Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance

Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance | bioRxiv /* */ /* */ <!-- <!-- /*! * 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-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance View ORCID Profile Andrew J. Balmer , View ORCID Profile Gemma G. R. Murray , View ORCID Profile Stephanie Lo , View ORCID Profile Olivier Restif , View ORCID Profile Lucy A. Weinert doi: https://doi.org/10.1101/2025.09.12.675231 Andrew J. Balmer 1 Department of Veterinary Medicine, University of Cambridge , Cambridge, CB3 0ES, UK 2 Genomic Surveillance Unit, The Wellcome Sanger Institute , Hinxton, Cambridge, CB10 1RQ, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Andrew J. Balmer For correspondence: ab69{at}sanger.ac.uk Gemma G. R. Murray 3 Department of Genetics, Evolution and Environment, University College London , Gower Street, London WC1E 6BT, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Gemma G. R. Murray Stephanie Lo 4 Parasites and Microbes, The Wellcome Sanger Institute, Wellcome Genome Campus , Hinxton, Cambridge, CB10 1RQ, UK 5 Milner Centre for Evolution, Department of Life Sciences, University of Bath , Bath, BA2 7AY, UK 6 The Great Ormond Street Institute of Child Health, University College London , London, WC1N 1EH, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Stephanie Lo Olivier Restif 1 Department of Veterinary Medicine, University of Cambridge , Cambridge, CB3 0ES, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Olivier Restif Lucy A. Weinert 1 Department of Veterinary Medicine, University of Cambridge , Cambridge, CB3 0ES, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Lucy A. Weinert Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract The global rise in antimicrobial resistance (AMR) has motivated large-scale surveillance studies of bacterial pathogens. Although these datasets contain rich information on drug resistance, they are both complex and multivariate, with each isolate’s susceptibility measured across several antibiotics. Current analytical methods for these datasets overwhelmingly focus on genotypes, while tools to study multivariate resistance phenotypes–the direct targets of selection and clinical intervention–remain severely limited. To address this gap, we develop “Antimicrobial Resistance Cartography”, a set of methods for interpreting high-dimensional drug resistance in large isolate collections. As a proof-of-concept, we analysed 3,628 Streptococcus pneumoniae isolates from the Active Bacterial Core (ABC) surveillance program, each with MICs measured for six beta-lactam antibiotics. We demonstrate how this toolkit simplifies visualisation of antibiotic resistance data across multiple drugs, resolves common issues such as missing or censored values, and identifies genetic mutations driving correlated resistance changes across multiple antibiotics. We characterise penicillin-binding protein (PBP) substitutions that increase minimum inhibitory concentration to multiple beta-lactam subclasses, revealing subtle differences in their correlated phenotypic effects, and identifying potential constraints on resistance evolution across genetic backgrounds. AMR cartography therefore provides a versatile framework for quantifying and visualising correlated resistance across antibiotics, enhancing the interpretability of complex susceptibility data for both research and public health. Introduction Antimicrobial resistance (AMR) was directly responsible for 1.27 million deaths globally in 2019, ranking among the World Health Organization’s ten most pressing threats to human and animal health ( 1 – 3 ). Controlling its spread requires careful monitoring of resistance frequencies in clinical settings. Consequently, over the past two decades, large-scale surveillance programs have amassed collections of thousands of bacterial samples ( 1 , 4 – 6 ), each tested for their sensitivity to multiple drugs, using phenotypic assays such as minimum inhibitory concentration (MIC) ( 7 – 10 ). Two observations repeatedly emerge from these studies: I) susceptibility profiles can vary widely–even within defined resistance categories, where MICs often span multiple dilution steps (e.g. 4, 8, 16, or 32 μg/mL), and II) susceptibility phenotypes across drugs are often correlated ( 11 – 14 ). This covariation can arise through several mechanisms: in some cases, one genotype confers resistance to multiple antibiotics; in others, epidemiological factors, such as sequential treatment with separate drugs, generate co-resistance ( 11 ). Understanding how these correlations arise and persist is crucial, because they can affect choice of drug combinations and influence bacterial evolution during exposure to subsequent antibiotics–with the potential to accelerate multidrug resistance and restrict treatment options ( 13 ). There is therefore a growing need for methods that can visualise and simplify these complex phenotypic patterns and understand how they relate to genetic and epidemiological data ( 11 , 15 ). However, to develop these tools, three main methodological challenges must be addressed. Firstly, MIC values are often reduced to binary resistant/susceptible categories based on clinical or epidemiological cut-offs ( 9 , 10 ). While useful for guiding patient treatment, these simplifications discard fine-scale quantitative information that is crucial in clinical, epidemiological and evolutionary contexts. For example, gradual MIC increases can reduce treatment efficacy, prolong infections, and confer selective advantages during low-level antibiotic exposure–common in both carriage sites and environmental reservoirs ( 16 – 18 ). Secondly, tools are needed to analyse joint susceptibility profiles for multiple drugs, rather than each drug independently ( 19 ). Clinically, this is essential for selecting effective combination therapies, but it is also important from an evolutionary perspective, because genetic changes that affect multiple phenotypes can shape subsequent trajectories ( 20 ). For instance, if a single mechanism confers resistance to multiple antibiotics, selection for resistance to any of those drugs may propagate multidrug resistance ( 11 , 21 , 22 ). Finally, there remain gaps in the standardisation of antibiotic susceptibility assays, making it difficult to compare variation across studies ( 23 – 25 ). Issues like missing MIC values, measurement errors, and multiple assay methods within a single collection often necessitate separate statistical analyses per drug, complicating clinical surveillance and reducing statistical power to identify subtle, but important evolutionary changes over time ( 23 , 26 , 27 ). Existing multivariate approaches–such as pairwise correlations, categorical classification rules, and multi-trait GWAS–have been developed to address some aspects of these problems. However, they are limited in their ability to capture fine-scale variation in continuous MIC data, identify interaction effects, and provide an intuitive representation of high-dimensional resistance in relation to metadata ( 11 , 19 , 28 ). Addressing all of these limitations requires bespoke methods to study patterns of phenotypic variation and covariation and test hypotheses on their causes. To overcome these challenges, we propose an approach inspired by ‘Antigenic Cartography’. Originally used to visualise antigenic evolution in H3N2 influenza, antigenic cartography has enabled the precise tracking of antigenic changes in viral pathogens–guiding vaccine design and improving understanding of immune responses to evolving pathogens ( 29 – 32 ). The tool is based on a dimensionality reduction method called multidimensional scaling (MDS)–a statistical technique that visualises high-dimensional data by placing isolates in a low-dimensional phenotypic ‘map’–where distances between points represent differences in measured assay data. This enables cartography to summarise complex antigenic data in a concise, visual format, reduce measurement error, and improve interpretation of phenotypic differences among viral strains. Since its assumptions do not restrict its application to antigenic phenotypes, we adapted it to map antimicrobial resistance phenotypes, using MIC profiles for multiple antibiotics. We call this tool ‘Antimicrobial Resistance Cartography’. To develop this toolkit, we focused on beta-lactam resistance in Streptococcus pneumoniae , a major human pathogen responsible for ∼1.2 million infections and 7000 deaths annually in the USA alone ( 33 ). The extent of beta-lactam use in treating Streptococcal infection has driven widespread resistance–in some cases repeatedly, over short timeframes and across diverse populations ( 33 – 35 ). In streptococci, beta-lactams bind to the active site of penicillin binding proteins (PBPs), inhibiting their role in peptidoglycan cell wall biosynthesis ( 36 ). Resistance primarily arises through the modification of three PBPs—PBP1A, PBP2B, and PBP2X—where changes in the transpeptidase domains reduce antibiotic binding affinity ( 37 ). The variants in these proteins that reduce antibiotic binding (i.e. causal variants) are contained within highly variable ‘mosaic’ blocks, formed via recombination with other streptococci species ( 38 – 40 ). However, within these blocks, many substitutions are merely hitchhikers (i.e. variants that do not directly affect resistance), and others only increase MIC in combination or alter the fitness effects of primary resistance mutations ( 38 , 41 – 43 ). While GWAS and machine learning methods reliably link changes in the PBPs with non-susceptible phenotypes, precisely identifying the phenotypic effects of individual substitutions across multiple beta-lactam subclasses remains challenging, particularly when GWAS analyses each antibiotic separately or as binary resistant/susceptible traits, as this limits the detection of correlated effects across drugs ( 26 , 37 , 38 , 44 ). Identifying these correlated effects is critical, as this would enable more accurate prediction of resistance frequencies, inform diagnostic assays, and allow targeted surveillance of emerging resistance. Here, we make full use of the fact that beta-lactam phenotypes are both continuous and multivariate. While all beta-lactams target PBPs, different subclasses bind to each PBP with distinct binding affinities–for example, penicillins (e.g. penicillin, amoxicillin) primarily bind to PBP2X and PBP2B, whereas cephalosporins bind predominantly to PBP2X and 1A ( 45 – 47 ). Because of this, a single PBP mutation can affect susceptibility to multiple beta-lactams in different ways ( 47 , 48 ). Multiple PBP alterations are therefore required for high MIC–each individually with minor effects to one or more drugs–but cumulatively leading to clinical resistance across several drugs ( 36 – 38 ). Although subclasses show positively correlated MICs (i.e. partial cross-resistance), these differences in binding affinity mean beta-lactam resistance is not a single phenotype, but rather a complex, interrelated set of traits. This makes S. pneumoniae an ideal model system for developing and applying methods to visualise multivariate drug resistance phenotypes, and to use them to interpret how mutations collectively confer high-level resistance. These tools would then allow us to investigate whether key PBP substitutions decrease susceptibility to one drug but increase others, and quantify how phenotypes differ across phylogenetic lineages (i.e. Multilocus Sequence Types - MLST) ( 49 , 50 ). As a proof of concept, we apply ‘AMR Cartography’ to a large collection of clinical isolates of S. pneumoniae for which joint phenotypic and genotypic data were available: specifically, we focused on MIC values against six beta-lactams and amino acid sequences of the transpeptidase regions of PBPs. We have three objectives: I) quantitatively visualise and interpret continuous covariation in susceptibility across six beta-lactams; II) examine how these phenotypes associate with specific phylogenetic lineages (MLSTs); and III) identify and rank PBP amino acid substitutions according to the strength and consistency of their association with increased beta-lactam MIC. In addition, we demonstrate that our approach mitigates common experimental issues, such as measurement error, censored MIC values, and missing data, and increases the resolution of genotype-phenotype associations, including the detection of epistatic effects. Methods Isolates and metadata We used a published dataset of 4,309 invasive pneumococcal isolates, collected through the Active Bacterial Core surveillance (ABC) programme between 1998 and 2015 ( 37 , 51 ). The ABC project is run by the Centers for Disease Control and Prevention (CDC) Emerging Infectious Disease programme ( 5 ). As part of this initiative, microbiology laboratories in participating hospitals identify cases of invasive pneumococcal disease, conduct laboratory testing, and share data with the programme. Isolates were defined as pneumococcal strains isolated from a normally sterile site and were collected from clinical cases within ten U.S. states, each representing different geographic regions. MICs were previously measured for six beta-lactam antibiotics using two-fold microbroth dilution methods by the ABC program (Supplementary Table S1). For our analysis, we focused on 3,628 isolates with MIC available for 6 beta-lactams, of three subclasses: penicillin, amoxicillin (penicillins), ceftriaxone, cefotaxime, cefuroxime (cephalosporins), and meropenem (a carbapenem). Whole genome sequencing for all isolates was performed using the Illumina HiSeq or MiSeq platforms ( 37 , 51 ), and sequencing data were used to determine capsular serotype, MLST, and transpeptidase sequences for penicillin binding proteins PBP1A, PBP2B and PBP2X. Additional metadata—including date, geographical location, and clinical symptoms—are publicly available, but were not used here ( 37 , 51 ). Phenotype map construction Phenotype maps visualise a Euclidean distance matrix derived from MIC data. To run multidimensional scaling (SMACOF) using MIC data, MIC values were log₂ transformed to approximate a standard distribution ( 52 , 53 ). Titres for the six drugs were then transformed into a multivariate distance matrix by calculating pairwise Euclidean distances between isolates ( D i , j ), therefore representing differences across all drugs simultaneously. Given this set of target distances, MDS finds a low-dimensional representation of those distances with minimal error ( 52 , 53 ). Distances in the representation are determined by minimising the differences between map distances ( d i , j ) and table distances (D i , j ) for each pair of isolates i and j, using an error function called stress ( S ): Here, ‘stress’ is the sum of the squared residuals between points on the map and their measured distances. The MDS algorithm searches for a minimum stress value by iteratively shifting the positions of isolates on the map. As MDS can be applied to either metric or rank data, both numeric and censored MIC values can be represented (Supplementary Section S1.6). We also developed “genetic maps”, which visualise genetic differences among isolates using a Hamming distance matrix based on an amino acid sequence alignment (Supplementary Figure S8) ( 30 ). We evaluated the goodness-of-fit of these maps using cross-validation, confidence intervals and dimensionality testing (Supplementary Section S1). Genotype-phenotype association methods To elucidate the molecular basis of multivariate phenotypic variation, we focused on amino acid substitutions in the transpeptidase regions of PBPs, given their known role in beta-lactam resistance. We used four methods to identify substitutions associated with changes in position on the phenotype map (Supplementary Section S2.5-8), each of which were originally developed to identify substitutions underlying antigenic change in H3N2 influenza ( 30 , 54 ). First, we identified pairs of isolates that varied by a single amino acid and calculated the phenotypic distances between them. Second, we used clustering algorithms to partition the phenotype map into distinct regions and identify substitutions associated with these regions. Third, we used multivariate linear mixed models (mvLMM) to test associations between amino acid substitutions and isolate position on the phenotype map, and estimate the phenotypic effect sizes of each substitution ( 55 , 56 ). Lastly, we extended the mvLMM to test for interaction effects between PBP substitutions, while including the individual markers as covariates (see Supplementary Section S2.9). To enhance statistical power relative to conventional genome-wide association approaches, we used several methods to reduce the correction for multiple testing burden, such as using amino acid changes instead of nucleotide changes, and restricting associations to PBP transpeptidase sequences (Supplementary Section S2.10) ( 54 ). Results Antimicrobial Resistance Cartography of Streptococcus pneumoniae We analysed 3,628 invasive pneumococcal isolates from the Active Bacterial Core (ABC) surveillance program, with MICs measured for six beta-lactam antibiotics using two-fold microbroth dilution. Across the entire collection, most isolates exhibited low MIC values for all six antibiotics, reflecting a high level of susceptibility to the drugs. Despite this, MIC values varied considerably across isolates–with penicillin MICs ranging from ≤0.03 to ≥8μg/mL. MIC values were positively correlated across all beta-lactams, although the strength of these correlations varied, particularly among drugs with many upper and lower censored values e.g. cefuroxime (Pearson correlation coefficients = 0.77-0.96). We visualised multivariate resistance profiles among the isolates using multidimensional scaling on a Euclidean distance matrix derived from log₂ transformed MIC data ( Figure 1A ). On the resulting phenotype ‘map’, each isolate is represented as a point, with distances between them reflecting differences in MIC values across the six antibiotics ( Figure 1B ). Instead of simplifying data into binary resistant/susceptible categories, this approach provides an intuitive representation of the full distribution of multivariate phenotypes, capturing both fine-scale MIC variation between isolates and the underlying correlations in resistance profiles across drugs. Most isolates displayed low MIC values for all antibiotics, and these are represented by the cluster in the lower-left region of the figure. The map also shows substantial variability in MIC profiles among the remaining isolates. For example, isolates with intermediate penicillin MIC (0.12–1μg/mL) varied widely in their MIC values for other drugs, sometimes differing by up to six two-fold dilution steps, yet remaining below the clinical breakpoint for penicillin (2μg/mL) ( Figure 1 ). Additionally, a subset of isolates in the top-right of the map had uniformly high penicillin MIC but varied in MIC to cephalosporins and carbapenems ( Figure 1C , Supplementary Figure S2). The overall diagonal distribution of isolates on the map reflects the strong positive correlations in MIC values across beta-lactams, but also demonstrates that correlations between different drugs are not perfectly linear. Download figure Open in new tab Figure 1. A) MIC measurements for 3628 isolates of S. pneumoniae across six beta-lactam drugs, visualised using a pairwise correlation plot (Pearson). Axes represent the dilution step in a MIC series for the 6 beta-lactam drugs, where red lines indicate clinical breakpoints for each drug. The upper triangle shows the pairwise correlation coefficients for each pair of drugs. Although multivariate MIC data is often visualised using pairwise correlation plots such as these, it is difficult to interpret susceptibility profiles across multiple drugs for any given isolate. For example, isolates might have varying resistance profiles e. g. S-S-S-R-R-R, or S-R-R-R-R-R, but it is not possible to interpret these nor their relative proportions. B) Antimicrobial Cartography map for the same 3628 isolates, coloured by median MIC dilution step (i.e. 1 = the lowest concentration measured for a given drug). Points represent isolates, while the two primary axes of the map represent phenotypic distance in MIC units. One grid space corresponds to one average two-fold dilution step across the six beta-lactam assays. As multiple isolates can occupy the same point in the map, density distributions (red area – 50%, orange - 80% and blue, 95% of isolates) and marginal histograms are used to highlight the distribution of isolates. Biplot axes for each axis indicate the direction of increasing MIC for the drugs. These axes indicate penicillin MIC strongly correlates with the vertical axis, amoxicillin correlates with both the vertical and horizontal axes, and the remaining drugs primarily align horizontally. C) Biplot axes calibrated to show the MIC distribution of isolates on the map, where isolates in line with a value on the scale have that MIC value. Isolates are coloured by Penicillin MIC. Clinically relevant cut-offs can be included to aid visualisation (red and pink lines indicate resistance cut-off for Penicillin in non-meningitis and meningitis infections respectively - 2μg/mL & 0.06μg/mL). To evaluate the accuracy of the map, we calculated the pairwise Euclidean distances between isolates in the original six-dimensional MIC space and compared them to their relative distances on the 2D map. This allowed us to quantify how well the 2D projection preserves the original multivariate relationships. Fewer than 0.3% of pairwise distances had a discrepancy greater than one log₂ dilution between their true multivariate MIC distance and their map-based distance ( Table 1 , Supplementary Figure 2). This indicates that a two-dimensional map is an accurate representation of this set of MIC values. The map also displayed high stability across a range of resampling, dimensionality, and cross-validation tests (Supplementary Sections S1.7-1.10). View this table: View inline View popup Download powerpoint Table 1. Goodness-of-fit metrics across MDS dimensionalities for the phenotype map. Projection error refers to the mean change in each unique phenotype’s position when adding an additional dimension to the map. Values therefore reflect the relative mean decrease in error when moving from 1D to 2D, 2D to 3D, etc. A higher value indicates a larger gain in information from adding an additional dimension. Note the sharp reduction in overall stress when moving from a 1D to 2D map (55.5%). Although isolates could, in theory, be positioned anywhere throughout the multidimensional space, they predominantly aligned along a near-linear straight line. Nevertheless, goodness-of-fit comparisons confirmed that a two-dimensional map provided a significantly better representation of fine-scale differences in MIC profiles among isolates than a one-dimensional projection, as indicated by the large decrease in error when moving from a one- to two-dimensional map ( Table 1 , Supplementary Sections S1.5 and S1.12). Notably, although mean pairwise error values appear low across all dimensions, this is due to the fact that many isolates exhibited uniformly low, censored MIC values, and were consequently well fit in a single dimension. However, phenotypes with elevated MIC contributed proportionally much higher stress when represented in a single dimension, as indicated by the higher pairwise error and decrease in stress values. This was further supported by a projection error analysis, which quantified the average change in position of each unique phenotype when adding additional dimensions, showing a large decrease in error when moving from one- to two-dimensions (1.28 MIC units). While higher-dimensional projections further reduced error ( Table 1 ), improvements beyond two dimensions were more modest, suggesting 2D maps adequately capture key phenotypic patterns for this dataset. Moreover, as 2D maps performed better in cross-validation analyses, and isolate distributions on the 2D and 3D maps remained directly comparable, the 2D maps are presented throughout this manuscript. However, future studies should explicitly evaluate dimensionality and goodness-of-fit, as other datasets or bacterial species may require additional dimensions. Robustness of AMR Cartography in addressing common assay issues Theoretically, by positioning isolates using data from multiple drugs, the map can use the associations between them to more accurately infer MIC values–even when underlying data is missing or inaccurate. In principle, this means the map can offer greater accuracy than using MIC values for any single drug individually. We tested whether this was the case by assessing whether the map could alleviate common measurement issues in resistance surveillance datasets, such as missing values, experimental error, and combining multiple assay methods. To test robustness in handling missing data, we randomly removed 10% of MIC values, reconstructed the map with the remaining data, and predicted the missing MICs using the relative distances between points (Supplementary Figure S4). Across 100 iterations, 86.8% of prediction errors for missing MIC values were within a single MIC dilution (mean error = 0.467)—the range of error expected if isolates were remeasured experimentally ( 7 , 9 , 10 )—demonstrating the robustness of the map in imputing missing data ( Table 2 , Supplementary Figure S4). Prediction error was low across the distribution of the map, though typically increased with the number of missing MIC values for a given isolate. This was because isolates with additional missing values had less information on where to position them, although error still typically remained below 2 MIC units (Supplementary Figure S4). View this table: View inline View popup Download powerpoint Table 2. Results of error-prediction analyses for the phenotype map. Each row summarises results from 100 datasets, each adjusted by either removing values (Missing), adding random measurement error (Noise), or integrating categorical susceptibility data with MIC data (Combining MIC and categorical sensitivity). For isolates with altered MIC values, prediction error was measured as the mean Euclidean distance between predicted and observed positions of the isolates on the map (in MIC units) (Supplementary Figures S4-6). The proportion of errors within specified thresholds (2 MIC units) as well as the mean relative change in each isolate’s stress contribution (as a % of total stress) were also calculated. Negative values indicate a reduction in stress contribution of the isolates compared to their contribution when values were not altered. The same imputation method also allowed multiple antimicrobial susceptibility assay results to be combined within a single analysis. For example, isolates characterised by categorical susceptibility only (e.g., susceptible/resistant profiles–as derived from disc-diffusion tests) could be accurately positioned on the phenotype map, enabling imputation of numeric MIC values within a single dilution (mean error = 0.320) (Supplementary Figure S5). Moreover, the framework provided tools to quantify and mitigate error (Supplementary Figure S1 and Section S1.10), delineate confidence intervals for isolate phenotypes (Supplementary Figure S3), and identify samples that have potential measurement issues based on their relative stress contribution (Supplementary Section S1.10). The method also improved interpretation of censored MIC values i.e. those beyond the dilution range of the assay, by positioning points using the information available for other drugs (Supplementary Section S1.11). Associating phenotypes with MLST lineages Overlaying the 12 most common MLST types (of those with non-sensitive isolates) onto the phenotypic map revealed that lineages occupy distinct regions of phenotypic space, highlighting between-lineage variation in beta-lactam MIC ( Figure 2A-B ). Isolates within the same MLST mostly cluster on the map, indicating similar resistance profiles. For example, isolates of MLST 695 and 1840 showed mean pairwise phenotypic distances of 1.68 and 1.02 MIC units (Supplementary Table S11). However, some MLSTs (63, 156, and 1092) exhibited broader variation, with mean pairwise distances over 2 MIC units. Download figure Open in new tab Figure 2. Phenotype map coloured by the 12 most common MLSTs among non-sensitive isolates. A) Coloured points highlight isolates from the 12 MLSTs; smaller grey points indicate isolates of all other MLST types. B) Isolates of each MLST type are shown separately. Shaded regions indicate the range of locations for each unique PBP sequence centroid within that MLST. Most MLSTs cluster tightly, though MLST 63, 199, and 558 show greater phenotypic variability, spanning both susceptible and intermediate MIC values. Three MLST types had extensive modification of their PBP-types but showed low-to-intermediate MIC values (338, 1373 and 3280). C) Relationship between genetic divergence in PBP transpeptidase regions and phenotypic distance on the map (MIC units), relative to the most common susceptible PBP-type (2-0-2). Small points represent the median distances for each unique PBP-type (coloured by MLST); larger points show the median distances across all PBP-types within each MLST. Previous findings have shown elevated MIC correlates with genetic divergence among PBPs ( 37 ). While this was also observed here, isolates of MLSTs 338, 1373, and 3280 had extensive modification of their PBPs relative to the most common sensitive PBP-type (mean 71.53, 64.85 and 93.52 amino acid changes respectively), but exhibited only low-to-intermediate MICs (4.52, 3.16 and 5.50 MIC units difference) ( Figure 2C ). This highlights that while overall divergence in PBP sequence is strongly correlated with phenotypic distance, the strength and direction of this relationship varied across lineages. However, substantial shifts on the phenotype map did not occur without a corresponding change in PBP sequence, highlighting genetic variation in PBPs as a prerequisite for change in MIC. Together, these observations suggest genetic changes in the PBPs correspond strongly with differences in MICs; yet, the precise phenotypic effects of these changes depend on the specific underlying molecular variants (see below), and these can vary across evolutionary lineages. Identifying and visualising PBP substitutions underlying variation in beta-lactam MI Multivariate methods increase statistical power to identify substitutions with effects on multiple drugs, potentially revealing trade-offs between beta-lactam subclasses ( 26 , 28 , 44 , 54 ). We therefore used these methods to identify specific PBP substitutions underlying variation in MIC. Variation in PBP transpeptidase sequences accounted for almost all variation on the phenotypic map (broad-sense heritability, H2 = 97.6% and 95.8%, narrow-sense heritability, h2 = 71% and 57.2% for axes 1 and 2 respectively), with PBP2X accounting for the majority (76% and 61.3%, respectively). To identify specific amino acid substitutions associated with changes in beta-lactam MIC, we applied four complementary approaches (Supplementary Section S2.5-9). First, we identified all pairs of isolates differing by a single amino acid substitution and quantified the corresponding shift in their positions on the phenotype map. Second, we applied clustering algorithms to define distinct groups of PBP-types and tested which substitutions segregated between regions of the phenotype map. Third, we used a multivariate linear mixed model (mvLMM) on the map axes, testing for an association between each amino acid substitution and changes in multivariate resistance phenotype. Fourth, we extended the mvLMM to test for epistatic (interaction) effects between substitutions while controlling for individual effects. Substitutions were considered to have the strongest evidence if they were supported by multiple methods. In total, we identified 172 amino acid changes, at 147 unique PBP locations, as associated with phenotypic effects on beta-lactam MIC, representing 51.6% of all variable positions within the transpeptidase regions ( Table 3 and Supplementary Section S2.11). Of these, 88 changes at 81 sites were supported by multiple methods (18 in PBP1A, 24 in PBP2B, 39 in PBP2X), with PBP2X I371T identified in all four analyses ( Figure 3 ). The effect sizes of these substitutions were within the range of 0.33-3.079 MIC units. Most of the 88 substitutions showed both additive and epistatic effects (89.8%), with interactions both within and across proteins (Supplementary Section S2.11). Download figure Open in new tab Figure 3. Visualising phenotypic effects of specific PBP substitutions on beta-lactam MICs A) A single substitution comparison, where isolates on the phenotype map are coloured by the amino acid at location PBP2X-371. B) A subset of isolates (larger points) have the same PBP genetic background except for a single substitution at location PBP2X-371. The shaded areas highlight the distribution of the groups that differ in their amino acid at this location, while the coloured crosses and black line represent the median centroid positions and the phenotypic distance between them respectively. C) Amino acid changes can be directly visualised on the map by colouring isolates by the amino acid at a given position. The hierarchical structuring of substitutions highlights some only appear on specific genetic backgrounds (left to right) (Supplementary Figure S11). View this table: View inline View popup Download powerpoint Table 3. PBP substitutions associated with changes in beta-lactam MIC (only the 20 substitutions with the largest effect sizes are shown). The strength of evidence for each substitution is categorised by the number of methods which identified a phenotypic effect: 4 methods - ‘Very Strong’ evidence, 3 – ‘Strong’, 2 - ‘Moderate’, 1 – ‘Weak’, and 0 – ‘No Evidence’. ‘-’ indicates an absence of evidence for a given method, such as positions without single substitution comparisons (full table in Supplementary File 1). The column ‘Sig. mvLMM/uvLMM’ compares the results of the multivariate and univariate LMMs. ‘Yes’ and ‘No’ indicates whether a substitution was significant in the mvLMM on map axes, and the uvLMM tests on individual MIC values analysis respectively. The numbers in brackets show the number of drugs which were significantly associated in the uvLMM. Typically, substitutions had specific effects on individual drug subclasses (59.08%), rather than common effects across multiple subclasses (13.63%) (Supplementary Section S3.1). However, several substitutions in PBP2X exhibited large correlated effects across multiple drugs, including I371T, N444S and S531Y. In contrast, many PBP2B variants showed contrasting phenotypic effects, by increasing penicillin MIC but reducing cephalosporin and meropenem MIC, e.g. P568S, L609S and G467L/S469N/G483A/A490S. Notably, isolates from MLSTs 338 and 1373 had distinct combinations of these PBP2B mutations, even though these were uncommon in other MLST types. Correspondingly, these MLSTs showed intermediate penicillin MICs but minimal increases in cephalosporin MICs, reflecting the contrasting phenotypic effects of their PBP substitutions ( Figure 2 ). Improved resolution of multivariate over univariate methods Using several combined multivariate methods (alongside techniques to reduce multiple testing burden) increased statistical power to detect associations between PBP substitutions and phenotypic effects. After assessing the PBP positions previously identified by GWAS, random forest and in vitro methods, we found the cartography framework demonstrated improved overlap across each of these methods relative to any single previous method (Supplementary Figure S19). Moreover, while our primary focus was on associating PBP substitutions with multivariate resistance profiles, we also compared the results of the multivariate mvLMM to conventional univariate methods applied to individual MICs for each drug. Multivariate mvLMMs detected 94 substitutions with phenotypic effects not identified by univariate LMMs on MICs for each drug individually ( Table 3 and Supplementary Section S3.3). These included several substitutions with contrasting subclass effects—such as those in PBP1A and PBP2B—many of which were also supported by both clustering and single-substitution analyses. Discussion In this study, we introduced AMR Cartography—a multivariate framework for analysing complex, continuous resistance phenotypes—and applied it to 3,628 Streptococcus pneumoniae isolates with MICs measured for six beta-lactam antibiotics. Our analysis revealed three main findings. First, a two-dimensional phenotype map accurately captured correlated changes in MIC values across six beta-lactam subclasses, providing simple visualisations of complex datasets with minimal information loss. Although susceptibility profiles were highly correlated across drugs, phenotypes deviated from a strictly linear path, as shown by the improved fit of a two-dimensional over a one-dimensional map. We hypothesise these divergences arise from intrinsic variation in drug-protein binding and the relative use of specific beta-lactam subclasses across populations ( 30 , 57 ). Second, genetic background was strongly associated with resistance phenotype, with many substitutions only appearing on the background of others, a result consistent with previous findings, but demonstrated here across multiple beta-lactam subclasses ( 37 , 41 , 58 ). While some PBP substitutions were common across MLST types, others were lineage-specific, emphasising the role of genetic context in resistance evolution. Third, the cartography framework enabled us to identify 88 PBP substitutions–primarily in PBP2X and PBP2B–associated with increased beta-lactam MIC, including several with contrasting effects across subclasses, highlighting the multivariate nature of beta-lactam resistance in this species. Together, these findings highlight the epidemiological relevance of sub-clinical MIC changes, the importance of multivariate genotype-phenotype comparisons, and the value of novel tools for phenotypic surveillance. Beyond these core findings, AMR Cartography addresses several technical challenges common in resistance monitoring. By leveraging correlations in MIC values across antibiotics, the framework can impute missing or censored measurements while mitigating experimental error. We also showed it can integrate categorical susceptibility data with MIC data in the same analysis, enabling imputation of numeric MICs from disc-diffusion assays. These features improve the curation and quality of phenotypic data, while facilitating integration with genotype-phenotype prediction methods ( 9 , 37 , 59 ). We also introduced methodological refinements to genotype-phenotype association testing in bacteria, such as fine-mapping within specific genomic regions and using amino acid changes rather than nucleotides ( 28 , 54 , 60 ). Combined with the incorporation of sub-clinical variation in MIC and multivariate methods, these refinements increased statistical power (by reducing multiple-testing burden) and improved overlap with computational and in vitro results ( 28 , 37 , 38 , 44 ). Together, these tools enhance the value of resistance surveillance data and provide a quantitative, generalisable framework for analysing phenotypes, analogous to the way phylogenetic methods are used for genetic data. We designed AMR Cartography to be broadly adaptable to other pathogens and drug classes, provided phenotypic correlations exist within a species. The tool can therefore be applied to study shared resistance mechanisms, epidemiological patterns, and co-selective pressures across a range of bacterial datasets. Beta-lactam resistance in S. pneumoniae offered an ideal case study for these methods, due to its complex genetic basis and phenotypic variation and covariation ( 11 , 14 , 38 ). Applying the framework to this species allowed us to pinpoint specific PBP substitutions with phenotypic effects and quantify how they collectively drive covariation in resistance. The resulting list of PBP substitutions—including effect sizes, specific vs. common effects, and evidence strength—will aid in tracking resistance among S. pneumoniae populations ( 31 , 58 ), particularly where phenotypic data are unavailable, or novel combinations of substitutions emerge ( 33 , 37 , 51 , 61 ). These markers may also help monitor resistance during clinical interventions, such as serotype displacement after vaccination ( 62 ). Our findings reaffirm that multiple PBP substitutions underlie increased beta-lactam MIC, each exerting small to moderate effects (0.33-3.079 MIC units). Remarkably, 147 of 285 variable positions (51.6%) within the PBP transpeptidase regions showed some evidence of phenotypic effects, many with both additive and epistatic effects. 81 of these positions had evidence from multiple methods, a proportion substantially higher than previous studies. For example, Chewapreecha et al. and Li. et al. highlighted 31 and 27 positions as associated with high-level resistance to beta-lactams, respectively ( 33 , 37 , 38 ). While this increase may be driven in part by differences across datasets, our results highlight the greater sensitivity of multivariate approaches in detecting subtle and correlated genotype-phenotype associations ( 28 , 38 , 44 ). Many substitutions altered MIC below clinical breakpoints, demonstrating the value of detecting variants which generate sub-breakpoint changes in MIC, in addition to those which cause transitions between clinical thresholds. The presence of such variants among bacterial populations suggests small, incremental MIC changes are epidemiologically relevant–either by directly contributing to treatment failure rates or by facilitating stepwise progression toward clinically significant resistance ( 27 , 63 ). Even if transitory, these changes may confer competitive advantages under variable drug exposure, such as in carriage sites or biofilm formations ( 16 , 64 , 65 ). Early detection of such changes is therefore essential for informing treatment guidelines and pre-empting the emergence of high-level resistance ( 18 , 27 , 63 , 64 ). Further validation of these substitutions in additional isolate collections will be required, along with exploration of other genomic regions known to affect beta-lactam MIC (e.g. murM ) ( 58 ). While some PBP2X substitutions (e.g. I371T, N444S and S531Y) had large, correlated effects across multiple beta-lactams, most substitutions showed subclass-specific effects (59.08%). Some had contrasting effects–for example, several changes in PBP2B increased penicillin MIC but reduced cephalosporin MIC (e.g. P568S and G467L/S469N/G483A/A490S). This indicates these substitutions decrease PBP2B binding affinity for penicillins, while simultaneously increasing it for cephalosporins ( 45 , 47 , 48 ). Such contrasting effects are reflected in the two-dimensional phenotype map, which shows that although PBP changes can exhibit correlated effects, they often diverge, resulting in multidimensional phenotypes. This suggests that selection from one drug subclass may not necessarily confer cross-resistance to others; instead, independent changes are possible. The precise multivariate phenotypes under selection likely depend on local drug usage, fitness dynamics, linkage disequilibrium, and the target-binding specificity of drug structures ( 43 , 47 , 48 , 65 , 66 ). In this case, the strong correlations in MIC between subclasses have likely emerged through co-occurring selective pressures, rather than direct selection for cross-resistance. This is an important distinction, as substitutions with effects on multiple drugs have a greater capacity to evade multiple antibiotics, and understanding how this in turn affects their relative selection pressure is crucial in determining how multidrug resistance might evolve. In future, these insights could inform therapeutic protocols that exploit collateral sensitivity or calibrate models of co-selection in response to interventions ( 22 , 67 , 68 ). Mapping MLSTs revealed substantial phenotypic differences between lineages, reflecting variation in selective pressures, genetic backgrounds, or recombination rate ( 66 , 69 – 71 ). Some MLSTs, including 320 and 156, shifted in prevalence following the introduction of pneumococcal conjugate vaccines–which coincided with increases in resistance frequencies due to serotype replacement ( 72 , 73 ). These MLSTs (along with 558 and 1451), exhibited uniformly high MICs, consistent with highly divergent PBPs. Other MLSTs displayed narrower MIC ranges, suggesting weaker selective pressure for increased MIC. Notably, MLSTs 338 and 1373 displayed extensive PBP divergence yet only intermediate penicillin MICs, likely due to the contrasting phenotypic effects of their PBP2B substitutions. Together, these findings suggest S. pneumoniae can achieve elevated MICs through multiple distinct mutational combinations, that these combinations can be specific to certain lineages, and that they are likely shaped by the extensive epistatic effects among PBPs ( 69 – 71 ). This has practical implications for clinical surveillance, as tailoring treatment to genetic backgrounds may help slow the emergence of high-level resistance ( 38 , 66 , 69 ). More broadly, this highlights the importance of genetic background in shaping evolutionary trajectories in this species, as it demonstrates the fitness and phenotypic outcomes of individual mutations are strongly context-dependent ( 41 , 43 , 58 , 66 ). In summary, AMR Cartography provides a powerful, quantitative method for analysing multivariate drug resistance. By visualising several continuous phenotypic traits simultaneously, these “maps” enable comparisons of resistance patterns among bacterial populations ( 11 , 30 , 74 ), facilitate modelling of full susceptibility profiles ( 19 ), and track phenotypic changes over evolutionary time ( 75 – 77 ). The dimensionality of these maps is shaped by intrinsic aspects of species biology ( 52 , 53 )–including molecular mechanisms, fitness dynamics, and population-level antibiotic usage. Studying these maps therefore offers insights into these factors, and provides a generalisable tool for interpreting variation and covariation among populations. The tool clarifies how resistance to individual drugs evolves into multidrug resistance, bridging the gap between single- and multidrug analyses, and helping to integrate several relevant perspectives on drug resistance (e.g. clinical, evolutionary and epidemiological). In doing so, it supports the development of informed surveillance and stewardship strategies. Because AMR Cartography relies solely on phenotypic data, it is readily adaptable to other species and drugs, provided phenotypic correlations exist. The method is therefore directly applicable to large-scale surveillance programs (e.g. NARMS, ABC, or GLASS) ( 1 , 4 , 5 ), and in future, could support comparative studies of S. pneumoniae lineages using Global Pneumococcal Sequencing (GPS) data ( 50 , 78 ). Beyond surveillance, the methods could be applied to experimental evolution studies, within-host treatment monitoring, or be used to integrate surveillance data with fitness landscapes or phylogenetics ( 79 ). More broadly, the technique could, in theory, be applied to any high-dimensional system–including phage-host coevolution, cancer-drug resistance, or virulence traits ( 80 – 82 ). Ultimately, AMR Cartography is both a valuable clinical surveillance tool, and a framework for understanding the evolutionary dynamics which govern multivariate drug resistance. Contributions Ideation - A.J.B., Conceptualisation - A.J.B., O.R. & L.A.W., Methodology - A.J.B., O.R. & L.A.W., Formal Analysis - A.J.B., Visualisations - A.J.B., Investigation - A.J.B., Writing code - A.J.B., Writing Initial Draft - A.J.B., Writing, Reviewing and Editing Subsequent Drafts - A.J.B., O.R. & L.A.W., G.G.M, S. L., Supervision - O.R. & L.A.W., Funding Acquisition - O.R. & L.A.W.. All authors provided input to the manuscript and reviewed the final version. Funding Sources This work was supported by the Biotechnology and Biological Sciences Research Council (BBSRC) through a Doctoral Training Partnership [Student Number 2113638]. Link to GitHub Repository All code and data required to reproduce this paper is available at: https://github.com/AndrewBalmer/AMR-cartography . Conflicts of interest The authors have declared that no competing interests exist. Declaration of generative AI and AI-assisted technologies in the writing process The authors declare that they used ChatGPT 4.0 to improve readability, grammar and syntax during writing of the manuscript, but did not use this tool for content generation. The authors reviewed and edited subsequent drafts and take full responsibility for its final contents. Acknowledgements We would like to thank both the CDC and ABC programme for making the data used in this manuscript publicly available. We would also like to acknowledge several current and former members of the Centre for Pathogen Evolution (Department of Zoology, University of Cambridge) for comments and feedback during development of the project. In particular, this includes David Pattinson for discussion of his PhD thesis–which inspired many of the genotype-phenotype methods used here, as well as Sam Wilks, Sarah Jones, Leah Katzelnick, Daniel Fabian, Longzhu Shen, and Derek Smith. We would also like to thank Yuan Li, Benjamin Metcalf, Lesley McGee of the CDC, the members of Stephen Bentley’s group (Wellcome Trust Sanger Institute) and Henrik Salje’s Group (Department of Genetics, University of Cambridge) for useful comments and feedback during the project. Lastly, we would like to thank Nicole Wheeler and Julian Parkhill, for examining the PhD thesis on which this work was based. Footnotes ↵ + Joint Senior Author References 1. ↵ Global antimicrobial resistance and use surveillance system (GLASS) report : 2022 [Internet]. [cited 2024 June 22]. Available from: https://www.who.int/publications/i/item/9789240062702 2. Murray CJ , Ikuta KS , Sharara F , Swetschinski L , Robles Aguilar G , Gray A , et al. Global burden of bacterial antimicrobial resistance in 2019: a systematic analysis . The Lancet . 2022 Feb 12; 399 ( 10325 ): 629 – 55 . OpenUrl CrossRef 3. ↵ O’Neill , J. Tackling drug-resistant infections globally: final report and recommendations . Lond HM Gov Wellcome Trust . 2016 ; 4. ↵ Karp BE , Tate H , Plumblee JR , Dessai U , Whichard JM , Thacker EL , et al. National Antimicrobial Resistance Monitoring System: Two Decades of Advancing Public Health Through Integrated Surveillance of Antimicrobial Resistance . Foodborne Pathog Dis . 2017 Oct 1; 14 ( 10 ): 545 – 57 . OpenUrl CrossRef PubMed 5. ↵ Schuchat A , Hilger T , Zell E , Farley MM , Reingold A , Harrison L , et al. Active bacterial core surveillance of the emerging infections program network . Emerg Infect Dis . 2001 ; 7 ( 1 ): 92 – 9 . OpenUrl CrossRef PubMed Web of Science 6. ↵ McCormick AW , Whitney CG , Farley MM , Lynfield R , Harrison LH , Bennett NM , et al. Geographic diversity and temporal trends of antimicrobial resistance in Streptococcus pneumoniae in the United States . Nat Med . 2003 Apr; 9 ( 4 ): 424 – 30 . OpenUrl CrossRef PubMed Web of Science 7. ↵ Andrews JM . Determination of minimum inhibitory concentrations . J Antimicrob Chemother . 2001 July 1; 48 ( Suppl 1 ): 5 – 16 . OpenUrl CrossRef PubMed Web of Science 8. Johnson AP . Surveillance of antibiotic resistance . Philos Trans R Soc Lond B Biol Sci . 2015 June 5; 370 ( 1670 ): 20140080 . OpenUrl CrossRef PubMed 9. ↵ Jorgensen JH , Ferraro MJ . Antimicrobial susceptibility testing: a review of general principles and contemporary practices . Clin Infect Dis . 2009 Dec 1; 49 ( 11 ): 1749 – 55 . OpenUrl CrossRef PubMed Web of Science 10. ↵ Michael A , Kelman T , Pitesky M . Overview of Quantitative Methodologies to Understand Antimicrobial Resistance via Minimum Inhibitory Concentration . Anim Basel . 2020 Aug 12; 10 ( 8 ): 1405 . OpenUrl 11. ↵ Chang HH , Cohen T , Grad YH , Hanage WP , O’Brien TF , Lipsitch M . Origin and proliferation of multiple-drug resistance in bacterial pathogens . Microbiol Mol Biol Rev . 2015 Mar; 79 ( 1 ): 101 – 16 . OpenUrl Abstract / FREE Full Text 12. Pérez-Trallero E , García-de-la-Fuente C , García-Rey C , Baquero F , Aguilar L , Dal-Ré R , et al. Geographical and ecological analysis of resistance, coresistance, and coupled resistance to antimicrobials in respiratory pathogenic bacteria in Spain . Antimicrob Agents Chemother . 2005 May; 49 ( 5 ): 1965 – 72 . OpenUrl Abstract / FREE Full Text 13. ↵ Link-Gelles R , Thomas A , Lynfield R , Petit S , Schaffner W , Harrison L , et al. Geographic and temporal trends in antimicrobial nonsusceptibility in Streptococcus pneumoniae in the post-vaccine era in the United States . J Infect Dis . 2013 Oct 15; 208 ( 8 ): 1266 – 73 . OpenUrl CrossRef PubMed 14. ↵ Colijn C , Cohen T , Fraser C , Hanage W , Goldstein E , Givon-Lavi N , et al. What is the mechanism for persistent coexistence of drug-susceptible and drug-resistant strains of Streptococcus pneumoniae ? J R Soc Interface . 2010 June 6; 6 ( 47 ): 905 – 19 . OpenUrl 15. ↵ Knight GM , Davies NG , Colijn C , Coll F , Donker T , Gifford DR , et al. Mathematical modelling for antibiotic resistance control policy: do we know enough? BMC Infect Dis . 2019 Nov 29; 19 ( 1 ): 1011 . OpenUrl CrossRef PubMed 16. ↵ Andersson DI , Hughes D . Microbiological effects of sublethal levels of antibiotics . Nat Rev Microbiol . 2014 July; 12 ( 7 ): 465 – 78 . OpenUrl CrossRef PubMed 17. Baquero F , Negri MC , Morosini MI , Blázquez J . Antibiotic-selective environments . Clin Infect Dis Off Publ Infect Dis Soc Am . 1998 Aug; 27 ( Suppl 1 ): 5 – 11 . OpenUrl 18. ↵ Gullberg E , Cao S , Berg OG , Ilbäck C , Sandegren L , Hughes D , et al. Selection of resistant bacteria at very low antibiotic concentrations . PLoS Pathog . 2011 July; 7 ( 7 ): e1002158 . OpenUrl CrossRef PubMed 19. ↵ Ryu S , Cowling BJ , Wu P , Olesen S , Fraser C , Sun DS , et al. Case-based surveillance of antimicrobial resistance with full susceptibility profiles . JAC Antimicrob Resist . 2019 Dec; 1 ( 3 ):dlz070. 20. ↵ Nichol D , Rutter J , Bryant C , Hujer AM , Lek S , Adams MD , et al. Antibiotic collateral sensitivity is contingent on the repeatability of evolution . Nat Commun . 2019 Jan 18; 10 ( 1 ): 334 . OpenUrl CrossRef PubMed 21. ↵ Pouwels KB , Muller-Pebody B , Smieszek T , Hopkins S , Robotham JV . Selection and co-selection of antibiotic resistances among Escherichia coli by antibiotic use in primary care: An ecological analysis . PLoS One . 2019 ; 14 ( 6 ): e0218134 . OpenUrl CrossRef PubMed 22. ↵ Pouwels KB , Freeman R , Muller-Pebody B , Rooney G , Henderson KL , Robotham JV , et al. Association between use of different antibiotics and trimethoprim resistance: going beyond the obvious crude association . J Antimicrob Chemother . 2018 June 1; 73 ( 6 ): 1700 – 7 . OpenUrl CrossRef PubMed 23. ↵ Seale AC , Hutchison C , Fernandes S , Stoesser N , Kelly H , Lowe B , et al. Supporting surveillance capacity for antimicrobial resistance: Laboratory capacity strengthening for drug resistant infections in low and middle income countries . Wellcome Open Res . 2017 ; 2 : 91 . 24. Iskandar K , Molinier L , Hallit S , Sartelli M , Hardcastle TC , Haque M , et al. Surveillance of antimicrobial resistance in low- and middle-income countries: a scattered picture . Antimicrob Resist Infect Control . 2021 Mar 31; 10 ( 1 ): 63 . OpenUrl PubMed 25. ↵ Argimón S , Masim MAL , Gayeta JM , Lagrada ML , Macaranas PKV , Cohen V , et al. Integrating whole-genome sequencing within the National Antimicrobial Resistance Surveillance Program in the Philippines . Nat Commun . 2020 June 1; 11 ( 1 ): 2719 . OpenUrl CrossRef PubMed 26. ↵ Lees JA , Mai TT , Galardini M , Wheeler NE , Horsfield ST , Parkhill J , et al. Improved Prediction of Bacterial Genotype-Phenotype Associations Using Interpretable Pangenome-Spanning Regressions . mBio . 2020 July 7; 11 ( 4 ): e01344 – 20 . OpenUrl CrossRef PubMed 27. ↵ Dhand A , Sakoulas G . Reduced vancomycin susceptibility among clinical Staphylococcus aureus isolates (‘the MIC Creep’): implications for therapy . F1000 Med Rep . 2012 Feb 1; 4 : 4 . OpenUrl 28. ↵ CRyPTIC Consortium . Quantitative measurement of antibiotic resistance in Mycobacterium tuberculosis reveals genetic determinants of resistance and susceptibility in a target gene approach . Nat Commun . 2024 Jan 12; 15 ( 1 ): 488 . OpenUrl CrossRef PubMed 29. ↵ Fonville JM , Wilks SH , James SL , Fox A , Ventresca M , Aban M , et al. Antibody landscapes after influenza virus infection or vaccination . Science . 2014 Nov 21; 346 ( 6212 ): 996 – 1000 . OpenUrl Abstract / FREE Full Text 30. ↵ Smith DJ , Lapedes AS , de Jong JC , Bestebroer TM , Rimmelzwaan GF , Osterhaus ADME , et al. Mapping the antigenic and genetic evolution of influenza virus . Science . 2004 July 16; 305 ( 5682 ): 371 – 6 . OpenUrl Abstract / FREE Full Text 31. ↵ Wilks SH , Mühlemann B , Shen X , Türeli S , LeGresley EB , Netzl A , et al. Mapping SARS-CoV-2 antigenic relationships and serological responses . Science . 2023 Oct 6; 382 ( 6666 ): eadj0070 . OpenUrl CrossRef PubMed 32. ↵ Katzelnick LC , Fonville JM , Gromowski GD , Arriaga JB , Green A , James SL , et al. Dengue viruses cluster antigenically but not as discrete serotypes . Science . 2015 Sept 18; 349 ( 6254 ): 1338 – 43 . OpenUrl Abstract / FREE Full Text 33. ↵ Dewé TCM , D’Aeth JC , Croucher NJ . Genomic epidemiology of penicillin-non-susceptible Streptococcus pneumoniae . Microb Genomics . 2019 Oct; 5 ( 10 ): e000305 . OpenUrl 34. Croucher NJ , Hanage WP , Harris SR , McGee L , van der Linden M , de Lencastre H , et al. Variable recombination dynamics during the emergence, transmission and ‘disarming’ of a multidrug-resistant pneumococcal clone . BMC Biol . 2014 June 23; 12 ( 1 ): 49 . OpenUrl CrossRef PubMed 35. ↵ Croucher NJ , Chewapreecha C , Hanage WP , Harris SR , McGee L , van der Linden M , et al. Evidence for soft selective sweeps in the evolution of pneumococcal multidrug resistance and vaccine escape . Genome Biol Evol . 2014 June 10; 6 ( 7 ): 1589 – 602 . OpenUrl CrossRef PubMed 36. ↵ Hakenbeck R , Brückner R , Denapaite D , Maurer P . Molecular mechanisms of β-lactam resistance in Streptococcus pneumoniae . Future Microbiol . 2012 Mar; 7 ( 3 ): 395 – 410 . OpenUrl CrossRef PubMed Web of Science 37. ↵ Li Y , Metcalf BJ , Chochua S , Li Z , Gertz RE , Walker H , et al. Penicillin-Binding Protein Transpeptidase Signatures for Tracking and Predicting β-Lactam Resistance Levels in Streptococcus pneumoniae . mBio . 2016 June 14; 7 ( 3 ) : doi: 10.1128/mbio.00756-16 . OpenUrl CrossRef 38. ↵ Chewapreecha C , Marttinen P , Croucher NJ , Salter SJ , Harris SR , Mather AE , et al. Comprehensive Identification of Single Nucleotide Polymorphisms Associated with Beta-lactam Resistance within Pneumococcal Mosaic Genes . PLOS Genet . 2014 Aug 7; 10 ( 8 ): e1004547 . OpenUrl CrossRef PubMed 39. Hakenbeck R . Discovery of β-lactam-resistant variants in diverse pneumococcal populations . Genome Med . 2014 Sept 25; 6 ( 9 ): 72 . OpenUrl CrossRef PubMed 40. ↵ Dowson CG , Hutchison A , Brannigan JA , George RC , Hansman D , Liñares J , et al. Horizontal transfer of penicillin-binding protein genes in penicillin-resistant clinical isolates of Streptococcus pneumoniae . Proc Natl Acad Sci U S A . 1989 Nov; 86 ( 22 ): 8842 – 6 . OpenUrl Abstract / FREE Full Text 41. ↵ Skwark MJ , Croucher NJ , Puranen S , Chewapreecha C , Pesonen M , Xu YY , et al. Interacting networks of resistance, virulence and core machinery genes identified by genome-wide epistasis analysis . PLOS Genet . 2017 Feb 16; 13 ( 2 ): e1006508 . OpenUrl CrossRef PubMed 42. Pensar J , Puranen S , Arnold B , MacAlasdair N , Kuronen J , Tonkin-Hill G , et al. Genome-wide epistasis and co-selection study using mutual information . Nucleic Acids Res . 2019 Oct 10; 47 ( 18 ): e112 . OpenUrl PubMed 43. ↵ Orio AGA , Piñas GE , Cortes PR , Cian MB , Echenique J . Compensatory Evolution of pbp Mutations Restores the Fitness Cost Imposed by β-Lactam Resistance in Streptococcus pneumoniae . PLOS Pathog . 2011 Feb 17; 7 ( 2 ): e1002000 . OpenUrl CrossRef PubMed 44. ↵ Mallawaarachchi S , Tonkin-Hill G , Croucher NJ , Turner P , Speed D , Corander J , et al. Genome-wide association, prediction and heritability in bacteria with application to Streptococcus pneumoniae . NAR Genomics Bioinforma . 2022 Mar; 4 ( 1 ): lqac011 . OpenUrl 45. ↵ Mouz N , Di Guilmi AM , Gordon E , Hakenbeck R , Dideberg O , Vernet T . Mutations in the active site of penicillin-binding protein PBP2x from Streptococcus pneumoniae . Role in the specificity for beta-lactam antibiotics . J Biol Chem . 1999 July 2; 274 ( 27 ): 19175 – 80 . OpenUrl Abstract / FREE Full Text 46. Zerfaß I , Hakenbeck R , Denapaite D . An Important Site in PBP2x of Penicillin-Resistant Clinical Isolates of Streptococcus pneumoniae : Mutational Analysis of Thr338 . Antimicrob Agents Chemother . 2009 Mar; 53 ( 3 ): 1107 – 15 . OpenUrl Abstract / FREE Full Text 47. ↵ Kocaoglu O , Tsui HCT , Winkler ME , Carlson EE . Profiling of β-lactam selectivity for penicillin-binding proteins in Streptococcus pneumoniae D39 . Antimicrob Agents Chemother . 2015 ; 59 ( 6 ): 3548 – 55 . OpenUrl Abstract / FREE Full Text 48. ↵ Coffey TJ , Daniels M , McDougal LK , Dowson CG , Tenover FC , Spratt BG . Genetic analysis of clinical isolates of Streptococcus pneumoniae with high-level resistance to expanded-spectrum cephalosporins . Antimicrob Agents Chemother . 1995 June; 39 ( 6 ): 1306 – 13 . OpenUrl Abstract / FREE Full Text 49. ↵ Enright MC , Spratt BG . A multilocus sequence typing scheme for Streptococcus pneumoniae : identification of clones associated with serious invasive disease . Microbiology . 1998 Nov; 144 (Pt 11 ): 3049 – 60 . OpenUrl CrossRef PubMed Web of Science 50. ↵ Gladstone RA , Lo SW , Goater R , Yeats C , Taylor B , Hadfield J , et al. Visualizing variation within global pneumococcal sequence clusters (GPSCS) and country population snapshots to contextualize pneumococcal isolates . Microb Genomics . 2020 Apr 30; 6 ( 5 ): 1 – 13 . OpenUrl 51. ↵ Li Y , Metcalf BJ , Chochua S , Li Z , Gertz RE , Walker H , et al. Validation of β-lactam minimum inhibitory concentration predictions for pneumococcal isolates with newly encountered penicillin binding protein (PBP) sequences . BMC Genomics . 2017 Aug 15; 18 ( 1 ): 621 . OpenUrl CrossRef PubMed 52. ↵ Mair P , Groenen PJF , De Leeuw J . More on Multidimensional Scaling and Unfolding in R: smacof Version 2 . J Stat Softw . 2022 May 13; 102 : 1 – 47 . OpenUrl 53. ↵ De Leeuw J , Mair P . Multidimensional Scaling Using Majorization: SMACOF in R . J Stat Softw . 2009 Aug 4; 31 : 1 – 30 . OpenUrl CrossRef PubMed 54. ↵ Pattinson DJ . Predicting the Antigenic Evolution of Influenza Viruses with Application to Vaccination Strategy [Internet] . [ Cambridge, UK ]: University of Cambridge ; 2020 [cited 2025 Aug 15]. Available from: https://www.repository.cam.ac.uk/items/0fd9bca6-c64b-47cd-92b4-34302f5e6bd6 55. ↵ Lippert C , Listgarten J , Liu Y , Kadie CM , Davidson RI , Heckerman D . FaST linear mixed models for genome-wide association studies . Nat Methods . 2011 Oct; 8 ( 10 ): 833 – 5 . OpenUrl CrossRef PubMed Web of Science 56. ↵ Lippert C , Casale FP , Rakitsch B , Stegle O. LIMIX: genetic analysis of multiple traits [Internet] . bioRxiv ; 2014 [cited 2025 Aug 15]. p. 003905 . Available from: https://www.biorxiv.org/content/10.1101/003905v1 57. ↵ Bedford T , Rambaut A , Pascual M . Canalization of the evolutionary trajectory of the human influenza virus . BMC Biol . 2012 Apr 30; 10 ( 1 ): 38 . OpenUrl CrossRef PubMed 58. ↵ Gibson PS , Bexkens E , Zuber S , Cowley LA , Veening JW . The acquisition of clinically relevant amoxicillin resistance in Streptococcus pneumoniae requires ordered horizontal gene transfer of four loci . PLOS Pathog . 2022 July 25; 18 ( 7 ): e1010727 . OpenUrl CrossRef PubMed 59. ↵ Batisti Biffignandi G , Chindelevitch L , Corbella M , Feil EJ , Sassera D , Lees JA . Optimising machine learning prediction of minimum inhibitory concentrations in Klebsiella pneumoniae . Microb Genomics . 2024 Mar; 10 ( 3 ): 001222 . OpenUrl 60. ↵ Schaid DJ , Chen W , Larson NB . From genome-wide associations to candidate causal variants by statistical fine-mapping . Nat Rev Genet . 2018 Aug; 19 ( 8 ): 491 – 504 . OpenUrl CrossRef PubMed 61. ↵ D’Aeth JC , van der Linden MP , McGee L , de Lencastre H , Turner P , Song JH , et al. The role of interspecies recombination in the evolution of antibiotic-resistant pneumococci . eLife . 2021 July 14; 10 : e67113 . OpenUrl CrossRef PubMed 62. ↵ Lekhuleni C , Ndlangisa K , Gladstone RA , Chochua S , Metcalf BJ , Li Y , et al. Impact of pneumococcal conjugate vaccines on invasive pneumococcal disease-causing lineages among South African children . Nat Commun . 2024 Sept 27; 15 ( 1 ): 8401 . OpenUrl PubMed 63. ↵ Colangeli R , Jedrey H , Kim S , Connell R , Ma S , Chippada Venkata UD , et al. Bacterial Factors That Predict Relapse after Tuberculosis Therapy . N Engl J Med . 2018 Aug 30; 379 ( 9 ): 823 – 33 . OpenUrl CrossRef PubMed 64. ↵ Negri MC , Lipsitch M , Blázquez J , Levin BR , Baquero F . Concentration-dependent selection of small phenotypic differences in TEM beta-lactamase-mediated antibiotic resistance . Antimicrob Agents Chemother . 2000 Sept; 44 ( 9 ): 2485 – 91 . OpenUrl Abstract / FREE Full Text 65. ↵ Opatowski L , Mandel J , Varon E , Boëlle PY , Temime L , Guillemot D . Antibiotic dose impact on resistance selection in the community: a mathematical model of beta-lactams and Streptococcus pneumoniae dynamics . Antimicrob Agents Chemother . 2010 June; 54 ( 6 ): 2330 – 7 . OpenUrl Abstract / FREE Full Text 66. ↵ Lehtinen S , Chewapreecha C , Lees J , Hanage WP , Lipsitch M , Croucher NJ , et al. Horizontal gene transfer rate is not the primary determinant of observed antibiotic resistance frequencies in Streptococcus pneumoniae . Sci Adv . 2020 May 20; 6 ( 21 ): eaaz6137 . OpenUrl FREE Full Text 67. ↵ Liakopoulos A , Aulin LBS , Buffoni M , Fragkiskou E , van Hasselt JGC , Rozen DE . Allele-specific collateral and fitness effects determine the dynamics of fluoroquinolone resistance evolution . Proc Natl Acad Sci U S A . 2022 May 3; 119 ( 18 ): e2121768119 . OpenUrl CrossRef PubMed 68. ↵ Hernando-Amado S , Laborda P , Martínez JL . Tackling antibiotic resistance by inducing transient and robust collateral sensitivity . Nat Commun . 2023 Mar 30; 14 ( 1 ): 1723 . OpenUrl CrossRef PubMed 69. ↵ Lehtinen S , Blanquart F , Croucher NJ , Turner P , Lipsitch M , Fraser C . Evolution of antibiotic resistance is linked to any genetic mechanism affecting bacterial duration of carriage . Proc Natl Acad Sci U S A . 2017 Jan 31; 114 ( 5 ): 1075 – 80 . OpenUrl Abstract / FREE Full Text 70. Croucher NJ , Harris SR , Fraser C , Quail MA , Burton J , van der Linden M , et al. Rapid pneumococcal evolution in response to clinical interventions . Science . 2011 Jan 28; 331 ( 6016 ): 430 – 4 . OpenUrl Abstract / FREE Full Text 71. ↵ Hanage WP , Fraser C , Tang J , Connor TR , Corander J . Hyper-recombination, diversity, and antibiotic resistance in pneumococcus . Science . 2009 June 12; 324 ( 5933 ): 1454 – 7 . OpenUrl Abstract / FREE Full Text 72. ↵ Dagan R , Klugman KP . Impact of conjugate pneumococcal vaccines on antibiotic resistance . Lancet Infect Dis . 2008 Dec; 8 ( 12 ): 785 – 95 . OpenUrl CrossRef PubMed Web of Science 73. ↵ Kyaw MH , Lynfield R , Schaffner W , Craig AS , Hadler J , Reingold A , et al. Effect of introduction of the pneumococcal conjugate vaccine on drug-resistant Streptococcus pneumoniae . N Engl J Med . 2006 Apr 6; 354 ( 14 ): 1455 – 63 . OpenUrl CrossRef PubMed Web of Science 74. ↵ Wildfire J , Waterlow NR , Clements A , Fuller NM , Knight GM . MIC distribution analysis identifies differences in AMR between population sub-groups . Wellcome Open Res . 2024 ; 9 : 244 . 75. ↵ Kenyon C , Laumen J , Van Den Bossche D , Van Dijck C . Where have all the susceptible gonococci gone? A historical review of changes in MIC distribution over the past 75 years . BMC Infect Dis . 2019 Dec 27; 19 ( 1 ): 1085 . OpenUrl PubMed 76. Katzelnick LC , Coello Escoto A , Huang AT , Garcia-Carreras B , Chowdhury N , Maljkovic Berry I , et al. Antigenic evolution of dengue viruses over 20 years . Science . 2021 Nov 19; 374 ( 6570 ): 999 – 1004 . OpenUrl CrossRef PubMed 77. ↵ Bedford T , Suchard MA , Lemey P , Dudas G , Gregory V , Hay AJ , et al. Integrating influenza antigenic dynamics with molecular evolution . eLife . 2014 Feb 4; 3 : e01914 . OpenUrl CrossRef PubMed 78. ↵ Gladstone RA , Lo SW , Lees JA , Croucher NJ , van Tonder AJ , Corander J , et al. International genomic definition of pneumococcal lineages, to contextualise disease, antibiotic resistance and vaccine impact . EBioMedicine . 2019 Apr 16; 43 : 338 – 46 . OpenUrl PubMed 79. ↵ Balmer A. Multivariate Methods for the Study of Beta-lactam Resistance in Streptococci [Internet] [PhD Thesis] . [ Cambridge, UK ]: University of Cambridge ; 2025 [cited 2025 Aug 15]. Available from: https://www.repository.cam.ac.uk/items/a20b3ee7-5ae5-417f-a95f-15299a52caff 80. ↵ Shaer Tamar E , Kishony R . Multistep diversification in spatiotemporal bacterial-phage coevolution . Nat Commun . 2022 Dec 28; 13 ( 1 ): 7971 . OpenUrl CrossRef PubMed 81. Emran TB , Shahriar A , Mahmud AR , Rahman T , Abir MH , Siddiquee MohdFR , et al. Multidrug Resistance in Cancer: Understanding Molecular Mechanisms, Immunoprevention and Therapeutic Approaches . Front Oncol . 2022 June 23; 12 : 891652 . OpenUrl CrossRef PubMed 82. ↵ Stromberg ZR , Phillips SMB , Omberg KM , Hess BM . High-throughput functional trait testing for bacterial pathogens . mSphere . 2023 Sept 13; 8 ( 5 ): e00315 – 23 . OpenUrl View the discussion thread. Back to top Previous Next Posted September 16, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. 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 Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv 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 Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance Andrew J. Balmer , Gemma G. R. Murray , Stephanie Lo , Olivier Restif , Lucy A. Weinert bioRxiv 2025.09.12.675231; doi: https://doi.org/10.1101/2025.09.12.675231 Share This Article: Copy Citation Tools Antimicrobial Resistance Cartography: A Generalisable Framework for Studying Multivariate Drug Resistance Andrew J. Balmer , Gemma G. R. Murray , Stephanie Lo , Olivier Restif , Lucy A. Weinert bioRxiv 2025.09.12.675231; doi: https://doi.org/10.1101/2025.09.12.675231 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 Genomics Subject Areas All Articles Animal Behavior and Cognition (7622) Biochemistry (17648) Bioengineering (13871) Bioinformatics (41880) Biophysics (21423) Cancer Biology (18561) Cell Biology (25461) Clinical Trials (138) Developmental Biology (13364) Ecology (19866) Epidemiology (2067) Evolutionary Biology (24290) Genetics (15590) Genomics (22475) Immunology (17713) Microbiology (40328) Molecular Biology (17148) Neuroscience (88473) Paleontology (666) Pathology (2827) Pharmacology and Toxicology (4816) Physiology (7635) Plant Biology (15114) Scientific Communication and Education (2044) Synthetic Biology (4286) Systems Biology (9815) Zoology (2268)