Predicting clinical outcome ofEscherichia coliO157:H7 infections using explainable Machine Learning

Predicting clinical outcome of Escherichia coli O157:H7 infections using explainable Machine Learning | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Predicting clinical outcome of Escherichia coli O157:H7 infections using explainable Machine Learning Julian A. Paganini , Suniya Khatun , Sean McAteer , Lauren Cowley , David R. Greig , David L. Gally , Claire Jenkins , Timothy J. Dallman doi: https://doi.org/10.1101/2025.06.05.25329036 Julian A. Paganini 5 Faculty of Veterinary Medicine, Institute for Risk Assessment Sciences (IRAS), Utrecht University , 3584 CL Utrecht, the Netherlands Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: j.a.paganini{at}uu.nl Suniya Khatun 1 Institute of Structural and Molecular Biology, Division of Biosciences, University College London , London, WC1E 6BT, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Sean McAteer 4 Division of Bacteriology, The Roslin Institute and Royal (Dick) School of Veterinary Studies, University of Edinburgh , UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Lauren Cowley 2 University of Bath, Biology and Biochemistry , Bath, BA2 7AY, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site David R. Greig 3 United Kingdom Health Security Agency , 61 Colindale Avenue, London, NW9 5EQ, United Kingdom 4 Division of Bacteriology, The Roslin Institute and Royal (Dick) School of Veterinary Studies, University of Edinburgh , UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site David L. Gally 4 Division of Bacteriology, The Roslin Institute and Royal (Dick) School of Veterinary Studies, University of Edinburgh , UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Claire Jenkins 3 United Kingdom Health Security Agency , 61 Colindale Avenue, London, NW9 5EQ, United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Timothy J. Dallman 5 Faculty of Veterinary Medicine, Institute for Risk Assessment Sciences (IRAS), Utrecht University , 3584 CL Utrecht, the Netherlands Find this author on Google Scholar Find this author on PubMed Search for this author on this site Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF ABSTRACT Background Shiga toxin-producing Escherichia coli (STEC) O157:H7 is a globally dispersed zoonotic pathogen capable of causing severe disease outcomes, including bloody diarrhoea and haemolytic uraemic syndrome. While variations in Shiga toxin subtype are well-recognised drivers of disease severity, many unexplained differences remain among strains carrying the same toxin profile. Results We applied explainable machine learning approaches—Random Forest and Extreme Gradient Boosting—to whole-genome sequencing data from 1,030 STEC O157:H7 isolates to predict patient clinical outcomes, using data collected over two years of routine surveillance in England. A phylogeny-informed cross-validation strategy was implemented to account for population structure and avoid data leakage, ensuring robust model generalizability. Extreme Gradient Boosting outperformed Random Forest in predicting minority classes and correctly predicted high-risk isolates in traditionally low-risk lineages, illustrating its utility for capturing complex genomic signatures beyond known virulence genes. Feature importance analyses highlighted phage-encoded elements, including potentially novel intergenic regulators, alongside established virulence factors. Moreover, key genomic regions linked to small RNAs and stress-response pathways were enriched in isolates causing severe disease. Conclusions These findings underscore the capacity of explainable ML to refine risk assessments, offering a valuable tool for early detection of high-risk STEC O157:H7 and guiding targeted public health interventions. BACKGROUND Shiga toxin-producing Escherichia coli (STEC) serotype O157:H7 emerged as a significant public health concern in the 1980s with outbreaks associated with gastrointestinal symptoms that ranged from mild to severe bloody diarrhoea( 1 ). Approximately 6% of affected individuals develop haemolytic uraemic syndrome (HUS)( 2 , 3 ), a potentially fatal condition mainly affecting children and the elderly. Antibiotics are contraindicated for STEC infection because Shiga toxins (Stx) are released in response to DNA damage and SOS response. Clinical management involves rehydration therapy and palliative care to reduce renal, cardiac and neurological complications. In the UK, ruminants, mainly cattle and sheep, have been identified as the zoonotic reservoir for STEC. Epidemiological and environmental investigations of outbreaks in England have shown transmission of STEC from ruminants to humans to occur mainly through direct or indirect contact with animals or with their contaminated environments, consumption of contaminated foods that are unwashed or undercooked and person-person contact of infected individuals( 4 , 5 ). The key virulence factor in STEC O157:H7 - which also defines STEC- is Stx( 6 ). Stx is an AB5 toxin, consisting of a pentamer of B subunits non-covalently bound to an enzymatically active A subunit. Upon entering the host cell, the A subunit catalyzes the cleavage of ribosomal RNA (rRNA), leading to ribosome inactivation and inhibition of protein synthesis. Consequently, cells undergo programmed cell death as the apoptosis signalling pathway is activated( 7 ). Clinical symptoms observed during STEC infection are instigated by Stx causing local damage to the colon and renal endothelial cells, disrupting the microvascular system via direct toxicity and through induction of local cytokine production to cause renal inflammation( 8 ). There are two main subgroups of Stx, encoded by the stx1 and stx2 genes, each comprising several subtypes with certain subtypes, specifically stx2a and stx2d , associated with more severe clinical outcomes( 9 , 10 ). Stx is encoded on lambdoid bacteriophages, which are released during phage-mediated lysis following the switch from the lysogenic to the lytic cycle. In this process, the bacterial cell is lysed, and both new bacteriophages and Stx are produced and released( 11 ). STEC O157:H7 genomes can harbor multiple Stx prophages as well as non-Stx prophages, with a total estimated prophage content ranging from 11.0 to 14.5 % of the genome( 9 ). Phylogenomic analyses have demonstrated that different lineages of STEC O157:H7 display distinct phage content, which correlates with variations in Stx subtypes, Stx production levels and clinical outcomes( 9 , 12 , 13 ). For instance, isolates in Lineage IIc typically carry stx1a and stx2c and are primarily associated with bloody diarrhea (BD) but not HUS. In contrast, Lineages Ic and I/II, strongly associated with the more virulent stx2a gene, are frequently linked to both BD and HUS. Interestingly, carrying multiple Stx-phages appears to influence toxin expression levels( 14 , 15 ) and overall strain pathogenicity. Isolates harboring both stx2a and stx2c may exhibit reduced virulence compared to those carrying only stx2a ( 10 ), suggesting that interactions between different prophages, as well as between phages and the chromosomal background, modulate virulence and impact clinical outcomes. Nevertheless, little is understood about the molecular mechanism at play and role of other putative STEC virulence factors and their respective clinical outcomes. Machine learning (ML) approaches are increasingly employed as powerful supervised learning tools in various scientific domains, including microbiology ( 16 , 17 ). These methods are particularly effective in handling complex, non-linearly correlated datasets and missing data, making them well-suited for predictive modelling in biological systems. In microbial genomics, ML has been successfully applied across various areas, such as predicting antimicrobial resistance from genome sequences( 18 , 19 ), to uncover genetic variants linked to pathogenicity and virulence in genome-wide association studies( 20 ), for source attribution of food-borne pathogens( 21 – 26 ) and in metagenomic studies, to classify microbial communities and predict microbiome functions( 27 ). Random forest (RF) is a machine learning algorithm, which is based on an ensemble technique that can perform classification tasks using multiple decision trees to determine an outcome( 28 ). A RF classifier which utilises bagging techniques and feature randomness has been proven to be more accurate and reliable than single classifiers due to their ability to handle non-linearly correlated data and being robust to noise( 29 ). RF is usually the most common model chosen for classifier problems where the data is discrete. Extreme Gradient Boosting (XGB) is a state-of-the-art machine learning algorithm renowned for its effectiveness and scalability. Its robust performance has led to widespread adoption in various biological classification problems( 30 – 33 ). Similarly to RF, XGB enhances model performance by combining the output of multiple decision trees. However, unlike RF, XGB constructs these trees sequentially, with each new tree specifically trained to correct the errors made by its predecessor( 34 ). Additionally, XGB incorporates regularization in its learning objectives to prevent overfitting and uses a sparsity-aware split finding algorithm to efficiently handle missing data. Importantly, both RF and XGB offer advantages in terms of interpretability, which is crucial for biological research. Feature importance in RF can be assessed with metrics such as Mean Decrease in Impurity and Mean Decrease in Accuracy( 28 ), while XGB uses Gain, Cover, and Frequency metrics to rank features based on their impact on model performance( 34 ). Both algorithms can also leverage SHAP (Shapley Additive Explanations) for assessing feature importance associated with each individual outcome( 35 ). This interpretability is essential when applying ML to biological data, as it facilitates interpretation of the underlying biological relevance of the model outputs. In this study, we employed RF and XGB to investigate the potential association between STEC O157:H7 genomes and clinical outcomes. Our findings revealed that XGB outperformed RF in accurately classifying minority classes, including the more severe disease outcome, HUS. Additionally, we utilized SHAP values to identify the most important genomic elements that XGB leverages to predict the pathogenicity of STEC isolates. RESULTS Extreme-Gradient Boosting outperforms Random Forest for classification of HUS cases We compared the performance of Extreme Gradient Boosting (XGB) and Random Forest (RF) models for classifying STEC infection outcomes: Diarrhea (D), Bloody Diarrhea (BD), and Hemolytic Uremic Syndrome (HUS). To address class imbalance, we explored the impact of various strategies, including random upsampling of the minority classes, Synthetic Minority Over-sampling Technique (SMOTE) and optimizing hyperparameters to increase balanced accuracy. For all models, we calculated precision, recall and F1-Score for each individual class. Additionally, we obtained a global value for each metric by averaging the values across all classes and calculated the overall accuracy. When evaluating the overall performance of all models on the test set, the RF-Accuracy model achieved the highest accuracy (0.737), closely followed by RF-Upsample and RF-Balanced (both at 0.732) (Supplementary Figure S2). Interestingly, the RF-SMOTE model had the lowest accuracy among all RF models (0.690). For the XGB models, accuracy varied between 0.723 (XGB-Balanced) and 0.694 (XGB-Upsample). To evaluate model performance on the training set, we aggregated the results from each cross-validation fold to compute the mean accuracy. XGB models generally showed higher accuracy in the training set, with XGB-Upsample reaching the highest mean accuracy (0.772), followed by XGB-Accuracy and XGB-SMOTE (both at 0.760). Notably, XGB-Balanced had the lowest mean accuracy of all models (0.703). In contrast, RF models demonstrated more consistent performances on the training set, with mean accuracies ranging from 0.753 (RF-Upsample) to 0.737 (RF-Balanced). These results are summarized in Supplementary Figure S2, which combines test and training set performance for a comprehensive overview. A more detailed analysis of individual clinical outcomes showed that RF models consistently attained the highest F1-Scores for Bloody Diarrhea (BD) in the test set. This was primarily driven by their high recall rates, ranging from 0.93 (RF-Accuracy) to 0.81 (RF-SMOTE) ( Figure 1 ). Notably, all RF models failed to correctly classify HUS-causing isolates, resulting in F1-Scores of 0 for this category. In contrast, XGB models demonstrated better performance in classifying the minority classes, D and HUS. Similar patterns were observed in the training set (Supplementary Figure S3). Download figure Open in new tab Figure 1. Comparative analysis of F1 Score, Precision, and Recall based on the test set (n=213) for Random Forest (RF) and Extreme Gradient Boosting (XGB) models under different class balancing strategies. Each bar represents a performance metric for classifying STEC infection outcomes: Diarrhea (D, brown bars), Bloody Diarrhea (BD, red bars), and Hemolytic Uremic Syndrome (HUS, blue bars). The gray bars indicate the macro average of all classes. The algorithms and strategies are displayed on the x-axis, segmented into RF (Accuracy, Balanced, SMOTE, Upsample) and XGB (Accuracy, Balanced, SMOTE, Upsample). The XGB-Balanced model emerged as the top performer in the test set, achieving the highest average values for F1-Score (0.611), recall (0.632), and precision (0.600) when considering all classes. Across all average metrics, XGB models consistently outperformed their RF counterparts ( Figure 1 ). In the training set, XGB models also recorded the highest averages for F1-Score and recall, although RF models generally surpassed XGB in terms of precision (Supplementary Figure S3). Given the superior performance of XGB-Balanced in classifying isolates from the test set, this model was selected for downstream analysis. We evaluated the accuracy of the XGB-Balanced classifier as a risk-assessment method and compared it against other traditional approaches. Risk assessment of STEC O157:H7 in terms of pathogenicity has generally focussed on the presence of specific virulence factors, such as the presence of the stx2a variant, to classify isolates as high-risk( 36 ). Alternatively, population structure has also been used, with isolates from lineages I/II, Ia, Ic, and IIc commonly categorized as high-risk. In our analysis, we defined isolates causing BD and HUS as ’high risk’ and those associated with diarrhea as ’low risk.’ Using this binary classification, the XGB-Balanced model achieved an accuracy of 0.784, outperforming the accuracy based on stx2a presence (0.606) and the accuracy associated with high-risk lineage membership (0.756) (Supplementary Figure S4 A). Notably, the XGB classifier’s primary advantage came from improved accuracy in predicting high-risk isolates, while all models showed similar performance in identifying low-risk cases (Supplementary Figure S4 B). This result underscores the XGB model’s capacity to capture complex genomic signatures that traditional methods might miss, which is critical for identifying emerging virulent subtypes and detecting high-risk isolates even within traditionally low-risk lineages (Supplementary Figure S4C ). Full prediction outcomes for all isolates in the test dataset are provided in Supplementary Dataset 1. Feature selection approach highlights known virulence factors and new genomic regions with potential impact over STEC clinical outcomes All models were trained using a total of 1,665 optimized features. Of these, only 283 (17%) exhibited a strong lineage effect, meaning that at least 90% of isolates within any given sub-lineage (e.g., IIa, IIb, IIc, I/I, Ia, Ib, Ic) possessed that feature. A significant proportion of features (85.6%, n = 1,426) originated from prophages, with 942 (56.6%) located within prophages potentially encoding a Stx gene (Supplementary Figure S5 A). The majority of features (61.3%, n=1,020) were identified within protein-coding sequences while the remaining (38.7%, n=345) were classified as non-coding. Among features within protein-coding sequences, only 31.2% (n = 318) aligned to 31 genes of known or predicted function. The distribution of these features varied significantly depending on their genomic origin: in phage-associated regions, only 13.3% of features aligned to known genes, whereas in non-phage regions, 53.6% of features aligned to known genes (Supplementary Figure S5 B). This disparity highlights the higher concentration of functionally-characterised elements within non-phage regions. Notably, among the features aligning to known genes ( Table 1 ), we identified two well-described virulence factors associated with STEC: 28 features aligned to the gene encoding for the subunit A of the Shiga toxin, STEC’s primary virulence factor. Additionally, 7 features aligned to the gene encoding the serine protease EspP, which can cleave several coagulation factors and components of the complement system, potentially impairing coagulation and immune responses in the host and potentially contributing to bloody diarrhea ( 37 – 41 ). View this table: View inline View popup Download powerpoint Table 1. Features annotations, origin, function and literature references Our feature selection approach also identified novel elements potentially implicated in pathogenesis. Notably, 41 features aligned to yraK , a gene within the yra operon that encodes a type-1 fimbriae. The yra operon has been shown to facilitate adhesion to bladder cells in E. coli K-12( 42 ). Similarly, four features aligned to ompX , which codes for an outer membrane protein known to mediate adhesion and invasion of E. coli into various mammalian cell types, including kidney epithelial cells. Interestingly, ompX mutants have also demonstrated reduced motility, likely due to diminished flagellar production, further underscoring its multifunctional role in pathogenesis( 43 – 45 ). Moreover, 20 features aligned to fliI, a gene that is central for flagellum assembly( 46 – 49 ). E. coli O157:H7 relies on the flagellum to reach and adhere to optimal colonization sites after entering into the host intestine ( 50 , 51 ). Additionally, 19 features aligned to a gene encoding a methylglyoxal reductase, an enzyme critical for reducing intracellular levels of the toxic metabolite methylglyoxal( 52 ). This metabolite is known to accumulate in E. coli cells during shifts from nutrient scarcity to abundance( 53 ), and macrophages have been shown to increase methylglyoxal production when challenged by pathogens such as Salmonella enterica ( 54 ) . Several genes associated with genome plasticity and reorganization were identified within predictive features. Among these, rusA , a resolvase of Holliday junctions, plays a key role in DNA repair following homologous recombination and in responding to DNA damage ( 55 , 56 ). Additionally, xerC , a site-specific recombinase, was identified. This enzyme is essential not only for the resolution of circular chromosomes prior to cell division but also for plasmid stability and the integration of certain prophages ( 57 ). Finally, multiple features aligned to various insertion sequences, including ISEc8 , which is hypothesized to play a significant role in generating small-scale structural polymorphisms in STEC O157:H7( 58 ). Notably, ISEc8 —along with other insertion sequences—has been observed to disrupt the stx gene( 59 ), thereby abolishing toxin production. This suggests that STEC may sometimes face selective pressures favoring the inactivation of toxin expression, potentially under conditions where toxin production incurs a fitness cost. These findings underscore the effectiveness of our feature selection method in isolating features relevant to disease severity and reveal potential novel elements that could play a role in pathogenesis. Detailed sequences, annotations, and classifications for all features are available in Table 1 and Supplementary Dataset 2. SHAP values identify the most important features for predicting each clinical outcome The contribution of each feature to predicting specific clinical outcomes was evaluated using SHapley Additive exPlanations (SHAP) values. Feature importance was analyzed based on their presence ( Figure 2 , Supplementary Figures S6 and S7) or absence (Supplementary Figure S8). To account for overlapping feature contributions effectively, and given the additive nature of SHAP values, features were grouped based on co-occurrence patterns, as described in the Methods section. Download figure Open in new tab Figure 2. Bar plots (top-right) show the mean SHAP values for the top 20 most important features (or feature clusters) that contribute to the prediction of HUS, when present. The networks (top-left) illustrate clusters of features (nodes) that co-occur in all isolates. Colors represent the potential origin of each feature: Stx-phage (purple), Possible Stx-phage(blue), non-Stx phage (green), or non-phage (yellow). Maximum likelihood tree based on core-genome SNPs (bottom), reflecting the phylogenetic relationships between STEC isolates included in this study (n=1030). Leafs are colored according to their lineage. Metadata blocks display: Stx variants, Symptom of patients, predictions made by the XGB-Balanced classifier. Presence of top 20 most important features for each clinical outcome are encoded in blue. A notable proportion of the 20 most-predictive features for each class mapped to intergenic regions (IR) (n=19, 31.6%) or hypothetical proteins (n=29,48.3%). Despite this, key genes and sequences critical for predicting HUS phenotype were identified ( Figure 2 ). Different variants of phage-encoded Lysozyme Feature 105, aligning to the Stx-phage-encoded lysozyme gene ( rrrD ), was one of the top four predictors of HUS. RrrD plays a critical role in host cell lysis, facilitating the release of Shiga toxin ( 65 – 67 ). This feature seems to be distributed across lineage Ic and I/II predominantly, so they might be associated with a particular phage-subtype. Surprisingly, the absence of feature 68, which also aligns to RrrD, emerged as an important predictor of HUS (Supplementary Figure S8). Feature68 is not encoded by a Stx-phage and BLASTP alignment of representative sequences of each Lysozyme variant revealed a low sequence identity (32.87%) among these (Supplementary Figure S9). These results suggest that carrying different variants of this lysozyme might affect the timing or efficiency of host cell lysis, thereby influencing toxin release. Variations in kilR and it’s upstream region Cluster75, aligning to the kilR gene on a Rac prophage, also emerged as a significant predictor ( Figure 2 ). Under normal conditions, kilR and most other Rac-prophage genes remain tightly repressed ( 75 , 76 ). However, in the presence of nalidixic acid or oxidative stress, kilR is transiently induced, causing a temporary growth arrest in an SOS-independent manner( 73 ), which allows time for DNA damage repair( 78 ). Only 41% (69/167) of kilR -harboring isolates possessed features from Cluster75, indicating multiple kilR variants. Indeed, clustering kilR sequences at 100% identity revealed seven variants and a MSA showed that Cluster75 features span positions –8 to +137 relative to the kilR translation start site, with Variant_1 linked to Cluster75 (Supplementary Figure S10A). Protein-sequence analysis similarly identified seven KilR variants in total (Supplementary Figure S10B), of which Variants 1 and 3 encompassed all Cluster75-containing isolates. An MSA of KilR protein-sequences indicated that Variants 1 to 4 differ between each other only by conservative mutations, suggesting minimal impact on overall KilR functionality (Supplementary Figure S10C). Features in Cluster74 also aligned to kilR and were predictive of HUS (Average SHAP value = 0.0172, Supplementary Dataset 3). Moreover, these features co-occurred with Cluster75 in 95% (66/69) of isolates. Cluster74 spans positions –21 to +88, suggesting variant sequences in the upstream region of kilR . When we compared the –25 regions of all kilR -containing isolates, they grouped into four distinct upstream variants (Supplementary Figure S11A), with Variant_0 corresponding to the Cluster74 features (Supplementary FIgure S11B). Evidence for a regulatory small RNA encoded downstream of stx2a The presence of feature Cluster 2 was the strongest predictor for HUS ( Figure 2 ), while its absence was associated with the least severe clinical outcome, D (Supplementary Figure S8). Cluster 2 was present in 97% (41/42) of lineage I/II isolates, 94% (241/256) of lineage Ic isolates, and 55% (10/18) of lineage Ia isolates, with sporadic occurrences in other lineages (Supplementary Figure S12). Features in cluster 2 (Feature 1 and Feature 3) overlap and are associated with Stx-carrying prophages. Alignment of these features against six representative Stx2-carrying prophages revealed that both features fall within an IR region, 40 bp downstream of the Stx B-subunit gene (Supplementary table S1). Similarly, Features 2 and 6—both strong predictors of HUS—align within the same IR region, positioned 12 bp and 25 bp downstream of the Stx B-subunit gene, respectively. Interestingly, Feature 1, 3, 2 and 6 were present in two stx2a -carrying prophages, associated with isolates frequently linked to HUS: E30228( 109 ) and 267849( 110 ). In contrast, the stx2a -carrying prophage from isolate 315176 lacked these features due to two specific mutations: a SNP (T↔C) located 90 bp downstream of the stx2 B-subunit gene and a T deletion located 103 bp downstream of the same gene ( Figure 3A ). This prophage, recently identified in lineage IIb and inserted at the sbcB site( 111 ), carries the stx2a variant; however, lineage IIb isolates are rarely associated with HUS. A recent study suggests that this phage evolved from an stx2c prophage backbone that acquired the stx2a variant through horizontal gene transfer( 111 ). In line with this hypothesis, we observed that a stx2c prophage with a highly similar backbone, obtained from isolate E116508( 111 ), also lacked the aforementioned HUS-predictive features ( Figure 3 A and B ). Download figure Open in new tab Figure 3. (A) Multiple sequence alignment of the genomic region downstream of the stx2 B-subunit gene in six representative Stx2-carrying prophages, highlighting Features 1, 3, 2, and 6. Numbers on the x-axis indicate relative positions with respect to the end of the stx2 B-subunit gene. Two mutations—a SNP (T↔C) at position +90 and a T deletion at position +103—are boxed in red and correspond to the absence of these HUS-predictive features in isolates 315176 and E116508. Computational and experimental annotations of small RNAs (sRNAs) in this region are shown: STnc100 (orange arrow) and EcOnc27 (blue arrow). (B) Comparative genomic analysis of stx2 -carrying prophages from four isolates (E116508, 315176, 267849, and E30228), illustrating the genomic context of the stx2 locus. Homology between prophages is represented by shaded regions. The stx2 A- and B-subunit genes are labeled in red. Notably, the stx2c -carrying prophage from isolate E116508 shares high structural similarity with the stx2a -carrying prophage from 315176, supporting the hypothesis that the latter evolved from an stx2c prophage backbone via horizontal gene transfer. Since IRs frequently encode regulatory elements, we used Infernal( 112 ) as implemented in Bakta( 113 ), to predict non-coding RNAs (ncRNAs) within this region. Mutations associated with Features 1, 3, 2 and 6 overlap with a predicted small RNA (sRNA), annotated as STnc100 ( Figure 3A ). Supporting this computational prediction, two independent experimental studies in E. coli O157:H7 strains Sakai ( 114 ) and EDL933 ( 115 ) identified an Hfq-binding sRNA immediately downstream of the stx2a B-subunit, termed EcOnc27 (sRNA103 in EDL933). Notably, EcOnc27 partially overlaps with STnc100 and the mutations defining Features 1, 3, 2 and 6 ( Figure 3A and Supplementary table S1). Interestingly, overexpression of EcOnc27 in EDL933 was previously shown to increase fimZ and espA transcript levels( 115 ). EspA, a key effector of the locus of enterocyte effacement (LEE), is exported via STEC’s type III secretion system and assembles into a large filamentous appendage that mediates direct contact between the bacterium and the host cell. This interaction is critical for the translocation of EspB into infected host cells ( 116 ), an essential step in the formation of attaching and effacing (A/E) lesions, a hallmark of STEC pathogenesis ( 117 ). FimZ is a response regulator that modulates the expression of 109 genes in E. coli K-12, including the upregulation of 10 SOS-responsive genes ( dinD, dinI, ibpA, ibpB, lexA, recA, recN, recX, yebF , and yebG ), the puu operon (putrescine biosynthesis) and the sfm operon (fimbriae formation)( 118 ). These results suggest a potential regulatory role for EcOnc27 in virulence-associated pathways. Stx2 production variation among lineage I/II isolates carrying features linked to the EcOnc27 sRNA To further assess the relationship between HUS-predictive features linked to EcOnc27 and Stx2 production, we analyzed publicly available data from 18 STEC O157:H7 lineage I/II (clade 8) isolates from Japan. This dataset includes both complete genomes and quantitative Stx2 production measurements, obtained using fluorescence resonance energy transfer (FRET) assays. As detailed in Supplementary table S2, BLASTN analysis confirmed that these isolates carried all EcOnc27 -associated HUS-predictive features (Feature 1, 3, 2, and 6), located downstream of the stx2 B-subunit gene—consistent with our earlier observations that this cluster is nearly ubiquitous in lineage I/II. Despite the universal presence of these features, Stx2 production levels in mitomycin C-induced cultures varied widely, ranging from 1.59 × 10 4 ng/ml (isolate 26_141088) to 4.21 × 10 5 ng/ml (isolate 93_161312), with the highest-producing strain exhibiting a 27-fold increase compared to the lowest producer (Supplementary Figure S13). This lack of correlation suggests that mutations in EcOnc27 alone do not dictate toxin output under the tested conditions. Instead, our findings reinforce the notion that, while these features are strong predictors of HUS in our machine-learning models, their precise role in Stx2 regulation remains context-dependent. Additional phage-encoded regulatory elements, host stress responses, or environmental triggers may influence toxin expression in ways that are not captured by measuring secreted toxin levels under standard laboratory conditions. Further investigation is needed to determine whether EcOnc27 -linked features contribute to intracellular toxin accumulation, phage induction dynamics, or bacterial survival strategies in clinically relevant environments. KilR knockout mutants show no significant effect on Shiga toxin production KilR, a Rac prophage-encoded protein, was identified as one of the Top-20 most important features for predicting HUS in our machine-learning analysis. Under oxidative stress conditions, KilR inhibits bacterial cell-division protein FtsZ, leading to cell-division arrest and cell elongation. To investigate whether KilR has an effect on Stx production in STEC O157:H7, we created kilR knockout (KO) mutants in two strains, designated 869 and 1813, isolated from human and cattle respectively. These strains were grown in LB or MEM medium, and Stx production was measured via ELISA. Parallel experiments were performed in which kilR was induced at a low level of IPTG (see Methods). Across all tested conditions, no significant differences in Stx production were detected between the WT and kilR KO strains ( Table 2 ). In strain 869, the kilR KO displayed a modest increase in Stx production in LB (1.25-fold, SD = 0.67) and a slight decrease in MEM (0.9-fold, SD = 0.33) relative to the WT. A similar pattern was seen for strain 1813 in LB (1.27-fold, SD = 0.36) and MEM (3.58-fold, SD = 3.6), although the high standard deviation in MEM indicates substantial variability in these measurements. Likewise, induction of kilR at low IPTG concentrations produced no significant changes (0.7–0.8-fold, SD 0.16–0.26) in Stx levels for both strains. Taken together, these results suggest that manipulating kilR alone does not substantially alter Stx production under the laboratory conditions tested. View this table: View inline View popup Download powerpoint Table 2. Relative Stx levels for strains 869 and 1813 comparing WT, kilR knockout, and kilR induction under LB or MEM growth conditions. DISCUSSION The ability to accurately predict the pathogenic potential of STEC O157:H7 strains has significant implications for public health, potentially enabling earlier identification of high-risk clones, improved clinical management, and more precisely targeted epidemiological interventions. STEC O157:H7 pathogenicity has been linked to Stx subtype composition, prophage diversity, toxin expression levels, and the presence of additional virulence determinants such as eae (intimin)( 10 , 13 , 15 , 119 – 121 ) . However, the complex interplay among these elements, as well as the potential contributions of poorly characterized genomic regions—such as regulatory sequences, small RNAs, and hypothetical proteins—remains incompletely understood. To address this complexity, we employed ML combined with SHAP-based interpretability to predict the clinical outcomes of STEC infections and to uncover novel genomic features influencing pathogenicity. Importantly, we ensured robust validation by implementing a phylogeny-informed dataset-split and cross-validation strategy to prevent data leakage and overfitting due to population structure( 122 ) , thereby reinforcing the generalizability of our model. Our findings demonstrate that XGB outperforms RF in predicting minority classes (notably HUS), likely due to XGB’s iterative boosting mechanism that corrects misclassifications sequentially( 34 ). Oversampling via SMOTE or random upsampling did not enhance predictive metrics for either algorithm, possibly because purely synthetic or duplicated data fail to capture the subtle genomic features associated with severe clinical outcomes. Alternatively, these results might indicate that class imbalance alone is not the primary factor limiting accuracy. Indeed, bacterial genomics alone cannot fully predict STEC clinical outcomes, as disease severity also depends on multiple factors, including host immunity, infectious dose, and gut microbiota composition( 123 – 132 ). Although our models highlight key genomic predictors, even genetically identical isolates can yield distinct clinical manifestations ( 5 , 110 , 125 , 133 ). Incorporating host-pathogen interaction data—such as immune response, microbiome composition, and gene expression levels—offers a promising way to enhance both the accuracy and clinical relevance of future predictive models. Despite its limitations, our model improves on traditional risk assessment, which relies on stx2a presence or lineage classification. The XGB classifier identified high-risk variants within low-risk lineages, highlighting ML’s potential for detecting emerging virulent clones. Given STEC’s dynamic genome, where phage-mediated recombination drives adaptation( 10 , 134 – 136 ), our analysis confirmed that most predictive features (85.6%, n = 1,426) map to prophages. This reinforces the role of phage-borne elements in virulence and underscores ML’s value in uncovering high-risk variants that standard surveillance might miss. SHAP analysis showed that over 30% of the top predictors mapped to intergenic regions, highlighting the importance of regulatory regions in STEC pathogenicity. Stx-carrying bacteriophages encode numerous regulatory sRNAs, many of which interact with Hfq, a conserved RNA chaperone that acts as a global post-transcriptional regulator in multiple species( 114 , 137 – 139 ). In STEC O157:H7 , hfq deletion increases Stx2AB expression in strains 86-24 and EDL933 but differentially impacts LEE effector regulation—repressing it in 86-24 while upregulating it in Sakai and EDL933( 140 – 143 ) . This variability suggests that Hfq-associated sRNAs fine-tune virulence pathways in a strain-dependent manner, potentially integrating with broader regulatory networks that respond to environmental or genetic contexts. In line with this, four of the most important features for predicting HUS mapped to EcOnc27, an Hfq-binding sRNA located immediately downstream of the stx2a gene cluster. These features corresponded to two specific SNPs strongly associated with high-risk isolates. Given that single-nucleotide changes in sRNAs can significantly alter their binding affinity to target mRNAs and Hfq( 144 , 145 ), these SNPs could potentially influence EcOnc27’s regulatory activity and thereby modulate STEC pathogenicity. Overexpression of EcOnc27 in STEC O157:H7 strain 86-24 led to increased levels of fimZ ( 115 ), a response regulator that appears to modulate the expression of over 100 genes, including the overexpression of 10 SOS-response genes( 118 ). In addition, EcOnc27 overexpression also elevates espA transcript levels, implying that it may impact both toxin regulation and the formation of attaching and effacing lesions( 115 ). However, analysis of publicly available data from 18 STEC O157:H7 lineage I/II isolates from Japan, revealed up to a 27-fold variation in Stx2a production levels, despite all isolates carrying the EcOnc27-associated HUS-predictive features. This suggests that EcOnc27 alone does not control toxin expression, consistent with the complexity of Hfq-dependent sRNA networks, where competition for Hfq binding and interactions with other regulatory elements shape gene regulation ( 146 – 148 ). Similarly, KilR, a Rac-prophage-encoded protein, emerged as a top HUS-predictive feature in our analysis. Induced by oxidative stress, KilR inhibits FtsZ, triggering transient growth arrest in an SOS-independent manner This mechanism allows STEC additional time to repair oxidative DNA damage( 78 ). Notably, oxidative stress is recognized as a major physiological challenge in the gut, where neutrophils generate hydrogen peroxide (H₂O₂) as a key defense mechanism( 149 – 152 ). Enhanced ability to withstand such oxidative assaults may thus correlate with a more severe clinical outcome. In line with this hypothesis, our finding that kilR knockouts or mild KilR induction did not substantially affect Stx levels suggests that, under our experimental conditions, KilR primarily supports bacterial survival rather than directing toxin regulation. However, co-culture of STEC O157:H7 with either H₂O₂ or neutrophils elevates Stx production( 153 , 154 ), likely by inducing Stx-carrying prophages in only a small fraction of the bacterial population—a process sometimes described as “bacterial altruism”( 153 , 155 , 156 ). This suggests that a delicate balance exists between oxidative stress-induced bacterial cell death (releasing toxin) and the survival of remaining bacteria. Our results raise the intriguing possibility that specific KilR variants identified by our machine-learning model may influence this balance, potentially through subtle differences in their binding affinity or interaction dynamics with FtsZ. Computational modeling of KilR–FtsZ interactions could clarify whether the SNPs highlighted by our predictive models overlap with critical binding regions. Furthermore, future studies evaluating Stx production in kilR knockout strains or variant isolates should be performed under physiologically relevant oxidative stress (e.g., neutrophil co-culture or H₂O₂ exposure) to elucidate whether KilR directly influences toxin regulation or primarily enhances bacterial survival while other mechanisms dominate Stx expression. Similarly, the finding that certain RrrD lysozyme variants correlate with HUS suggests another axis by which STEC may finely regulate lysis and toxin release—particularly in lineages Ic and I/II. Yet, recent work indicates that phage-encoded lytic genes alone are not essential for STEC virulence( 157 ), possibly because both Stx1 and Stx2 are routinely exported on membrane vesicles under aerobic and anaerobic conditions across diverse STEC lineages( 158 – 160 ). Paradoxically, the absence of a distinct, non–Stx-phage-encoded RrrD variant also correlates with HUS, suggesting that encoding multiple variants of lysozymes could have an impact on cell lysis regulation and subsequent toxin release. RrrD function may therefore depend on particular phage subtypes, lineage backgrounds, or environmental pressures. Future studies systematically examining these diverse rrrD variants in physiologically relevant models will be essential to clarify when and how RrrD-mediated lysis meaningfully contributes to pathogenicity. We also identified several well-characterized virulence factors (e.g., EspP, FliI, YraK) among selected features( 40 – 42 , 50 ). In contrast, eae (intimin) did not appear as a key predictor, likely due to near-universal conservation across STEC O157:H7( 161 ). Intriguingly, multiple hypothetical proteins also ranked highly, pointing to underexplored virulence determinants. Future structural predictions via AlphaFold( 162 ) could guide deeper functional characterization of these proteins, tying them to specific pathogenic mechanisms or regulatory pathways. CONCLUSIONS Our findings underscore the value of explainable ML in dissecting microbial pathogenicity and reinforce the central role of phage elements in STEC O157:H7 virulence. From a practical standpoint, an XGB-based surveillance pipeline could supplement or replace traditional risk indicators (e.g., stx2a presence), offering earlier detection of high-risk clones for targeted epidemiological follow-up. Nonetheless, factors such as host variability, immune responses, and environmental influences still limit predictive performance. Moving forward, unraveling the mechanistic interplay between phage-borne regulators and host-pathogen interactions remains a pivotal challenge, one that must be addressed to develop refined STEC risk models and advance precision public health interventions. METHODS Data Selection In England, STEC O157:H7 isolated from faecal specimens from hospitalised and community cases with symptoms of gastrointestinal disease, are submitted to the Gastrointestinal Bacteria Reference Unit (GBRU) within United Kingdom Health Security Agency (UKHSA) where they undergo whole-genome sequencing (WGS). The dataset in this study included 1030 STEC O157:H7 isolates received by UKHSA for routine typing in 2017 and 2018. All human cases of confirmed STEC O157:H7 in England were requested to complete an enhanced surveillance questionnaire to ascertain clinical presentation. Clinical outcomes were classed into 3 categories: diarrhoea (D), bloody diarrhoea (BD) and HUS. For cases with multiple clinical outcomes recorded the most severe outcome was used. In total 599 cases reported bloody diarrhoea, 387 cases reported diarrhoea and 44 reported HUS. Full metadata of these isolates and SRA accessions are available in Supplementary table S3. NGS data processing FASTQs from each isolate were quality trimmed with Trimmomatic( 163 ) (v0.36) with the following parameters: ILLUMINACLIP 2:20:10:3, LEADING 3, TRAILING 3, SLIDINGWINDOW 4:15, MINLEN 36. Trimmed reads were assembled using SPAdes (v3.12.0) with default parameters( 164 ). Dataset split The dataset was split into training (n=817) and validation (n=213) sets. Isolates in the validation set were removed from all efforts of model optimization, and predictions of clinical outcome were performed only once with each model. To have a fair representation of the phylogenetic diversity in the validation set, an individual train/validation split was performed within each of the STEC sub-lineages (Ia, Ib, Ic, I/II, IIa, IIb and IIc). Each split was performed by using the StratifiedGroupKFold function from the scikit-learn library( 165 ) (v1.3.0). We further stratified by clinical outcome frequency, in order to approximate the distribution of the training set. Moreover, to avoid ‘data leakage’, groups of isolates displaying a SNP distance equal or less than five were kept together in either the training or validation set. Feature extraction and engineering Using the SPAdes assemblies included the training set, all unique k-mers (with sizes ranging from 31 to 100) were identified using fsm-lite( 166 ) (v1.0-stable) and binary encoded to values of 1 or 0 to indicate presence or absence of each k- mer in each genome. To reduce computational complexity and increase interpretability of the features k-mers of length smaller than 80 base pairs were discarded. Further feature reduction was performed via a two step process. First, a chi-sq test was performed to retain features that showed dependency with disease outcome (p-value<=0.05), which yielded 1,665,645 features. Subsequently, the most relevant of these features were selected by using the minimally biased features selection algorithm( 167 ), as implemented in the py-MUVR package( 168 ) (v1.0.1). MUVR was performed for 10 iterations with 5 outer segments, 4 inner segments and a feature dropout rate of 0.9. The remaining set of features (n=1665) were used to build the classifiers. Training Machine Learning Classifiers Random Forest (RF) and Extreme Gradient Boosting (XGB) classifiers were implemented in Python using the scikit-learn( 165 ) (v1.3.0) and xgboost( 169 ) (v2.0.3) libraries. Hyperparameters for both models were optimized using a random search within a predefined search space. To assess the quality of different hyperparameter combinations, we employed 10-fold grouped stratified cross-validation, utilizing the RandomizedSearchCV function from the scikit-learn library. Optimal hyperparameter values were selected based on either model accuracy or balanced accuracy (Supplementary datasets 4 & 5). A detailed flowchart on the training process can be found in Supplementary Figure S1. To address class imbalance in the training data, we applied oversampling techniques to the minority classes in each cross-validation fold. This was done using either the RandomOversampler or SMOTE( 170 ) function from the imblearn (v0.11.0) library( 171 ). Additionally, to evaluate the impact of these oversampling methods, we also optimized each model without up-sampling the training data, allowing for a comprehensive comparison of different approaches to handling class imbalance. Each model was named to reflect its specific configuration and the strategy used for addressing class imbalance: models labeled as “RF-Accuracy” or “XGB-Accuracy” were optimized purely for overall accuracy without any oversampling; “RF-SMOTE” and “XGB-SMOTE” models employed the SMOTE technique to generate synthetic samples for minority classes; “RF-Upsample” and “XGB-Upsample” models utilized random oversampling of minority classes; and “RF-Balanced” and “XGB-Balanced” models maximize overall balanced accuracy, without applying any oversampling techniques for class imbalance. Evaluating the performance of the classifiers The overall performance of the classifier was evaluated using accuracy, recall, precision, and F 1 -score,calculated as: Where TP = True positive, TN = True negative, FP = False positive, FN = False negative Feature importance analysis Feature importances were assessed using SHapley Additive exPlanations (SHAP) values( 35 ), calculated with the SHAP Python package( 172 ) (v0.42.1). SHAP values offer advantages with respect to other methods for calculating feature importances due to their consistency, additive properties, and ability to provide both global and local interpretability( 173 ). For each isolate in the training set, three SHAP values were computed for every feature. These SHAP values quantify the contribution of the presence (value=1) or absence (value=0) of each feature to each of the potential model outcomes (D, BD, and HUS), offering insights into how individual genomic features influence the classification. Global SHAP values for each feature (as depicted in Figure 2 and Supplementary Figures S6-S8) were obtained by calculating the mean SHAP value across all isolates, stratified by whether the feature was present or absent. Given that k-mers frequently overlap within genomes, resulting in highly correlated features, we employed a clustering approach to improve the estimation of the importance of each genomic region. Specifically, features were clustered based on their co-occurrence across genomes. Only features that co-occurred in all genomes were clustered together, and their SHAP values were summed in the global SHAP value calculation. Feature Annotation All STEC isolate genomes were annotated using Prokka( 174 ) (v1.14.6). The resultant GFF files were used to define the pangenome using Roary( 175 ) (v3.11.2) with the ‘-e’, ‘-n’, ‘-g 100000’ options. The pan_genome_reference.fa (Supplementary Dataset 6) file returned by Roary was converted into BLAST (Basic Local Alignment Search Tool)( 176 ) database in a FASTA file format using the makeBLASTdb application within BLASTN of BLAST 2.11.0+ package. All features obtained after MUVR selection were searched against this pangenome-database using BLASTN with the following options: evalue 1e-20, max_hsps 1, outfmt = 5 (XML BLAST output). Features were classified as belonging to a coding-region if at least 25% of its length aligned to one of the genes predicted by Prokka. Otherwise, features were classified as “non-coding”. A BLAST database comprising 263 prophages from complete STEC O157:H7 genomes was constructed and grouped based on their encoding of a Shiga toxin (Supplementary Dataset 7). High-ranking features were compared against this prophage database and classified based on identical matches. Each feature was assigned to one of four categories: ( 1 ) Stx-phage, if it aligned exclusively to Stx-carrying prophages; ( 2 ) Non-Stx phage, if it aligned only to non-Stx-carrying prophages; ( 3 ) Possible Stx-prophage, if it aligned to both Stx and non-Stx prophages; and ( 4 ) Non-phage, if it did not produce any hits against the database. Multiple sequence alignment of HUS-predictive features to prevalent stx-carrying phages FASTA files corresponding to the closed genomes of six isolates carrying stx2-containing phages were downloaded from the NCBI nucleotide archive using the following accession numbers: NC_002695.2 (Sakai), AE005174.2 (EDL933), VXJR01000001.1 (267849), VXJQ01000001.1 (315176), XJO01000001.1 (E30228), and VXJP01000001.1 (E116508). BLASTN alignments of HUS-predictive features (Features 1, 2, 3, and 6) against these genomes revealed that, when present, the features were consistently located downstream of the stx2a B-subunit gene, as summarized in Supplementary table S1. BLASTN was executed with the parameters -perc_identity 100 and -qcov_hsp_perc 100 to ensure perfect sequence identity and full query coverage. To compare the DNA sequences downstream of the stx2 gene, Samtools ( 177 ) (v1.21) was used to extract 270 bp downstream of the stx2 cluster from all isolates. A multiple sequence alignment (MSA) was then generated using Clustal Omega( 178 ) (v1.2.4) with default parameters, incorporating sequences from Features 1, 2, 3, and 6. Visualization of the MSA was performed using the ggmsa( 179 ) (v1.0.2) package in R. In silico predicted sRNA STnc100 coordinates were obtained using BAKTA( 113 ) (v1.8.2) with default parameters and manually added to Figure 2 . Additionally, in vitro predicted sRNA EcOnc27 coordinates for the Sakai genome were retrieved from Supplementary table 2 of Tree et al( 114 ), and manually annotated in Figure 2 . Stx2-carrying phages comparative genomics Prophage coordinates in complete reference genomes were detected using Phastaf (v0.1.0). Any detected prophages separated by less than 4 kbp were conjoined into a single phage, as described elsewhere( 9 , 111 ). Prophage regions were extracted using Samtools (v1.21) and annotated using Pharokka( 180 ) (v1.7.2). Gene cluster comparison and alignment visualization was performed using Clinker( 181 ). Phylogenetic Analysis Trimmed FASTQs were aligned to the STEC O157:H7 reference genome Sakai( 133 ) using Snippy( 182 ) and core genome alignment produced with the 24 prophage and prophage-like elements masked. Recombinant regions in the alignment with filtered using Gubbins( 183 ) (v3.3) and the phylogenetic tree produced using IQ-TREE2 ( 184 ) (v.2.3.0) using the ‘AUTO’ function to choose the best evolutionary model with polytomies collapsed. Lineage and sub-lineage assignments were performed based on discriminatory SNPs, extracted directly from SnapperDB( 185 ) (v0.2.5), that define the population structure, as described previously( 10 ). Shiga toxin subtyping were performed as previously described ( 186 ) . Construction of the kilR -KO mutants and kilR -inducible strains Two E. coli O157:H7 PT21/28 strains that harbour both stx2a and stx2c — bovine isolate 869 (SRA: SRX11678658, SRX11678653) and human isolate 1813 (SRA: SRX11678629) — were used in all toxin-release experiments ( Table 3 ). View this table: View inline View popup Download powerpoint Table 3. Strains used in this study To generate kilR deletion mutants, regions flanking the kilR gene were PCR-amplified from strain 1813 (identical in 869) using primer pairs No/Ni (5′ flank) and Co/Ci (3′ flank) ( Table 4 ). These fragments were fused by overlap-extension PCR and cloned into the suicide vector pKNG-SceI using NEBuilder®. The plasmid was first propagated in E. coli DH5α(pir) to maintain the R6K origin and was sequence-verified by PlasmidsNG®. View this table: View inline View popup Table 4. Oligonucleotides used in this study pKNG-SceI is an unpublished derivative of pKNG101( 187 ), constructed by Matthieu Haudiquet and kindly provided by Olaya Rendueles and Eduardo Rocha (Institut Pasteur, Paris). The verified construct was mobilised into E. coli S17-MFD(pir)( 188 ) and conjugated into strains 869 and 1813. Double-crossover recombinants were selected by sacB-based counter-selection and confirmed by PCR using primers N-ext/C-ext ( Table 4 ), yielding the Δ kilR knockout mutants. To generate KilR-inducible derivatives, the kilR coding sequence was amplified with primers Nt-pTrc-kilR/Ct-pTrc-kilR ( Table 4 ) and cloned into pTrc99a (NcoI/HindIII), placing it under the IPTG-inducible trc promoter. Correct constructs were confirmed by whole-plasmid sequencing (Eurofins WPS®) and by the characteristic IPTG-dependent growth inhibition (lethality) caused by KilR over-production. Measuring Shiga-toxin production levels Overnight cultures in Luria–Bertani (LB) broth or Minimal Essential Medium (MEM) were diluted 1 : 100 and incubated at 37 °C with shaking. For KilR-inducible strains, media were supplemented with ampicillin (100 µg ml −1 ) and 20 µM IPTG (higher IPTG concentrations prevented growth). At equivalent cell density (OD 600 ), cultures were passed through 0.45 µm polyethersulfone filters to obtain cell-free supernatants. Extracellular Stx was quantified with the RIDASCREEN® Verotoxin ELISA (R-Biopharm, C2201) according to the manufacturer’s instructions. Absorbance values were normalised to the corresponding wild-type grown in the same medium, and results are reported as fold change ( Table 2 ). Each condition was tested in at least three independent cultures, with duplicate ELISA measurements for every sample. DATA AVAILABILITY Supplementary datasets can be obtained from the following zenodo repository: https://doi.org/10.5281/zenodo.15576327 All codes necessary to reproduce the result of this work and the ML models can be found in the following git repository: https://github.com/jpaganini/rf_0157 Phylogenetic tree displayed in Figure 2 can be accessed at: https://microreact.org/project/uk-stec-tree AUTHOR CONTRIBUTIONS T.J.D conceived and designed the study; C.J. collected the dataset; J.A.P, T.J.D. and S.K. constructed and validated the ML models and completed downstream bioinformatic analysis; D.G. and S.M. designed and executed Stx expression experiments, J.A.P and S.K. wrote the paper with input from all other authors. All authors read and approved the manuscript. FUNDING This work was supported by the BBSRC London Interdisciplinary Doctoral Programme and Public Health of England. J.A.P and T.J.D. were founded by the HealthHolland TKI-LSI grant, project number: LSHM23021. D.G and S.M. supported by funding from BBSRC: BBS/E/RL/230002C. CONFLICT OF INTEREST There are no conflicts of interest. ACKNOWLEDGEMENTS We would like to thank all the staff at Gastrointestinal Bacteria Reference Unit and Health Protection Research Unit at Public Health of England for their support and guidance during this project. REFERENCES 1. ↵ Wells JG , Davis BR , Wachsmuth IK , Riley LW , Remis RS , Sokolow R , et al. Laboratory investigation of hemorrhagic colitis outbreaks associated with a rare Escherichia coli serotype . J Clin Microbiol . 1983 Sep ; 18 ( 3 ): 512 . OpenUrl Abstract / FREE Full Text 2. ↵ Byrne L , Jenkins C , Launders N , Elson R , Adak GK . The epidemiology, microbiology and clinical impact of Shiga toxin-producing Escherichia coli in England, 2009-2012 . Epidemiol Infect . 2015 Dec; 143 ( 16 ): 3475 – 87 . OpenUrl CrossRef 3. ↵ Launders N , Byrne L , Jenkins C , Harker K , Charlett A , Adak GK . Disease severity of Shiga toxin-producing E. coli O157 and factors influencing the development of typical haemolytic uraemic syndrome: a retrospective cohort study, 2009-2012 . BMJ Open . 2016 Jan 29; 6 ( 1 ): e009933 . OpenUrl Abstract / FREE Full Text 4. ↵ Wilson D , Dolan G , Aird H , Sorrell S , Dallman TJ , Jenkins C , et al. Farm-to-fork investigation of an outbreak of Shiga toxin-producing Escherichia coli O157 . Microb Genomics . 2018 Mar ; 4 ( 3 ): e000160 . OpenUrl 5. ↵ Cowley LA , Dallman TJ , Fitzgerald S , Irvine N , Rooney PJ , McAteer SP , et al. Short-term evolution of Shiga toxin-producing Escherichia coli O157:H7 between two food-borne outbreaks . Microb Genomics . 2016 Sep ; 2 ( 9 ): e000084 . OpenUrl 6. ↵ Krüger A , Lucchesi PMA . Shiga toxins and stx phages: highly diverse entities . Microbiol Read Engl . 2015 Mar ; 161 (Pt 3 ): 451 – 62 . OpenUrl 7. ↵ Tesh VL . Induction of apoptosis by Shiga toxins . Future Microbiol . 2010 Mar ; 5 ( 3 ): 431 – 53 . OpenUrl CrossRef PubMed Web of Science 8. ↵ Melton-Celsa A , Mohawk K , Teel L , O’Brien A . Pathogenesis of Shiga-toxin producing escherichia coli . Curr Top Microbiol Immunol . 2012 ; 357 : 67 – 103 . OpenUrl CrossRef PubMed 9. ↵ Shaaban S , Cowley LA , McAteer SP , Jenkins C , Dallman TJ , Bono JL , et al. Evolution of a zoonotic pathogen: investigating prophage diversity in enterohaemorrhagic Escherichia coli O157 by long-read sequencing . Microb Genomics . 2016 Dec ; 2 ( 12 ): e000096 . OpenUrl 10. ↵ Dallman TJ , Ashton PM , Byrne L , Perry NT , Petrovska L , Ellis R , et al. Applying phylogenomics to understand the emergence of Shiga-toxin-producing Escherichia coli O157:H7 strains causing severe human disease in the UK . Microb Genomics . 2015 ; 1 ( 3 ): e000029 . OpenUrl 11. ↵ Gyles CL . Shiga toxin-producing Escherichia coli: an overview . J Anim Sci . 2007 Mar ; 85 ( 13 Suppl ): E45 – 62 . OpenUrl CrossRef PubMed Web of Science 12. ↵ Pinto G , Sampaio M , Dias O , Almeida C , Azeredo J , Oliveira H . Insights into the genome architecture and evolution of Shiga toxin encoding bacteriophages of Escherichia coli . BMC Genomics . 2021 May 19; 22 ( 1 ): 366 . OpenUrl CrossRef PubMed 13. ↵ Miyata T , Taniguchi I , Nakamura K , Gotoh Y , Yoshimura D , Itoh T , et al. Alteration of a Shiga toxin-encoding phage associated with a change in toxin production level and disease severity in Escherichia coli . Microb Genomics . 2023 ; 9 ( 2 ): 000935 . OpenUrl 14. ↵ Serra-Moreno R , Jofre J , Muniesa M . The CI Repressors of Shiga Toxin-Converting Prophages Are Involved in Coinfection of Escherichia coli Strains, Which Causes a Down Regulation in the Production of Shiga Toxin 2 . J Bacteriol . 2008 Jul; 190 ( 13 ): 4722 – 35 . OpenUrl Abstract / FREE Full Text 15. ↵ Ogura Y , Mondal SI , Islam MR , Mako T , Arisawa K , Katsura K , et al. The Shiga toxin 2 production level in enterohemorrhagic Escherichia coli O157:H7 is correlated with the subtypes of toxin-encoding phage . Sci Rep . 2015 Nov 16; 5 ( 1 ): 16663 . OpenUrl CrossRef PubMed 16. ↵ Jiang Y , Luo J , Huang D , Liu Y , Li D dan . Machine Learning Advances in Microbiology: A Review of Methods and Applications . Front Microbiol [Internet]. 2022 May 26 [cited 2024 Oct 31]; 13 . Available from: https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2022.925454/full 17. ↵ Machine learning and applications in microbiology | FEMS Microbiology Reviews | Oxford Academic [Internet]. [cited 2024 Oct 31]. Available from: https://academic.oup.com/femsre/article/45/5/fuab015/6174022?login=false 18. ↵ Kaya DE , Ülgen E , Kocagöz AS , Sezerman OU . A comparison of various feature extraction and machine learning methods for antimicrobial resistance prediction in streptococcus pneumoniae . Front Antibiot [Internet]. 2023 Mar 24 [cited 2024 Oct 31]; 2 . Available from: https://www.frontiersin.org/journals/antibiotics/articles/10.3389/frabi.2023.1126468/full 19. ↵ Assessing computational predictions of antimicrobial resistance phenotypes from microbial genomes | Briefings in Bioinformatics | Oxford Academic [Internet] . [cited 2024 Oct 31]. Available from: https://academic.oup.com/bib/article/25/3/bbae206/7665136 20. ↵ Allen JP , Snitkin E , Pincus NB , Hauser AR . Forest and Trees: Exploring Bacterial Virulence with Genome-wide Association Studies and Machine Learning . Trends Microbiol . 2021 Jul 1; 29 ( 7 ): 621 – 33 . OpenUrl PubMed 21. ↵ Castelli P , De Ruvo A , Bucciacchio A , D’Alterio N , Cammà C , Di Pasquale A , et al. Harmonization of supervised machine learning practices for efficient source attribution of Listeria monocytogenes based on genomic data . BMC Genomics . 2023 Sep 22; 24 ( 1 ): 560 . OpenUrl PubMed 22. Guzinski J , Tang Y , Chattaway MA , Dallman TJ , Petrovska L. Development and validation of a random forest algorithm for source attribution of animal and human Salmonella Typhimurium and monophasic variants of S. Typhimurium isolates in England and Wales utilising whole genome sequencing data . Front Microbiol [Internet] . 2024 Mar 12 [cited 2024 Oct 31]; 14 . Available from: https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2023.1254860/full 23. Mughini-Gras L , Paganini JA , Guo R , Coipan CE , Friesema IHM , van Hoek AHAM , et al. Source attribution of Listeria monocytogenes in the Netherlands . Int J Food Microbiol . 2024 Oct 29; 110953 . 24. Lupolova N , Dallman TJ , Matthews L , Bono JL , Gally DL . Support vector machine applied to predict the zoonotic potential of E. coli O157 cattle isolates . Proc Natl Acad Sci . 2016 Oct 4; 113 ( 40 ): 11312 – 7 . OpenUrl Abstract / FREE Full Text 25. Lupolova N , Lycett SJ , Gally DL . A guide to machine learning for bacterial host attribution using genome sequence data . Microb Genomics . 2019 ; 5 ( 12 ): e000317 . OpenUrl 26. ↵ Chalka A , Dallman TJ , Vohra P , Stevens MP , Gally DL . The advantage of intergenic regions as genomic features for machine-learning-based host attribution of Salmonella Typhimurium from the USA . Microb Genomics . 2023 ; 9 ( 10 ): 001116 . OpenUrl 27. ↵ Mathieu A , Leclercq M , Sanabria M , Perin O , Droit A. Machine Learning and Deep Learning Applications in Metagenomic Taxonomy and Functional Annotation . Front Microbiol [Internet]. 2022 Mar 14 [cited 2024 Oct 31]; 13 . Available from: https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2022.811495/full 28. ↵ Breiman L . Random Forests . Mach Learn . 2001 Oct 1; 45 ( 1 ): 5 – 32 . OpenUrl CrossRef PubMed Web of Science 29. ↵ Khalilia M , Chakraborty S , Popescu M . Predicting disease risks from highly imbalanced data using random forest . BMC Med Inform Decis Mak . 2011 Jul 29; 11 ( 1 ): 51 . OpenUrl CrossRef PubMed 30. ↵ Arning N , Sheppard SK , Bayliss S , Clifton DA , Wilson DJ . Machine learning to predict the source of campylobacteriosis using whole genome data . PLOS Genet . 2021 Oct 18; 17 ( 10 ): e1009436 . OpenUrl CrossRef PubMed 31. Nguyen M , Elmore Z , Ihle C , Moen FS , Slater AD , Turner BN , et al. Predicting variable gene content in Escherichia coli using conserved genes . mSystems . 2023 Jun 14; 8 ( 4 ): e00058 – 23 . OpenUrl PubMed 32. Yang S , Ha K , Song W , Fujita M , Kübler K , Polak P , et al. COOBoostR: An Extreme Gradient Boosting-Based Tool for Robust Tissue or Cell-of-Origin Prediction of Tumors . Life . 2023 Jan ; 13 ( 1 ): 71 . OpenUrl 33. ↵ Zhang H , Wang K , Zhou J , Chen J , Xu Y , Wang D , et al. VariFAST: a variant filter by automated scoring based on tagged-signatures . BMC Bioinformatics . 2019 Dec 30; 20 ( 22 ): 713 . OpenUrl PubMed 34. ↵ XGBoost | Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet] . [cited 2024 Oct 31]. Available from: https://dl.acm.org/doi/10.1145/2939672.2939785 35. ↵ Lundberg S , Lee SI . A Unified Approach to Interpreting Model Predictions [Internet] . arXiv; 2017 [cited 2024 Jun 18]. Available from: http://arxiv.org/abs/1705.07874 36. ↵ Veneti L , Lange H , Brandal L , Danis K , Vold L . Mapping of control measures to prevent secondary transmission of STEC infections in Europe during 2016 and revision of the national guidelines in Norway . Epidemiol Infect . 2019 Jan ; 147 : e267 . OpenUrl 37. ↵ Tontanahal A , Sperandio V , Kovbasnjuk O , Loos S , Kristoffersson AC , Karpman D , et al. IgG Binds Escherichia coli Serine Protease EspP and Protects Mice From E. coli O157:H7 Infection . Front Immunol [Internet] . 2022 Feb 18 [cited 2024 Nov 5]; 13 . Available from: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2022.807959/full 38. Brunder W , Schmidt H , Karch H . EspP, a novel extracellular serine protease of enterohaemorrhagic Escherichia coli O157:H7 cleaves human coagulation factor V . Mol Microbiol . 1997 May ; 24 ( 4 ): 767 – 78 . OpenUrl CrossRef PubMed Web of Science 39. Dziva F , Mahajan A , Cameron P , Currie C , McKendrick IJ , Wallis TS , et al. EspP, a Type V-secreted serine protease of enterohaemorrhagic Escherichia coli O157:H7, influences intestinal colonization of calves and adherence to bovine primary intestinal epithelial cells . FEMS Microbiol Lett . 2007 Jun 1; 271 ( 2 ): 258 – 64 . OpenUrl CrossRef PubMed 40. ↵ Kuo KHM , Khan S , Rand ML , Mian HS , Brnjac E , Sandercock LE , et al. EspP, an Extracellular Serine Protease from Enterohemorrhagic E. coli, Reduces Coagulation Factor Activities, Reduces Clot Strength, and Promotes Clot Lysis . PLoS ONE . 2016 Mar 2; 11 ( 3 ): e0149830 . OpenUrl PubMed 41. ↵ EspP, a Serine Protease of Enterohemorrhagic Escherichia coli, Impairs Complement Activation by Cleaving Complement Factors C3/C3b and C5 | Infection and Immunity [Internet] . [cited 2024 Nov 5]. Available from: https://journals.asm.org/doi/10.1128/iai.00488-10 42. ↵ Korea CG , Badouraly R , Prevost MC , Ghigo JM , Beloin C . Escherichia coli K-12 possesses multiple cryptic but functional chaperone–usher fimbriae with distinct surface specificities . Environ Microbiol . 2010 ; 12 ( 7 ): 1957 – 77 . OpenUrl CrossRef PubMed Web of Science 43. ↵ Roles of OmpX, an Outer Membrane Protein, on Virulence and Flagellar Expression in Uropathogenic Escherichia coli [Internet] . [cited 2024 Dec 27]. Available from: https://journals.asm.org/doi/epdf/10.1128/iai.00721-20 44. Meng X , Liu X , Zhang L , Hou B , Li B , Tan C , et al. Virulence characteristics of extraintestinal pathogenic Escherichia coli deletion of gene encoding the outer membrane protein X . J Vet Med Sci . 2016 ; 78 ( 8 ): 1261 – 7 . OpenUrl CrossRef PubMed 45. ↵ Vogt J , Schulz GE . The structure of the outer membrane protein OmpX from Escherichia coli reveals possible mechanisms of virulence . Structure . 1999 Oct 15; 7 ( 10 ): 1301 – 9 . OpenUrl CrossRef PubMed 46. ↵ Claret L , Calder SR , Higgins M , Hughes C . Oligomerization and activation of the FliI ATPase central to bacterial flagellum assembly . Mol Microbiol . 2003 Jun ; 48 ( 5 ): 1349 – 55 . OpenUrl CrossRef PubMed Web of Science 47. Vogler AP , Homma M , Irikura VM , Macnab RM . Salmonella typhimurium mutants defective in flagellar filament regrowth and sequence similarity of FliI to F0F1, vacuolar, and archaebacterial ATPase subunits . J Bacteriol . 1991 Jun ; 173 ( 11 ): 3564 – 72 . OpenUrl Abstract / FREE Full Text 48. Dreyfus G , Williams AW , Kawagishi I , Macnab RM . Genetic and biochemical analysis of Salmonella typhimurium FliI, a flagellar protein related to the catalytic subunit of the F0F1 ATPase and to virulence proteins of mammalian and plant pathogens . J Bacteriol . 1993 May ; 175 ( 10 ): 3131 – 8 . OpenUrl Abstract / FREE Full Text 49. ↵ Chen S , Beeby M , Murphy GE , Leadbetter JR , Hendrixson DR , Briegel A , et al. Structural diversity of bacterial flagellar motors . EMBO J . 2011 Jun 14; 30 ( 14 ): 2972 – 81 . OpenUrl Abstract / FREE Full Text 50. ↵ Mahajan A , Currie CG , Mackie S , Tree J , McAteer S , McKendrick I , et al. An investigation of the expression and adhesin function of H7 flagella in the interaction of Escherichia coli O157 : H7 with bovine intestinal epithelium . Cell Microbiol . 2009 ; 11 ( 1 ): 121 – 37 . OpenUrl CrossRef PubMed Web of Science 51. ↵ Fernandez-Brando RJ , McAteer SP , Montañez-Culma J , Cortés-Araya Y , Tree J , Bernal A , et al. Mechanisms involved in the adaptation of Escherichia coli O157:H7 to the host intestinal microenvironment . Clin Sci . 2020 Dec 21; 134 ( 24 ): 3283 – 301 . OpenUrl 52. ↵ Ozyamak E , Almeida C , Moura APS de , Miller S , Booth IR . Integrated stress response of Escherichia coli to methylglyoxal: transcriptional readthrough from the nemRA operon enhances protection through increased expression of glyoxalase I . Mol Microbiol . 2013 May 5; 88 ( 5 ): 936 . OpenUrl CrossRef PubMed 53. ↵ Tötemeyer S , Booth NA , Nichols WW , Dunbar B , Booth IR . From famine to feast: the role of methylglyoxal production in Escherichia coli . Mol Microbiol . 1998 ; 27 ( 3 ): 553 – 62 . OpenUrl CrossRef PubMed 54. ↵ Eriksson S , Lucchini S , Thompson A , Rhen M , Hinton JCD . Unravelling the biology of macrophage infection by gene expression profiling of intracellular Salmonella enterica . Mol Microbiol . 2003 Jan ; 47 ( 1 ): 103 – 18 . OpenUrl CrossRef PubMed Web of Science 55. ↵ Mahdi AA , Sharples GJ , Mandal TN , Lloyd RG . Holliday junction resolvases encoded by homologous rusA genes in Escherichia coli K-12 and phage 82 . J Mol Biol . 1996 Apr 5; 257 ( 3 ): 561 – 73 . OpenUrl CrossRef PubMed Web of Science 56. ↵ Chan SN , Harris L , Bolt EL , Whitby MC , Lloyd RG . Sequence Specificity and Biochemical Characterization of the RusA Holliday Junction Resolvase of Escherichia coli * . J Biol Chem . 1997 Jun 6; 272 ( 23 ): 14873 – 82 . OpenUrl Abstract / FREE Full Text 57. ↵ Castillo F , Benmohamed A , Szatmari G. Xer Site Specific Recombination: Double and Single Recombinase Systems . Front Microbiol [Internet] . 2017 Mar 20 [cited 2024 Nov 5]; 8 . Available from: https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2017.00453/full 58. ↵ Ooka T , Ogura Y , Asadulghani M , Ohnishi M , Nakayama K , Terajima J , et al. Inference of the impact of insertion sequence (IS) elements on bacterial genome diversification through analysis of small-size structural polymorphisms in Escherichia coli O157 genomes . Genome Res . 2009 Oct ; 19 ( 10 ): 1809 – 16 . OpenUrl Abstract / FREE Full Text 59. ↵ Fitzgerald SF , Beckett AE , Palarea-Albaladejo J , McAteer S , Shaaban S , Morgan J , et al. Shiga toxin sub-type 2a increases the efficiency of Escherichia coli O157 transmission between animals and restricts epithelial regeneration in bovine enteroids . PLOS Pathog . 2019 Oct 3; 15 ( 10 ): e1008003 . OpenUrl CrossRef PubMed 60. Mehta P , Casjens S , Krishnaswamy S . Analysis of the lambdoid prophage element e14 in the E. coli K-12 genome . BMC Microbiol . 2004 Jan 20; 4 ( 1 ): 4 . OpenUrl CrossRef PubMed 61. North OI , Davidson AR . Phage Proteins Required for Tail Fiber Assembly Also Bind Specifically to the Surface of Host Bacterial Strains . J Bacteriol . 2021 Jan 11; 203 ( 3 ) : doi: 10.1128/jb.00406-20 . OpenUrl CrossRef 62. Groth AC , Calos MP . Phage Integrases: Biology and Applications . J Mol Biol . 2004 Jan 16; 335 ( 3 ): 667 – 78 . OpenUrl CrossRef PubMed Web of Science 63. Anso I , Zouhir S , Sana TG , Krasteva PV . Structural basis for synthase activation and cellulose modification in the E. coli Type II Bcs secretion system . Nat Commun . 2024 Oct 11; 15 ( 1 ): 8799 . OpenUrl PubMed 64. Omadjela O , Narahari A , Strumillo J , Mélida H , Mazur O , Bulone V , et al. BcsA and BcsB form the catalytically active core of bacterial cellulose synthase sufficient for in vitro cellulose synthesis . Proc Natl Acad Sci . 2013 Oct 29; 110 ( 44 ): 17856 – 61 . OpenUrl Abstract / FREE Full Text 65. ↵ Srividhya KV , Krishnaswamy S . Sub classification and targeted characterization of prophage-encoded two-component cell lysis cassette . J Biosci . 2007 Aug 1; 32 ( 1 ): 979 – 90 . OpenUrl CrossRef PubMed Web of Science 66. Toba FA , Thompson MG , Campbell BR , Junker LM , Rueggeberg KG , Hay AG . Role of DLP12 lysis genes in Escherichia coli biofilm formation . Microbiol Read Engl . 2011 Jun ; 157 (Pt 6 ): 1640 – 50 . OpenUrl 67. ↵ Pasqua M , Zennaro A , Trirocco R , Fanelli G , Micheli G , Grossi M , et al. Modulation of OMV Production by the Lysis Module of the DLP12 Defective Prophage of Escherichia coli K12 . Microorganisms . 2021 Feb 12; 9 ( 2 ): 369 . OpenUrl PubMed 68. Felczak MM , Kaguni JM . The rcbA Gene Product Reduces Spontaneous and Induced Chromosome Breaks in Escherichia coli . J Bacteriol . 2012 May ; 194 ( 9 ): 2152 – 64 . OpenUrl Abstract / FREE Full Text 69. Gulmezian M , Hyman KR , Marbois BN , Clarke CF , Javor GT . The role of UbiX in Escherichia coli coenzyme Q biosynthesis . Arch Biochem Biophys . 2007 Nov 15; 467 ( 2 ): 144 – 53 . OpenUrl CrossRef PubMed Web of Science 70. Zhang H , Javor GT . Regulation of the isofunctional genes ubiD and ubiX of the ubiquinone biosynthetic pathway of Escherichia coli . FEMS Microbiol Lett . 2003 Jun 6; 223 ( 1 ): 67 – 72 . OpenUrl CrossRef PubMed 71. Meganathan R . Ubiquinone biosynthesis in microorganisms . FEMS Microbiol Lett . 2001 Sep 25; 203 ( 2 ): 131 – 9 . OpenUrl CrossRef PubMed Web of Science 72. Spielmann A , Baumgart M , Bott M . NADPH-related processes studied with a SoxR-based biosensor in Escherichia coli . MicrobiologyOpen . 2018 Dec 25; 8 ( 7 ): e00785 . OpenUrl PubMed 73. ↵ Conter A , Bouché JP , Dassain M . Identification of a new inhibitor of essential division gene ftsZ as the kil gene of defective prophage Rac . J Bacteriol . 1996 Sep ; 178 ( 17 ): 5100 – 4 . OpenUrl Abstract / FREE Full Text 74. Marepalli A , Nandhakumar M , Govindarajan S. KilR of E. coli Rac prophage is a dual morphogenetic inhibitor of bacterial cell shape [Internet] . bioRxiv; 2025 [cited 2025 Jan 10]. p. 2025.01.07.631774. Available from: https://www.biorxiv.org/content/10.1101/2025.01.07.631774v1 75. ↵ Cardinale CJ , Washburn RS , Tadigotla VR , Brown LM , Gottesman ME , Nudler E . Termination factor Rho and its cofactors NusA and NusG silence foreign DNA in E. coli . Science . 2008 May 16; 320 ( 5878 ): 935 – 8 . OpenUrl Abstract / FREE Full Text 76. ↵ Krishnamurthi R , Ghosh S , Khedkar S , Seshasayee ASN . Repression of YdaS Toxin Is Mediated by Transcriptional Repressor RacR in the Cryptic rac Prophage of Escherichia coli K-12 . mSphere . 2017 Nov 22; 2 ( 6 ) : doi: 10.1128/msphere.00392-17 . OpenUrl CrossRef 77. Wang X , Kim Y , Ma Q , Hong SH , Pokusaeva K , Sturino JM , et al. Cryptic prophages help bacteria cope with adverse environments . Nat Commun . 2010 ; 1 : 147 . OpenUrl CrossRef PubMed 78. ↵ Barshishat S , Elgrably-Weiss M , Edelstein J , Georg J , Govindarajan S , Haviv M , et al. OxyS small RNA induces cell cycle arrest to allow DNA damage repair . EMBO J . 2018 Feb ; 37 ( 3 ): 413 – 26 . OpenUrl Abstract / FREE Full Text 79. Poole SJ , Diner EJ , Aoki SK , Braaten BA , t’Kint de Roodenbeke C , Low DA , et al. Identification of functional toxin/immunity genes linked to contact-dependent growth inhibition (CDI) and rearrangement hotspot (Rhs) systems . PLoS Genet . 2011 Aug ; 7 ( 8 ): e1002217 . OpenUrl CrossRef PubMed 80. Koskiniemi S , Lamoureux JG , Nikolakakis KC , t’Kint de Roodenbeke C , Kaplan MD , Low DA , et al. Rhs proteins from diverse bacteria mediate intercellular competition . Proc Natl Acad Sci . 2013 Apr 23; 110 ( 17 ): 7032 – 7 . OpenUrl Abstract / FREE Full Text 81. Vlazny DA , Hill CW . A stationary-phase-dependent viability block governed by two different polypeptides from the RhsA genetic element of Escherichia coli K-12 . J Bacteriol . 1995 Apr ; 177 ( 8 ): 2209 – 13 . OpenUrl Abstract / FREE Full Text 82. Missiakas D , Raina S . Signal transduction pathways in response to protein misfolding in the extracytoplasmic compartments of E. coli: role of two new phosphoprotein phosphatases PrpA and PrpB . EMBO J . 1997 Apr 1; 16 ( 7 ): 1670 – 85 . OpenUrl Abstract / FREE Full Text 83. Shi L , Kehres DG , Maguire ME . The PPP-family protein phosphatases PrpA and PrpB of Salmonella enterica serovar Typhimurium possess distinct biochemical properties . J Bacteriol . 2001 Dec ; 183 ( 24 ): 7053 – 7 . OpenUrl Abstract / FREE Full Text 84. Zhang Y , Laing C , Zhang Z , Hallewell J , You C , Ziebell K , et al. Lineage and Host Source Are Both Correlated with Levels of Shiga Toxin 2 Production by Escherichia coli O157:H7 Strains . Appl Environ Microbiol . 2010 Jan ; 76 ( 2 ): 474 – 82 . OpenUrl Abstract / FREE Full Text 85. Pedersen K , Gerdes K . Multiple hok genes on the chromosome of Escherichia coli . Mol Microbiol . 1999 Jun ; 32 ( 5 ): 1090 – 102 . OpenUrl CrossRef PubMed Web of Science 86. Gerdes K , Gultyaev AP , Franch T , Pedersen K , Mikkelsen ND . ANTISENSE RNA-REGULATED PROGRAMMED CELL DEATH . Annu Rev Genet . 1997 Dec 1;31(Volume 31 , 1997 ): 1 – 31 . OpenUrl CrossRef PubMed Web of Science 87. Rothe M , Alpert C , Loh G , Blaut M . Novel insights into E. coli’s hexuronate metabolism: KduI facilitates the conversion of galacturonate and glucuronate under osmotic stress conditions . PloS One . 2013 ; 8 ( 2 ): e56906 . OpenUrl CrossRef PubMed 88. Rothe M , Alpert C , Engst W , Musiol S , Loh G , Blaut M . Impact of nutritional factors on the proteome of intestinal Escherichia coli: induction of OxyR-dependent proteins AhpF and Dps by a lactose-rich diet . Appl Environ Microbiol . 2012 May ; 78 ( 10 ): 3580 – 91 . OpenUrl Abstract / FREE Full Text 89. Tubeleviciute A , Teese MG , Jose J . Escherichia coli kduD encodes an oxidoreductase that converts both sugar and steroid substrates . Appl Microbiol Biotechnol . 2014 Jun 1; 98 ( 12 ): 5471 – 85 . OpenUrl 90. Huang YH , Lin MJ , Huang CY . DnaT is a single-stranded DNA binding protein . Genes Cells . 2013 ; 18 ( 11 ): 1007 – 19 . OpenUrl CrossRef PubMed Web of Science 91. Masai H , Bond MW , Arai K . Cloning of the Escherichia coli gene for primosomal protein i: the relationship to dnaT, essential for chromosomal DNA replication . Proc Natl Acad Sci U S A . 1986 Mar ; 83 ( 5 ): 1256 – 60 . OpenUrl Abstract / FREE Full Text 92. Amitai S , Kolodkin-Gal I , Hananya-Meltabashi M , Sacher A , Engelberg-Kulka H . Escherichia coli MazF leads to the simultaneous selective synthesis of both “death proteins” and “survival proteins .” PLoS Genet . 2009 Mar ; 5 ( 3 ): e1000390 . OpenUrl CrossRef PubMed 93. Amitai S , Yassin Y , Engelberg-Kulka H . MazF-mediated cell death in Escherichia coli: a point of no return . J Bacteriol . 2004 Dec ; 186 ( 24 ): 8295 – 300 . OpenUrl Abstract / FREE Full Text 94. Zorzini V , Mernik A , Lah J , Sterckx YGJ , De Jonge N , Garcia-Pino A , et al. Substrate Recognition and Activity Regulation of the Escherichia coli mRNA Endonuclease MazF♦ . J Biol Chem . 2016 May 20; 291 ( 21 ): 10950 – 60 . OpenUrl Abstract / FREE Full Text 95. Control of Acid Resistance inEscherichia coli [Internet] . [cited 2025 Jan 13]. Available from: https://journals.asm.org/doi/epdf/10.1128/jb.181.11.3525-3535.1999?src=getftr&utm_source=sciencedirect_contenthosting&getft_integrator=sciencedirect_contenthosting 96. Jung IL , Kim IG . Polyamines and Glutamate Decarboxylase-based Acid Resistance in Escherichia coli * . J Biol Chem . 2003 Jun 20; 278 ( 25 ): 22846 – 52 . OpenUrl Abstract / FREE Full Text 97. Mates AK , Sayed AK , Foster JW . Products of the Escherichia coli acid fitness island attenuate metabolite stress at extremely low pH and mediate a cell density-dependent acid resistance . J Bacteriol . 2007 Apr ; 189 ( 7 ): 2759 – 68 . OpenUrl Abstract / FREE Full Text 98. Cianfanelli FR , Monlezun L , Coulthurst SJ. Aim, Load, Fire: The Type VI Secretion System, a Bacterial Nanoweapon . Trends Microbiol . 2016 Jan 1; 24 ( 1 ): 51 – 62 . OpenUrl CrossRef PubMed 99. Zhong H , Wang P , Chen Y , Wang H , Li J , Li J , et al. ClpV1 in avian pathogenic Escherichia coli is a crucial virulence factor contributing to meningitis in a mouse model in vivo . Vet Microbiol . 2021 Dec 1; 263 : 109273 . OpenUrl PubMed 100. Ma J , Bao Y , Sun M , Dong W , Pan Z , Zhang W , et al. Two Functional Type VI Secretion Systems in Avian Pathogenic Escherichia coli Are Involved in Different Pathogenic Pathways . Infect Immun . 2014 Aug 12; 82 ( 9 ): 3867 – 79 . OpenUrl Abstract / FREE Full Text 101. Makino K , Kim SK , Shinagawa H , Amemura M , Nakata A . Molecular analysis of the cryptic and functional phn operons for phosphonate use in Escherichia coli K-12 . J Bacteriol . 1991 Apr ; 173 ( 8 ): 2665 – 72 . OpenUrl Abstract / FREE Full Text 102. Chen CM , Ye QZ , Zhu ZM , Wanner BL , Walsh CT . Molecular biology of carbon-phosphorus bond cleavage. Cloning and sequencing of the phn (psiD) genes involved in alkylphosphonate uptake and C-P lyase activity in Escherichia coli B . J Biol Chem . 1990 Mar 15; 265 ( 8 ): 4461 – 71 . OpenUrl Abstract / FREE Full Text 103. Jochimsen B , Lolle S , McSorley FR , Nabi M , Stougaard J , Zechel DL , et al. Five phosphonate operon gene products as components of a multi-subunit complex of the carbon-phosphorus lyase pathway . Proc Natl Acad Sci U S A . 2011 Jul 12; 108 ( 28 ): 11393 – 8 . OpenUrl Abstract / FREE Full Text 104. Imber R . Regulation of expression of the cloned repE gene from the F plasmid of Escherichia coli . Gene . 1987 ; 52 ( 1 ): 1 – 9 . OpenUrl PubMed 105. Masson L , Ray DS . Mechanism of autonomous control of the Escherichia coli F plasmid: purification and characterization of the repE gene product . Nucleic Acids Res . 1988 Jan 25; 16 ( 2 ): 413 – 24 . OpenUrl CrossRef PubMed 106. Shimada T , Yamazaki K , Ishihama A . Novel regulator PgrR for switch control of peptidoglycan recycling in scherichia coli . Genes Cells . 2013 ; 18 ( 2 ): 123 – 34 . OpenUrl CrossRef PubMed 107. Choi U , Park YH , Kim YR , Seok YJ , Lee CR . Increased expression of genes involved in uptake and degradation of murein tripeptide under nitrogen starvation in Escherichia coli . FEMS Microbiol Lett . 2016 Jul 1; 363 ( 14 ): fnw136 . OpenUrl CrossRef PubMed 108. Paradis-Bleau C , Kritikos G , Orlova K , Typas A , Bernhardt TG . A genome-wide screen for bacterial envelope biogenesis mutants identifies a novel factor involved in cell wall precursor metabolism . PLoS Genet . 2014 Jan ; 10 ( 1 ): e1004056 . OpenUrl CrossRef PubMed 109. ↵ Scotland SM , Willshaw GA , Smith HR , Rowe B . Properties of strains of Escherichia coli belonging to serogroup O157 with special reference to production of Vero cytotoxins VT1 and VT2 . Epidemiol Infect . 1987 Dec ; 99 ( 3 ): 613 – 24 . OpenUrl CrossRef PubMed 110. ↵ Gobin M , Hawker J , Cleary P , Inns T , Gardiner D , Mikhail A , et al. National outbreak of Shiga toxin-producing Escherichia coli O157:H7 linked to mixed salad leaves, United Kingdom, 2016 . Euro Surveill Bull Eur Sur Mal Transm Eur Commun Dis Bull . 2018 May; 23 ( 18 ): 17 – 00197 . OpenUrl 111. ↵ Yara DA , Greig DR , Gally DL , Dallman TJ , Jenkins C . Comparison of Shiga toxin-encoding bacteriophages in highly pathogenic strains of Shiga toxin-producing Escherichia coli O157:H7 in the UK . Microb Genomics . 2020 Mar ; 6 ( 3 ): e000334 . OpenUrl 112. ↵ Infernal 1.1: 100-fold faster RNA homology searches | Bioinformatics | Oxford Academic [Internet] . [cited 2025 Feb 7]. Available from: https://academic.oup.com/bioinformatics/article/29/22/2933/316439?login=false 113. ↵ Schwengers O , Jelonek L , Dieckmann MA , Beyvers S , Blom J , Goesmann A . Bakta: rapid and standardized annotation of bacterial genomes via alignment-free sequence identification . Microb Genomics . 2021 ; 7 ( 11 ): 000685 . OpenUrl 114. ↵ Tree JJ , Granneman S , McAteer SP , Tollervey D , Gally DL . Identification of Bacteriophage-Encoded Anti-sRNAs in Pathogenic Escherichia coli . Mol Cell . 2014 Jul ; 55 ( 2 ): 199 – 213 . OpenUrl CrossRef PubMed 115. ↵ Gruber CC , Sperandio V . Global Analysis of Posttranscriptional Regulation by GlmY and GlmZ in Enterohemorrhagic Escherichia coli O157:H7 . Infect Immun . 2015 Mar 17; 83 ( 4 ): 1286 – 95 . OpenUrl Abstract / FREE Full Text 116. ↵ Knutton S , Rosenshine I , Pallen MJ , Nisan I , Neves BC , Bain C , et al. A novel EspA-associated surface organelle of enteropathogenic Escherichia coli involved in protein translocation into epithelial cells . EMBO J . 1998 Apr 15; 17 ( 8 ): 2166 – 76 . OpenUrl Abstract / FREE Full Text 117. ↵ Iizumi Y , Sagara H , Kabe Y , Azuma M , Kume K , Ogawa M , et al. The Enteropathogenic E. coli Effector EspB Facilitates Microvillus Effacing and Antiphagocytosis by Inhibiting Myosin Function . Cell Host Microbe . 2007 Dec 13; 2 ( 6 ): 383 – 92 . OpenUrl CrossRef PubMed 118. ↵ Ogawa A , Kojima F , Miyake Y , Yoshimura M , Ishijima N , Iyoda S , et al. Regulation of constant cell elongation and Sfm pili synthesis in Escherichia coli via two active forms of FimZ orphan response regulator . Genes Cells . 2022 Nov ; 27 ( 11 ): 657 – 74 . OpenUrl CrossRef PubMed 119. ↵ Zangari T , Melton-Celsa AR , Panda A , Smith MA , Tatarov I , De Tolla L , et al. Enhanced virulence of the Escherichia coli O157:H7 spinach-associated outbreak strain in two animal models is associated with higher levels of Stx2 production after induction with ciprofloxacin . Infect Immun . 2014 Dec ; 82 ( 12 ): 4968 – 77 . OpenUrl Abstract / FREE Full Text 120. de Sablet T , Bertin Y , Vareille M , Girardeau JP , Garrivier A , Gobert AP , et al. Differential expression of stx2 variants in Shiga toxin-producing Escherichia coli belonging to seropathotypes A and C . Microbiol Read Engl . 2008 Jan ; 154 (Pt 1 ): 176 – 86 . OpenUrl 121. ↵ The Locus of Enterocyte Effacement and Associated Virulence Factors of Enterohemorrhagic Escherichia coli | Microbiology Spectrum [Internet] . [cited 2025 Mar 11]. Available from: https://journals.asm.org/doi/10.1128/microbiolspec.ehec-0007-2013 122. ↵ Roberts DR , Bahn V , Ciuti S , Boyce MS , Elith J , Guillera-Arroita G , et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure . Ecography . 2017 ; 40 ( 8 ): 913 – 29 . OpenUrl CrossRef 123. ↵ Teunis P , Takumi K , Shinagawa K . Dose Response for Infection by Escherichia coli O157:H7 from Outbreak Data . Risk Anal . 2004 ; 24 ( 2 ): 401 – 7 . OpenUrl CrossRef PubMed Web of Science 124. Tilden J , Young W , McNamara AM , Custer C , Boesel B , Lambert-Fair MA , et al. A new route of transmission for Escherichia coli: infection from dry fermented salami . Am J Public Health . 1996 Aug ; 86 ( 8 ): 1142 – 5 . OpenUrl CrossRef PubMed Web of Science 125. ↵ Tuttle J , Gomez T , Doyle MP , Wells JG , Zhao T , Tauxe RV , et al. Lessons from a large outbreak of Escherichia coli O157:H7 infections: insights into the infectious dose and method of widespread contamination of hamburger patties . Epidemiol Infect . 1999 Apr ; 122 ( 2 ): 185 – 92 . OpenUrl CrossRef PubMed 126. Goswami K , Chen C , Xiaoli L , Eaton KA , Dudley EG . Coculture of Escherichia coli O157:H7 with a Nonpathogenic E. coli Strain Increases Toxin Production and Virulence in a Germfree Mouse Model . Infect Immun . 2015 Nov ; 83 ( 11 ): 4185 – 93 . OpenUrl Abstract / FREE Full Text 127. Nawrocki EM , Mosso HM , Dudley EG . A Toxic Environment: a Growing Understanding of How Microbial Communities Affect Escherichia coli O157:H7 Shiga Toxin Expression . Appl Environ Microbiol . 2020 Nov 24; 86 ( 24 ): e00509 – 20 . OpenUrl PubMed 128. Russo LM , Abdeltawab NF , O’Brien AD , Kotb M , Melton-Celsa AR . Mapping of genetic loci that modulate differential colonization by Escherichia coli O157:H7 TUV86-2 in advanced recombinant inbred BXD mice . BMC Genomics . 2015 Nov 16; 16 ( 1 ): 947 . OpenUrl CrossRef PubMed 129. Rauw KD , Buyl R , Jacquinet S , Piérard D . Risk determinants for the development of typical haemolytic uremic syndrome in Belgium and proposition of a new virulence typing algorithm for Shiga toxin-producing Escherichia coli . Epidemiol Infect . 2019 Jan ; 147 : e6 . OpenUrl CrossRef 130. Isolation of Vero cytotoxin-producing Escherichia coli serotypes O9ab:H- and O101:H-carrying VT2 variant gene sequences from a patient with haemolytic uraemic syndrome - PubMed [Internet] . [cited 2025 Mar 11]. Available from: https://pubmed.ncbi.nlm.nih.gov/7889973/ 131. Stephan R , Untermann F . Virulence factors and phenotypical traits of verotoxin-producing Escherichia coli strains isolated from asymptomatic human carriers . J Clin Microbiol . 1999 May ; 37 ( 5 ): 1570 – 2 . OpenUrl Abstract / FREE Full Text 132. ↵ Fasel D , Mellmann A , Cernela N , Hächler H , Fruth A , Khanna N , et al. Hemolytic uremic syndrome in a 65-Year-old male linked to a very unusual type of stx2e- and eae-harboring O51:H49 shiga toxin-producing Escherichia coli . J Clin Microbiol . 2014 Apr ; 52 ( 4 ): 1301 – 3 . OpenUrl Abstract / FREE Full Text 133. ↵ Watanabe H , Wada A , Inagaki Y , Itoh K , Tamura K. Outbreaks of enterohaemorrhagic Escherichia coli O157:H7 infection by two different genotype strains in Japan, 1996 . Lancet Lond Engl . 1996 Sep 21; 348 ( 9030 ): 831 – 2 . OpenUrl 134. ↵ Iguchi A , Iyoda S , Terajima J , Watanabe H , Osawa R . Spontaneous recombination between homologous prophage regions causes large-scale inversions within the Escherichia coli O157:H7 chromosome . Gene . 2006 May 10; 372 : 199 – 207 . OpenUrl CrossRef PubMed 135. Wick LM , Qi W , Lacher DW , Whittam TS . Evolution of Genomic Content in the Stepwise Emergence of Escherichia coli O157:H7 . J Bacteriol . 2005 Mar ; 187 ( 5 ): 1783 – 91 . OpenUrl Abstract / FREE Full Text 136. ↵ Hayashi T , Makino K , Ohnishi M , Kurokawa K , Ishii K , Yokoyama K , et al. Complete genome sequence of enterohemorrhagic Escherichia coli O157:H7 and genomic comparison with a laboratory strain K-12 . DNA Res Int J Rapid Publ Rep Genes Genomes . 2001 Feb 28; 8 ( 1 ): 11 – 22 . OpenUrl 137. ↵ Sy BM , Tree JJ . Small RNA Regulation of Virulence in Pathogenic Escherichia coli . Front Cell Infect Microbiol . 2021 Jan 27; 10 : 622202 . 138. Sy BM , Lan R , Tree JJ . Early termination of the Shiga toxin transcript generates a regulatory small RNA . Proc Natl Acad Sci . 2020 Oct 6; 117 ( 40 ): 25055 – 65 . OpenUrl Abstract / FREE Full Text 139. ↵ Vogel J , Luisi BF . Hfq and its constellation of RNA . Nat Rev Microbiol . 2011 Aug ; 9 ( 8 ): 578 – 89 . OpenUrl CrossRef PubMed 140. ↵ Kendall MM , Gruber CC , Rasko DA , Hughes DT , Sperandio V . Hfq Virulence Regulation in Enterohemorrhagic Escherichia coli O157:H7 Strain 86-24 ▿ . J Bacteriol . 2011 Dec ; 193 ( 24 ): 6843 – 51 . OpenUrl Abstract / FREE Full Text 141. Hansen AM , Kaper JB . Hfq affects the expression of the LEE pathogenicity island in enterohaemorrhagic Escherichia coli . Mol Microbiol . 2009 Aug ; 73 ( 3 ): 446 – 65 . OpenUrl CrossRef PubMed Web of Science 142. Shakhnovich EA , Davis BM , Waldor MK . Hfq negatively regulates type III secretion in EHEC and several other pathogens . Mol Microbiol . 2009 Oct ; 74 ( 2 ): 347 – 63 . OpenUrl CrossRef PubMed 143. ↵ Wang D , McAteer SP , Wawszczyk AB , Russell CD , Tahoun A , Elmi A , et al. An RNA-dependent mechanism for transient expression of bacterial translocation filaments . Nucleic Acids Res . 2018 Apr 20; 46 ( 7 ): 3366 – 81 . OpenUrl PubMed 144. ↵ Kawamoto H , Koide Y , Morita T , Aiba H . Base-pairing requirement for RNA silencing by a bacterial small RNA and acceleration of duplex formation by Hfq . Mol Microbiol . 2006 Aug ; 61 ( 4 ): 1013 – 22 . OpenUrl CrossRef PubMed Web of Science 145. ↵ Małecka EM , Stróżecka J , Sobańska D , Olejniczak M . Structure of bacterial regulatory RNAs determines their performance in competition for the chaperone protein Hfq . Biochemistry . 2015 Feb 10; 54 ( 5 ): 1157 – 70 . OpenUrl CrossRef PubMed 146. ↵ Hussein R , Lim HN . Disruption of small RNA signaling caused by competition for Hfq . Proc Natl Acad Sci U S A . 2011 Jan 18; 108 ( 3 ): 1110 – 5 . OpenUrl Abstract / FREE Full Text 147. Santiago-Frangos A , Kavita K , Schu DJ , Gottesman S , Woodson SA . C-terminal domain of the RNA chaperone Hfq drives sRNA competition and release of target RNA . Proc Natl Acad Sci . 2016 Oct 11; 113 ( 41 ): E6089 – 96 . OpenUrl Abstract / FREE Full Text 148. ↵ Moon K , Gottesman S . Competition among Hfq-binding small RNAs in Escherichia coli . Mol Microbiol . 2011 Dec ; 82 ( 6 ): 1545 – 62 . OpenUrl CrossRef PubMed 149. ↵ Fournier BM , Parkos CA . The role of neutrophils during intestinal inflammation . Mucosal Immunol . 2012 Jul ; 5 ( 4 ): 354 – 66 . OpenUrl CrossRef PubMed Web of Science 150. Maloy KJ , Powrie F . Intestinal homeostasis and its breakdown in inflammatory bowel disease . Nature . 2011 Jun ; 474 ( 7351 ): 298 – 306 . OpenUrl CrossRef PubMed Web of Science 151. Segal AW . HOW NEUTROPHILS KILL MICROBES . Annu Rev Immunol . 2005 Apr 23; 23 (Volume 23, 2005): 197 – 223 . OpenUrl CrossRef PubMed Web of Science 152. ↵ Winterbourn CC , Kettle AJ , Hampton MB . Reactive Oxygen Species and Neutrophil Function . Annu Rev Biochem . 2016 Jun 2; 85 (Volume 85, 2016): 765 – 92 . OpenUrl CrossRef PubMed 153. ↵ Loś JM , Loś M , Wegrzyn A , Wegrzyn G . Hydrogen peroxide-mediated induction of the Shiga toxin-converting lambdoid prophage ST2-8624 in Escherichia coli O157:H7 . FEMS Immunol Med Microbiol . 2010 Apr ; 58 ( 3 ): 322 – 9 . OpenUrl CrossRef PubMed 154. ↵ Wagner PL , Acheson DW , Waldor MK . Human neutrophils and their products induce Shiga toxin production by enterohemorrhagic Escherichia coli . Infect Immun . 2001 Mar ; 69 ( 3 ): 1934 – 7 . OpenUrl Abstract / FREE Full Text 155. ↵ Loś JM , Loś M , Wegrzyn G , Wegrzyn A . Differential efficiency of induction of various lambdoid prophages responsible for production of Shiga toxins in response to different induction agents . Microb Pathog . 2009 Dec ; 47 ( 6 ): 289 – 98 . OpenUrl CrossRef PubMed 156. ↵ Licznerska K , Nejman-Faleńczyk B , Bloch S , Dydecka A , Topka G , Gąsior T , et al. Oxidative Stress in Shiga Toxin Production by Enterohemorrhagic Escherichia coli . Oxid Med Cell Longev . 2016; 2016 : 3578368 . 157. ↵ Atitkar RR , Hauser JR , Melton-Celsa AR . Shiga Toxin (Stx) Phage-Encoded Lytic Genes Are Not Required for the Mouse Virulence of O157:H7 Escherichia coli Stx2-Producing Clinical Isolates . Microbiol Spectr . 11 ( 3 ): e00372 – 23 . 158. ↵ Yokoyama K , Horii T , Yamashino T , Hashikawa S , Barua S , Hasegawa T , et al. Production of Shiga toxin by Escherichia coli measured with reference to the membrane vesicle-associated toxins . FEMS Microbiol Lett . 2000 Nov 1; 192 ( 1 ): 139 – 44 . OpenUrl CrossRef PubMed Web of Science 159. Bielaszewska M , Rüter C , Bauwens A , Greune L , Jarosch KA , Steil D , et al. Host cell interactions of outer membrane vesicle-associated virulence factors of enterohemorrhagic Escherichia coli O157: Intracellular delivery, trafficking and mechanisms of cell injury . PLOS Pathog . 2017 Feb 3; 13 ( 2 ): e1006159 . OpenUrl CrossRef PubMed 160. ↵ Rueter C , Bielaszewska M . Secretion and Delivery of Intestinal Pathogenic Escherichia coli Virulence Factors via Outer Membrane Vesicles . Front Cell Infect Microbiol [Internet ]. 2020 Mar 6 [cited 2025 May 28]; 10 . Available from: https://www.frontiersin.org/journals/cellular-and-infection-microbiology/articles/10.3389/fcimb . 2020.00091/full 161. ↵ Blanco M , Blanco JE , Mora A , Dahbi G , Alonso MP , González EA , et al. Serotypes, Virulence Genes, and Intimin Types of Shiga Toxin (Verotoxin)-Producing Escherichia coli Isolates from Cattle in Spain and Identification of a New Intimin Variant Gene (eae-ξ) . J Clin Microbiol . 2004 Feb ; 42 ( 2 ): 645 – 51 . OpenUrl Abstract / FREE Full Text 162. ↵ Abramson J , Adler J , Dunger J , Evans R , Green T , Pritzel A , et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3 . Nature . 2024 Jun ; 630 ( 8016 ): 493 – 500 . OpenUrl CrossRef PubMed 163. ↵ Trimmomatic: a flexible trimmer for Illumina sequence data | Bioinformatics | Oxford Academic [Internet] . [cited 2024 Oct 31]. Available from: https://academic.oup.com/bioinformatics/article/30/15/2114/2390096 164. ↵ Bankevich A , Nurk S , Antipov D , Gurevich AA , Dvorkin M , Kulikov AS , et al. SPAdes: A New Genome Assembly Algorithm and Its Applications to Single-Cell Sequencing . J Comput Biol . 2012 May ; 19 ( 5 ): 455 . OpenUrl CrossRef PubMed 165. ↵ Pedregosa F , Varoquaux G , Gramfort A , Michel V , Thirion B , Grisel O , et al. Scikit-learn: Machine Learning in Python . J Mach Learn Res . 2011 ; 12 ( 85 ): 2825 – 30 . OpenUrl CrossRef PubMed 166. ↵ Välimäki N. nvalimak/fsm-lite [Internet] . 2023 [cited 2024 Oct 31]. Available from: https://github.com/nvalimak/fsm-lite 167. ↵ Shi L , Westerhuis JA , Rosén J , Landberg R , Brunius C . Variable selection and validation in multivariate modelling . Bioinformatics . 2019 Mar 15; 35 ( 6 ): 972 – 80 . OpenUrl CrossRef PubMed 168. ↵ datarevenue-berlin/py-MUVR [Internet] . Data Revenue; 2024 [cited 2024 Oct 31]. Available from: https://github.com/datarevenue-berlin/py-MUVR 169. ↵ dmlc/xgboost: Scalable, Portable and Distributed Gradient Boosting (GBDT, GBRT or GBM) Library, for Python, R, Java, Scala, C++ and more . Runs on single machine, Hadoop, Spark, Dask, Flink and DataFlow [Internet] . [cited 2024 Oct 31]. Available from: https://github.com/dmlc/xgboost 170. ↵ Chawla NV , Bowyer KW , Hall LO , Kegelmeyer WP . SMOTE: synthetic minority over-sampling technique . J Artif Int Res . 2002 Jun 1; 16 ( 1 ): 321 – 57 . OpenUrl 171. ↵ scikit-learn-contrib/imbalanced-learn [Internet]. scikit-learn-contrib ; 2024 [cited 2024 Oct 31]. Available from: https://github.com/scikit-learn-contrib/imbalanced-learn 172. ↵ shap/shap: A game theoretic approach to explain the output of any machine learning model . [Internet]. [cited 2024 Oct 31]. Available from: https://github.com/shap/shap 173. ↵ Lundberg SM , Erion G , Chen H , DeGrave A , Prutkin JM , Nair B , et al. From local explanations to global understanding with explainable AI for trees . Nat Mach Intell . 2020 Jan ; 2 ( 1 ): 56 – 67 . OpenUrl PubMed 174. ↵ Prokka: rapid prokaryotic genome annotation | Bioinformatics | Oxford Academic [Internet] . [cited 2024 Oct 30]. Available from: https://academic.oup.com/bioinformatics/article/30/14/2068/2390517 175. ↵ Page AJ , Cummins CA , Hunt M , Wong VK , Reuter S , Holden MTG , et al. Roary: rapid large-scale prokaryote pan genome analysis . Bioinformatics . 2015 Nov 15; 31 ( 22 ): 3691 – 3 . OpenUrl CrossRef PubMed 176. ↵ Altschul SF , Gish W , Miller W , Myers EW , Lipman DJ . Basic local alignment search tool . J Mol Biol . 1990 Oct 5; 215 ( 3 ): 403 – 10 . OpenUrl CrossRef PubMed Web of Science 177. ↵ Li H , Handsaker B , Wysoker A , Fennell T , Ruan J , Homer N , et al. The Sequence Alignment/Map format and SAMtools . Bioinforma Oxf Engl . 2009 Aug 15; 25 ( 16 ): 2078 – 9 . OpenUrl 178. ↵ Sievers F , Higgins DG . The Clustal Omega Multiple Alignment Package . Methods Mol Biol Clifton NJ . 2021 ; 2231 : 3 – 16 . OpenUrl 179. ↵ ggmsa: a visual exploration tool for multiple sequence alignment and associated data | Briefings in Bioinformatics | Oxford Academic [Internet] . [cited 2025 Mar 5]. Available from: https://academic.oup.com/bib/article/23/4/bbac222/6603927 180. ↵ Bouras G , Nepal R , Houtak G , Psaltis AJ , Wormald PJ , Vreugde S . Pharokka: a fast scalable bacteriophage annotation tool . Bioinformatics . 2023 Jan 1; 39 ( 1 ): btac776 . OpenUrl CrossRef PubMed 181. ↵ Gilchrist CLM , Chooi YH . clinker & clustermap.js: automatic generation of gene cluster comparison figures . Bioinformatics . 2021 Aug 25; 37 ( 16 ): 2473 – 5 . OpenUrl CrossRef PubMed 182. ↵ Seemann T. tseemann/snippy [Internet] . 2024 [cited 2024 Oct 30]. Available from: https://github.com/tseemann/snippy 183. ↵ Croucher NJ , Page AJ , Connor TR , Delaney AJ , Keane JA , Bentley SD , et al. Rapid phylogenetic analysis of large samples of recombinant bacterial whole genome sequences using Gubbins . Nucleic Acids Res . 2015 Feb 18; 43 ( 3 ): e15 . OpenUrl CrossRef PubMed 184. ↵ IQ-TREE 2: New Models and Efficient Methods for Phylogenetic Inference in the Genomic Era | Molecular Biology and Evolution | Oxford Academic [Internet] . [cited 2024 Oct 31]. Available from: https://academic.oup.com/mbe/article/37/5/1530/5721363 185. ↵ SnapperDB: a database solution for routine sequencing analysis of bacterial isolates | Bioinformatics | Oxford Academic [Internet] . [cited 2024 Oct 30]. Available from: https://academic.oup.com/bioinformatics/article/34/17/3028/4961427 186. ↵ Ashton PM , Perry N , Ellis R , Petrovska L , Wain J , Grant KA , et al. Insight into Shiga toxin genes encoded by Escherichia coli O157 from whole genome sequencing . PeerJ . 2015 Feb 17; 3 : e739 . OpenUrl CrossRef 187. ↵ Kaniga K , Delor I , Cornelis GR . A wide-host-range suicide vector for improving reverse genetics in Gram-negative bacteria: inactivation of the blaA gene of Yersinia enterocolitica . Gene . 1991 Dec 20; 109 ( 1 ): 137 – 41 . OpenUrl CrossRef PubMed Web of Science 188. ↵ Ferrières L , Hémery G , Nham T , Guérout AM , Mazel D , Beloin C , et al. Silent Mischief: Bacteriophage Mu Insertions Contaminate Products of Escherichia coli Random Mutagenesis Performed Using Suicidal Transposon Delivery Plasmids Mobilized by Broad-Host-Range RP4 Conjugative Machinery . J Bacteriol . 2010 Dec 15; 192 ( 24 ): 6418 – 27 . OpenUrl Abstract / FREE Full Text View the discussion thread. Back to top Previous Next Posted June 06, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Predicting clinical outcome of Escherichia coli O157:H7 infections using explainable Machine Learning Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Predicting clinical outcome of Escherichia coli O157:H7 infections using explainable Machine Learning Julian A. Paganini , Suniya Khatun , Sean McAteer , Lauren Cowley , David R. Greig , David L. Gally , Claire Jenkins , Timothy J. Dallman medRxiv 2025.06.05.25329036; doi: https://doi.org/10.1101/2025.06.05.25329036 Share This Article: Copy Citation Tools Predicting clinical outcome of Escherichia coli O157:H7 infections using explainable Machine Learning Julian A. Paganini , Suniya Khatun , Sean McAteer , Lauren Cowley , David R. Greig , David L. Gally , Claire Jenkins , Timothy J. Dallman medRxiv 2025.06.05.25329036; doi: https://doi.org/10.1101/2025.06.05.25329036 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Infectious Diseases (except HIV/AIDS) Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (297) Cardiovascular Medicine (4421) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (606) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15212) Forensic Medicine (30) Gastroenterology (1121) Genetic and Genomic Medicine (6581) Geriatric Medicine (667) Health Economics (996) Health Informatics (4520) Health Policy (1366) Health Systems and Quality Improvement (1611) Hematology (539) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15906) Intensive Care and Critical Care Medicine (1103) Medical Education (620) Medical Ethics (144) Nephrology (667) Neurology (6580) Nursing (345) Nutrition (998) Obstetrics and Gynecology (1141) Occupational and Environmental Health (956) Oncology (3324) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1689) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5432) Public and Global Health (9212) Radiology and Imaging (2193) Rehabilitation Medicine and Physical Therapy (1368) Respiratory Medicine (1194) Rheumatology (593) Sexual and Reproductive Health (709) Sports Medicine (529) Surgery (709) Toxicology (99) Transplantation (288) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff5ece1e804d04e',t:'MTc3OTM5MTA4OA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();