspecies differences in prophage integrity. Both L. plantarum strains harbored two intact prophages: L22F (44.9 kb and 42.8 kb) and L25F (55.1 kb and 47.3 kb), with closest similarity to temperate Lactobacillus phages phi1 and Sha1 previously characterized in L. plantarum74. In contrast, L. salivarius AF2319 lacked intact prophages entirely, carrying only six degraded remnants (7.9–19.2 kb each; total 80.2 kb) distributed across two contigs and showing fragmented similarity to Lactobacillus phage jlb1 and Clostridium phages phiCD211 and CDKM975. These incomplete regions lacked canonical prophage modules including intact lysis cassettes, complete structural genes, and functional integrases, indicating extensive decay through deletion or mutation. The maintenance of intact prophages in L. plantarum versus their degradation in AF2319 likely reflects divergent evolutionary trajectories: environmental/food-associated L. plantarum strains retain prophage-mediated gene exchange capacity, whereas the host-adapted AF2319 lineage has undergone genome streamlining with prophage loss, consistent with genome reduction in specialized lactobacilli75. Given the documented capacity of prophages to mediate horizontal gene transfer in L. plantarum and L. salivarius, we systematically examined whether the antimicrobial resistance determinants identified were co-localized with prophage regions. Genomic mapping revealed clear spatial separation: in both L. plantarum strains, lnu(A) was located on small extrachromosomal contigs (L22F: contig_17, position 1065–1550; L25F: contig_13, position 1065–1550) distinct from the main chromosome harboring intact prophages (contig_1), while In AF2319, erm(C) (contig_4, position 601–1332) and the tet genes (tet(M) at contig_3 positions 2567–4492 and tet(L) at positions 4686–6062) were localized to extrachromosomal contigs separate from the six degraded prophage remnants on contigs 1 and 2. These findings indicate that in these strains, plasmid- and IS-element-mediated horizontal transfer, rather than prophage transduction, has been the predominant mechanism underlying AMR gene acquisition. However, the presence of intact prophages in the L. plantarum strains indicates ongoing potential for prophage-mediated gene exchange, which could theoretically mobilize antimicrobial resistance determinants in future evolutionary scenarios or through co-infection with AMR-carrying phages, a risk that would require experimental transduction assays or longitudinal metagenomic surveillance to fully assess76. Both L. plantarum strains encoded complete Type I-E CRISPR-Cas systems (24.5 kbp; eight cas genes) yet simultaneously harbored intact prophages, a coexistence pattern paradoxical given that CRISPR-Cas evolved to defend against phage infection77. This may reflect prophage integration predating CRISPR acquisition, anti-CRISPR protein expression by the prophages, or fitness benefits (lysogenic conversion, superinfection exclusion) that outweigh immune pressure78. In contrast, L. salivarius AF2319 lacks functional CRISPR-Cas (retaining only orphan arrays) and harbors only degraded prophage remnants, suggesting genome streamlining in a host-adapted lineage where neither phage defense nor prophage-mediated gene acquisition provided fitness advantages75. These contrasting profiles highlight divergent evolutionary strategies for balancing genomic defense with horizontal gene exchange potential, with implications for AMR dissemination risk in probiotic safety assessment. Although AMR profiling identified potential safety risks associated with the strains, virulence factor analysis confirmed the absence of pathogenic determinants, supporting their suitability as potential probiotics. None of the strains carried classical virulence genes (e.g., gelE, esp, hyl), which are considered exclusion criteria by EFSA57. Genes flagged during VFDB screening (tuf, clpB, cps4I, and adhesins) encode housekeeping or beneficial functions79,80. Surface adhesins facilitate mucosal adhesion and pathogen exclusion as key probiotic mechanisms81,82. All strains exhibited α-hemolysis, a nonpathogenic phenotype typical of GRAS-status LAB6,83. Thus, despite varying AMR profiles, their overall virulence landscape aligns with satisfactory probiotic safety standards84. In addition to safety profiling, WGS provided insights into probiotic functional specialization. L. salivarius AF2319 harbored multiple bacteriocin systems, including Salivaricin P, supporting strong antimicrobial activity. On the other hand, L. plantarum strains encoded Plantaricin J, with L22F containing a complete operon and L25F only the structural gene, explaining differences in bacteriocin potency85. Despite fewer bacteriocins, L. plantarum exhibited greater metabolic versatility than L. salivarius, encoding ~ 1.8× more carbohydrate metabolism genes and ~ 2.1× more amino acid metabolism genes, along with higher counts of stress response and environmental signaling genes86,87. Functional and comparative genomic analyses revealed distinct metabolic strategies among the three LAB strains. As mentioned above, L. plantarum L22F and L25F exhibited broader carbohydrate and amino acid metabolism, consistent with their known ecological versatility in plant- and food-associated environments1,88. The enrichment of transcription-related genes (COG-K) suggests enhanced regulatory capacity for metabolic adaptation, a hallmark of L. plantarum strains reported to thrive under nutrient-variable and acidic conditions87. In contrast, L. salivarius AF2319 showed higher representation of replication and repair functions, indicating stronger genomic maintenance mechanisms that may support stability in mucosal or host-associated niches89. Although AF2319 possessed a smaller genome, its higher annotation density per protein implies compact but functionally specialized coding potential. KEGG profiling highlighted that both L. plantarum strains encode extensive transporter systems, including MFS, ABC, and PTS families, which underpin efficient carbohydrate uptake and energy generation in diverse substrates90. Such transporter enrichment aligns with reports linking the ecological success of L. plantarum to its metabolic plasticity and nutrient-scavenging capacity91. The higher abundance of transposase genes in these strains further suggests elevated genomic plasticity, potentially facilitating rapid adaptation through horizontal gene exchange. Conversely, AF2319 harbored unique restriction-modification, recombination, and chemotaxis systems, implying a defensive genomic strategy and adaptation to mucosal environments where competitive and immune pressures are high66. Together, these findings illustrate two contrasting ecological paradigms: L. plantarum strains exhibit a metabolically expansive and regulatory-flexible lifestyle suited for nutrient-rich and variable environments, while L. salivarius AF2319 displays a streamlined genome emphasizing stability, defense, and mucosal specialization. Further comparative genomic analysis highlighted several rare and strain-specific genetic features underpinning probiotic resilience. All strains encoded clpE, a seldom-reported Clp ATPase implicated in heat shock regulation via the CtsR system92. Uniquely, the L. plantarum strains possessed hrcA, completing the full CtsR-regulated heat shock regulon93, suggesting an enhanced capacity for thermal stress control. In contrast, L. salivarius AF2319 uniquely carried relA, encoding a (p)ppGpp synthetase responsible for a stringent response under nutrient limitation94, implying superior adaptation to nutrient scarcity. Moreover, L. plantarum L22F harbored litA, a rarely annotated cold-shock gene associated with temperature fluctuation tolerance95, while exhibiting a complete opuCB–opuCD osmo-protectant system, distinct from proV-dependent osmotic regulation found in AF231996. Together, these findings reveal complementary stress adaptation strategies: L. plantarum lineages reinforce thermal and osmotic tolerance through expanded regulons, whereas L. salivarius AF2319 emphasizes metabolic endurance under nutritional and environmental constraints. Although our findings align with EFSA’s conservative position that the presence of acquired AMR genes affecting medically important antibiotics is sufficient to disqualify a strain from QPS status, there is ongoing debate regarding the clinical relevance of silent or unexpressed resistance genes in non-pathogenic LAB. Several studies have suggested that probiotic or starter culture strains carrying ARGs rarely act as direct sources of clinically significant resistance in humans, particularly when genes are chromosomally encoded and not readily mobilizable, whereas others argue that any additional reservoir of ARGs in the food chain should be minimized as a matter of precaution. Our data contributes to this discussion by showing that even strains with favorable phenotypic susceptibility profiles may harbor acquired ARGs in genomic contexts that regulators currently deem unacceptable, highlighting the tension between strictly precautionary regulatory frameworks and more nuanced risk assessments based on gene expression, mobility, and clinical exposure scenarios. We acknowledge that the analysis of only three strains (two L. plantarum and one L. salivarius) represents a major limitation of this work. Although these strains provide informative case examples of genotype–phenotype discordance and illustrate the value of WGS in complementing conventional safety assessment, the findings cannot be extrapolated to represent species-wide characteristics. The observed differences between strains in this study should therefore be interpreted strictly as strain-specific observations within this limited panel, rather than definitive species-level traits. Comprehensive species-level conclusions would require systematic analysis of larger and more diverse strain collections, including pangenome-scale comparisons and evaluation of publicly available genomes to determine the prevalence and genomic context of the ARG–MGE configurations identified here. Such analyses were beyond the scope of this strain-focused study. Future investigations incorporating broader strain diversity and complementary metagenomic analyses will be necessary to establish robust species-level safety profiles and to better understand the distribution and mobility of antimicrobial resistance determinants in livestock-associated LAB. Conclusion Whole-genome sequencing (WGS) of the three Thai LAB strains revealed acquired antimicrobial resistance genes that would preclude all strains from obtaining EFSA QPS status under current regulatory criteria, despite their favorable phenotypic safety profiles. L. plantarum L25F and L22F harbor extrachromosomal lnu(A) genes (99.79% identity) conferring lincosamide resistance to clinically important antimicrobials, which EFSA classifies as acquired resistance incompatible with QPS approval irrespective of the predicted mobility context. L. salivarius AF2319 carries acquired erm(C), conferring resistance to erythromycin (a WHO “highly important” macrolide), together with tet(M) and tet(L) on extrachromosomal DNA, a combination that is likewise not acceptable for QPS status and warrants caution even though the associated contigs lack identifiable conjugative backbones and show only indirect signatures of historical mobilization. Accordingly, although none of the three strains carries classical virulence factors and all exhibit benign in vitro safety phenotypes, they cannot be considered suitable candidates for authorized probiotic applications within the current EFSA regulatory framework. These findings underscore the critical role of WGS-based antimicrobial resistance gene profiling in uncovering acquired resistance determinants that override otherwise favorable phenotypic indicators and species-level QPS designations, while also contributing to the broader debate on how to balance precautionary regulatory approaches with nuanced assessments of gene expression, mobility and clinical risk in non-pathogenic LAB. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1