Quantitative Release Assessment of mcr-mediated Colistin-resistant Escherichia Coli from Japanese Pigs

Abstract Colistin is a critically important antibiotic for humans. The Japanese government withdrew colistin growth promoter and shifted therapeutic colistin to a second-choice drug for pigs in 2017. A quantitative release assessment of mcr-mediated colistin-resistant Escherichia coli (E. coli) in Japanese finisher pigs was conducted under the World Organisation for Animal Health (OIE) risk assessment framework. Input data included colistin resistance and mcr-1-5 test results for E. coli isolates in the Japan Veterinary Resistance Monitoring System (JVARM), postal survey results regarding indication disease occurrence and colistin use by swine veterinarians in 2017 and 2018, and colistin resistance and mcr monitoring experiments at four pig farms in 2017-2018. An individual-based model was developed to assess the risk: the proportion of Japanese finisher pigs with mcr-1-5-mediated colistin-resistant E. coli dominant in the gut on an arbitrary day. Before implementing risk management measures, the risk was estimated to be 5.5% (95% CI: 4.2%-10.1%). At 12 months after stopping colistin growth promoter, the proportion of pigs with plasmid-mediated colistin-resistant E. coli declined by 52.5% on the experiment farms (95% CI: 8.7%-80.8%). The probability of therapeutic colistin use at the occurrence of bacterial diarrhea declined from 37.3% (95% CI: 30.3%-42.5%) in 2017 to 31.4% (95% CI: 26.1%-36.9%), and that of edema disease declined from 55.0% (95% CI: 46.0%-63.7%) to 44.4% (95% CI: 36.9%-52.0%). After risk management implementation, the risk was estimated to have declined to 2.3% (95% CI: 1.8%-4.3%; 58.2% reduction). Scenario analyses showed that pen-level colistin treatment effectively reduces the risk from 5.5% to 4.7% (14.5% reduction), an effect similar to stoppage of therapeutic colistin (16.4% reduction to 4.6%).


Introduction
Colistin is a critically important antibiotic called an antibiotic of last resort 1) in light of the rapid global rise of multi-drug-resistant Enterobacteriaceae. Colistin sulfate, a polypeptide antibiotic, has been used in Japan since the 1950s for the treatment of gram-negative gastrointestinal infections and as a feed additive to promote healthy development in food-producing animals (cattle, swine, and chickens) 2,3) . In human medicine, the use of injection formulas, which had been suspended due to the frequent adverse effects such as renal dysfunction, was re-approved in 2015 in response to the global rise of multi-drug-resistant gram negative bacterial infections 3) .
Polymyxins (polymixin B and colistin) modify the lipopolysaccharide (LPS) of gram-negative bacteria by decreasing the negative charge of the lipid A moiety of LPS. Chromosomal colistin resistance is caused by the activation of two-component systems involving PhoP/PhoQ and PmrA/ PmrB via mutation, which results in the overexpression of LPS-modifying genes 4) . Prior to 2015, when a mobile colistinresistance gene, mcr-1, was reported in China 5) , this was the only known mechanism of colistin resistance. The mcr gene, which is harbored on a plasmid, can be transmitted between bacteria, which poses a significant threat to humans, as important Enterobacteriaceae pathogens such as multi-drugresistant Pseudomonas aeruginosa (MDRP), multi-drugresistant Acinetobacter (MDRA), and carbapenem-resistant Enterobacteriaceae (CRE) can acquire colistin resistance as well. Since the first discovery of mcr-1 in China, identification of different mcr genes has continued globally, and as of January 2020, mcr-1 to -10 have been reported [6][7][8][9][10] . In Japan, a high prevalence of mcr-1 (30.0%), -3 (8.3%), and -5 (28.3%) was reported among 120 isolates from diseased pigs 11) , and a low proportion (1.9%, 39/2052 isolates) of mcr-1 and the absence of mcr-2 was reported among healthy pigs 2) .
The Food Safety Commission of Japan (FSCJ) immediately conducted a qualitative risk assessment for colistin resistance after the discovery of mcr-1 5) , which determined the risks of release, exposure, and consequence to be medium, low, and high, respectively 3) . Based on these risk assessment results, reported in January 2017, the Ministry of Agriculture, Forestry, and Fisheries (MAFF) of Japan announced a stoppage of market sales of colistin growth promoter and shifted therapeutic colistin from a first-choice to secondchoice drug in December 2017. The actual withdrawal of colistin growth promoter from the market and the shift to second-choice drug took effect on July 1, 2018, and April 1, 2018, respectively.
The objectives of this study were to quantitatively assess the current risk of producing finisher pigs harboring mcrmediated colistin-resistant Escherichia coli (E. coli) at farms just before sending the animals to the slaughterhouse and estimate the effects of potential control measures, including those already implemented via the risk management measures instituted by the MAFF.

Framework of the Risk Assessment
This study employed an World Organisation for Animal Health (OIE) risk assessment framework 12) , which comprised a release (entry) assessment, exposure assessment, and consequence assessment. Release, in this case, is the use of colistin in pigs and selection of mcr-mediated colistin-resistant E. coli; exposure refers to a consumer ingesting mcr-mediated colistin-resistant E. coli due to consumption of pork derived from pigs administered colistin; and consequence refers to the effect of treatment failure when using colistin to treat an illness caused by mcr-mediated colistin-resistant bacteria, including those that obtained mcr genes via plasmids from mcr-mediated colistin-resistant E. coli. Among these steps, this study focused on the release assessment.
The risk was defined as the proportion on a given day of Japanese finisher pigs with mcr-mediated colistin-resistant E. coli dominating the gut, among all Japanese finisher pigs just before sending the animals to the slaughterhouse. Dominance in the gut by mcr-mediated colistin-resistant E. coli was defined as a concentration of mcr-mediated colistin-resistant E. coli in the gut higher than 10 5.08 CFU/g, following setting of the cut-off point as described in the Results section. Release was defined as both the use of colistin as a feed-additive growth promoter and therapeutic use of colistin, including metaphylaxis, mass medication of healthy animals when the disease of interest is present within the group/flock/herd 13) , at an occurrence of either edema disease or bacterial diarrhea during the weaning period.
Colistin resistance in E. coli was defined as a minimal inhibitory concentration (MIC) of ≥ 4 μg/mL, according to the European Committee on Antimicrobial Susceptibility Testing breakpoints for Enterobacteriaceae, MIC > 2 μg/mL (http://www.eucast.org/clinical_breakpoints/). In Japan, the presence of mcr-1-harboring E. coli with an MIC of 2 μg/ mL has been reported 2) , and these bacteria were considered susceptible to colistin in our study.
As of January 2019, when a risk assessment was conducted for 1,315 E. coli isolates collected between 2006 and 2015, the Japan Veterinary Antimicrobial Resistance Monitoring System (JVARM) of the National Veterinary Assay Laboratory (NVAL), MAFF of Japan, had tested for mcr-1 through mcr-5 among mcr-1 to -10. Of these, 59 isolates had an MIC ≥ 4 μg/mL, and 41 isolates (41/59, 69.5%) had either mcr-1, -3, or -5, suggesting the remaining 30.5% involved either chromosomal or other mcr-mediated resistance (no isolate had mcr-4). As our study defined plasmid-mediated colistin-resistant E. coli as those harboring mcr-1 to -5, the results may underestimate the actual risk for mcr-mediated colistin-resistant E. coli dominating the gut of pigs in Japan.

Data Collection
The colistin resistance test results and detection of mcr-1 through -5 in E. coli isolates collected between 2006 and 2015, in which mcr genes were detected throughout the period, were provided by the JVARM. In-depth discussions regarding the mechanism of selection of plasmid-mediated colistin-resistant E. coli were conducted with the NVAL, university researchers examining antimicrobial resistance, and field swine veterinarians to ensure the quality of the risk assessment model in terms of both scientific and field aspects.
Two postal surveys were conducted providing the structured questionnaires ( Table 1) to veterinarians belonging to the Japan Pig Veterinary Society in 2017 and 2018 in the same calendar period (November to December). The reason two surveys were conducted was to compare differences in the frequencies of edema disease and diarrhea in the weaning period and the probability of therapeutic use of colistin upon the occurrence of these diseases, between before and after the stoppage of feed-additive use of colistin as a growth promoter and the change in categorization of therapeutic colistin use from first to second choice in 2018. The representativeness of the responses was measured using Spearman's correlation test for the numbers of farrow-tofinisher and reproduction farms for which information was collected in the postal surveys and the numbers registered in the Statistical Survey on Livestock of Japan by prefectures as of February 2017 14) . The ethics of the questionnaire studies were assessed and approved for exemption from ethical examination on October 30, 2018, by Research Ethics Committee of the Rakuno Gakuen University.

Risk Assessment Model
An individual-based simulation model was developed using RStudio, version 1.1.456 (RStudio, Inc., Boston, MA, USA), to run in the statistics software R, version 3.5.1 15) . The default setting models the feeding situation as of 2017, before stoppage of feed-additive use of colistin as a growth promoter. In total, 1,000 pig farms were generated in the model, representing Japanese farrow-to-finisher and reproduction farms in terms of the number of sows (212 small   scale with 11-50 sows; 474 medium scale with 51-200 sows;  and 314 large scale with 201-600 sows) 14) . The numbers of reproduction and farrow-to-finisher farms in Japan as of 2017 February were 379, and 3,260, respectively; however, the output of the risk assessment is the proportion of finisher pigs with mcr-mediated colistin-resistant E. coli dominating the gut, and the risk can be correctly estimated. The number of sows in each of these 1,000 farms was randomly assigned by drawing from uniform distributions. In the model, all of the sows would give birth to 12 piglets, according to the expert opinions from swine medicine practitioners. All these piglets were monitored until finisher pigs. The model used probability distributions where necessary, and the types of distributions, parameters, and their sources are shown in Table 2 and Supplemental Table 9.
Out of 1,000 farms, pigs with mcr-harboring E. coli in the gut would be present in a proportion of farms, and the proportion of pigs with mcr-harboring E. coli dominating the gut in these farms was determined stochastically. In addition, 93% of pig farms administer feed-additive colistin growth promoter to weaning-period pigs, according to the above-mentioned questionnaire results. With or without the selection pressure of the growth promoter, bacterial diarrhea and/or edema disease would typically occur during 1 month of the weaning period with different probabilities between growth promoter-using and non-using farms, and veterinarians would use therapeutic colistin by adding it to the feed tank of the pigsty at a certain probability.
In the case of farm occurrence of bacterial diarrhea, the model ignored the death of pigs, and two scenarios (metaphylaxis using colistin, and no use of colistin) were considered (Fig. 1). Regardless of diarrhea disease status, in pigs with mcr-harboring E. coli exhibiting a colistin MIC ≥ 4 μg/mL, at any concentration of E. coli in the gut, the colistin-resistant E. coli will be selected and become dominant, and they will remain dominant at a given maintenance probability (default: 80%) until the time of harvesting. This maintenance is a function of the unknown fitness conferred on E. coli by mcr-harboring plasmid. In contrast, selection of plasmid-mediated colistin-resistant E. coli will not occur if the pigs do not have mcr-harboring E. coli in the gut. In the scenario in which therapeutic colistin is not used, regardless of disease status, the proportions of mcr-harboring colistinresistant E. coli in dominating and non-dominating pigs and those that do not harbor the E. coli in the gut follow the field situation at farms without intensive colistin selection pressure due to treatment, which will be explained in more detail below.
Regarding edema disease occurrence on a farm, all of the diseased animals die in the model, and again, two sce-narios (metaphylaxis or no use of colistin) were considered (Fig. 2). When metaphylaxis was used on a farm in which a proportion of pigs have mcr-harboring E. coli in the gut, all of the non-diseased pigs having mcr-harboring colistinresistant E. coli at any concentration of E. coli will exhibit dominance of colistin-resistant E. coli in the gut. According to the function of the unknown fitness conferred on E. coli by mcr-harboring plasmid, a proportion of pigs in which mcr-harboring colistin-resistant E. coli was selected will continue to have resistant E. coli dominant in the gut. In the scenario in which therapeutic colistin is not used, pigs in which mcr-harboring colistin-resistant E. coli does and does not dominate and those that do not have this E. coli in the gut among non-diseased pigs follow the field situation at farms without intensive colistin selection pressure due to treatment, identical to bacterial diarrhea cases.
In the model, bacterial diarrhea and edema disease occur on randomly selected farms, and within these farms, based on the steps explained in Figs. 1 and 2, the number of pigs with mcr-harboring colistin-resistant E. coli dominant in the gut will be calculated. Fig. 3 shows the Venn diagrams for the categories of swine farms based on the use of colistin growth

Attribute questions
Attributes of veterinarian Affiliation; association/academic society

Supervising farms
The number of farrow-to-finisher and reproduction farms supervising, by prefecture

Disease occurrence
Frequency of bacterial diarrhea during weaning period The number of farms falling into size categories based on the number of sows (≤50, 51-100, 101-200, 201-500, and ≥501) and frequency categories (almost no occurrence, once in 2-3 years, once in 7-12 months, once in 4-6 months, once in 2-3 months, and more than once a month) Proportion of weaning-period pigs having diarrhea at an occurrence Changes in the frequencies of weaning period diarrhea and edema disease (increased, no change, decreased, don't know)

Colistin use
Feed additive use of colistin as a growth promoter (only in the first questionnaire) Proportion of farms administering colistin-free feeds to weaning-period pigs in 2017 before stoppage Probability of using therapeutic colistin The probability of using therapeutic colistin at the occurrence of weaning-period diarrhea or edema disease Change in the probability of therapeutic colistin use (only in the second questionnaire) Change in the probability of therapeutic colistin use at the occurrence of weaning-period diarrhea or edema disease (increased, no change, decreased, don't know)

Colistin resistance cases
Probability of encountering an event of colistin resistance The probability of encountering an event in which colistin is not effective when used Change in the frequency of encountering an event of colistin resistance (only in the second questionnaire) Change in the frequency of encountering an event in which colistin is not effective when used (increased, no change, decreased, don't know) promoter (left and right panels), occurrences of bacterial diarrhea and edema disease (overlapped circles), and the use of therapeutic colistin (shaded and non-shaded areas within the circles). The areas A i and C i indicate the farms in which bacterial diarrhea and edema disease occur, respectively, and therapeutic colistin is used, where i = 1 represents farms at which colistin is used as a growth promoter; i = 2 represents non-colistin-feeding farms. The areas B i and D i are similar to A i and C i , and the difference is that therapeutic colistin is not used in these farms. On the farms in the areas E i and F i , both bacterial diarrhea and edema disease occur, and therapeutic colistin is used in the farms in areas E i , while not used in F i . The calculation is implemented in three separate farm-size categories, j (small, medium, and large) (Equation 1). A ijres to E ijres in Equation 1 indicate the number of pigs in which plasmid-mediated colistin-resistant E. coli dominate in the gut among the farm categories Ai to Ei in Fig. 3, and the total number of finisher pigs with mcr-mediated colistinresistant E. coli dominating in the gut among the 1,000 farms is denoted as N mcr . As shown in Fig. 3, both bacterial diarrhea 57.8%* 52.5%* (54.8%*)   1). Of all pigs born on the 1,000 farms, pigs with edema disease die, and the total number of pigs slaughtered, not including the number of pigs with edema disease, was calculated (T slaughtered in Equation 1). The risk of Japanese finisher pigs with mcr-mediated colistin-resistant E. coli dominating in the gut among all Japanese finisher pigs just prior to sending the animals to the slaughterhouse, on an arbitrary day, was calculated using Equation 1. The model was run for 5,000 iterations using the for-loop in R software. A sensitivity analysis was performed to ascertain the unknown probability of maintaining the dominance of mcr-1-5-mediated colistin-resistant E. coli in the gut of a pig after selection associated with therapeutic colistin use for the options of 40, 60, and 100% maintenance (default 80%). In the following sections, estimations of probability distributions of the variables used in the model are explained.

Estimation of the Proportion of Pigs with mcr-1-5-harboring Colistin-resistant E. Coli Dominant in the Gut on a Farm that Used Colistin as a Growth-promoting Feed Additive
If mcr-1-5-harboring colistin-resistant E. coli cultured from swine feces formed colonies that could be picked up on a non-selective bacterial agar, it was defined as being dominant in the gut. To determine the bacterial concentration in this situation, an experiment was carried out in early 2017 at four swine farms where colistin was used as a growth promoter (but not for therapeutic purposes) and there were pigs present with mcr-harboring E. coli. In the four farms, treatment using colistin had never been done before and during the experiment. Feces were sampled from 30 pigs, and three colonies of E. coli cultured on non-selective deoxycholate hydrogen sulfide lactose (DHL) agar were puri-   fied; isolates exhibiting an MIC ≥ 4 μg/mL were classified as colistin-resistant. When at least one E. coli colony was colistin-resistant, the pig of origin was classified as having colistin-resistant E. coli dominant in the gut (one qualitative result for each pig). The same samples were cultured using colistin-supplemented CHROMagar TM COL-APSE (CHRO-Magar, Paris, France), and the E. coli concentration in feces was determined from the number of suspected colonies on the agar based on the color of the colonies (one quantitative result for each pig). Using the test results of 22 weaningperiod or fattening pigs examined (colistin-resistant E. coli grew on colistin-supplemented CHROMagar for all of the 22 samples; the results of other suckling pigs and sows were not used in this analysis), receiver operating characteristic (ROC) analysis was performed to determine the cut-off value of the bacterial concentration to best differentiate between dominance of colistin-resistant E. coli in the gut or not, by maximizing both sensitivity and specificity using the ROCR package 16) in R software, version 3.5.1 15) .
Under the uniform distribution, 100 pairs of non-selective DHL and CHROMagar results were randomly sampled from the results of the 22 pigs in 2017, and the proportion of samples exceeding the cut-off value was calculated. This process was iterated 5,000 times, and a beta distribution was fit to the simulated values to solve the parameters using the fitdist() function of the fitdistrplus package 17) . This provided the probability, P dom_gp , that mcr-1-5-harboring E. coli would dominate in the gut of a pig on a farm that fed colistin as a growth promoter but did not use therapeutic colistin. The proportion of pigs with colistin-resistant E. coli dominance in the gut after therapeutic colistin use (P selected_gp ) on a farm feeding colistin as a growth promoter was modeled using the beta distribution with the parameters specified by the number of E. coli samples that grew on colistin-supplemented CHROMagar at any bacteria concentrate, 22 of 22 samples ( Table 2).

Estimation of the Proportion of Pigs with mcr-1-5-harboring Colistin-resistant E. Coli Dominant in the Gut on a Farm That Did Not Use Colistin as a Growth Promoter Feed Additive
As it was difficult to find farms not using colistin as a growth promoter feed additive, four farms that participated in the experiment described in section 2.4 and stopped use of colistin as a growth promoter immediately after the sampling in 2017 were studied again 12 months later. In both experiments in 2017 and 2018, 30 pigs each (in total 60 pigs) were used, and three E. coli isolates isolated from each feces sample (90 isolates in each year) cultured on DHL agar were tested for colistin resistance and mcr-1-8. The 1-year reduc-tion rate in the animal-level prevalence of mcr-mediated colistin resistance (Red mcr ) was calculated using Equation 2. where P mcr2017 represents the proportion of E. coli isolates that were colistin-resistant and had any of mcr-1 to -8 in the 2017 experiment, and P mcr2018 represents that proportion in the 2018 experiment (all the colistin-resistant E. coli isolates had at least one of mcr-1 to -8). In the simulation model, a point estimate of Red mcr was used, but for the purpose of presentation of the effect of stoppage, it was simulated stochastically separately using beta distributions. The reason we tested for mcr-1 to -8 rather than mcr-1 to -5 was that the objective of this experiment was different, and the results will be published elsewhere. It was assumed that the animallevel prevalence of mcr-1-5-harboring colistin-resistant E. coli dominating in the gut of pigs on a farm that never used colistin as a growth promoter would be similar to that observed 12 months after stoppage, as there was no actual relevant information available in Japan. The probability that mcr-1-5-harboring E. coli dominates in the gut of a pig on a farm that never fed growth promoter colistin or stopped feeding growth promoter colistin 12 months previously and had not used therapeutic colistin, P dom_nogp , was modeled by multiplying P dom_gp and a complement of Red mcr to 1 ( Table 2). The proportion of pigs with colistin-resistant E. coli dominance in the gut after therapeutic colistin use on a farm that never fed growth promoter colistin or stopped feeding growth promoter colistin 12 months previously but did use therapeutic colistin (P selected_nogp ) was modeled by multiplying P selected_gp and a complement of Red mcr to 1 ( Table 2).

Estimation of the True Proportion of Japanese Farrow-to-finisher and Reproduction Swine Farms Having Pigs with mcr-harboring Colistin-resistant E. Coli in the Gut
Our study relies on the diagnosis of colistin resistance in E. coli by the JVARM, which collected only one sample from a farm; however, as described in the previous section, a proportion of negative samples might be collected from swine farms actually having pigs with mcr-1-5-harboring colistinresistant E. coli in the gut. For this reason, the true proportion of Japanese farrow-to-finisher and reproduction swine farms having pigs with mcr-1-5-harboring colistin-resistant E. coli in the gut (P TPF ) was estimated. The probability that colistinresistant E. coli harboring mcr-1-5 will be isolated from one sample of feces from a finisher swine just before harvesting on a farm in the sampling of the JVARM program, P JVARM , can be calculated as the product of P TPF and P dom_gp (Fig. 4). Therefore, P TPF is calculated using Equation 3. A previous report described the isolation of mcr-1-5-harboring E. coli exhibiting an MIC of 2 μg/mL 2 . In addition to the true farm-level prevalence of mcr-1-5-harboring colistin-resistant E. coli, the true prevalence of farms with pigs having mcr-harboring E. coli including susceptible ones (an MIC of 2 μg/mL) was estimated using Equation 3, but in this case, P JVARM indicated the proportion of fecal samples having E. coli isolates harboring mcr-1-5 including susceptible isolates.

Estimation of 1-month Incidence Rates of Edema Disease and Bacterial Diarrhea Among Weaning Pigs at the Farm Level
Therapeutic colistin is used at the occurrence of edema disease or bacterial diarrhea, particularly during the 1-month weaning period. The incidence rates for edema disease and bacterial diarrhea at the farm level (IR dis k ) were estimated separately and among different farm size categories (k) using results of the 2017 questionnaire for farms feeding colistin as a growth promoter and those of the 2018 questionnaire for farms not feeding colistin.
In the questionnaires, the number of farms falling into several categories of disease frequency (l) and size (k) were asked ( Table 1). The number of farms in each category based on the veterinarian responses was summed to n lk . In modeling, for each farm size category (k), a set of n lk disease events within a 1-month period on farm m in disease frequency category (l) was drawn from a Poisson distribution with a rate parameter: the reciprocal of the between-disease-events period, which was drawn from a uniform distribution between a l and b l (e.g., 2 and 3 for the period category 2 to 3 months), was summed to calculate the total number of disease events occurring within a 1-month period in frequency category l. The number of disease events in the five disease frequency categories were summed and divided by the total number of farms in farm size category k (T Farm k ) to obtain the 1-month incidence rate (IR dis k ) for that farm size category (Equation 4). The calculation of IR dis k was iterated 5000 times, and a beta distribution was fit to the values using matching moment estimation in the fitdist() function of the R fitdis-

Estimation of the Proportion of Weaningperiod Pigs Affected by Edema Disease Or Bacterial Diarrhea within the Farm during An Occurrence
According to interviews with field swine veterinarians, at an occurrence of edema disease, almost 100% of the diseased pigs die, and in our model, these dead pigs must be excluded from the swine population. Moreover, in considering the mode of metaphylaxis (treating the entire herd or only affected pens), it is important to know the proportion of diseased weaning-period pigs at the occurrence of edema disease and bacterial diarrhea.
In the questionnaires, for edema disease and bacterial diarrhea and farm size categories separately, respondents were asked to allocate (to a total of 100%) weaning-period pigs affected by the disease into proportion categories, based on their clinical experience in 2017 and 2018 ( Table 1). To estimate the proportion of affected pigs among weaningperiod pigs on a farm, a proportion category for pigs affected was first selected, based on the weight given by the averaged percentage allocations of the categories using the sample() function in R. The random proportion was then assigned by drawing from a uniform distribution (c, d), where c represents the smaller range and d the larger range of the proportion category (e.g., for the 10.1-30% category, c = 10.1% and d = 30%). This process was iterated 5,000 times, and a beta distribution was fit to the sampled results using the fitdist() function in R, for 2017 and 2018 and different farm size categories. The weight matrixes used in these simulations are shown in Supplemental Tables 5 through 8. The distributions fit using 2017 data represented the situation in which colistin was used as a growth promoter, as 93% of farms were feeding colistin, whereas those fit using 2018 data represented the situation in which colistin was not used as a growth promoter.

Use of Colistin at the Occurrence of Edema Disease or Bacterial Diarrhea
In the questionnaires used in 2017 and 2018, respondents were asked to make point estimates of the probability of therapeutic use of colistin at the occurrence of edema disease or bacterial diarrhea ( Table 1). For respective years, a set of 100 random samples from the pool of responses was drawn, and the mean was calculated. This process was iter-ated 5,000 times, and a beta distribution was fit to the means using the fitdist() function in R to calculate the probability of therapeutic colistin use given the indication disease occurred (P use|dis ).

Assessment of the Effects of Stoppage of Growth Promoter Colistin Use and Shift of Colistin to a Second-choice Drug on the Occurrence of Indication Diseases and Frequency of Therapeutic Colistin Use
To compare the incidence rates of bacterial diarrhea and edema disease between 2017 and 2018, 50 samples each were drawn from the probability distributions of incidence rates for both years and logit transformed and compared using Welch's t-test for both diseases. The sample size, 50, was determined by calculating the minimum sample size for a comparison of two means, so that the size exceeded the requirement for all farm size categories.
Even after shifting colistin from first-choice drug to second, if the frequency of indication diseases was increased, the frequency of therapeutic colistin use may not decline. Therefore, for bacterial diarrhea and edema disease, respectively, the probability of therapeutic colistin use in a given 1-month period on a farm of size category k (P use k ) was calculated using Monte Carlo simulation of 5,000 iterations by multiplying the samples drawn from the posterior distributions of the 1-month incidence rate of disease (IR dis k ) and probability of therapeutic colistin use at the occurrence (P use|dis ) (Equation 5). A set of 50 values was sampled from the posterior probability distributions of therapeutic colistin use, P use k , in a given 1-month period for 2017 and 2018, respectively, and logit transformed and compared using Welch's t-test.

Scenario Analyses
Scenarios prepared for assessing potential intervention options included reduction of bacterial diarrhea and edema disease cases (50% and 80% reduction, respectively), reduction of the probability of therapeutic colistin use (50% and 80% reduction, respectively), reduction of the number of target pigs by pen-unit colistin use (20% of all weaning-period pigs therapy). For pen-unit use, the proportion 20% of all weaning pigs was chosen based on the proportion of diseased pigs at the occurrence of bacterial diarrhea or edema disease. The proportion of pigs with mcr-1-5-mediated colistin-resistant E. coli dominant in the gut after pen-unit therapy using colistin was calculated for farms where colistin as a growth promoter feed additive was used (P selected_ pen_gp ) and for farms where it was not used (P selected_ pen_nogp ), using Equations 6 and 7, respectively.

Equation 7
The primary purpose of this risk assessment was to characterize the risk of mcr-1-5-mediated colistin-resistant E. coli during a period of time when a majority of swine farmers were using colistin as a growth promoter. In addition, the risk at 12 months after stoppage of growth promoter colistin use and the shift of colistin to a second-choice therapeutic drug was assessed using the questionnaire survey results for 2018 on disease occurrence and therapeutic colistin use.  Table 2 shows the estimation results for variables associated with the within-and between-farm prevalence of mcr-1-5-mediated colistin-resistant E. coli. For the within-farm prevalence, the mean proportion of non-colistin-treated pigs with mcr-1-5-harboring colistin-resistant E. coli dominant in the gut in growth promoter colistin feeding farms (P dom_gp , 31.0% [95% credible interval, CI: 18.0%-45.6%, median 30.6%], Table 2) was estimated based on the dominance cutoff threshold of 10 5.08 CFU/g, with an accuracy score of 0.77, sensitivity 55.6%, and specificity 92.3%, determined using ROC curve analysis (Fig. 6). In contrast, colistin-resistant E. coli was cultured from all 22 samples collected at 4 farms where mcr-harboring E. coli was detected in the range of 10 3 to 1.12 × 10 8 CFU/g on colistin-supplemented CHRO-Magar, and the probability of selecting resistant E. coli after therapeutic colistin use, in other words, the proportion of mcr-mediated colistin-resistant E. coli dominant pigs when therapeutic colistin was used, in growth promoter feeding farms (P selected_gp ), was estimated to be 95.9% (95% CI: 85.2%-99.9%, median 97.0%, Table 2). In growth promoter colistin non-feeding farms, the proportion of mcr-mediated colistin-resistant E. coli dominant pigs was much lower: 13.1% (P dom_nogp , 95% CI: 7.6%-19.2%, median 12.9%) when therapeutic colistin was not used, and 40.5% (P selected_nogp , 95% CI: 36.0%-42.2%, median 40.9%) when it was used ( Table 2).

Risk Estimation as of 2017
The mean proportion of Japanese finisher pigs with mcr-1-5-mediated colistin-resistant E. coli dominating in the gut just prior to sending the animals to the slaughterhouse was estimated to be 5.5% (95% CI: 4.2%-10.1%; median 5.2%, Fig. 7a, Table 3) as of 2017, when colistin was fed to pigs as a growth promoter on 93% of farms, according to the results of the questionnaire survey. The risk was sensitive to the unknown probability of maintenance of colistin resistance in E. coli after selection due to therapeutic colistin use; a change in the probability of maintenance from 80% to 20% resulted in a 20.0% change ([5.5%-4.4%]/5.5%) in the mean overall risk ( Table 3).

Effects of Stoppage of Growth Promoter Colistin Use and Shift of Colistin to a Secondchoice Drug for Treatment
In the farm experiment, the proportion of mcr-mediated colistin-resistant E. coli among all E. coli isolates declined from 17.8% (16/90) in 2017 to 6.7% (6/90) in 2018. At the animal level, the mean reduction rate, Red mcr , was estimated to be 52.5% (95% CI: 8.7%-80.8%, median 54.8%, Table 2). Table 4 shows the change between 2017 and 2018 in the 1-month incidence rates of bacterial diarrhea and edema disease in the weaning period (IR dis k ). For both diseases, the overall rate increased significantly, and this change was due to the increased disease events on small-and medium-scale farms (P = 0.02 for small-scale farms, otherwise P < 0.01, Table 4). In contrast, the incidence rates for both diseases decreased significantly on large-scale farms (P < 0.01). The incidence rate was the lowest on small-scale farms and highest on large-scale farms in both years and for both diseases.
The mean probability of colistin use at the occurrence   Table 5 shows comparisons of the probability of therapeutic colistin use in a given 1-month period, P use k . On smallscale farms, P use did not differ for bacterial diarrhea between 2017 and 2018 but increased significantly on medium-scale farms (P < 0.01) and decreased significantly on large-scale farms (P < 0.01) in 2018. P use for edema disease increased significantly on small-and medium-scale farms in 2018 (P < 0.01, respectively) but decreased significantly on large-scale farms (P < 0.01). Table 6 shows a comparison of the risks estimated between  Table 3. Proportion of Japanese finisher pigs just prior to being sent to slaughterhouses with mcr-1-5-mediated colistin-resistant E. coli dominant in the gut as of 2017, at different probabilities of maintenance of resistance after selection associated with therapeutic colistin use (mean, median, and 95% credible interval)  Fig. 7b). However, the animal-level reduction rate of mcr-mediated colistinresistant E. coli in previously colistin growth promoter feeding farms (Red mcr ) had wide credible interval, and the overall risk in 2018 was estimated to be 1.0% (95% CI: 0.8%-1.8%, median = 0.9%) and 4.8% (95% CI: 3.7%-8.8%, median = 4.5%), when Red mcr took 80.8%, and 8.7%, respectively. Table 7 shows the change in the proportion of finisher pigs with mcr-1-5-mediated colistin-resistant E. coli dominant in the gut by several intervention options using the 2017 model. Compared with the default scenario, which was estimation of the risk in 2017, the risk did not decline with the reduction in edema disease occurrence at the farm level. In contrast, an 80% reduction in the occurrence of bacterial diarrhea at the farm level reduced the overall mean risk by 9% ([5.5%-5.0%]/5.5%), and the reduction was greatest on large-scale farms (12% reduction [5.8%-5.1%]/5.8%). A decrease in the probability of therapeutic colistin use exhibited an even greater reduction; an 80% reduction in colistin use reduced the risk by 12.7% ([5.5%-4.8%]/5.5%). When the probability of therapeutic colistin use was not changed but pen-unit therapy was applied, an even greater reduction in risk was observed (14.5% reduction, [5.5%-4.7%]/5.5%), which exhibited an effect similar to stoppage of therapeutic colistin use (16.4% reduction to 4.6%). The reduction rate was greatest on large-scale farms, whereas the risk did not change on small-scale farms for all intervention options.

Probability of maintenance of resistance
The distributions of 1-month incidence rates, proportion of weaning pigs affected at occurrence of indication diseases, and probability of therapeutic colistin use that were used in the simulations are provided in Supplemental Table 9.

Discussion
This study used an individual-based model for quantitative release assessment of the selection of mcr-1-5-mediated colistin-resistant E. coli in Japanese pigs just before slaughtering associated with growth promoting and therapeutic uses of colistin. To the best of our knowledge, this is the first study in the world to have taken this approach.
The mean proportion of pigs with mcr-1-5-mediated colistin-resistant E. coli dominating in the gut just before slaughtering was estimated at 5.5% as of 2017, and mcr genes were assessed as being widely spread in Japan: approximately one-fourth (23.7%) of reproduction and farrow-to-finisher swine farms, including those that did not use colistin, were estimated to have pigs with mcr-harboring E. coli.
In this assessment, parameters of probability distributions were determined based on JVARM data, questionnaire surveys, and farm experimental data, and were not solved using observed JVARM data by fitting approaches such as maximum-likelihood estimate, Markov-Chain Monte Carlo simulation, or approximate Bayesian computation 18) . Additional validation process may be needed for the model in future. However, the estimated risk was within the 95% CI of the proportion of positive samples for mcr-1-5-mediated colistin-resistant E. coli in the JVARM results ( Table 2); thus, the model assumption is plausible. Moreover, the purpose of the assessment included evaluating potential intervention programs, which this study achieved. However, the model has several limitations: (1) already reported mcr-6-9 [6][7][8][9][10] , and chromosomal-associated colistin resistance 4) were not considered; (2) information on edema disease and bacterial diarrhea was based on questionnaire surveys, and actual clinical records were not used; (3) the probability of maintenance of colistin resistance after selection remains unknown; (4) transmission of mcr-harboring E. coli or transmission of plasmid-harbored mcr genes between pigs, between pens, and between farms was not modeled; and (5) detailed within-farm hygiene practices were not modeled.
Regarding limitation (1) above, the actual risk associated with mcr is higher for the unknown proportion of mcr-6 to -10 that can cause colistin resistance in E. coli, and our assessment underestimated this risk. Chromosomal colistinresistant E. coli does not transmit resistance to other bacteria and was therefore outside the scope of this study. However, future completion of testing for mcr-6 to -10 or the potential discovery of other novel mcr genes using JVARM E. coli isolates would enable re-evaluation of the mcr risk and even the risk associated with chromosomal colistin-resistant E. coli using our simulation model, as our model is designed to select colistin-resistant E. coli regardless of the type of resistance, whether plasmid mediated or chromosomal.
Regarding limitation (2), in addition to a lack of accurate information from clinical records, questions in the postal questionnaire survey of 2017 related to bacterial diarrhea were phrased to refer to "weaning-period diarrhea". Some veterinarians suggested that the questions should have referred to "bacterial diarrhea", as the focus of the study was colistin-resistant E. coli. In the questionnaire provided in 2018, 16 of 28 respondents who participated in the 2017 survey responded in 2018 as well. A half of the respondents (50.0%, 8/16) answered about bacterial diarrhea, and one respondent (6.3%, 1/16) included diarrhea of a cause other than bacterial in 2017, whereas seven (43.8%, 7/16) could not remember (results not shown). However, considering the increase in 1-month incidence among small-and mediumscale farms in 2018, it is unlikely that the incidence rate of bacterial diarrhea in 2017 was substantially overestimated. Moreover, even if our estimates of incidence rates were accurate, the change in incidence rate might have been due to factors other than risk management, such as pure variability (e.g. purely random variation of disease occurrence). Analysis of the maintenance rate of colistin resistance after selection due to therapeutic use of colistin showed moderate sensitivity. In the United Kingdom, an outbreak of mcr-harboring colistin-resistant E. coli has been reported only on a pig farm, and by stopping therapeutic colistin use, mcr was eliminated from the farm after 20 months 19) . In Spain, by reducing therapeutic colistin use, the proportion of positive samples for colistin-resistant Salmonella in swine feces declined from 60% in 2015 to 35% in 2017, and that for mcr-1 in feces also declined, from 70% in 2015 to 53% in 2017 20) . According to our farm experiment estimate, by stopping the use of colistin as a growth promoter in feed, 52.3% of pigs with mcr-mediated colistin-resistant E. coli dominance in the gut would lose colistin-resistant E. coli in 12 months.
Biologically, both the transmission and maintenance of mcr genes are affected by the type and size of the host plasmid 18) . Therefore, our risk estimate of the post-risk management situation in 2018 is sensitive to variability in the characteristics of plasmids harboring mcr genes, which was not considered in the simulation model. To understand the dynamics of within-farm clearance of mcr genes, the relationship between the full genome sequence of mcr-harboring plasmids and the speed of clearance should be studied, and mathematical modeling could be suitable for this purpose, as it has been applied to model transmission elsewhere 21) .
Scenario analyses provided several clear insights. First, stoppage of colistin use as a growth promoter may be the most effective means of reducing the risk of producing pigs with mcr-mediated colistin-resistant E. coli dominant in the gut. Second, controlling bacterial diarrhea and reducing therapeutic colistin use have instantaneous effects on risk reduction, although the degree of reduction is not particularly high when compared with stoppage of colistin use as a growth promoter. Comparing the results of the questionnaire surveys for 2017 and 2018 showed reductions in both the incidence of bacterial diarrhea on large-scale farms and therapeutic colistin use in 2018. Pig veterinary clinicians appeared to respond well to the change by the implementation of risk management. Third, limited use of therapeutic colistin for affected pens was more effective than reducing the therapeutic use of colistin in entire weaning pig herds by 80%. According to the interviews with pig veterinary clinicians, metaphylaxis involving colistin administration via feed tanks was the most common mode, and the default model takes this option. As mcr genes pose health risks in humans, selective and pludent colistin use would reduce these risks in Japan more rapidly.
This study involved only release assessments at pig farms. The qualitative risk assessment conducted by FSCJ described the risk pathways for transmission of mcr genes to MDRP, MDRA, and CRE in the human gut via foods contaminated with mcr-harboring bacteria 22) . Colistin is the first choice for treating infections with MDRP, MDRA, or CRE, but it will not work if these pathogens have obtained mcr genes. More detailed experiment-based information related to the transmission of mcr genes between bacteria within the human gut and the associated clinical consequences is needed. In the future, it would be worthwhile to conduct a complete quantitative risk assessment of colistin resistance.
In conclusion, the mean probability of releasing pigs with mcr-1-5-mediated colistin-resistant E. coli dominant in the gut to slaughterhouses in Japan was estimated to be 5.5% in 2017 and 2.3% in 2018, after stoppage of use of colistin as a growth promoter and shifting therapeutic colistin to second-choice drug. Scenario analyses confirmed that these risk management options were well targeted. Pen-unit treatment and reduction of bacterial diarrhea via hygiene improvements, including the use of E. coli vaccines 23) , would further reduce the risk. Monitoring of mcr-mediated colistin-resistant bacteria in pigs should be continued, and whole-genome sequencing of mcr-harboring plasmids would provide more-accurate knowledge that could be used to further reduce the risk of mcr-mediated colistin-resistant bacteria in Japan. Supplemental Tables  Supplemental Table S5. Averaged percentage allocations on the proportion of weaning period pigs diseased at the occurrence of bacterial diarrhea in clinical cases as of 2017