Analysis of the microbiota of raw commercial feline diets to prioritize food safety investigations

Abstract

Microbiota sharing between people and their companion animals is a concern for development of antimicrobial resistance. To assess the risks associated with feeding raw products to cats, with an emphasis on previously understudied freeze-dried products, a collection of 112 conventional and raw products was purchased and investigated using a combination of cultivation and high-throughput sequencing techniques. Here we show that bacterial cultures were exclusively isolated from raw foods. A total of 19 genera were cultured including Salmonella, Clostridium, Escherichia, Klebsiella, Enterobacter, and Cronobacter. Carbapenem-resistant Pseudomonas aeruginosa and Pseudomonas fulva, and Stenotrophomonas lactitubi were isolated from frozen raw products, and 6 Bacillus strains harbored carbapenemase gene bla₂. Multidrug efflux pumps were highly abundant in frozen raw isolates. Clostridium sensu stricto I genus detection predicted a raw, freeze-dried product with 95% sensitivity and 78% specificity. Genera Pseudomonas, Paraclostridium and Peptostreptococcus were associated with frozen raw food products while the Bacillus genus was associated with conventional processing. Parasite genes were exclusively detected in raw foods. The presence of pathogenic species and high load of resistance genes in raw commercial food products, particularly those sold on shelves at room temperature, suggests a considerable health risk to cats and the families who care for them.

Predictive microbial ecology sheds light on human health risks of freeze-dried and frozen raw foods for cats.

Introduction

For more than 30 years, there have been attempts to return companion animal nutrition to its ancestral roots with bone and raw food diets (BARF)¹. For clarity, we will refer to BARF diets as raw diets in this article. While there is potential evidence for the benefits of raw diets, further studies have underscored the risks of these diets and their One Health implications^2–4. As such, it is important to understand the true risk of raw diets for all household members, meaning owners and pets. Despite news articles and even regulatory actions indicating that raw diets may be contaminated with zoonotic bacteria, viruses, and parasites when compared to conventionally processed foods, American and European pet owners continue to pursue this dietary endeavor in pursuit of better animal health⁵. As the dietary focus gains popularity, online communities have begun describing the benefits of this lifestyle for pets. A review article from 2013 by Freeman et al. identified how online proponents deliver anecdotal evidence of better digestion, immune health, and overall healthier appearance⁶. The article also emphasized that pet owners are easily influenced by the desire to do better for their companion animals. On the nutritional side, Buff et al. analyzed the metabolic needs for both domestic dogs and cats compared to their non-domestic relatives⁷. Their data support that raw diets adequately meet the nutritional needs of companion animals and offer more bioavailable nutrition compared to conventional foods. Subsequently, consumer conformity has led to the rapid spread of the tenets of raw diets throughout online spaces, leading to larger groups of adherents⁸. Approximately two-thirds of survey respondents indicated they would gather pet health information from the internet or other online resources as opposed to veterinarians, including food and diet recommendations⁵. The lack of proper information gathering also belies increased risk for both animal and owner from using raw diets.

Several studies have approached consumers who feed their pets non-conventional diets to determine their level of knowledge about the foods that they are handling. In particular, two separate studies by Anturaniemi et al. and Thomas and Feng aimed to evaluate the knowledge of owners who feed their animals raw diets^9,10. Consumer use and handling of food represent critical control points for food safety because they have the opportunity to reduce hazards to an acceptable level¹¹. The aforementioned studies indicate that American consumers lack adequate education on the proper handling of foods, especially those intended to be served raw^9,10. As such, they place themselves and their pets at a higher risk for infection. A 2019 case study by O’Halloran et al. identified a cluster of thirteen cats in the United Kingdom who suffered from a Mycobacterium bovis infection (a member of the Mycobacterium tuberculosis complex)¹². The investigation revealed that the six index cases in the five geographically isolated clusters were each indoor-only cats that ate frozen raw food products. Further, these index cases were associated with seven other contact cases who were all indicated as M. bovis positive, suggesting the highly communicable nature of this infection. In addition to M. bovis, a number of other pathogenic or antimicrobial resistant species have been attributed to raw diets. For instance, Baede et al. reported finding extended-spectrum β-lactamase (ESBL) producing enteric species from frozen raw pet food³ and Nemser et al. investigated the presence of Listeria, Salmonella, and toxigenic Escherichia species in several pet foods¹³. The latter finding 10% of the raw foods (all frozen products) tested positive for one of the listed pathogens, as opposed to less than 0.05% in conventional dry and semimoist foods. Ahmed et al. brought to light the heightened risks of parasitic infections that come with raw diets². Species groups such as Echinococcus cestodes and Sarcocystis protozoa rely on the predator-prey interactions for the persistence of their lifecycles¹⁴. Additionally, Toxoplasma gondii commonly infects cats in rural regions¹⁴. However, Ahmed et al. noted that pet animals will encounter these species more frequently in raw feed than in conventional feed, thus increasing their risk of colonization and subsequent human infection².

In addition to the presence of viable pathogenic microbes in raw meat, either from the original animal(s) or from cross-contamination during processing, another concern is the spreading of antimicrobial resistance genes (ARGs). A study by Nilsson in 2015 found evidence of transferrable ESBL genes in E. coli isolated from frozen raw poultry diets for dogs in Europe, with 100% of products tested containing viable bacteria¹⁵. Raw foods also harbor ARGs in viable microbes not considered foodborne (e.g., Klebsiella pneumoniae)¹⁶. As reviewed by Verraes et al., interspecies ARG transfer by conjugation has been experimentally demonstrated in a number of food matrices¹⁷. The ARGs could be transferred among the bacterial species present in the products and incorporated into the gut microbiota of household members via horizontal gene transfer. The addition of bacteriophages, lactic acid bacteria, and live “probiotics” such as Enterococcus sp. during processing may exacerbate the risk by adding additional potential vehicles for horizontal ARG transfer^15,17.

While there is mounting evidence that the increased risk of foodborne diseases or presence of antimicrobial-resistant (AMR) pathogens in animals fed frozen raw diets poses health concerns to both humans and animals, new types of formulations (e.g., freeze-dried raw treats) have come on the market that may be perceived as safer because they are sold as shelf-stable products. Our aims for this study were to expand the scope of known species and AMR mechanisms present in non-conventional cat foods to understand the risks involved with the various raw diets within a One Health perspective. In particular, we sought to identify microbial taxa of interest for guiding future regulatory actions involving different product formulations.

Results

To complete this study, a total of 112 commercial cat food products were selected for testing (Table 1; Supplementary Data 1). To reflect what an average USA-based pet owner would purchase, all food products were purchased from easily accessible major online vendors and grocery store chains. Products were then categorized by processing method (raw or conventional) and texture (canned, freeze-dried raw, frozen raw, kibble or refrigerated). Overall, 27 conventional foods were purchased and tested, and 85 raw (including raw and partially cooked) products were tested. The food type was also noted as “refrigerated” (n = 3), “frozen raw” (n = 25), “canned” (n = 22), “freeze-dried raw” (n = 49), or “kibble” (n = 13).

Table 1.

Summary of foods purchased and studied

Food type	Number of products studied		Total
	Raw	Conventional
Refrigerated	3	0	3
Frozen raw	25	0	25
Canned	0	22	22
Freeze-dried raw	49	0	49
Kibble	8	5	13
Total	85	27	112

Commercial foods were stratified by “raw or partially cooked” and “conventional”.

Protein source verification

We started our study by carrying out mitochondrial 16S rRNA gene amplification in order to verify the labeled protein source of our selected 112 products. Two products, FR08 and C18, did not show results of amplification and one product FD17 was removed from the study (see methods). Using ASV (Amplicon Sequence Variant) relative abundance, we first assigned primary, secondary and tertiary protein sources for the 109 other products when possible before grading the results as “matched” (protein sources detected in accordance with labeled ingredients), “neutral” (protein sources detected partially in accordance with labeled ingredients without major discrepancies), “discrepant” (major discrepancies between protein sources detected and labeled) for each of them. The full results are given in Supplementary Data 2 and summarized in Table 2. In brief, 59% of the food products were marked as “matched” and 28% as “neutral”. Matches were particularly high in raw products (66%) while conventional products were more distributed (~33% each grade). More precisely, freeze-dried raw (75%) and kibble (62%) were the food textures with the highest concordance, while freeze-dried raw products were the least discrepant (4%). We found that pork was the major (>75% ASV abundance) protein source detected for three canned samples (C12–14) with the ingredient “Meat By-Products” despite those products being labeled as chicken or fish. Overall, chicken and turkey were the most common unlabeled ingredients.

Table 2.

Summary of protein source comparison between product label and mitochondrial barcoding

	Canned	Frozen	Freeze-dried	Kibble	All raw	All conventional	Total
Matched	6	12	36	8	55	9	64
Neutral	7	10	10	2	22	8	30
Discrepant	8	2	2	3	6	9	15

Numbers shown are the counts of products tested.

Despite adequately confirming protein source compositions in our samples, we chose for this study an unbiased approach and decided to use labeled ingredients and not detected protein sources (Supplementary Data 1). Based on those labeled protein sources, we then assigned each sample a protein source category: mammalian (e.g., beef, pork, rabbit, deer, and lamb; n = 17), avian (e.g., turkey, chicken, and duck; n = 49), aquatic (all fish; n = 10) and mix (any product with protein source from multiple categories; n = 36).

16S rRNA gene metabarcoding

Diversity analyses

To characterize the microbial composition of the different sources of food, we performed 16S rRNA gene sequencing on all the samples (Fig. 1). Taxonomic analysis of the resultant ASV revealed the presence of 68 bacterial genera in all food products (Fig. 1A, B; Supplementary Fig. S2). Bacillus was by far the most abundant genus for both processing methods (Fig. 1B). The detected genera included Pseudomonas, Escherichia-Shigella, Enterococcus, Staphylococcus, and, in smaller abundance, Acinetobacter, Serratia and Aeromonas (Supplementary Fig. S2).

Fig. 1. Microbial diversity of raw and conventional food samples based on 16S rRNA gene amplicon sequencing data.

A Relative abundance of the 15 most abundant bacterial genera between conventional and raw food samples at the genus level. B Relative abundance as in (A), summarized by food processing method. All other genera were grouped under the Other category. C–E Comparison of alpha diversity indexes in conventional (n = 27) and raw (n = 84) food samples. C Shannon Diversity index (p = 0.00000077, Wilcoxon rank sum test; effect size large, 0.510). D Faith’s Phylogenetic Diversity index (p = 0.00000316, Wilcoxon rank sum test; effect size moderate, 0.481). E Pielou Evenness index (p = 0.000216, Wilcoxon rank sum test; effect size moderate, 0.384). (****p < 0.0001; ***p < 0.001). All boxplots display 1st quartile (Q1) to 3rd quartile (Q3) with median (Q2) as the black horizontal line.

We first characterized alpha aiversity metrics to compare the richness of bacterial species between raw food samples and conventional food samples. Raw food samples exhibited higher Shannon diversity (Fig. 1 C), higher Faith’s Phylogenetic Diversity (Fig. 1D), and higher Pielou’s evenness diversity (Fig. 1E) than conventional food samples. These results support the hypothesis that raw food products have a phylogenetically diverse variety of microbes. In addition, beta diversity analyses on rarefied ASV data revealed that microbial community structures are statistically dissimilar between raw food samples and conventional food samples (weighted UniFrac: p = 0.0001; R² = 0.06924, unweighted UniFrac: p = 0.0001; R² = 0.06272 and Bray–Curtis dissimilarity: p = 0.0001; R² = 0.04746; pairwise PERMANOVA tests, 9999 permutations, controlling for sequencing run effect).

Differential abundance analysis/indicator species analysis applied to raw vs conventional

We next sought the particular genera that were most likely to vary by food processing method. Differential abundance analysis (DAA) of our ASV table aggregated at the genus level highlighted several genera for which relative abundances were significantly different between raw and conventional products (Supplementary Fig. S3). Three Bacilli genera (Bacillus, Paenibacillus and Aneurinibacillus) were found with higher relative abundances in conventional products, while a total of 17 genera were higher for raw food products. These genera included Clostridium sensu stricto 1 (CSS1), Enterococcus, Esherichia-Shigella, Serratia and Pseudomonas, which were found in significantly higher abundance in raw products (p < 0.05, internal Corncob R package test result).

To investigate further, we applied an indicator species analysis (ISA) (Fig. 2A). This analysis showed a strong association between the detection of genera CSS1, Enterococcus, and Pseudomonas, as well as other potentially pathogenic genera such as Paraclostridium and Peptostreptococcus and the presence of raw ingredients in the given food product (high specificity). Of these, only CSS1 and Enterococcus had also a high sensitivity, i.e., a >90% probability to be detected in a raw food product. The genus Bacillus was the only highly sensitive (100% of the time) taxonomic group significantly associated with conventional food products. Despite that, the specificity of Bacillus was quite low, with only a 70% chance that a product will be conventional if Bacillus is detected. This result is in accordance with other results showing Bacillus as the most detected genus in both pet food processing methods (Fig. 1A, B). Every other genus significantly associated with conventional products was also a very good predictor (high specificity).

Fig. 2. Statistical analysis of key microbial genera linked to types of food using 16S rRNA gene amplicon sequencing data.

Indicator Species Analysis (ISA) sensitivity and specificity for genera found significantly linked to A processing methods (n = 111) and B raw food textures (n = 73). Bars were colored to reflect the processing method or the food texture for which a genus was found significantly associated with, p values are reported in parenthesis.

Investigating raw product food texture associations (freeze-dried raw vs frozen raw)

Given that our results indicate that raw pet food products are more likely to contain harmful microbial taxonomic groups, we next investigated the differences within raw pet foods divided by food texture. We performed ISA to compare frozen raw and freeze-dried raw products (Fig. 2B). We found that 5 genera were significantly associated with freeze-dried raw products, while 23 were for frozen raw products. Of the genera enriched in freeze-dried raw products, Bacillus genera had high sensitivity (100%) and high specificity (96%). Genus CSS1, previously strongly associated with raw products, seems to be an even better signal of freeze-dried raw products (95% sensitivity, 78% specificity). Of the genera enriched in frozen raw products, Pseudomonas, Lactobacillus, and Carnobacterium were predicted with a high sensitivity (>75%). All three also had high specificity (>80%).

Overall, considering the pathogenic genera strongly associated with raw products by the ISA of processing methods (Fig. 2A), CSS1 is a marker of freeze-dried raw products, while Pseudomonas, Paraclostridium and Peptostreptococcus are associated with frozen raw food products. Finally, one of the most important genus predictors was not found significantly associated with either raw texture and it is therefore likely that Enterococcus was widely distributed across raw products and not specific to either frozen raw or freeze-dried raw products.

Investigating CSS1

Given the importance of the genus CSS1 as a marker of freeze-dried products, we investigated all 43 CSS1 ASVs to determine the species-level proportions. We found that 23 ASVs (53%) were Clostridium perfringens, but accounted for ~85% of the total CSS1 relative abundance (Supplementary Fig. S4). On the other end, 20 ASVs were classified as other Clostridium genera but accounted for less than 15% of the total relative abundance of CSS1 genus (Supplementary Fig. S4). These findings included C. botulinum DNA detected in 6 frozen raw (FR2-7) and one freeze-dried product (FD11).

Bacterial culture and subtyping

Gram-negative enrichments

Given that we established that there are traces of microbial pathogen genomes strongly associated with raw food products, we aimed to evaluate the direct risks of poor handling of such products. Indeed, given the extensive usage of raw food diets, we sought to verify if the food products contained viable pathogens. We first aimed at enriching Gram-negative bacteria from all food products using a protocol from the Food and Drug Administration’s Bacteriological Analytical Manual (see method for details). In total, 92 good-quality bacterial isolates were obtained from the foods using general Gram-negative enrichment methods (Supplementary Data 3). Raw foods were found statistically more likely to harbor viable and cultivable Gram-negative bacteria than conventional food (Fig. 3A). Within raw foods, we found that cultivable Gram-negative bacteria were more prevalent in the frozen raw foods compared to the freeze-dried raw foods (Fig. 3B). Finally, we found that the mammalian protein source category was the only category for which the proportion of samples yielding or not yielding bacterial growth were comparable (Fig. 3C; Supplementary Fig. S5).

Fig. 3. Gram-negative bacterial isolate summary for non-selective culture.

Bar plots representing the proportion (%) of samples for which at least one isolate was cultured. A by processing methods (n = 111, p = 0.0001359) B by raw food textures (n = 73, p = 0.0001779) (Pearson’s Chi-squared test with Yates’ continuity correction. P value significance: ***p < 0.001). C Distribution of food samples by protein source categories (aquatic, avian, mammalian, mix). Darker colors represent samples for which at least one bacterium was isolated using FDA BAM procedures, while lighter colors represent no bacteria isolated (the mix category gathers non-unique protein sources, full list of categories in Supplementary Fig. S5).

As highlighted by Fig. 3A, all samples that yielded bacterial isolates were from raw processing methods. A major foodborne pathogen, Salmonella enterica, was detected in 16% (4/25) of frozen raw samples, all single-protein diets (mammalian, aquatic and avian*2). Other genera of human health concern identified included Cronobacter, Enterobacter, Enterococcus, Escherichia and Klebsiella. Escherichia coli isolates were, like Salmonella enterica, exclusively found in frozen raw products (n(isolates)=5;n(samples)=4). Enterobacter and Klebsiella isolates were grown from numerous products from both freeze-dried raw (n(isolates)=3;n(samples)=2 and n(isolates)=1;n(samples)=1) and frozen raw (n(isolates)=5;n(samples)=5 and n(isolates)=4;n(samples)=2) textures. The only Cronobacter sakazakii isolate was from a raw kibble product. Finally, Enterococcus isolates were cultured mostly from freeze-dried raw (n(isolates)=5;n(samples)=5) but were also isolated from one kibble. Other genera identified in our pool of isolates included Citrobacter, Hafnia, and Morganella (Supplementary Data 3).

Anaerobic cultures

While analyzing 16S rRNA gene amplicon data, we highlighted the very high relative abundance in many samples of DNA from genus CSS1, mostly comprised of the species Clostridium perfringens. Thus, we attempted to isolate anaerobic microbes to grow C. perfringens. We chose all 11 samples from our cohort with more than 35% of their total amplicon reads belonging to genus CSS1 (Supplementary Fig. S6). While we failed to grow C. perfringens retrospectively using the moderately anaerobic conditions (on frozen homogenates), we successfully grew, isolated, and sequenced 6 Bacillus species, 3 Enterococcus species, 2 Macrococccoides caseolyticum, 1 Clostridium tertium, and 1 Weizmannia agrestimuris. Untargeted sequencing yielded two samples (FD9 and FD15) with more than half the microbial DNA composition belonging to C. perfringens (642,631 and 641,547 reads per million).

Carbapenem cultures

We isolated carbapenem-resistant bacteria using carbapenem-selective enrichment on the food products for which carbapenemase genes were detected using the AMR panel direct sequencing (See Section ARGs). Those enrichments yielded cultures from 3 frozen raw products. Ultimately, four good-quality genomes were sequenced, and the species identified were Pseudomonas aeruginosa, Pseudomonas lundensis, Pseudomonas fulva, and Stenotrophomonas lactitubi. Of the four P. aeruginosa clones sequenced, two had carbapenem-associated point mutations (oprD_V359L or cmrA_A68V). The predicted Stenotrophomonas mechanism was bla_L1, resulting in an imipenem MIC > 8. Multidrug efflux pumps were predicted in 3 isolates. Strain FR19-erta1a (P. aeruginosa) had an imipenem MIC of 4, with ARGs mexA, mexE and mexX, FR24-erta1 (P. fulva) had only mexE predicted (imipenem MIC of 1), while FR24-erta2 (S. lactitubi) had multidrug efflux genes emrA, emrB and emrC predicted on one contig.

Functional genomics on bacterial isolate genomes

In total, whole-genome sequencing of pure culture yielded 89 non-clonal bacterial isolates that passed quality thresholds (48 from gram-negative enrichment, four from carbapenem enrichment and 37 from anaerobic conditions). With the idea to highlight the broad range of ARGs and virulence genes in our entire pool of isolates, we performed functional genomic analyses on all 89 isolates. All isolates except seven were predicted with at least one ARG (Fig. 4). In total, 69 unique ARGs were detected in isolates, predicting resistance to 15 different antimicrobial classes. Ten of those genes were predicted in at least ten isolates, including three that encoded efflux pumps. Eight predicted genes are known to be part of resistance mechanisms against multiple drug classes. In particular, two Bacillus genus-specific genes with predicted resistance to three antimicrobial classes, clbA and vmlR, were found in our anaerobic isolates from freeze-dried raw foods and kibbles. The most common genes found were emrD, msr(C) and aac(6’)-I, respectively granting resistance to the Efflux, Macrolide/Streptogramin and Aminoglycoside classes. emrD was detected mostly in frozen raw foods and in three different genera (Escherichia, Klebsiella and Shigella). On the contrary, msr(C) and aac(6’)-I were not found in any of the frozen raw foods isolates and were all predicted on genomes from the genus Enterococcus. In fact, msr(C) is known to be a chromosomal gene of E. faecium, and all nine products yielding live cultures of this species were freeze-dried treats or raw-coated kibbles. Here, we detected msr(C) in multiple E. lactis isolates, a closely related species to E. faecium, also used as a probiotic. Efflux class-resistant isolates were mostly found in frozen raw foods, with the exception of one freeze-dried raw Klebsiella isolate. One Enterobacter hormaechei isolate from raw product FD40 carried mcr-9.1, a mobile resistance gene against colistin. No Glycopeptide, Lincosamide (except lsa(A)), Macrolide or Streptothricin ARGs were detected in the Frozen raw isolates. Also, only one gene conferring resistance to the Glycopeptide antimicrobial class was found in our isolates. This gene, vanR-A, was predicted in a Bacillus strain (closest predicted species B. cereus, ANI = 96.5%) from a freeze-dried raw product. The same isolate was also isolated in another isolation condition (Supplementary Data 3). Similarly, only one gene conferring resistance to sulfonamide was predicted in our isolate genomes. This gene, sul2, was predicted in three isolates from three different frozen raw samples. Interestingly, for two of these isolates, the gene was predicted on putative plasmid contig with high confidence and on the genome itself with low confidence for the third isolate (Fig. 4; Supplementary Data 4). A total of three Fosfomycin resistance genes were found, but mostly one was widespread (fosA). Last but not least, despite not being able to grow isolates with the carbapenem genes detect in the AMR panel analyses on carbapenem selective media (see carbapenem culture section) we have successfully grown 6 Bacillus strains from 6 different samples with the gene bla₂.

Fig. 4. Presence-absence heatmap representing ARGs detected in isolates using WGS data.

The colors on the heatmap are indicative of the sample processing method when an ARG was detected or blank otherwise. Presence of “P” indicates that the corresponding ARG was detected on a putative plasmid. The numbers beside the boxes on the right describe the number of isolates in which an ARG was found and the drug class(es) this ARG confers resistance to. Bold text highlights ARGs found in 10 or more isolates. The upper legend indicates ANI best hit with the corresponding ANI value for each isolate at the genus level.

Stress response genes were abundant in isolates with 54 different genes representing four categories of stress resistance: Acid (n = 2), Biocide (n = 8), Heat (n = 10) and Metal (n = 34). In addition to being the most diverse with 34 different genes predicted, resistance to metals was the highest category predicted on isolates with 276 total predictions, including 71 genes encoding copper and silver resistance, 66 genes encoding arsenic resistance and 60 genes to help resist copper only. Resistance against heat (35 predictions), acid tolerance (24 predictions) and biocide (20 predictions) were observed. Furthermore, 19 isolates (of the genera Citrobacter, Enterococcus, Escherichia coli, and Klebsiella) across all raw food textures were predicted with one contig harboring at least a heat stress gene. For 10/19 isolates, those predictions (15 heat stress genes) were found on putative plasmid contigs (confidence in prediction >0.745). The other 9/19 isolates had those heat stress genes predicted on genomic contigs, but with very low confidence (<0.3). Particularly, one Escherichia coli isolate of concern, FR9-TG1-1 was predicted with 5 different heat stress-related genes on the same putative plasmid contig. Also, that contig did not yield any results when comparing it to the plsdb database¹⁸. Virulence genes (n(total ARGs)=245; n(unique ARGs)=42) were also detected, including cytolysins, hemolytic enterotoxins, and adhesins.

Clustering of bacterial isolate genomes with strains in NCBI pathogen detection

Bacterial genomes from this study were uploaded to the US National Center for Biotechnology Information (NCBI) Pathogen Detection Database¹⁹ as they were generated. The associated Pathogen Browser (https://www.ncbi.nlm.nih.gov/pathogens) matched 24 SNP clusters from this study as of August 3, 2024. Of these, 16 contained only isolates from this study, and the other eight connected strains were submitted by other submitters. The connected SNP clusters were all for Salmonella enterica, Enterococcus faecium, or Bacillus cereus. The isolated S. enterica serovars were Typhimurium (two different strains, one shown in Fig. 5), Kentucky (two different strains), and Montevideo. The Typhimurium isolates from Frozen raw product FR9 were within one single nucleotide polymorphism (SNP) of a presumed human clinical sample from the USA, submitted by the PulseNet program (strain PNUSAS214272). Eleven other presumed human clinical strains were located in the overall SNP cluster. The next closest strain (eight SNPs) came from a chicken carcass in Indiana, USA, tested by the US Department of Agriculture in 2020 (FSIS12032032), the same labeled protein source as the FR9 food product (produced in 2021). All other non-clinical strains in this cluster came from chicken sources and included antimicrobial resistance determinants for tetracyclines, sulfonamides, and multidrug efflux. The other S. enterica clusters had a similar source composition, with primarily either chicken or beef interspersed with clinical strains. One included several recent clinical strains, from 2023, and harbored the bla_CMY-2 gene encoding resistance to cefoxitin (Fig. 5). The E. faecium cluster consisted of one clinical strain from India and one bovine isolate from the USA.

Fig. 5. Evidence of zoonotic infection from raw cat food isolates using NCBI Pathogen Detection.

Snapshot of an NCBI SNP cluster issued from Pathogen Detection containing a Salmonella enterica serovar Typhimurium strain from FR19, a frozen raw cat food diet. The genome from this study (sample FR19) is highlighted in red. The scale bar indicates SNP distance, and the dates indicated are the collection dates submitted. Isolates labeled clinical are presumed to be from human infections.

Direct amplification of ARGs

ARGs screening by deep amplicon sequencing

In addition to highlighting the drivers of our different microbial communities and isolating the potential species with One Health concern, we aimed at inventorying the presence of ARGs in our samples using deep amplicon sequencing directly from food homogenates. This approach found that 91 food samples had at least one ARG detected (Fig. 6). We investigated the effect of processing methods and found that raw products have significantly higher ARGs detected than conventional food products (Fig. 7A). Following that result, we tested if the food texture was an important factor of detection of ARGs and found that Frozen raw products are expected to have larger ARG repertoire than freeze-dried products (Fig. 7B). With the idea that raw products may harbor ARGs from the protein source they are based on, we first carried out a one-way ANOVA to test for the effect of protein categories on ARG detection counts in each sample. We found a small, significant effect of protein categories through the ANOVA (p = 0.021, effect size = 0.114). To examine the effects of protein category ARG detection in raw foods, we conducted pairwise Wilcoxon rank sum tests between our 4 protein source categories (Fig. 7C). Despite visual clues showing that mammalian-based feline raw foods would tend to harbor more ARGs, we found only one significant difference between mammalian and avian-based raw food products (p = 0.007, p.adj = 0.041).

Fig. 6. Presence-absence heatmap representing ARGs detected directly in cat foods by deep amplicon sequencing.

Presence-absence heatmap representing ARGs detected directly in cat foods by deep amplicon sequencing. Only the main eight antimicrobial classes are displayed, as well as an “Other” category grouping all other classes included on the panel. The colors on the heatmap are indicative of the sample processing method when a gene was detected or blank otherwise. Categories at the bottom represent food textures. Categories on the right represent antimicrobial classes, sorted by the number of genes detected.

Fig. 7. Distribution of ARG counts per food sample using deep amplicon sequencing.

Boxplots representing the distribution of ARG counts per sample A by processing method (n = 111, p = 0.00234), B Raw food textures (n = 73, p = 0.0111) and C Protein sources categories of raw samples (n = 84). Statistical test results testing for a difference between categories are displayed (Wilcoxon rank sum tests, **p < 0.01; *p < 0.05; ns = non-significant p > 0.05). All boxplots display the 1st quartile (Q1) to the 3rd quartile (Q3) with the median (Q2) as the black horizontal line.

Overall, between 15% and 53% of the total number of genes present in the panel for each antimicrobial class was detected, with the tetracycline antibiotic class being the highest (Supplementary Fig. S7). Using this panel, we aimed at measuring the effect of cat food processing method on the presence of ARG counts per class (Fig. 8A). As expected from previous results (Fig. 7A), raw foods were found with a higher count of ARGs than conventional foods for three antimicrobial classes, including aminoglycoside and tetracycline, with very high confidence. All other antimicrobial classes were statistically similar in regard to the processing method. Following results displayed in Fig. 7B, we aimed at looking at differences between frozen raw and freeze-dried raw products. To do so, we carried out ARG counts comparison for every antimicrobial class in regard to the two raw food textures of interest (Fig. 8B). In total, seven out of the eight major antimicrobial classes were found significantly higher in frozen raw products than in freeze-dried raw products, with the only exception being the beta-lactam. Of note, the artificial “Other” class values were also found not to be significantly different between the two raw food textures. These results using direct amplification of antimicrobial resistance genes suggest that foods containing raw ingredients are bigger reservoirs of ARGs than purely conventional foods. This is especially true for the antibiotic classes aminoglycoside and tetracycline, and the frozen raw food products.

Fig. 8. Distribution of ARG counts per antimicrobial class using deep amplicon sequencing data.

Boxplots representing the distribution of ARG counts per antimicrobial class. Results were grouped and colored by A Processing method (n = 111) and B Raw food textures (n = 73). Statistical test results testing for a difference between both categories are displayed for each class (Wilcoxon rank sum test, ****p < 0.0001; ***p < 0.001; **p < 0.01; *p < 0.05; ns = non-significant p > 0.05). All boxplots display 1st quartile (Q1) to 3rd quartile (Q3) with median (Q2) as the black horizontal line.

Eight carbapenemase genes were detected directly from food homogenates. Six canned (conventional) foods, two conventional kibble products, and 13 raw products had bla₂ detected. This gene is typically found in the chromosome of Bacillus sp. The other seven genes (cphA, bla_b, bla_BIC, bla_MUS, bla_OXA.24, bla_OXA.134, and bla_OXA.51-like) were exclusively detected in raw foods. The ambler class B cphA was the next most common gene detected (n = 7, across all raw textures including kibble).

Eukaryotic pathogen 18S rRNA

Through an 18S rRNA gene metabarcoding approach, we aimed at detecting eukaryotic pathogens within food products. In order to broaden our chances of detecting parasites, we applied two bioinformatic pipelines to the reads dataset: an ASV approach through Qiime2²⁰ and a pathogen-oriented workflow using CZID²¹. Despite most of the data being taxonomically assigned to plants or fungi, we successfully detected the presence of parasite DNA in 16 samples. All of these samples were raw food products, with six being freeze-dried raw and ten frozen raw. Colpodella sp. (genus Cryptosporidium) was the most detected in both pipelines (see Fig. 9). Two species of Sarcocystis, Sarcocystis cruzi and Sarcocystis tenella, were detected by both approaches in multiple Frozen raw food products. Additionally, fish and aquatic invertebrate pathogen genera Calyptosporidae and Goussia were detected by the Qiime approach in one Frozen raw sample labeled with an aquatic protein source. On the same note, but using the CZID approach, the intracellular protozoan parasite species Eimeria maxima, causing avian coccidiosis, was found in two freeze-dried samples labeled avian. Also detected by a single approach (Qiime), Candida parapsilosis was found in an avian frozen raw product. Notably, both pipelines agreed on only four samples (FR16, FR17, FR19, FR20). All five are frozen raw and from the same manufacturer. Only FR16, a frozen raw mammalian protein source-based product, was found with two different potential parasites. One of the parasites is Sarcocystis cruzii, and the other is from genus Theileria. Sarcosystis cruzi was also detected in FR17. Another species of Sarcocystis, S. tenella, was found by both approaches in FR20, while FR19 was found with Colpodella sp./genus Cryptosporidium. No helminth parasites were found in our dataset while looking for phyla Platyhelminthes, Nematoda, Cestoda or Trematoda. As well, no pathogenic fungi other than Candida parapsilosis were detected by 18S rRNA gene amplicon sequencing. By untargeted RNAseq, the C. parapsilosis in this sample (FR15) was confirmed, and also detected in sample FR10. Additionally, Aspergillus flavus was found at fairly high levels in FR10 (458 reads per million) by RNAseq.

Fig. 9. Presence-absence heatmap of major parasite taxa based on 18S rRNA gene amplicon sequencing data.

Heatmap of parasite taxa detected in our 18S rRNA gene ASV dataset. Detections were carried out using two pipelines. We report here only the samples with hits and parasites found at least once in a food product. Hits are reported in green if detected locally with Qiime2 using reference databases Silva138Ref and PR2v5, in red if detected online using CZID or yellow if detected using both pipelines.

Influenza A screening

Due to reported outbreaks of highly pathogenic Influenza A linked to raw cat food in Poland²² and Korea²³ while this study was ongoing, we undertook RNA purification from all products to assess the potential risk of viral RNA. The overall RNA yield ranged from 0-70 ng/µL and was significantly higher for raw products (p = 0.0008). No Influenza A RNA, however, was detected in any sample. We performed untargeted RNAseq on the five samples with the highest RNA yield (>50 ng/µL) and integrity to assess if the genomes of any other human viruses could be detected. Only traces of livestock viruses were seen, including Rotavirus F, Bovine alphaherpesvirus 5, and Avian leukosis virus. This sequencing also contained results that corroborated the amplicon-based microbial findings from those food products.

Discussion

By expanding the ‘Bacteriological Analytical Manual Salmonella spp. procedures²⁴ to characterize all growth, we sought to determine the prevalence of other related Gram-negative organisms in addition to the primary objective of culturing Salmonella. Although using selective media likely biased the results towards certain Gram-negative species, we decided not to discard the non-Salmonella isolates in order to examine the potential context for gene transfer. The high abundance of Gram-positives of public health significance appreciated through molecular barcoding, particularly CSS1 and Carnobacterium, was an unexpected finding. We attempted anaerobic culture retrospectively on a subset of samples, but conditions may not have been ideal due to the products being exposed to oxygen and frozen in BPW prior to those attempts. Future studies and potential regulatory investigations should include cultures specifically for these organisms from the initial processing.

Our focus here was primarily on AMR organisms and AMR genes. The presence of ARGs in a bacterial community allows for the movement of these genes into sensitive recipients. Thus, while not all a concern for foodborne transmission per se, the co-infection of live Klebsiella, E. coli, P. aeruginosa, S. enterica, S. maltophilia, E. cloacae, and/or E. faecium in a feline food product poses real concern for horizontal gene transfer of ARGs between these species. The close genetic proximity of some of these strains with clinical cases in the NCBI Pathogen Browser¹⁹ also highlights that such transfers would not be limited to infecting cats. Two of our Salmonella isolates (serovars Typhimurium and Montevideo) were in fact nearly identical to presumed human case isolates, implying a very close common source and a Public Health concern.

The application of some bacterial and bacteriophage species in the food industry as bioprotective agents has been proven effective against pathogens such as Listeria monocytogenes, Candida albicans or Pseudomonas aeruginosa, as recently been reviewed by Evangelista et al.²⁵. One notable case is the genus Carnobacterium, associated with raw products with very high specificity (99.6%) by our ISA analysis and significantly detected with higher abundance by one of the DAA methods used (corncob, p-value = 1.236549e-07). The presence of this genus could be due either to intentional addition as a bioprotectant or due to spoilage. While not as well documented, we can also expect genera Lactococcus and Vagococcus to be part of a Lactic Acid Bacteria (LAB) solution sprayed on Raw food products to reduce pathogen loads due to their antimicrobial properties²⁶. Considering the potential risk to immunocompromised persons, adding these to the label would be prudent. A number of products did have “probiotics” on the label, including, in some cases, Enterococcus faecium. This organism, although an opportunistic pathogen, is permitted as a direct-fed microorganism for animals in the USA if it is a non-toxigenic strain²⁷. The toxigenic potential of this species was only recently recognized²⁸, and the plasmid-borne genes encoding enterococcus toxins could easily be transferred between strains. Given our previous findings of multidrug-resistant E. faecium colonization in dogs²⁹, constantly feeding even a benign strain to companion animals could serve as a vehicle for both resistance and toxin genes to transfer into the same strains.

The last genera significantly associated with raw products were not found as theoretically protective. In fact, most of them contain species that are associated with diseases and/or conditions, even in humans. Genus Paraclostridium, for example, contains species Paraclostridium bifermentans, which has recently been characterized³⁰ as a potential emerging human pathogen while genus Peptostreptococcus has been associated with multiple infections³¹ despite being a normal inhabitant of human microbiomes. Finally, three genera of the family Enterobacteriaceae, often used to determine the quality of food products, were found significantly linked to raw food products: Enterobacteriaceae family, Escherichia-Shigella and Hafnia-Obesumbacterium.

The food resistomes not only threaten to seed other bacteria present with those mechanisms, but also to introduce these genes into the gut microbiota of the animals eating them. For example, K. pneumoniae, found in a number of our Raw products, is known as a vector for interspecies AMR gene transfer via outer membrane vesicles³². These vesicles are resistant to bacteriophage treatments commonly used for knocking down bacterial loads in the raw pet food industry. A 2008 study demonstrated the ability of AMR K. pneumoniae species to transfer their AMR genes to sensitive E. coli species colonizing in the same digestive tract of mice³³. Another study³⁴ found the presence of transposons Tn4451/Tn4453 that carry ARGs in several commensal bacterial hosts, such as Clostridioides, indicating the potential transferability of ARGs from food animals to host via foodborne route. At present, the US Food and Drug Administration (FDA) only routinely inspects pet foods for STEC, Salmonella, and L. monocytogenes^35,36. Another species, Cronobacter sakazakii, is also now under greater scrutiny³⁷ due to contamination events in powdered infant formula, and our results suggest that freeze-dried raw pet treats could be another important reservoir for this species. Reportable foods under the regulation of the US FDA are defined as “an article of food/feed for which there is a reasonable probability that the use of, or exposure to, such article of food will cause serious adverse health consequences or death to humans or animals”³⁶. Because K. pneumoniae, non-STEC E. coli, and P. aeruginosa are not amongst the diseases to qualify as reportable, they would largely go undetected in commercial food products. This likely explains the lack of NCBI Pathogen Detection database matches in regard to the strains of those species sequenced for this study. However, the unregulated species could introduce ARGs into the larger microbiome of the food and gut, not only of the animals eating the food but also the humans in the household.

The detection of ESBL genes in the food products aligns with the current literature that has identified ESBL-producing species in the feces of animals fed raw diets^15,38,39. This may suggest that ARGs are persistent in most animal-source foods regardless of the antimicrobial exposure history of the food animal. A study from 2015 on the microbiomes of an isolated Amerindian population showed evidence of AMR genes protecting against new-age synthetic antimicrobials such as quinolones and sulfonamides⁴⁰. The presence of ARGs in a population isolated from Western medicine suggests a persistence or high rate of horizontal transfer, even without the selective pressures of antimicrobial usage. This data may also corroborate an increased rate of gene shedding into the environment, putting humans, animals, and plants all at an increased risk of contact with a resistant species⁴¹. Encountering ARGs in both cooked and uncooked foods underscores the importance of the intersection of animal, human, and environmental health outcomes, and further studies are needed to explore the health implications of ARG consumption.

We detected a higher ARG load in food samples with mammalian protein source, which could reflect efforts to regulate antimicrobials in the US poultry industry. The lack of significance for other comparisons might be due to the limited raw sample size of the category aquatic and the fact that the mix includes products with mammalian ingredients (Supplementary Fig. S5). We chose a targeted method for ARG detection, rather than untargeted metagenomic sequencing, to maximize sensitivity for genes in low abundance. The breadth of resistance genes identified directly from food was limited to those included in the panel⁴² utilized here, which lacks any genes discovered since 2019, and does not include any of the mobilized colistin resistance (mcr) genes. Due to this limitation, we supplemented this analysis with whole-genome sequencing of all cultured isolates, which could be screened with the latest database of known ARGs in addition to virulence and stress response genes. One mcr gene was identified with this approach, mcr-9.1⁴³, which has the potential to confer colistin resistance when expressed with an active promoter. Resistance mechanisms for glycopeptides, considered “last resort” antimicrobials to combat gram-positive bacteria, were found, as well as for the broad-spectrum fosfomycin. A perspective paper in 2017⁴⁴ raised the question about the use of Fosfomycin from UTI infections to other applications, but, to date, it remains mostly used for human-based infections, especially UTIs. Nonetheless, it has been used in Central and South American countries in chicken and swine farms, and increasing resistance has been shown in Asia, as the drug is used against ESBL infections⁴⁵. Additionally, finding putative mobile heat stress response genes was also concerning given that some of these products are claimed to be “partially” or “gently” cooked.

Another limitation of our approach was that the carbapenem-selective enrichment was only performed retrospectively on a small subset of food products with carbapenem ARGs detected by this panel. Additionally, the method we used may not have recovered Acinetobacter that was likely present in the foods containing blaOXA_51-like, blaOXA_2.4, and blaOXA₁₃₄. Carbapenem-resistant bacteria pose a substantial threat to mortality worldwide, and the finding of ARGs that have been associated with carbapenem-resistant Acinetobacter baumannii (CRAB) is possibly the finding of highest public health significance from this study. The World Health Organization’s 2023 analysis of antibacterial agents in development and their 2024 bacterial priority pathogens report categorically position carbapenem-resistant strains of Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacteriaceae -- all of which are Gram-negative bacterial species -- as occupying the highest level of urgency, denoted as “critical” priority^46,47. In the USA, carbapenem-resistant Enterobacteriaceae (CRE) are primarily identified among patients with healthcare exposure, but there is potential for CRE to spread outside of healthcare settings, given that Enterobacteriaceae are a common cause of community-associated infections^48,49. The detection of several carbapenemase genes, such as bla₂ (also in isolated strains) and bla_OXA-51, in both raw and conventionally processed foods cannot be ignored either. This suggests that the prevalence of carbapenem-resistant bacterial species in companion animals may be higher than previously estimated. Past research has identified the presence of carbapenemase-producing Enterobacteriaceae in food-producing animals⁵⁰. Notably, these animals are a principal protein source in the cat foods examined in this investigation. Our research implies that carbapenemase-encoding genes from zoonotic or commensal bacteria could be transferred to companion animals through the foods ingested by these animals. Given the mounting evidence of shared gut microbiome carriage between people and their pets^51,52, this would also pose a risk to humans in the household.

Metabarcoding approaches for Eukaryotic pathogens are far less well developed compared to bacteria. We opted for an 18S rRNA gene-based approach⁵³ here in order to get a snapshot of potential eukaryotic pathogen presence. No culture was attempted for parasites or fungi, limiting interpretation of those results to the presence of the genome only. The results presented, therefore, do not imply that parasites were viable. The majority of sequences identified belonged to plant components of the foods, e.g., corn, squash, rice, etc. The presence of Sarcocystis sp. DNA was of interest, identified in raw products marketed for both cats and dogs. This genus comprises intracellular, apicomplexan protozoans. Dogs are the definitive host for both of the Sarcocystis species identified (S. cruzi and S. tenella), and cattle or sheep are the intermediate hosts. It could be that the carcasses used for these food products were downgraded towards pet food due to associated lesions observed during meat inspection. Although we did not evaluate infectivity, it would be prudent to keep dogs fed these products away from livestock so as not to perpetuate the life cycle. The finding of Aspergillus flavus, known for the production of mycotoxins, is also a potential indicator of suboptimal or possibly spoiled grain ingredients. Finally, Candida parapsilosis is a human pathogen of particular concern for immunocompromised individuals^54,55.

The ongoing outbreak of Highly Pathogenic Avian Influenza A(H5N1) Clade 2.3.4.4b in US dairy cattle and cats⁵⁶ raises the possible concern of foodborne viral transmission. Infection in mammalian wildlife could be from contact with mucous membranes during bird predation. The feline infections in the US reported by Burrough et al.⁵⁶ however, were linked to consumption of raw milk. Findings from recent outbreaks in Korea²³ and Poland²² indicate that Influenza A is also transmissible to cats from commercial foods. We did not detect Influenza A in any products tested here, which were all purchased prior to the Clade 2.3.4.4b outbreak. Our finding of substantially higher total RNA yield in raw products supports a focus on these products for future surveillance efforts.

The principle of One Health considers the intersection of animal, plant, and environmental health as facets of any individual’s health in the ecosystem⁵⁷. The close interactions of humans and their companion animals is an important and often overlooked component of this principle. For example, there is now ample evidence (reviewed by Singla et al.⁵⁸) that SARS-CoV-2 can be transmitted to animals by humans and vice versa. Just as easily, there is concern over parasitic species for which humans are accidental hosts and animals are definitive hosts. Ahmed et al. cite five specific parasites and eight different bacterial genera which present a zoonotic hazard². Great concern lies within the risk of introducing these species into populations that would not normally encounter them. Specifically, urban developments would not typically encounter domestic animals infected with most parasites. However, the feeding of raw diets to companion animals opens a portal of entry for urban populations².

Close-quarters living poses a significant risk of transmission between companion animals and their families, as well as other animals. A study conducted in the UK among pet owners who chose raw diets for their pets revealed low perceived susceptibility to food-borne illness associated with the diet and limited knowledge on safe raw pet food handling⁵⁹. Eighty-nine percent of participants were not aware that raw pet food poses a risk of food-borne illness to themselves and their family. In addition, improper raw pet food handling, including rinsing raw meat and no segregation between utensil and kitchen preparation areas, was reported in 27% and 52% of participants, respectively. Subsequently, another UK study associated raw diets for dogs with shedding of multidrug-resistant E. coli⁶⁰. When pet food products are contaminated with pathogenic microbiomes, unsafe raw food handling, including sharing utensils and the absence of cleaning kitchen surfaces after food preparation open the route of infection for pet owners, their family, and other pets in the household³⁸.

Households with immunocompromised members and young children should take particular care in selecting pet foods. One of the most striking and unexpected findings of our study was our observation that not a single freeze-dried product we purchased was fully cooked prior to freeze-drying. This determination was made after extensive efforts to contact the manufacturers. Out of 49 freeze-dried products purchased, all containing uncooked meat in the form of treats or coated kibble, only a small minority indicated some place on the packaging that it was raw. Freeze-drying slows microbial growth and preserves viability^61,62. This is consistent with reports of dried pig ear treats for dogs having very high Salmonella prevalence⁶³ and causing large outbreaks⁶⁴. Given that these products are sold on store shelves at ambient temperature, a consumer could easily purchase them without being aware of the risks or even realizing that they are purchasing an uncooked product containing viable microbes. Easily legible label warnings on packaging would potentially reduce morbidity and mortality in both humans and animals.

In this study, we identified raw food products as potential vehicles for the lateral gene transfer of AMR mechanisms into pathogens of significant consequence to One Health. Also, we highlighted freeze-dried products as a concern to the health of companion animals and owners from viable pathogens. Finally, we have highlighted general concerns regarding the lack of information and clear feeding safety guidance on the raw products. Overall, this study expands the evidence base for investigation of illness in relation to cat foods.

Methods

Food selection

A total of 112 commercial food products were selected and purchased for testing between 2021 and 2023 (Table 1). Food products were purchased from three major online vendors based in the USA and three chain grocery stores with locations in NY, USA. Direct-to-consumer commercial products were also purchased from the manufacturer’s websites after identifying them by web search for “raw cat food”. A total of 31 different manufacturers were included (identities intentionally withheld from publication). Products from the same manufacturer were different formulations, with an emphasis on diversity of animal protein sources and textures. The foods were classified based on their labeling on the food package and website materials for the processing method: “raw” (not fully cooked) or “conventional” (fully cooked). Texture was classified as either canned, freeze-dried raw, frozen raw, kibble, or fresh (refrigerated paté). “Partially” or “gently” cooked products were classified as raw, and shelf-stable kibble products with freeze-dried raw coating were classified as raw kibble. For instances when the packaging was not clear, the manufacturer was anonymously contacted to confirm. One sample (FD17) was removed from the study due to apparent cross-contamination of the homogenate with another sample.

DNA extraction

One ml of food homogenate (as described below for bacterial culture), or broth culture from a single colony, was centrifuged at 8000 × g for 5 min to pellet the bacterial cells for DNA extraction (Supplementary Fig. S1). DNA was extracted using the MagMAX Microbiome Ultra Nucleic Acid Isolation kit and associated consumables from Thermo Fisher Scientific USA (A42354). In brief, samples were suspended in lysis buffer and lysed using a prepared silicon bead tube at 4 m/s for 10 minutes. After centrifugation, the supernatant was transferred to a deep-well KingFisher sample plate, and Proteinase K was added to each sample. The plate was loaded into the KingFisher Apex instrument (Thermo Fisher Scientific USA) and run with the manufacturer’s MagMAX_Microbiome_Soil_Flex program. Extracted DNA was quantified with fluorometry (Qubit 2.0; Thermo Fisher Scientific USA).

Protein source detection using 16S mitochondrial rRNA gene amplicon sequencing

For protein source detection, amplification of the host 16S mitochondrial rRNA gene was performed using the following untagged primers: UnivMamF: 5’- GACGAGAAGACCCTATGGAGC-3’, UnivMamR: 5’- TCCGAGGTCGCCCCAACC-3⁶⁵. Amplification was performed using Platinum II Hot-Start Green PCR Master Mix (Invitrogen USA) with 1 µM final concentration of each primer in 50 µL reactions using the following cycling conditions: 94 °C for 2 minutes; 35 cycles of denature 94 °C 15 seconds, anneal 55 °C for 15 seconds, extend 68 °C for 30 seconds; final extension 72 °C for 10 minutes. Those samples with a band of 98–120 bp (visualized on a 2% agarose gel) were moved forward for sequencing. Purification was performed with Ampure XP beads (Beckman Coulter USA) with a 1.85X bead to sample ratio, followed by an abbreviated DNA Prep library preparation (Illumina USA), replacing the final cleanup with the Ampure XP cleanup as before. Cleaned, barcoded amplicons were adjusted to 4 nM, pooled, and sequenced with 2 × 100 bp Illumina MiSeq v3 chemistry. Controls included remnant diagnostic tissue from a rabbit and negative setup controls at the extraction and PCR stages.

Analysis of sequences was performed using Qiime2²⁰. First, sequences were denoised into amplicon sequence variants (ASVs) using the ‘dada2 denoised-paired’ plugin with trimming options set to 6 and truncating options to 70. Taxonomy predictions were assigned to the ASVs with plugin ‘feature-classifier classify-sklearn’ using a classifier based on complete mitochondrial genomes downloaded from NCBI. A reference classifier was adapted to our primer set using plugin ‘feature-classifier extract-reads’ options ‘--p-f-primer GACGAGAAGACCCTATGGAGC --p-r-primer TCCGAGGTCGCCCCAACC --p-min-length 70 --p-max-length 150’. Then those trimmed sequences were trained into a classifier using plugin ‘feature-classifier fit-classifier-naïve-bayes’. The final ASV matrix was exported using plugin ‘tools export’ (ASV table in biom format) followed by ‘biom convert’ from biom-format⁶⁶ v2.1.0 to create a tsv file. Taxonomy was downloaded from the online view (qiime2 view) of the qzv taxonomy file made with ‘metadata tabulate’ plugin. ASV tsv file and taxonomy tsv file downloaded were combined and investigated manually to detect protein sources. A major discrepancy was defined as an expected category (Mammalian, Avian, or Aquatic) entirely missing or an unexpected category identified. Figures and panels were made with R (script isolate_presence_vs_metadata.R) using ggplot2⁶⁷ and statistical tests with R base and ggpubr⁶⁸. Final figures were made using svglite⁶⁹, patchwork⁷⁰ and Inkscape (https://inkscape.org).

Sample preparation and bacterial culture

Using the US Food and Drug Administration’s Bacteriological Analytical Manual (BAM), food products were processed based on the Salmonella protocols (Supplementary Fig. S1)²⁴. Products were weighed into 25 g aliquots, added to 225 ml of 11% (m/v) in buffered peptone water (BD#218105), and blended for 2 minutes. Aside from aliquots saved for direct DNA extraction (or other follow up studies) and frozen neat at −80 °C, the homogenate was transferred to a sterile screw-cap jar and incubated 1 hour at room temperature with jar loosely closed. The sample jars were tightly closed, shaken, and then incubated overnight at 35 °C with a loose lid.

Jars were removed from the incubator and gently shaken again with the lid tightened. 100 µL of the homogenate was transferred to 10 mL of Rappaport-Vassiliadis (RV) medium (Becton Dickinson USA #218581), and 1 mL of the homogenate transferred to 10 mL of tetrathionate (TT) broth (Becton Dickinson USA #210430). Liquid cultures were incubated overnight at 42 °C. Liquid cultures were removed from the incubator, and each liquid culture was streaked onto both xylose lysine desoxycholate (XLD) agar (Becton Dickinson USA #221284) and Hektoen enteric (HE) agar (Becton Dickinson USA #221366). Plates were incubated overnight at 35 °C. Plates were examined the next day for growth, and all unique and distinct colonies were selected for isolation. Selected colonies were inoculated into fresh Luria-Bertani (LB) broth (Becton Dickinson USA #241420) and incubated overnight at 35 °C. After incubation, broth cultures were transferred to a 15% glycerol-LB solution and saved at −80 °C. Isolates were also grown for 24 hours on 5% sheep’s blood trypticase soy agar plates from BBL (Becton Dickinson USA #221261) and identified by matrix-assisted laser desorption/ionization (MALDI) mass spectrometry (Bruker Corporation USA) for the Gram-negative enrichments, or by Whole-genome sequencing (WGS) for the carbapenem and anaerobic isolates as described below.

Carbapenem-selective enrichment was attempted retrospectively on the food products harboring carbapenem resistance genes (based on applying annotations from the AMRFinder database⁷¹ to the Illumina Ampliseq AMR panel output). Frozen (−80 °C) food homogenates in BPW were thawed, and 100 µl was inoculated separately into 2.5 ml MacConkey broth tubes with a 10 µg ertapenem or meropenem disc for 17 hours at 37 °C. A loop was dipped into the broth culture and streaked on carbapenem-resistant Enterobacteriaceae (CRE) chromogenic agar (HardyCHROM, Hardy Diagnostics USA). Positive colonies were re-streaked for isolation on TSA + 5% sheep blood and cultured overnight at 35 °C. Antimicrobial susceptibility testing was performed on the Sensititre Companion Animal Gram Negative COMPGN1F Vet AST Plate (Thermo Fisher USA).

Anaerobic culture was attempted retrospectively on samples with more than 3.5% of their total reads in 16S rRNA gene amplicon data belonging to genus Clostridium sensu stricto 1. Homogenates were streaked on MacConkey agar (Becton Dickinson USA #212123) and brain heart infusion agar (BHIS, Becton Dickinson USA #211059) with 5 g/L yeast extract (Becton Dickinson USA #212750)) and incubated in an anaerobe container system (GasPak, Becton Dickinson USA #260678) at 42 °C for up to 48 hours. Individual colonies were inoculated into 3 mL brain heart infusion broth (supplemented with 5 g/L yeast extract as above) in sealed tubes and incubated at 37 °C overnight in the anaerobe container with three of the manufacturer’s carbon-based packets.

WGS and bioinformatics analyses

DNA extractions of bacterial isolates were processed for WGS (Supplementary Fig. S1). Genomic libraries were prepared and barcoded using the DNA Prep Kit (Illumina USA), then sequenced on the Illumina MiSeq platform using the MiSeq Reagent Kit v3 (Illumina USA) with 2 × 250 bp chemistry. Sequencing quality for each isolate was checked using an adapted standalone version of the GalaxyTrakr MicroRunQC pipeline⁷², which is based on de novo assembly of trimmed reads with SKESA v2.4.0⁷³ followed by computation of coverage, N50 and reads QC30 as well as multilocus sequence typing (MLST) using the PubMLST database schemes, to which we added ANI prediction with fastANI⁷⁴ v1.31 against a database from NCBI RefSeq sequences. WGS data were submitted to the NCBI Pathogen Detection database (https://www.ncbi.nlm.nih.gov/pathogens) under BioProject Accession PRJNA832331 for comparison with other genomes with Pathogen Detection¹⁹ and are globally accessible. AMR gene detection for bacterial isolates was performed using AMRFinderPlus v3.12.8^19,71. The tool Deeplasmid⁷⁵ vDecember_22_2022 was used to determine if contigs matching known AMR mechanisms were likely to belong to a plasmid or chromosomal sequence. Contigs of interest that were predicted to be plasmidic were further analyzed with plsdb API¹⁸. Finally, we carried out all-vs-all ANI analysis using fastANI⁷⁴ on the good-quality isolates and used NCBI Pathogen Detection to screen and remove duplicate clones (based on SNP distance) (Supplementary Data 3). Analyses were carried out in R⁷⁶ (v4.3.1) (script AMR_isolates.R) using dplyr⁷⁷, stringr⁷⁸, tidyr⁷⁹, tibble⁸⁰ and gtools⁸¹. The heatmap was made using ComplexHeatmap⁸² and svglite⁶⁹ R packages.

Bacterial 16S and parasite 18S rRNA gene microbiome profiles

Extracted food homogenate DNA was sequenced to target amplification of the V3 and V4 regions of the 16S rRNA gene⁸³ (Supplementary Fig. S1) with Illumina adapter tags (underlined) added to the amplification primers as follows: 16S rRNA gene amplicon PCR Forward Primer = 5’ TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGCCTACGGGNGGCWGCAG 16S rRNA gene amplicon PCR Reverse Primer = 5’ GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAGGACTACHVGGGTATCTAATCC. DNA fragments were amplified using the Platinum Hot Start PCR Master Mix (2X) kit (Invitrogen USA) with all primers at a final concentration of 0.2 μM. The PCR cycling condition was an initial denaturation at 95 °C for 5 minutes; 25 cycles of 95 °C for 30 seconds, 55 °C for 30 seconds, and 72 °C for 30 seconds; and a final extension at 72 °C for 5 minutes⁸⁴. Amplicons were cleaned with Ampure XP beads (Beckman Coulter USA), barcoded with Nextera UD Indexes (Illumina USA), then purified again. These libraries were normalized to 2 nM and sequenced using Illumina 2×300 bp chemistry. Half the samples were pooled and sequenced using a NextSeq Instrument, the other half using a MiSeq instrument (both manufactured by Illumina USA).

The V4 region of the 18S rRNA gene from DNA extracted from homogenates was also amplified to explore the relative risk of parasite exposures between different diets. PCR was carried out using Platinum II Hot Start (Invitrogen USA #14001013) with the following primers⁵³: EK-565F (GCAGTTAAAAAGCTCGTAGT), 18S-EUK-1134-R-UNonMet (TTTAAGTTTCAGCCTTGCG). The cycling conditions were: 94 °C for 3 min, 35 cycles of 94 °C for 1 min, 60 °C for 1 min, 72 °C for 1 min, and final extension of 10 min at 72 °C. Size selection (~550 bp) and PCR product cleanup were done with Ampure XP Beads (Beckman Coulter USA #A63881) at 1.8x concentration. Libraries were prepared with the Illumina DNA Prep Kit (Illumina USA #20060059) and sequenced with Illumina 2 × 300 bp chemistry (MiSEQ v3).

Reads produced with 16S rRNA gene sequencing were processed using QIIME2²⁰ version ‘qiime2-amplicon-2023.9-py38’ in a conda environment (detailed command lines in doCommands_16SrRNA_QiimePipelineCommandLines.sh file). First, for each run, ASV tables were created by denoising the reads trimmed of primers via cutadapt trim-paired command. Denoising happened with dada2 plugin with option ‘--p-min-fold-parent-over-abundance’ set to 4 and, options ‘--p-trunc-len-f’ and ‘--p-trunc-len-r’ set based on their own sequencing qualities, respectfully 0 and 235 for the MiSeq and 0 and 0 for the NextSeq (0 means no size selection). The option ‘p-trim-left’ was also used in both cases set to 0 in order to merge the denoising results of both runs into one ASV table for downstream analyses. ASVs were then filtered using the ‘filter-features-conditionally’ plugin to retain ASVs with at least 0.25% relative abundance in one sample as suggested by Reitmeier et al.⁸⁵. A phylogenetic tree using the ‘sepp’ plugin was done to place our filtered ASVs into a reference phylogenetic tree based on SILVA128 generated and provided by QIIME to later compute phylogenetic distance-based diversity metrics. Taxonomy was assigned to the ASVs using the ‘classify-sklearn’ plugin using a reference database adapted from SILVA138 (available from QIIME) using our primers via the plugin ‘extract-reads’. Other plugins like ‘tabulate’, ‘import’ or ‘export’ were used to navigate between the analyses, display plugin results along twith he overall data processing.

Diversity analyses were carried out in R⁷⁶ (v4.3.1 – script 16SrRNA_diversity_analyses.R) by importing QIIME2 outputs (relative abundance and sampling depth filtered ASV table, phylogenetic placement tree and taxonomy classifications) into a phyloseq⁸⁶ object using the package Qiime2R. ASVs were cleaned from Mitochondrial and Chloroplast sequences, and abundances were adjusted for batch effect using ComBat-seq⁸⁷ from the sva R package⁸⁸ v. 3.21. ASVs were then grouped at the genus level and plotted with packages microViz⁸⁹ (‘tax_select’, ’tax_agg’, ‘ps_get’, ’comp_barplot’) and ggplot2⁶⁷. Alpha diversity metrics were computed using ‘ps_calc_diversity’ and ‘ps_calc_richness’ from microViz⁸⁹ (Shannon and Pielou), ‘phyloseq_phylo_div’ from metagMisc⁹⁰ (FaithPD). Boxplots were made using ggplot2⁶⁷ and statistical significance computed using ‘wilcox_test’ from rstatix⁹¹ and displayed using ggpubr⁶⁸. The final figure was crafted using the patchwork⁷⁰ package. Beta diversity was computed on a rarefied matrix (using ‘phyloseq_mult_raref_avg’ from metagMisc) using ‘dist_calc’ and then permanova was conducted using ‘dist_permanova’ (both from microViz⁸⁹ R package). DAA on non-rarefied data was investigated using ‘differentialTest’ from R package corncob⁹². ISA was performed at the genus level on the rarefied ASV table using ‘multipatt’ from indicspecies⁹³. Barplots of Sensitivity and Specificity were produced with ggplot2⁶⁷ with data crafted with dplyr⁷⁷, tibble⁸⁰, and tidyr⁷⁹ R packages. The final figure was made with the patchwork⁷⁰ R package (script ISA-DA.R). Clostridium Sensu Stricto 1 ASVs were manually checked in the sepp tree made with Qiime (see above), and manual taxonomy annotation was chosen. Pie charts were made with ggplot2⁶⁷ (script CSS1.R).

Analysis of 18S rRNA gene amplicons was performed using Qiime2²⁰ version ‘qiime2-amplicon-2023.9-py38’ in a conda environment. First, sequences were denoised using plugin ‘dada2 denoised-paired’ with trimming options set to 0, truncating options adapted to the quality of each sequencing run individually and option --p-min-fold-parent-over-abundance set to 4. Then, denoised sequences (ASV) from each sequencing run were merged (‘feature-table merge’ plugin) to create only one ASV dataset for further analyses (116 samples and 25,969 ASVs). Next, a fixed filter of 10 was applied using the plugin ‘feature-table filter-features’ option –p-min-frequency (106 samples and 13,277 ASVs). Finally, two taxonomy predictions were assigned to the ASVs with plugin ‘feature-classifier classify-sklearn’ using classifiers from reference databases PR2 and Silva138Ref. Reference classifiers derived from the reference database sequences that were trimmed in accordance with our primer sets to fit our data better using plugin ‘feature-classifier extract-reads’ options ‘--p-f-primer GCAGTTAAAAAGCTCGTAGT --p-r-primer TTTAAGTTTCAGCCTTGCG --p-min-length 300 --p-max-length 650. Then those trimmed sequences were trained into a classifier using plugin ‘feature-classifier fit-classifier-naïve-bayes’. The final ASV matrix was exported using plugin ‘tools export’ (ASV table in biom format), followed by ‘biom convert’ from biom-format⁶⁶ v2.1.0 to create a tsv file. Taxonomy was downloaded from the online view (qiime2 view) of the qzv taxonomy file made with the’metadata tabulate’ plugin. The ASV tsv file and taxonomy tsv file downloaded were used manually to detect parasites of interest.

Deep amplicon sequencing for AMR gene (ARG) detection

All food homogenates were directly tested for the presence of ARGs using an 815-primer ARG detection panel. To amplify ARGs from the food homogenates, the AmpliSeq for Illumina AMR Panel was used with the AmpliSeq Library PLUS kit (Illumina USA). All samples were first diluted to 4 ng/µl and amplified with each of the two provided primer pools separately for 16 cycles with an annealing time of 4 minutes. Amplicons were then pooled, digested, and barcoded with Unique Dual indexes following the manufacturer’s instructions. The final amplification cleanup was performed with AMPure XP beads (Beckman Coulter USA). Libraries were normalized to 2 nM and sequenced using Illumina 2 × 150 bp chemistry on the MiSeq platform (three different runs). Alignment to the reference gene sequences provided with the kit was performed using the DNA Amplicon application in the Basespace platform (https://basespace.illumina.com/apps/), which outputs coverage.csv files (one per sample). For each sample, the maximum coverage found in a negative control from the same run was subtracted, and any resultant value below zero was set to zero. Then the coverage values were compiled into a matrix and transformed into presence/absence to remove amplification bias and only keep gene detection information. Antibiotic class annotation of the genes included in this panel was conducted primarily using AMRFinderPlus^19,71 (v2024-07-22.1), as well as the Comprehensive Antibiotic Resistance Database⁹⁴ (CARD) when necessary (file available in the GitLab repository). For genes annotated to confer resistance to multiple antibiotic classes, coverage metrics were duplicated during data analysis to reflect multi-drug resistance when necessary (analyses focusing on antibiotic class). Analyses were carried out in R⁷⁶ (v4.3.1) with the following packages: dplyr⁷⁷, stringr⁷⁸, tidyr⁷⁹, forcats⁹⁵, gtools⁸¹ and tibble⁸⁰ (script AMR_panel.R). AMR results were displayed with R packages ggplot2⁶⁷, patchwork⁷⁰ and ComplexHeatmap⁸². Statistical tests were carried out with ggpubr⁶⁸ and rstatix⁹¹. Figures were exported from R in a vectorial image using svglite⁶⁹.

Influenza A screening and RNAseq

RNA purification from food homogenates was performed using the MagMAX mirVana Total RNA Isolation Kit (Thermo Fisher USA). Real-time PCR was performed using the CDC Universal Influenza A⁹⁶ assay using FastVirus master mix on the QuantStudio 3 platform (both Thermo Fisher USA). Untargeted RNAseq on a subset of samples was performed by the Cornell Transcriptional Regulation and Expression Facility. Library preparation was performed using the NEB directional library preparation kit. Sequencing was performed with PE 2 × 150 bp read length on the NovaSeq 6000 instrument (Illumina USA) with 100 M reads per sample. Taxonomic analysis of RNAseq data was performed using the CZID²¹ Illumina mNGS Pipeline v8.3, NCBI Index Date: 2021-01-22.

Statistics and reproducibility

The Shapiro-Wilk test for normality was used to assess the normality of the distribution of the investigated parameters. Neither parameter (cultured bacteria count and AMR gene count) was distributed normally. Differences in medians were tested by the Wilcoxon Rank Sum Test. Pearson’s Chi-squared tests with Yates’ continuity correction were used to verify associations between growing isolates and groups of samples. One-way ANOVA tests were performed separately to evaluate the effect of protein source categories on the number of cultured bacteria species and ARG count. All analyses were performed using R Statistical Software⁷⁶ (v4.3.1), and the R packages cited throughout the methods were managed with renv⁹⁷ (versions in renv.lock file).

Author contributions

L.B.G. obtained funding. L.B.G. and G.R. designed the study with help from A.C.M. (cultures), Y.T.Y. (AMR panel), K.S. (cultures), and B.J.S. (sample selection). A.C.M. performed sample homogenization. J.M.U. and K.S. performed protein source lab work. Y.T.Y., G.R. and J.M.U. analyzed and interpreted the protein source results. A.C.M., K.S., L.B.G., B.W. performed cultures, glycerol stocks and whole-genome sequencing of isolates. R.F.G. performed MALDI-TOF isolate identifications. G.R. performed bioinformatics analyses on isolate genomic data and interpretation. A.C.M., Y.T.Y., L.B.G. performed ampliseq lab work. G.R., Y.T.Y., A.C.M., L.B.G. analyzed and interpreted ampliseq data. Y.G. performed parasite lab work. G.R., Y.G., L.B.G. analyzed and interpreted parasite data. Y.J., Z.C., K.S., L.B.G. performed Influenza lab work and interpreted the data. A.C.M., Y.T.Y., Y.G. drafted portions of the manuscript. G.R., L.B.G. wrote the final manuscript. G.R., L.B.G., A.C.M. uploaded data to NCBI. All authors read and approved the final manuscript.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Dr Sabina La Rosa and Dr Ophelia Bu.

Data availability

All sequences generated as part of this study are available on NCBI in the umbrella Project PRJNA1088075. The cultured bacterial isolate genomes are under this umbrella in BioProject PRJNA832331. Amplicon and RNAseq data are in BioProject PRJNA1129155. Specific format files compiling raw data required for analyses and summary tables can be found on the GitLab: https://gitlab.com/goodmanlab/cat-food page. All other source data can be obtained from Supplementary data 1–5:

File name: Supplementary Data 1

Description: Metadata for all samples tested

File name: Supplementary Data 2

Description: Protein source analysis results

File name: Supplementary Data 3

Description: Bacterial isolate genome quality metrics

File name: Supplementary Data 4

Description: AMR and mobility annotations for bacterial isolate genomes

File name: Supplementary Data 5

Description: Direct deep AMR amplicon sequencing results from food samples.

Code availability

All command lines used to generate QIIME data from raw sequences and R scripts written to analyze them are available on GitLab: https://gitlab.com/goodmanlab/cat-food.

Competing interests

The authors declare the following competing interests: L.B.G. held an expert witness contract with the US Food and Drug Administration at the time the experiments were performed. The authors declare no other competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-025-08756-8.