ColTapp, an automated image analysis application for efficient microbial colony growth dynamics quantification

Phenotypic heterogeneity occurs in a population of genetically identical bacteria due to stochastic molecular fluctuations and environmental variations. In extreme cases of phenotypic heterogeneity, a fraction of the bacterial population enters dormancy, and these metabolically inactive or non-dividing bacteria persist through most antibiotic challenges. These subpopulations of persister cells are difficult to study in patient samples. However, the proportion of persisters in a sample can be accessed by physically separating bacteria on a plate measuring the time until colonies become visible as dormant bacteria resume growth later than their active counterparts and form smaller colonies. Here, we present ColTapp (Colony Time-lapse app), an application dedicated to bacterial colony growth quantification, freely available for download together with its MATLAB source code or as a MacOS/Windows executable. ColTapp’s intuitive graphical user interface allows users without prior coding knowledge to analyze endpoint or time-lapse images of colonies on agar plates. Colonies are detected automatically, and their radius can be tracked over time. Downstream analyses to derive colony lag time and growth rate are implemented. We demonstrate here the applicability of ColTapp on a dataset of Staphyloccocus aureus colony time-lapse images. Colonies on dense plates reached saturation early, biasing lag time estimation from endpoint images. This bias can be reduced by considering the area available to each colony on a plate. By facilitating the analysis of colony growth dynamics in clinical settings, this application will enable a new type of diagnostics, oriented towards personalized antibiotic therapies.


33
Delayed and insufficient clearance of bacterial infections leading to treatment failure is associated 34 with antibiotic resistance as well as antibiotic persistence. In antibiotic persistence, a subpopulation of 35 bacteria, termed persister cells, can survive antibiotic challenges due to their phenotypic state of 36 metabolic inactivity and then reconstitute the population by resuming growth when the stress is 37 relieved (Balaban et al., 2004). This manifestation of phenotypic heterogeneity exists even in 38 homogeneous liquid cultures, where each bacterial cell experiences the same local conditions 39 (Ackermann, 2015). On top of this, bacteria often aggregate in dense communities such as biofilms, 40 where environmental conditions are highly heterogeneous and further promote wide phenotypic 41 distributions (Stewart & Franklin, 2008). 42 The clinical relevance of antibiotic persistence remains poorly understood. The main reason for this 43 is that the existence and extent of persistence is difficult to assess because it is a transient phenotype 44 and only concerns a small fraction of the bacterial population. However, these inactive bacteria are 45 lagging longer than already actively dividing bacteria when plated on nutrient-rich agar medium, and 46 the bacterial colonies' macroscopic appearance time is a good proxy of single-cell lag time (Guillier et  In clinical settings, the direct observation of colony growth dynamics from bacteria recovered from 51 infection sites is rare (Barr et al., 2016, Vulin et al., 2018. Typically, in clinical microbiology laboratories 52 the patient's samples are plated and observed at a single timepoint, e.g. after 18 hours of incubation. 53 Colonies which are smaller than the bulk at that timepoint have been associated with antibiotic 54 tolerance, persister cells and relapsing infections (Proctor et al., 2006). Small colonies can result either 55 from mutations that affect growth rate (small colony variants (Proctor et al., 1995, Kahl et al., 2016) 56 or from heterogeneous growth resumption (Jõers et al., 2010, Levin-Reisman et al., 2010 access to colony growth dynamics quantification and enable diagnostic microbiology laboratories to 68 further improve their routines. ColTapp is an image analysis pipeline embedded in an intuitive 69 graphical user interface, which allows any user to derive colony sizes, growth rates and appearance 70 times from time-lapse images of bacterial colonies. It also includes the possibility to estimate colony 71 growth parameters from endpoint images. Additionally, it can report metrics characterizing color, 72 shape and proximity of neighboring colonies. We speculate that ColTapp will prove useful in a broad 73 spectrum of microbiology applications, such as species identification in environmental samples as 74 shown by Ernebjerg & Kishony (2012). 75 A technical problem common to both clinical and environmental samples is the difficulty to assess 76 their initial bacterial density: samples can be accidentally plated at high or uneven densities. Yet, 77 colonies compete for nutrients on the agar plates and thus the total number of colonies and their 78 spatial distribution impacts their size at an endpoint (Chacón et al., 2018). This results in difficult 79 appearance time estimation. To address this problem, as well as demonstrate the applicability of 80 ColTapp, we present a dataset of Staphylococcus aureus bacterial colony time-lapse images and 81 explore their response to density. The colonies' Voronoi cell areas, one of the spatial metrics computed 82 by ColTapp and previously explored in the context of colony growth (Chacón et al., 2018), can be taken 83 into account to minimize the bias in appearance time estimation due to density. 84

85
A user-friendly application implementing image-and downstream analysis 86 ColTapp takes images as input and implements image analysis functions to detect microbial 87 colonies and track colony radius over time when images are part of a time-lapse sequence. Common 88 formats (png, tiff and jpeg) and either color (Red Green Blue) or grayscale images are supported. 89 ColTapp also includes downstream analysis steps, to extract biologically relevant data such as colony 90 growth rate and appearance time. A graphical user interface (GUI) provides access to all functionalities 91 and displays the images, with the possibility to visualize certain data, such as the circles around the 92 detected colonies (Fig. 1A, S1 and S2). The interface enables early visual evaluation of the results by 93 including simple data visualization options (SI text 9). The generated data can be exported in a standard 94 csv file (Fig. S3). 95 The application's workflow from raw images to exported data is illustrated in Fig. 1B. The program 96 operates in two different modes, depending on the user input data: endpoint images or time-lapse 97 images. In the first mode, Time-Lapse (TL), images in a folder are considered as an ordered time-lapse 98 image series of a single plate. In the second mode, Endpoint (EP), all images in a folder are considered 99 independent from each other. Typically, they are an aggregate of images from several plated samples 100 or from replicates of the same sample, captured at a single timepoint. 101

102
In the following sections, implementation and performance of the two main image analysis 103 algorithms (colony detection and radii tracking over time) are described, referring to the 104 supplementary information for more detailed directives to the user concerning interaction with the 105 interface, quality control functionalities and parameters optimization. 106 Then downstream analyses to derive growth parameters from the resulting data are demonstrated 107 on a dataset of 22 time-lapse movies of S. aureus colonies plates, designed to include both 108 homogeneous and heterogenous colony lag time and various colony densities. In addition to growth 109 parameters, a palette of colony characteristics may be exported by the user for further analysis, 110 including shape and color metrics and spatial metrics (describing the colony neighbors' proximity, 111 SI text 3). 112 In conclusion, an example is given to show how the ColTapp-derived measurements of radius, 113 appearance time and spatial metrics can be used to assess the influence of colony density on 114 appearance time estimation from endpoint images and how spatial metrics can be used to correct for 115 density effects. 116

124
(EP) mode depending on input data, illustrated by the two different folders (turquoise and yellow respectively). The turquoise 125 highlighted functionalities are specific to the Time-lapse mode, while the yellow highlighted ones are specific to Endpoint 126 mode. In the middle, the green highlighted functionalities are common to both modes.

Algorithm implementation 129
Colony detection uses a series of image analysis operations starting from a grayscale image (Fig.  130 2A) and depends on user-specified minimal and maximal expected radius of colonies (see SI text for 131 additional pre-processing possibilities). ColTapp uses top-hat filtering with a disk-shaped structuring 132 element based on the minimal expected radius to reduce lighting gradients and other inhomogeneities 133 of the agar background (Fig. 2B). Local adaptive thresholding is then used to create a binary image 134 (Fig. 2C). This image is cleaned from artifacts and unwanted objects by discarding small and big isolated 135 objects (threshold defined by minimal and maximal expected radius, respectively). Objects with a low 136 extent (foreground/background pixel ratio), are discarded as well (Fig. 2D). Next, distance 137 transformation is used to derive local minima, which are subsequently filtered by minima imposition. 138 Overlapping colonies are separated with watershed segmentation (Fig. 2E). Sequentially, ColTapp 139 extracts images containing an isolated object (imgcrop) from the top-hat filtered grayscale image and 140 further performs circular filtering and contrast enhancement ( can be tuned to potentially improve performance (SI text 3.2). In addition, the user may intervene 151 directly to correct the results (SI text 3.3). 152 In Endpoint mode, colony detection is done independently on each frame. In Time-lapse mode, it is 153 done on a single frame, as the colony radius can subsequently be tracked over time (see below). 154   Table S1. 176 Generally, ColTapp accurately detected colonies when the area of interest, expected radius range 177 and the method for conversion of a color image into a grayscale image were set for each image 178 separately (SI text 2.1). Note, that although colonies are usually lighter than the background, phage 179 plaques are darker than the bacterial lawn: an option to find darker-than-background circles is 180 available. 181 Overall, false positive rate varied substantially, ranging from 0% to 55.6% (median = 2.8%) (Fig. 3A). 182 Almost all images with false positive rates exceeding 5% had most of their wrongly detected circles in 183 clusters (marked with blue symbols in Fig. 3A), usually in areas with lighting artifacts. These clusters 184 can be cleared with a few clicks and do not pose a problem in our opinion. The one other case with a 185 high false positive rate was an image displaying colonies of the marine bacterial species Alteromonas 186 macleodii and a lot of chitinous debris, forming particles which were only marginally different in size 187 as compared to the bacterial colonies. Therefore, false positive rate should not exceed 5% if ideal 188 imaging conditions are maintained. Additionally, we did not observe any correlation between false 189 positive rate and total number of colonies on a plate (Fig. 3A). 190 The false negative rate varied between 0% and 17.8% (median = 3.7%) and slightly increased with 191 total number of colonies on the plate (

209
Note that computational speed is inversely proportional with image size, average colony size and 210 number of colonies. A high proportion of plate area covered with colonies with high amount of overlap 211 will generally yield poorer detection results and require manual correction (Table S1). In conclusion 212 ColTapp successfully detects colonies from various bacterial species on distinctly colored agar plates. 213 Adapting imaging conditions towards homogeneous light and reduced reflections can reduce the 214 amount of manual correction required. 215 216 Tracking colony radius over time 217

Algorithm implementation 218
In order to track colony radii over time, applying the previously described colony detection method 219 on each frame would be slow and would require extensive manual correction since it is prone to 220 misclassifications, especially at early timepoints when colonies are not yet visible. Therefore, ColTapp 221 proposes to initially detect all colonies on a frame corresponding to a late timepoint, which is set as 222 the reference frame. Once colonies have been detected on a late frame of the time-lapse series, their 223 radius is tracked on the previous frames. angle ranges corresponding to adjacent colonies for kymograph creation (SI text 6.1). 237 In order to detect the kymograph edge to derive colony radius over time, two methods are 238 implemented: Global thresholding and Edge detection. Both methods first apply an automatic contrast 239 enhancement function, then a pixel-wise adaptive low-pass Wiener filter (Lim, 1990) followed by a 240 circular averaging filter (pillbox) to the kymograph to derive a smoother image. Additionally, a low-241 pass threshold for maximal intensity may be applied. 242  (Otsu, 1979) to derive a global threshold from which a user-defined constant is subtracted to 254 prevent biasing towards background (Fig. 4C).The Edge detection method uses a series of edge 255 detection and morphological operations (Fig. 4D). In brief, the method by Canny (1986) is used for 256 initial edge detection. Subsequently, small isolated foreground pixels are removed with area opening 257 operations. Straight vertical and horizontal lines, typically originating from light artifacts and particles 258 on the agar respectively, are removed. Two iterative morphological closing operations are used with 259 line-and disk-shaped kernels, respectively. Additionally, active contouring using the Chan-Vese method (Chan & Vese, 2001) is applied to the pre-processed kymograph to derive a second binary 261 image with a filled region of probable colony area. The two binary images are merged, and 262 morphological closing is performed to close potential gaps to successfully fill connected regions 263 afterwards. A final morphological opening is applied to remove jagged edges. The fully connected 264 object located in the lower right side of the kymograph (corresponds to end of time-lapse and close to 265 colony center) is kept as the only foreground object in the binary image. 266 ColTapp includes automatic error detection and manual correction tools for the kymograph derived 267 radius growth curves (SI text 6.2, Fig. S5). frames and colonies to process, because each frame and colonies are processed sequentially 273 (Table S2). Therefore, computational time per colony and per frame is most representative of the 274 computational efficiency of the colony radius tracking algorithm (average = 0.079 s, SD = 0.047 s). 275 Additionally, the computational time was observed to increase with the size of the colonies on the 276 reference frame, as the sub-images used for radius tracking are bigger. 277 We assessed the accuracy of the algorithm by manually evaluating the quality of the 278 1411 kymograph derived radial growth curves. The Global thresholding method usually yields correct 279 binary images except for complex kymographs resulting from high amount of colony overlap and/or 280 lighting artifacts. When using the default Global thresholding method to derive radial growth curves, 281 some curves of our dataset were classified as incorrect (mean = 21%, SD = 15%) (Table S2). ColTapp 282 inbuilt automatic quality assessment function detected on average 83% (SD = 9%) of these incorrect 283 growth curves (Table S2). The Edge detection method is more complex and is suggested when the 284 default Global thresholding method fails. Switching to the Edge detection method for the incorrect 285 growth curves resulted in a reduction of the number of growth curves requiring further manual 286 correction to 3.1% (SD = 2.7%) (Table S2). 287 288 Appearance time and growth rate determination from time-lapse images 289 Tracking the colony radius over time using time-lapse imaging makes it possible to directly 290 determine colony appearance time and growth rate. Colonies formed by non-swimming bacteria pass 291 through three growth phases (Fig. 5A). Initially, when all bacteria forming the colony can access 292 nutrients and thus contribute to its expansion, the colony radius increases exponentially while the 293 colony also increases in height. Eventually, bacteria at the center of the colony do not have access to 294 nutrients diffusing from the edge of the colony. The zone of growth at the edge of the colony becomes 295 constant and the radial growth rate becomes linear (at this point radius = Rlin, see Fig. 5B) (Pirt, 1967). 296 Finally, when the nutrients become scarce, or when growth byproducts reach an inhibitory 297 concentration, the colony enters the saturation phase, during which the radial growth rate starts 298 decreasing until plateauing (Be'er et al., 2009). Fig. S6 shows the exponential phase of S. aureus 299 colonies grown on sheep blood agar captured by time-lapse imaging with a macro lense. However, in 300 typical settings where the entire plate is captured, the resolution is not high enough to observe the 301 exponential phase. We define a detectable size threshold (Rthresh), and the time at which a colony 302 reaches this threshold as the appearance time (tapp). The minimal possible Rthresh depends on the image 303 quality and needs to be set at the same value for comparisons of tapp in different experiments. In our 304 analysis setting, we assume that colonies reaching this size are already in the linear growth phase 305 (Rlin < Rthresh, Fig. 5B, C). ColTapp determines appearance time and growth rate by detecting the first of 6 consecutive frames 319 (to avoid noise) for which a colony radius is bigger than Rthresh (default: 200 µm). From the detected 320 frame, the radius values measured on the following frames, corresponding to a user-specified 321 timespan in hours (Frlin, default: 10 h), are used for a linear regression to determine linear radial growth 322 rate and the exact time a colony reached Rthresh (Fig 5C). with Rthresh defining the radius (R) at time ti = tapp and a reference growth rate (GRmax). 351 Note that using a linear regression with a reference growth rate to estimate the appearance time 352 has two strong assumptions: first, that all colonies are still in the linear phase of their growth and 353 second, that they all have the same linear radial growth rate. 354 To allow users to evaluate the validity of these two assumptions, ColTapp proposes a functionality 355 to analyze a sequence of endpoint images (SI text 8.2). By taking endpoint images at multiple 356 timepoints and performing colony detection at each timepoint one can create a timeseries of colony 357 radius. We exemplify this functionality by analyzing two series of 4 frames extracted from time-lapse 358 image sequences of the demonstration dataset (Fig. 6A). ColTapp computes the slope between 359 colonies' radii at different timepoints (Eq. 1), which results in an individual colony radial growth rate 360 for each time interval (Fig. 6B). 361 This enables assessment of the entry time to saturation phase, identified by a decrease in growth 362 rate. Moreover, should colonies have different growth rates, the individual colony radial growth rate 363 can be used to estimate the appearance time, rather than a fixed reference growth rate for all colonies 364  between appearance times determined from time-lapse images and appearance times estimated from 378 endpoint images (SI Fig. 7), it could not accurately predict appearance time for colonies past the linear 379 phase (i.e. colonies on plate 18 at 24h in Fig. 6). 380 Note that using individual colony growth rates for appearance time estimation is appropriate to 381 correct for differences in size independent from saturation, for example when a subpopulation on the 382 plate has mutations affecting growth rate or mixed-species plates. Should colonies be past the linear 383 growth phase, one needs to take saturation into account. We introduce an approach to density-based 384 corrections of saturation in the next section. with neighboring colonies become strong enough, a colony enters the saturation phase of its growth, 393 meaning that its radial growth rate decreases over time until plateauing (Fig. 5). Therefore, colonies 394 on denser plates are expected to be smaller than colonies on less dense plates at a given point (Fig. 7A). 395 On top of this, if the plating is heterogeneous, not all colonies have access to the same amount of 396 resources. They therefore differ in size within a plate as soon as they enter the saturation phase. Thus, 397 experiments aiming to compare either colony size or appearance time are biased by colony density. 398 We explored the response to density using our purposely designed demonstration dataset of 399 colonies formed by the S. aureus lab strain Cowan I growing on Columbia sheep blood agar. Of the 400 22 time-lapse movies, eight were plated from a liquid culture at exponential phase to obtain a 401 homogeneously growing population (control dataset, Fig. 7A). The other fourteen were plated after 402 exposure to the antibiotic rifampicin, which arrests bacterial cell growth by stopping protein synthesis, 403 resulting in a bacterial population with a wide appearance time distribution (further referred to as rifa 404 dataset, Fig. 7A) (Kwan et al., 2013). In order to observe the impact of density on colony radius over 405 time and elaborate a method to correct this bias, the replicates were plated at different densities over 406 two orders of magnitude (ranging from 16 to 1509 colonies per plate) (SI Fig. 8, 9). 407 The effect of high densities on the colony growth rate is shown on Fig. 7BD displaying the colonies' 408 radii against their observed appearance times at a given timepoint. One may follow the evolution of 409 this relationship over time (SI movie 2). At low densities, colonies grow at the maximal linear radial 410 growth rate, GRmax, which depends on strain and culture conditions (Fig. 5). Colonies that deviate from 411 the GRmax appear below the expected linear regression line (Fig. 7C). Therefore, when performing a 412 linear regression using GRmax, the appearance time of colonies which are already in the saturation 413 phase are overestimated (Fig. 6D). The systematic error (E) introduced by saturation corresponds to 414 the difference between tapp, GRmax  Voronoi cell area is time dependent and strongest at a late growth stage, as colonies approach their 438 maximal size (Chacón et al., 2018). 439 Aiming to use the Voronoi cell area (Va) to approximate the error in appearance time estimation 440 introduced by saturation (E), we evaluated the mathematical relation between log(Va) and E in the 441 control dataset. At 24 h, E was close to 0 for colonies with large Va and increased for colonies with 442 smaller Va (Fig. 8A). The relation E(Va) was fitted with a modified logistic model (Fig. 8A): 443 where Emin is set to 0, and Emax, b and c are fitted parameters using the least square method. Emax, 444 exists because even for infinitely small values of Va, colonies still grow by using nutrients immediately 445 below them. b and c are shape parameters for the transition from Emin to Emax. 446 As expected, the relation E(Va) was time dependent. When fitted through time (SI movie 3, 447 SI Fig. 10A, B, C), a linear increase of Emax was observed starting as soon as the first colonies enter 448 saturation phase. In addition to being time-dependent, E(Va) was observed to depend on the general 449 lag time of the plate: at a given timepoint, colonies originating from populations with a large median 450 lag time (e.g. rifa dataset) were in an earlier phase of their growth as compared to colonies originating 451 from populations without lag (e.g. control dataset) (Fig. 8B). Ideally, one would adapt the relation E(Va) 452 to take into account the median lag time of the plate, so that the correction applied to the rifa dataset 453 observed at a given timepoint (tobs) corresponds to E(Va) fitted on the control dataset at an earlier 454 timepoint, when the colonies were still at that stage of growth. 455 The difference in appearance time between the control dataset (median appearance time: 9 h) and 456 the rifa dataset (median appearance time: 17.4 h) was Δtapp = 8.4 h. The correction applied to the rifa 457 dataset when observing colonies at tobs = 24 h should thus correspond to E(Va) fitted on the control 458 dataset at tobs-Δtapp = 15.6 h (Fig. 8C). However, the median appearance time is not known a priori 459 when observing colonies on a plate at an endpoint. 460 Therefore, we used an iterative process to estimate appearance time. An initial guess of the 461 appearance time (tapp,i) was obtained with Eq. 2 where E(Va) was fitted on the control dataset at tobs . 462 Then the difference between the known median tapp of the control dataset and the median tapp,i of the 463 observed plate (Δtapp,i) was calculated. A new tapp,i was estimated with Eq. 2 where E(Va) was fitted on 464 the control dataset at tobs -Δtapp,i, which in turn gave a new Δtapp,i. This process was iterated until 465 stabilization for a final estimate of tapp ( Fig. 8E to H, Fig. S10 A to H and L). 466 Note that the shown approach tended to overcorrect the appearance time of colonies on dense 467 plates due to the nature of the data: fully saturated colonies have little information left regarding their 468 appearance time since the radius does not change with time anymore. Thus, this correction cannot 469 perform well if colonies have reached late saturation phase (SI Fig. 10E to F)

492
We developed ColTapp, a user-friendly application with a graphical interface. This application 493 extracts biologically relevant data from images of microbial colonies including size, growth rate and 494 appearance time as well as color, shape and spatial metrics. Colony growth dynamics parameters are 495 derived from time-lapse images. As it might not always be feasible to utilize high-throughput time-496 lapse imaging, ColTapp also includes a framework to estimate appearance time from endpoint images. 497 Colony size at an endpoint is affected by colony density, resulting in a bias in appearance time 498 estimation. Thus, we propose an iterative correction utilizing local plate density to moderate this bias. 499 The correction is based on a control dataset and applied to a dataset with different growth    With ColTapp, we propose an easy to use graphical user interface, so that any laboratory, regardless 538 of the technical equipment and image analysis expertise accessible, can approach single cell growth 539 phenotypes by monitoring colony growth dynamics. We qualitatively compare ColTapp with other 540 tools available in Table 1 below. Usage of colony size determined by image analysis to assess stressor 541 effects is a long-established method (Dykes, 1999) and various end-user applications to facilitate 542 colony counting and size assessment were developed over the years (e.g. a standalone GUI by Clarke application including a graphical user interface currently available for this purpose. Therefore, by 557 bringing together the standard endpoint analysis with time-lapse image analysis, our user-friendly 558 ColTapp fills a gap in the collection of existing dedicated image analysis tools. 559 560 As for any image analysis tool, the quality of the analysis will highly depend on the quality of the 561 images themselves. We recommend users to maintain homogenous lighting, avoid light flares and 562 reflections and create good contrast using background modifications and proper focusing. ColTapp is 563 flexible enough to allow processing of colored and grayscale images acquired either through dedicated 564 platforms or with simple camera solutions. For endpoint images, this can be best achieved using a 565 dedicated white box with diffused, indirect light, but recent phone cameras can also give decent results 566 on a bench. 567 We propose an analysis of the linear radial growth phase, assuming a colony has already reached 568 this stage upon appearance. It is not difficult to observe the initial exponential part of the colony 569 growth (occurring before a colony reaches a radius of ~100µm) by using commercially available 570 photographic macro lenses (Fig. S6). However, one should be aware that defining appearance time 571 with a linear regression is incorrect in this case as the exponential growth phase is actually observed 572 and should be taken into account. Similarly, our estimation of appearance time from endpoint images 573 assumes that colonies are in the linear growth phase. 574 As soon as the steady flow of nutrients sustaining colony growth decreases (timing is dependent on 575 proximity of competing neighbors and the total number of colonies on the plate), the colony enters 576 saturation phase, which impacts the linear appearance time estimation. Density may even affect the 577 observed appearance time if colonies enter saturation phase before they are visible (Levin-Reisman et 578 al., 2010). However, in the presented S. aureus dataset (including extremely dense plates), only a 579 marginal correlation was observed between colony appearance time and density (SI Fig. 11). Colony 580 growth is a complex biological process and different species on different growth substrates will 581 respond differently to density (Chacón et al., 2018). For this reason, it is impossible to propose a 582 universally valid method to take density into account. Nevertheless, the explored correction based on 583 spatial metrics, which can easily be exported with ColTapp, might be used as a starting point to develop 584 suitable correction methods for the investigated species. 585 We wrote the ColTapp application using the classical programming language MATLAB, and the code 586 is designed as a modular shell that can host further image analysis methods that may meet specific 587 needs, while benefiting from the easy to operate graphical interface. We enable the user to export all 588 generated data, including radius, appearance time, growth rate, spatial metrics, and colony 589 characteristics such as shape and color. We envision that more measurements can subsequently be 590 added to facilitate the further study of intra-and inter-species colony interactions. 591

592
Software 593 We programmed the ColTapp application with MATLAB 2020a (MathWorks). Preliminary code of 594 the GUI and time-lapse radius tracking was proposed earlier (Vulin et al., 2018). All algorithm accuracy 595 and computational efficiency tests were performed on a computer with Windows 10 OS, i7-8700K 596 @4.70GHz, 32GB RAM. 597 We used external code acquired from the File Exchange server of MathWorks for certain functions. 598 Namely, these were a Voronoi tessellation function by Sievers (2020) ColTapp does not rely on a specific image format or quality. Achieving homogenous lighting is 612 undoubtedly critical for successful image analysis, but image acquisition itself can be operated using 613 any kind of device. For correct spatial calibration, the camera should be positioned perpendicular to 614 the plate. For time-lapse imaging, acquisition triggering at defined time intervals is necessary. Typical 615 systems include scanners, dedicated applications and cameras or commercially available consumer-616 grade camera, provided they include a time-lapse mode or can be triggered by an external controller. 617 To generate the data presented in this paper, the time-lapse images were acquired either using a 618 Canon EOS 1200D reflex camera triggered through an Arduino controller ( Figure 1A) or Basler acA5472-619 17um 20MP cameras with Fujinon Objectives CF16ZA-1S 16mm/1.8M37.5x0.5 triggered with Basler's 620 pylon software suite (Version: 6.0.1). The plates were rested on a black background surface below the 621 cameras and the whole system was set inside an incubator at 37° C. To prevent drying out of agar 622 plates and airborne contamination, lids were kept on plates used for time-lapse imaging. Best image 623 analysis quality was observed when special attention was brought to the illumination method: indirect, 624 homogeneous white illumination was obtained by covering the incubator walls with white paper and using a LED based light source emitting minimal heat. Endpoint images were acquired manually with a 626 Canon EOS 1200D reflex camera and lids were temporarily removed. Plates were placed in a custom-627 built box with light diffusing sides and a fixed plate-and camera-holder. 628 629

630
Staphylococcus aureus strain Cowan I was grown in shaking test tubes containing tryptic soy broth 631 (Difco) inoculated at an optical density of 0.2 with or without rifampicin (1 µg/ml). The control dataset 632 was generated by plating appropriate dilutions of the non-treated cultures at mid-exponential phase. 633 For the test dataset, the rifampicin exposed cultures were recovered after 24 hours, pelleted (10'000 g, 634 3 minutes), resuspended twice in phosphate buffered saline solution and then plated at appropriate 635 dilutions. 636 Blood agar plates (Columbia, 5% sheep blood) were purchased at BioMérieux. Other strains and 637 media used are described in the main text. 638