Expanding Tuberculosis Drug Resistance Prediction beyond binary: Deep Learning for Minimum Inhibitory Concentration prediction

Expanding Tuberculosis Drug Resistance Prediction beyond binary: Deep Learning for Minimum Inhibitory Concentration prediction | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (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],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Expanding Tuberculosis Drug Resistance Prediction beyond binary: Deep Learning for Minimum Inhibitory Concentration prediction Linfeng Wang, Susana Campino, Taane G. Clark, Jody E. Phelan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7621453/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 24 You are reading this latest preprint version Abstract Tuberculosis (TB), caused by Mycobacterium tuberculosis , remains a major global health concern, with case numbers increasing since 2021. In 2024, over 10 million new cases were reported, along with 1.1 million deaths 1 . At the same time, the widespread adoption of whole genome sequencing (WGS) has made it possible to predict drug resistance from genetic data, allowing for faster diagnostic workflows. Several tools have been developed to classify isolates as "resistant" or "susceptible". However, drug resistance is not always binary and often exists along a continuum. Cryptic resistance is one example, where phenotypic resistance occurs in the absence of known resistance mutations. The CRyPTIC dataset, which includes over 12,000 isolates and MIC (minimum inhibitory concentration) measurements across 12 anti-TB drugs, provides a valuable resource for moving beyond binary classification. MIC values quantify the lowest concentration of a drug needed to inhibit bacterial growth, offering a more detailed picture of drug susceptibility. In this study, we build models that extend binary classification to directly predict MIC levels from genomic features. To handle the severe class imbalance in completeness across MIC distributions, especially for newer or second-line drugs, we use oversampling and label-aware training techniques. We compare two modelling approaches: XGBoost, which is well-suited to structured data, and convolutional neural networks (CNNs), which can capture spatial and hierarchical relationships within genomic inputs. XGBoost demonstrated more consistent performance in the presence of imbalance, while CNNs achieved higher resolution when the MIC classes were more evenly distributed. Feature importance analysis revealed that some variants previously thought to cause resistance were linked to lower MIC values, suggesting they may only contribute to low drug resistance instead, where the isolate can be killed with higher doage of the same drug. These insights open the door to more tailored treatment strategies, including the use of higher doses of first-line drugs, which could reduce toxicity, improve patient adherence, and slow the emergence of resistance to newer therapies. Figures Figure 1 Figure 2 INTRODUCTION Tuberculosis disease (TB), caused by Mycobacterium tuberculosis , continues to pose a pressing global health challenge. While existing treatment regimens are generally effective, the emergence and increasing prevalence of drug-resistant strains are rapidly complicating the landscape. In 2024 alone, over 10 million new TB cases were reported, leading to 1.1 million deaths worldwide, including among individuals co-infected with HIV 1 . Multidrug-resistant tuberculosis (MDR-TB), where there is resistance to isoniazid (INH) and rifampicin (RIF), accounts for a growing proportion of these cases, yet only 40% of patients with MDR-TB receive proper treatment, underlining an urgent need for rapid, precise, and accessible diagnostics. Worryingly, extensively drug-resistant TB strains (XDR; MDR + fluoroquinolones + group A resistance) and pre-XDR forms exist, providing a progression of resistance and limiting treatment options. In response, recent WHO guidelines suggest a 6-month regimen (BPaLM) comprising bedaquiline (BDQ), pretomanid, linezolid (LNZ) and moxifloxacin (MXF) to reduce both treatment duration and non-compliance levels due to drug toxicity 2 , 3 . Genotypic drug resistance prediction based on whole-genome sequencing (WGS) has shown strong promise for rapid, culture-free prediction of binary resistance phenotypes. Tools such as TB-Profiler 4 can identify canonical resistance mutations and deliver results in hours rather than weeks. However, binary classification reduces the complexity of resistance to a simple “resistant” or “susceptible” label, overlooking intermediate or cryptic resistance patterns. In laboratory phenotypic drug susceptibility testing (pDST), the minimum inhibitory concentration (MIC) is defined as the lowest concentration of an antimicrobial that prevents visible growth of a microorganism under defined laboratory conditions, and it offers a more detailed view 5 . In TB, MIC values quantify the bacterial susceptibility to each antibiotic, revealing gradations that binary labels obscure. MIC values are derived from a series of drug dilutions, where growth inhibition is assessed. This approach not only confirms resistance but also highlights intermediate susceptibility and subtle shifts in drug response, both of which are critical for detecting emerging or cryptic resistance. By capturing this continuum, MIC-based prediction models can extend the capabilities of current binary tools and support more precise, individualised treatment strategies. Machine learning algorithms applied to WGS data make it possible to infer this richer information directly from genomic sequences 6 , 7 , 8 . Among these methods, deep learning models, particularly convolutional neural networks (CNNs) 9 , 10 , with their layered architecture and spatially aware kernels, are well-suited to identifying latent hierarchies and complex non-linear patterns in genomic data 10 . Building on existing binary prediction frameworks, CNN-based and other machine learning approaches can be adapted to predict MIC values, offering an enhanced resolution of drug resistance profiles. In this study, we build on existing binary drug resistance prediction frameworks by developing models capable of producing more refined, quantitative predictions in the form of MIC values across 12 anti-TB drugs. Our approach is designed to capture the full spectrum of M. tuberculosis susceptibility, moving beyond simple “resistant” or “susceptible” classifications toward predictions that can inform more nuanced, precision-guided therapy. At the same time, we address one of the major challenges in applying MIC-based prediction to real-world data: the severe class imbalance within the CRyPTIC dataset, particularly in cryptic resistance subsets where canonical resistance mutations are absent. By imbalance, we refer to the disproportionate representation of susceptible isolates compared to resistant ones, as well as the uneven distribution of MIC values across the full dilution range. This imbalance is not random but arises because resistant isolates, especially those carrying rare or novel mutations, are much less frequently observed in clinical settings. It is further exacerbated for second-line and third-line drugs, where resistance is comparatively rare and fewer patient samples are available, making it difficult to capture the complete diversity of resistance phenotypes. To achieve these goals, we design and compare a range of machine learning models, including CNNs and XGBoost classifiers 11 , two widely used architectures in computational drug resistance research. RESULTS Drug resistance phenotypes The MIC distributions across 12 anti-TB drugs for 12,285 Mtb isolates from the CRyPTIC 12 dataset are summarised. These drugs are INH, RIF, Ethambutol (EMB), Ethionamide (ETH), Injectables (Amikacin (AMI), Kanamycin (KAN)), Fluoroquinolones (FQs; Levofloxacin (LEV), Moxifloxacin (MXF)), BDQ, Clofazimine (CFZ), Delamanid (DLM), and LZD (Table 1 , Fig. 1 ). Based on ECOFF thresholds, each drug can be grouped into two or three resistance categories. This preliminary analysis highlights substantial class imbalance across most drugs, with full breakdowns of MIC category counts provided ( Supplementary Fig. 1) . The dataset includes three first-line drugs: INH, RIF, and EMB, among which INH and RIF display high resistance prevalence at 48% and 38%, respectively. In contrast, EMB shows a greater degree of imbalance, with only 8% of isolates exhibiting resistance. This pattern of imbalance extends across most drugs, with MXF showing the highest non-susceptible proportion at 14%. In comparison, the newer TB drugs, BDQ and DLM, show the lowest resistance rates at 0.9% and 1.5%, respectively. Figure 1 b illustrates the distribution of MIC values, where the majority of samples over 80% fall into the susceptible or low-resistance range, indicated by high counts near the lower MIC endpoints. Missing values are generally rare across all drugs, though data for BDQ and CFZ show the highest missingness at 221 and 240 samples out of 12,285, respectively (Table 1 ). Table 1 The dataset (n = 12,289) Drug N MIC median (range) MIC Imbalance Susceptible Resistant Binary imbalance AMI 11,738 0.25 (0.25-16.0) 222.2 10316 1422 7.25 BDQ 11,743 0.03125 (0.015-2.0) 16.8 11400 343 33.24 CFZ 11,718 0.125 (0.062-4.0) 1.5 10003 1715 5.83 DLM 11,594 0.0078 (0.0078-1.0) 2.1 11177 417 26.80 EMB 11,823 2.0 (0.5–16.0) 46.0 8207 3616 2.27 ETH 11,800 1.0 (0.5–16.0) 1.6 9291 2509 3.70 INH 11,736 0.05 (0.025-3.2) 19.5 5817 5919 0.98 KAN 11,797 2.0 (1.0–32.0) 57.7 9103 2694 3.38 LEV 11,829 0.5 (0.125-16.0) 21.6 9055 2774 3.26 LZD 11,854 0.5 (0.016-8.0) 1.1 10112 1742 5.80 MXF 11,860 0.25 (0.125-8.0) 38.0 9716 2144 4.53 RIF 11,769 0.125 (0.031-8.0) 4.7 7132 4637 1.54 AMI: Amikacin, BDQ: Bedaquiline, CFZ: Clofazimine, DLM: Delamanid, EMB: Ethambutol, ETH: Ethionamide, INH: Isoniazid, KAN: Kanamycin, LEV: Levofloxacin, LZD: Linezolid, MXF: Moxifloxacin, RIF: Rifampicin Initial modelling using XGBoost and CNNs trained directly on the imbalanced dataset yielded poor predictive performance, skewed toward susceptible classifications (e.g., AUC = 0.5, sensitivity = 0, specificity = 1). To address this, all subsequent models were trained on an oversampled dataset, where minority MIC categories were boosted to at least 50% of the size of the most represented category, mitigating the skew and enabling more informative learning. All MIC values across drugs are nested and predicted in the same model. Pairwise correlations between MIC values across drugs highlighted KAN and AMI (both injectables) and MXF and LEV (both fluoroquinolones) cross-resistance (Fig. 2 ). Mono-resistance was rare, with isoniazid being the highest (5%). Model Performance Across the full drug panel, substantial binary imbalance was observed between resistant and susceptible isolates (Table 1 ). This was most pronounced in BDQ and DLM, where fewer than 2% of samples were resistant. However, this binary skew does not always align with inter-MIC class imbalance. For example, AMI displayed highly uneven MIC category distributions but retained a relatively balanced binary resistance profile. This highlights an added layer of complexity in the dataset, categorical imbalance within MIC bins may not correspond directly to resistance thresholds, complicating classification. Model comparisons underscore how these imbalances impact performance (Table 2 ). Binary-supported CNNs outperformed conventional CNNs in double dilution accuracy across 8 of 12 drugs, including low-resistance drugs such as BDQ and CFZ, where standard CNNs failed to generalise (0.05 and 0.01 accuracy, respectively). Binary support consistently rescued CNN performance in these cases, suggesting that label-aware sampling plays a critical role in mitigating skew. In contrast, XGBoost had greater stable performance with or without binary support. Gains from binary-aware sampling were modest, such as in BDQ where AUC improved from 0.51 to 0.53. Notably, XGBoost achieved higher AUC than CNNs in 7 of 12 drugs, pointing to its stronger class separation ability. However, in terms of double dilution, i.e., accurate MIC bin prediction, CNNs often had the edge when balanced. Table 2 Model performance across binary-supported CNN, conventional CNN, and XGBoost, assessed using Doubling Dilution Accuracy and AUC-OVR (one-vs-rest strategy). Double dilution AUC-OVR Binary supported CNN CNN XGB Binary supported XGB Binary supported CNN CNN XGB Binary supported XGB AMI 0.91 0.10 0.84 0.91 0.62 0.54 0.54 0.60 BDQ 0.59 0.05 0.90 0.90 0.53 0.50 0.51 0.53 CFZ 0.94 0.01 0.85 0.85 0.53 0.50 0.50 0.53 DLM 0.54 0.46 0.60 0.60 0.56 0.56 0.57 0.55 EMB 0.81 0.87 0.88 0.89 0.76 0.76 0.76 0.76 ETH 0.88 0.84 0.84 0.89 0.64 0.65 0.63 0.62 INH 0.88 0.86 0.87 0.89 0.70 0.73 0.75 0.71 KAN 0.96 0.78 0.90 0.95 0.71 0.55 0.56 0.71 LEV 0.87 0.87 0.86 0.85 0.58 0.71 0.73 0.60 LZD 0.90 0.11 0.89 0.89 0.57 0.49 0.49 0.58 MXF 0.69 0.71 0.81 0.82 0.71 0.74 0.73 0.72 RIF 0.86 0.85 0.86 0.89 0.77 0.80 0.82 0.71 AMI: Amikacin, BDQ: Bedaquiline, CFZ: Clofazimine, DLM: Delamanid, EMB: Ethambutol, ETH: Ethionamide, INH: Isoniazid, KAN: Kanamycin, LEV: Levofloxacin, LZD: Linezolid, MXF: Moxifloxacin, RIF: Rifampicin, CNN: Convolutional Neural Network, XGB: Extreme Gradient Boost Taken together, these results show that binary-supported CNNs are highly effective when class balance is enforced, offering fine-grained MIC prediction. Yet they remain fragile under imbalance. In contrast, XGBoost is robust under skewed distributions, but may lack the resolution for precise MIC classification. This trade-off between discrete classification and probabilistic ranking is central to model selection in computational DST. Four drugs are presented (AMI, BDQ, RIF, EMB), and capture a diverse range of resistance profiles and modelling complexities. AMI and BDQ exhibit pronounced MIC-level imbalance, with BDQ also showing extreme binary skew due to its low resistance prevalence. RIF, by contrast, offers a relatively balanced dataset, while EMB sits between the two extremes with moderate imbalance. This selection enables a more comprehensive evaluation of model behaviour under different data distribution scenarios. Both binary-supported CNN and XGBoost models produce similar proportions of correct and incorrect predictions across all four drugs ( Supplementary Fig. 2 ). In highly imbalanced cases such as BDQ, both models misclassify large portions of the data, often redirecting predictions to more populated categories ( Supplementary Fig. 3 ). Interestingly, the nature of these misclassifications differs by model, with CNN and XGBoost tending to err in different directions. Notably, XGBoost fails to predict certain low-frequency MIC bins (e.g., 0.5 and 1), effectively ignoring them during classification. A similar pattern is observed for AMI, another imbalanced case. In contrast, RIF and EMB, which have more even distributions, show prediction errors primarily in minor classes, but all MIC bins remain represented in the model outputs. SNP feature importance in rifampicin resistance To illustrate the interpretability of the binary-supported CNN model, feature ablation was applied to assess the relative contribution of individual SNPs and auxiliary features to RIF resistance prediction ( Supplementary Table 1 ). RIF was selected because its resistance mechanisms are well characterised. Ranked by absolute integrated gradients (IG abs ), which reflect a feature's influence on the model output, the most dominant signal was the binary drug resistance label predicted by TB-Profiler, vastly exceeding the contributions of individual SNPs. This highlights the strength of combining curated rule-based predictions with learned representations. Among the top 20 genomic variants, the majority fall within the rpoA, rpoB , and rpoC genes, consistent with established mechanisms of rifampicin resistance. However, several variants, including rpoC Leu1245Arg and rpoA Ile3Val, while not currently annotated by WHO, exhibited strong positive IG direction and high ΔAUC, indicating both directional and predictive importance. Interestingly, canonical rifampicin resistance mutations such as rpoB Ser450Leu were not among the top-ranked features in this model, suggesting they may contribute more to classification certainty rather than to MIC shifts in this regression framework, in this case their influence may have been covered in the binary drug resistance input. Several mutations, including rpoB Val970Met and Rv2752c Gly28Ser, displayed negative IG direction, implying an inverse contribution to resistance probability in the modelled output range. This directional signal reflects the model’s interpretation of these variants as associated with lower MIC values, despite their presence in resistant isolates. In contrast, ΔAUC captures the impact of ablation on predictive performance across the test set, revealing variants such as rpoA c.-141C > G and rpoC Val1039Gly as enhancing the discriminative capacity of the model. DISCUSSION The CRyPTIC dataset offers a valuable foundation for advancing computational drug susceptibility testing in Mtb , but it also highlights limitations and significant challenges. In addition to the binary imbalance between susceptible and resistant isolates, there is substantial skew across the MIC spectrum, with certain MIC categories, particularly for newer and repurposed drugs such as BDQ and DLM, being sparsely represented. This imbalance is most pronounced in cryptic resistance cases, where canonical resistance mutations are absent and phenotypic resistance remains rare. Such distributional bias complicates both training and evaluation, especially for deep learning architectures that depend on balanced examples to generalise effectively. Our work reframes prediction from a binary classification task to a more granular MIC regression problem, building on binary tools while addressing the imbalance inherent to cryptic subsets. To improve model stability and accuracy under these conditions, we evaluated strategies such as ensemble learning, sample reweighting, and feature dropout. CNNs demonstrated strong performance for fine-grained MIC estimation when guided by these mitigation strategies, benefiting from their capacity to detect complex, non-linear genomic patterns. XGBoost, while less sensitive to imbalance, delivered competitive results with fewer architectural adjustments, suggesting complementary strengths between the two approaches. The analysis of feature attribution uncovered several mutations with negative integrated gradient (IG) direction, indicating that their presence systematically lowers the predicted MIC. Notably, mutations such as Rv2752c Gly28Ser and rpoC c.339T > C, despite being enriched in resistant isolates and previously catalogued as drug-resistance-associated, appear to shift model predictions toward susceptibility. This discrepancy suggests these variants may play a role in low drug resistance or tolerance rather than conferring full resistance. Such mutations may permit effective treatment through intensified dosing of first-line agents like rifampicin, potentially reducing reliance on newer drugs with greater toxicity profiles. This raises important clinical implications: better stratification of resistant subtypes could optimise regimen design, enhance adherence, and slow the emergence of resistance to next-generation therapeutics. Our results highlight two key advances. First, MIC-based modelling can meaningfully extend binary prediction tools, offering richer phenotypic resolution for clinical decision-making. Second, careful handling of data imbalance, particularly in cryptic subsets, can stabilise training and preserve performance in deep learning models. Beyond improving predictive accuracy, integrating interpretable machine learning into this workflow enhances transparency and opens avenues for hypothesis generation, guiding experimental validation of both known and novel variants. Lastly, to fundamentally cope with the data imbalance problem, more clinical samples or lab-generated samples with various levels of MIC need to be generated across different lineages. More generally, increased MIC characterisation combined with WGS will lead to improvements in TB clinical decision-making, thereby improving patient outcomes. METHODOLOGY Dataset The CRyPTIC dataset 12 encompasses minimum inhibitory concentration (MIC) phenotypes for 12,289 Mtb complex isolates from 23 different countries, spanning lineages L1 to L4, and L6. A total of 6,814 isolates show resistance to at least one drug. Among these, 2,129 (31.2%) isolates met the clinical criteria for being categorised as rifampicin-resistant (RR), multidrug-resistant (MDR), pre-extensively drug-resistant (pre-XDR), or extensively drug-resistant (XDR). The MIC data was gathered using a 96-well broth microdilution plate, covering a spectrum of anti-TB drugs including three primary ones (isoniazid, rifampicin, ethambutol), alongside rifabutin, two fluoroquinolones (moxifloxacin, levofloxacin), two injectables (amikacin, kanamycin), ethionamide, and four newer or repurposed medications (bedaquiline, linezolid, clofazimine, delamanid). The edge case MIC values that signify a boundary were replaced by the halving or doubling the lower or upper edge case MIC values to follow double dilution as shown in Supplementary Table 2 . All other edge cases equal or smaller than this value is merged into the stated MIC edge value. The specific cut-offs used are the same as The CRyPTIC consortium publication in 2022 14 , according to the ECOFF values 13 . Data Preprocessing In this study, to reduce computational load and memory usage, we adopted a streamlined representation of genetic variation by constructing drug-specific resistance vectors for each isolate. For each antibiotic, only SNPs located within gene regions with well-established links to resistance were included. These genes were as follows: inhA , katG , kasA , oxyR-ahpC , and fabG1 for isoniazid; ethA and inhA for ethionamide; embA , embB , and embC for ethambutol; rpoA , rpoB , and rpoC for rifampicin; gyrA and gyrB for levofloxacin and moxifloxacin; rrs and eis for kanamycin and amikacin; Rv0678 , atpE , and pepQ for bedaquiline; Rv0678 and pepQ for clofazimine; fgd1 , fbiA , fbiB , fbiC , fbiD , and ddn for delamanid; and rrl and rplC for linezolid. Each resistance vector was composed of binary indicators representing the presence or absence of SNPs in these loci, derived from TB-Profiler variant calls. This targeted encoding strategy reduced the input dimensionality per model while preserving biological relevance to the resistance phenotype being predicted. To ensure compatibility with batch processing and model input requirements, sequence padding and normalisation were applied where necessary. All datasets were stratified into training, validation, and testing subsets to allow for consistent and rigorous performance evaluation. CNN model The convolutional neural network (CNN) architecture ( Supplementary Table 3 ) used for MIC prediction comprised an initial feature extraction block with 64 filters and kernel size of 25, followed by two stacked convolutional modules with dynamic filter scaling. Each convolutional layer was followed by batch normalisation, Rectified Linear Unit (ReLU) activation, and dropout regularisation at a rate of 0.4. Downstream processing was performed by two dense layers, each with 128 neurons and dropout at a rate of 0.7. A prediction head merged latent features with an additional scalar input, followed by a sequence of fully connected layers and non-linear activations to yield multi-class predictions. Kaiming normal initialisation was applied to all linear layers. Training proceeded for up to 500 epochs. A grid search was performed to optimise model hyperparameters. Oversampling and category weighting were incorporated into a custom cross-entropy loss function to address class imbalance. The Adam optimiser was used to update network weights, with a learning rate fixed at 1x10 − 4 and weight decay (L2 regularisation) set at 1x10 − 4 . For the binary supported CNN architecture, the binary predictions for drug resistance from TB-profiler was fed into the final layer of the CNN. The CNN was developed and implemented using Pytorch 1.12.1 in Python (v3.9.15) with CUDA 11.3. Model training was performed on a Tesla v100-pcie-32gb graphics processing unit (GPU). Benchmarking To benchmark the results, the binary supported CNN model was compared against a normal CNN of the same architecture, as well as XGBoost. Two performance evaluation metrics were calculated, namely the Doubling Dilution Accuracy and the AUC-OVR. Doubling Dilution Accuracy reflects how closely predicted MICs align with actual laboratory-measured values on a two-fold dilution scale, while AUC-OVR quantifies overall classification performance across all MIC levels Feature importance To assess the contribution of individual features to drug resistance prediction, we applied an integrated attribution framework combining model ablation with gradient-based interpretation. Three complementary metrics were computed: delta AUC (ΔAUC), integrated gradients direction (IG dir ), and integrated gradients magnitude (IG abs ). ΔAUC quantifies the drop in predictive performance when a feature is replaced with baseline values (default: 0). For each SNP or categorical input, we masked the feature across a test batch and measured the resulting change in area under the ROC curve, averaged over multiple batches. A larger positive ΔAUC indicates greater loss of predictive power upon feature ablation. In parallel, integrated gradients were computed with respect to each input feature. IG dir captures the signed attribution, indicating whether a feature increases or decreases the model's prediction for the target class, while IG abs reflects the overall magnitude of influence. Together, these metrics provide complementary insights into model behaviour. ΔAUC assesses contribution to class discrimination, IG dir reveals the direction of predictive influence, and IG abs ranks the salience of features regardless of direction. Declarations ETHICS APPROVAL AND CONSENT TO PARTICIPATE This study did not involve the collection of new human or animal data. All analyses were conducted on de-identified, publicly available data from the CRyPTIC project (https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3001721), an open-access resource. The CRyPTIC dataset is curated in accordance with ethical standards, including removal of personal identifiers, and is made available for research under the program’s data-sharing policies. As no direct patient contact or intervention was performed by the authors, additional institutional ethical approval was not required. This study was conducted in accordance with the ethical principles set out in the Declaration of Helsinki. CONSENT TO PARTICIPATE DECLARATION Not applicable. CONSENT FOR PUBLICATION All authors have read and approved the final manuscript. They consent to the publication of this work and confirm that the content is original and has not been published or submitted for publication elsewhere. AVAILABILITY OF DATA AND MATERIAL Not data collection was done. All data used is from online database TB portal. All data is available upon request from TB portal: https://tbportals.niaid.nih.gov/ The code and model can be found in the author’s Github: https://github.com/linfeng-wang/TBpt DATA AVAILABILITY Not data collection was done. All data used is from online database TB portal. CRyPTIC data used in the study can be found in: https://ftp.ebi.ac.uk/pub/databases/cryptic/release_june2022/reuse/ The code and model can be found in the author’s Github: https://github.com/linfeng-wang/tb_dr_MIC3 COMPETING INTERESTS The authors declare that they have no conflicts of interest. AUTHOR CONTRIBUTIONS JEP and LW conceived and directed the project. LW developed the models under the supervision of SC, TGC and JEP. LW wrote the first draft of the manuscript. All authors commented and edited various versions of the draft manuscript and approved the final manuscript. LW, TGC and JEP compiled the final manuscript. FUNDING LW is funded by a BBSRC LIDO studentship (Ref. BB/T008709/1). TGC and SC are funded by the UKRI (BBSRC BB/X018156/1; MRC MR/R020973/1, MRC MR/X005895/1; EPSRC EP/Y018842/1). The funders had no role in the study design, data collection and analysis, the decision to publish, or preparation of the manuscript. The authors declare no conflicts of interest. References Global Tuberculosis Report. 2024. https://www.who.int/teams/global-programme-on-tuberculosis-and-lung-health/tb-reports/global-tuberculosis-report-2024 Vanino E, et al. Update of drug-resistant tuberculosis treatment guidelines: A turning point. Int J Infect Dis. 2023;130:S12–5. WHO consolidated guidelines on tuberculosis. Module 4: treatment - drug-resistant tuberculosis treatment, 2022 update. https://www.who.int/publications/i/item/9789240063129 Phelan JE, et al. Integrating informatics tools and portable sequencing technology for rapid detection of resistance to anti-tuberculous drugs. Genome Med. 2019;11:1–7. Popov S, Kuzmin A, Sabgayda T, Vedenina N. Minimum inhibitory concentrations (MIC) determination of TB drugs and broad-spectrum antibiotics in M.tuberculosis with M/X/TDR. Eur Respir J 46, (2015). Deelder W, et al. Using deep learning to identify recent positive selection in malaria parasite sequence data. Malar J. 2021;20:270. Libiseller-Egger J, et al. TB-ML—a framework for comparing machine learning approaches to predict drug resistance of Mycobacterium tuberculosis. Bioinform Adv. 2023;3:vbad040. Wang L, Campino S, Phelan J, Clark TG. Mixed infections in genotypic drug-resistant Mycobacterium tuberculosis. Sci Rep. 2023;13:17100. Krizhevsky A, Sutskever I, Hinton GE. ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems. Volume 25. Curran Associates, Inc.; 2012. Green AG et al. A convolutional neural network highlights mutations relevant to antimicrobial resistance in Mycobacterium tuberculosis. bioRxiv 2021.12.06.471431 (2021) 10.1101/2021.12.06.471431 Chen T, Guestrin C, XGBoost:. A Scalable Tree Boosting System. in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794Association for Computing Machinery, New York, NY, USA, (2016). 10.1145/2939672.2939785 Consortium T, Cr. A data compendium associating the genomes of 12,289 Mycobacterium tuberculosis isolates with quantitative resistance phenotypes to 13 antibiotics. PLoS Biol. 2022;20:e3001721. CRyPTIC Consortium. Epidemiological cut-off values for a 96-well broth microdilution plate for high-throughput research antibiotic susceptibility testing of M. tuberculosis. Eur Respir J. 2022;60:2200239. Consortium T, Cr. Quantitative drug susceptibility testing for Mycobacterium tuberculosis using unassembled sequencing data and machine learning. PLoS Comput Biol. 2024;20:e1012260. Additional Declarations No competing interests reported. Supplementary Files SUPPLEMENTARYINFORMATION.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 12 Nov, 2025 Reviews received at journal 12 Nov, 2025 Reviewers agreed at journal 11 Nov, 2025 Reviews received at journal 11 Nov, 2025 Reviews received at journal 10 Nov, 2025 Reviewers agreed at journal 10 Nov, 2025 Reviewers agreed at journal 10 Nov, 2025 Reviews received at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviews received at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers agreed at journal 02 Nov, 2025 Reviewers agreed at journal 31 Oct, 2025 Reviewers agreed at journal 30 Oct, 2025 Reviewers agreed at journal 29 Oct, 2025 Reviewers invited by journal 28 Oct, 2025 Editor invited by journal 26 Oct, 2025 Editor assigned by journal 17 Sep, 2025 Submission checks completed at journal 17 Sep, 2025 First submitted to journal 15 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7621453","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":539270897,"identity":"0caf987f-f4a4-4226-93e9-03c924fd2c52","order_by":0,"name":"Linfeng Wang","email":"","orcid":"","institution":"London School of Hygiene \u0026 Tropical Medicine","correspondingAuthor":false,"prefix":"","firstName":"Linfeng","middleName":"","lastName":"Wang","suffix":""},{"id":539270899,"identity":"3eb64e36-8809-47d3-af3f-b04d6ae26f71","order_by":1,"name":"Susana Campino","email":"","orcid":"","institution":"London School of Hygiene \u0026 Tropical Medicine","correspondingAuthor":false,"prefix":"","firstName":"Susana","middleName":"","lastName":"Campino","suffix":""},{"id":539270901,"identity":"9e9d974d-b547-49a1-911b-cc6ccb407f31","order_by":2,"name":"Taane G. Clark","email":"","orcid":"","institution":"London School of Hygiene \u0026 Tropical Medicine","correspondingAuthor":false,"prefix":"","firstName":"Taane","middleName":"G.","lastName":"Clark","suffix":""},{"id":539270903,"identity":"d17facd6-b570-4d57-879c-ec4c1d248424","order_by":3,"name":"Jody E. Phelan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAElEQVRIiWNgGAWjYBAC9gYGAyB1gIGPgYHxwQMgAwoScGphhGlhY2BgNkhgOCBBkhY2CeK0tDdvk2DccUeOjf3ssYqEP3fqdBuYH35gbEvDraXnWJkE45lnxmw8eWk3EtueSZgdYDOWYGzLwa1lRo4ZUMHhxDaGHLMbiQ2HgVoYzBgY2yoIaqlv439jVpDwB6SF/RteLYJQLQlsEjlmDAlsIC08IFtwO0ya51ixReKZw4ZtEm+MJRLbDktuO8xTLJFwDrf3+dibN974uOOwPD9/juGHD38O85sdb9/44UNZMk4tQMAikdiAzGdmwBcrECUfGBvwqxgFo2AUjIIRDgD6/FaYoZFAAgAAAABJRU5ErkJggg==","orcid":"","institution":"London School of Hygiene \u0026 Tropical Medicine","correspondingAuthor":true,"prefix":"","firstName":"Jody","middleName":"E.","lastName":"Phelan","suffix":""}],"badges":[],"createdAt":"2025-09-15 14:08:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7621453/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7621453/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95500900,"identity":"1cf6f7d4-c64f-4c1d-ba4d-48878f761b05","added_by":"auto","created_at":"2025-11-10 05:23:46","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":277503,"visible":true,"origin":"","legend":"","description":"","filename":"MICDraft3sub.docx","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/6fb2cb748fed3f843855982c.docx"},{"id":95500908,"identity":"770ff632-5d49-4cff-87e6-d381092f9691","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7015,"visible":true,"origin":"","legend":"","description":"","filename":"96415ee0d4fd4dc191e9ec80f3ae125e.json","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/41d7521539cf84e7e4578a07.json"},{"id":95500904,"identity":"3267ff8c-840d-4057-80f6-11dce50b4068","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3264329,"visible":true,"origin":"","legend":"","description":"","filename":"SUPPLEMENTARYINFORMATION.docx","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/d67f8ad20dc45778872edc9f.docx"},{"id":95500907,"identity":"e698e114-dcff-4612-98c2-32d487b9e5cd","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77127,"visible":true,"origin":"","legend":"","description":"","filename":"96415ee0d4fd4dc191e9ec80f3ae125e1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/2da05a46cb2a2b8b28acef83.xml"},{"id":95500897,"identity":"6b37ea31-fc54-461f-b5d6-2f9d0797726d","added_by":"auto","created_at":"2025-11-10 05:23:46","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52172,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/e08d344a17d67e69c95a73a4.png"},{"id":95500899,"identity":"9ca25388-58ab-43b5-9853-e1440eaaca54","added_by":"auto","created_at":"2025-11-10 05:23:46","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16160,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/1ebb281b500fd9c04d824453.png"},{"id":95500905,"identity":"001d9fd7-bd38-457b-8377-f0650a7e88fe","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17258,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/4dfc886ec5dc63e44bb5154c.png"},{"id":95500874,"identity":"bd8e91f6-6530-4c2e-8687-96f96c0194f2","added_by":"auto","created_at":"2025-11-10 05:23:44","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5538,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/6a201978dc82d6979a82806b.png"},{"id":95500901,"identity":"348b780e-cb2d-4e25-9247-8b8cec6f1dc4","added_by":"auto","created_at":"2025-11-10 05:23:47","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":27776,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/46b921a3f929ff6c7f0d691a.png"},{"id":95500896,"identity":"ae62ba1a-7365-440b-9509-82f0459e1cc2","added_by":"auto","created_at":"2025-11-10 05:23:46","extension":"xml","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":75049,"visible":true,"origin":"","legend":"","description":"","filename":"96415ee0d4fd4dc191e9ec80f3ae125e1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/6047dee784e063c25b7addd4.xml"},{"id":95500903,"identity":"6f2f3512-784f-4449-89c3-cfdc9ba990aa","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"html","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83443,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/519b59975feade7fae65e43f.html"},{"id":95500906,"identity":"ce20f1bd-ac99-4c57-8ec6-2131c598278f","added_by":"auto","created_at":"2025-11-10 05:23:48","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":84542,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePhenotypic drug resistance (n=12,289)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/7715543cdf7095e79835b16c.png"},{"id":95500902,"identity":"9c0ab782-b85b-49a6-96e5-941d48c91557","added_by":"auto","created_at":"2025-11-10 05:23:47","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":130980,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeatmap of correlations between MIC values across 12 drugs (n=12,289).\u003c/strong\u003e Pair-wise resistance co-occurrence in off-diagonal elements, representing poly-resistant isolates resistant to at least two drugs.\u003c/p\u003e\n\u003cp\u003eAMI: Amikacin, BDQ: Bedaquiline, CFZ: Clofazimine, DLM: Delamanid, EMB: Ethambutol, ETH: Ethionamide, INH: Isoniazid, KAN: Kanamycin, LEV: Levofloxacin, LZD: Linezolid, MXF: Moxifloxacin, RIF: Rifampicin\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/0f1dbbd7591e350d776dad3a.png"},{"id":95531674,"identity":"ea3ebb8e-9954-4a7f-9e7a-acefbd6e7d7b","added_by":"auto","created_at":"2025-11-10 10:23:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":967027,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/22510f82-532a-4680-8d62-ca56d7dc9f4a.pdf"},{"id":95528614,"identity":"1c40101f-efb7-468c-9409-2f588ae81f96","added_by":"auto","created_at":"2025-11-10 10:16:20","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":3264329,"visible":true,"origin":"","legend":"","description":"","filename":"SUPPLEMENTARYINFORMATION.docx","url":"https://assets-eu.researchsquare.com/files/rs-7621453/v1/35488b92ab104a29565dae4b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Expanding Tuberculosis Drug Resistance Prediction beyond binary: Deep Learning for Minimum Inhibitory Concentration prediction","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eTuberculosis disease (TB), caused by \u003cem\u003eMycobacterium tuberculosis\u003c/em\u003e, continues to pose a pressing global health challenge. While existing treatment regimens are generally effective, the emergence and increasing prevalence of drug-resistant strains are rapidly complicating the landscape. In 2024 alone, over 10\u0026nbsp;million new TB cases were reported, leading to 1.1\u0026nbsp;million deaths worldwide, including among individuals co-infected with HIV\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Multidrug-resistant tuberculosis (MDR-TB), where there is resistance to isoniazid (INH) and rifampicin (RIF), accounts for a growing proportion of these cases, yet only 40% of patients with MDR-TB receive proper treatment, underlining an urgent need for rapid, precise, and accessible diagnostics. Worryingly, extensively drug-resistant TB strains (XDR; MDR\u0026thinsp;+\u0026thinsp;fluoroquinolones\u0026thinsp;+\u0026thinsp;group A resistance) and pre-XDR forms exist, providing a progression of resistance and limiting treatment options. In response, recent WHO guidelines suggest a 6-month regimen (BPaLM) comprising bedaquiline (BDQ), pretomanid, linezolid (LNZ) and moxifloxacin (MXF) to reduce both treatment duration and non-compliance levels due to drug toxicity\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eGenotypic drug resistance prediction based on whole-genome sequencing (WGS) has shown strong promise for rapid, culture-free prediction of binary resistance phenotypes. Tools such as TB-Profiler\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e can identify canonical resistance mutations and deliver results in hours rather than weeks. However, binary classification reduces the complexity of resistance to a simple \u0026ldquo;resistant\u0026rdquo; or \u0026ldquo;susceptible\u0026rdquo; label, overlooking intermediate or cryptic resistance patterns. In laboratory phenotypic drug susceptibility testing (pDST), the minimum inhibitory concentration (MIC) is defined as the lowest concentration of an antimicrobial that prevents visible growth of a microorganism under defined laboratory conditions, and it offers a more detailed view\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. In TB, MIC values quantify the bacterial susceptibility to each antibiotic, revealing gradations that binary labels obscure. MIC values are derived from a series of drug dilutions, where growth inhibition is assessed. This approach not only confirms resistance but also highlights intermediate susceptibility and subtle shifts in drug response, both of which are critical for detecting emerging or cryptic resistance. By capturing this continuum, MIC-based prediction models can extend the capabilities of current binary tools and support more precise, individualised treatment strategies.\u003c/p\u003e\u003cp\u003eMachine learning algorithms applied to WGS data make it possible to infer this richer information directly from genomic sequences\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Among these methods, deep learning models, particularly convolutional neural networks (CNNs)\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, with their layered architecture and spatially aware kernels, are well-suited to identifying latent hierarchies and complex non-linear patterns in genomic data\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Building on existing binary prediction frameworks, CNN-based and other machine learning approaches can be adapted to predict MIC values, offering an enhanced resolution of drug resistance profiles.\u003c/p\u003e\u003cp\u003eIn this study, we build on existing binary drug resistance prediction frameworks by developing models capable of producing more refined, quantitative predictions in the form of MIC values across 12 anti-TB drugs. Our approach is designed to capture the full spectrum of \u003cem\u003eM. tuberculosis\u003c/em\u003e susceptibility, moving beyond simple \u0026ldquo;resistant\u0026rdquo; or \u0026ldquo;susceptible\u0026rdquo; classifications toward predictions that can inform more nuanced, precision-guided therapy. At the same time, we address one of the major challenges in applying MIC-based prediction to real-world data: the severe class imbalance within the CRyPTIC dataset, particularly in cryptic resistance subsets where canonical resistance mutations are absent. By imbalance, we refer to the disproportionate representation of susceptible isolates compared to resistant ones, as well as the uneven distribution of MIC values across the full dilution range. This imbalance is not random but arises because resistant isolates, especially those carrying rare or novel mutations, are much less frequently observed in clinical settings. It is further exacerbated for second-line and third-line drugs, where resistance is comparatively rare and fewer patient samples are available, making it difficult to capture the complete diversity of resistance phenotypes. To achieve these goals, we design and compare a range of machine learning models, including CNNs and XGBoost classifiers\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, two widely used architectures in computational drug resistance research.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eDrug resistance phenotypes\u003c/h2\u003e\u003cp\u003eThe MIC distributions across 12 anti-TB drugs for 12,285 \u003cem\u003eMtb\u003c/em\u003e isolates from the CRyPTIC\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e dataset are summarised. These drugs are INH, RIF, Ethambutol (EMB), Ethionamide (ETH), Injectables (Amikacin (AMI), Kanamycin (KAN)), Fluoroquinolones (FQs; Levofloxacin (LEV), Moxifloxacin (MXF)), BDQ, Clofazimine (CFZ), Delamanid (DLM), and LZD (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Based on ECOFF thresholds, each drug can be grouped into two or three resistance categories. This preliminary analysis highlights substantial class imbalance across most drugs, with full breakdowns of MIC category counts provided (\u003cb\u003eSupplementary Fig.\u0026nbsp;1)\u003c/b\u003e. The dataset includes three first-line drugs: INH, RIF, and EMB, among which INH and RIF display high resistance prevalence at 48% and 38%, respectively. In contrast, EMB shows a greater degree of imbalance, with only 8% of isolates exhibiting resistance. This pattern of imbalance extends across most drugs, with MXF showing the highest non-susceptible proportion at 14%. In comparison, the newer TB drugs, BDQ and DLM, show the lowest resistance rates at 0.9% and 1.5%, respectively. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb illustrates the distribution of MIC values, where the majority of samples over 80% fall into the susceptible or low-resistance range, indicated by high counts near the lower MIC endpoints. Missing values are generally rare across all drugs, though data for BDQ and CFZ show the highest missingness at 221 and 240 samples out of 12,285, respectively (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eThe dataset (n\u0026thinsp;=\u0026thinsp;12,289)\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003eDrug\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMIC median (range)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMIC Imbalance\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSusceptible\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eResistant\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eBinary imbalance\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAMI\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,738\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.25 (0.25-16.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e222.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e10316\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1422\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e7.25\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBDQ\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,743\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.03125 (0.015-2.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e16.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e11400\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e343\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e33.24\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCFZ\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,718\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.125 (0.062-4.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e10003\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1715\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e5.83\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDLM\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,594\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.0078 (0.0078-1.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e2.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e11177\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e417\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e26.80\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eEMB\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,823\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.0 (0.5\u0026ndash;16.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e46.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e8207\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e3616\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e2.27\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eETH\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,800\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.0 (0.5\u0026ndash;16.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e9291\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2509\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e3.70\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eINH\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,736\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.05 (0.025-3.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e19.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e5817\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e5919\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eKAN\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,797\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.0 (1.0\u0026ndash;32.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e57.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e9103\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2694\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e3.38\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLEV\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,829\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.5 (0.125-16.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e21.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e9055\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2774\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e3.26\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLZD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,854\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.5 (0.016-8.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e1.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e10112\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e1742\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e5.80\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMXF\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,860\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.25 (0.125-8.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e38.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e9716\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e2144\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e4.53\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRIF\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e11,769\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.125 (0.031-8.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e4.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e7132\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e4637\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e1.54\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"7\"\u003eAMI: Amikacin, BDQ: Bedaquiline, CFZ: Clofazimine, DLM: Delamanid, EMB: Ethambutol, ETH: Ethionamide, INH: Isoniazid, KAN: Kanamycin, LEV: Levofloxacin, LZD: Linezolid, MXF: Moxifloxacin, RIF: Rifampicin\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eInitial modelling using XGBoost and CNNs trained directly on the imbalanced dataset yielded poor predictive performance, skewed toward susceptible classifications (e.g., AUC\u0026thinsp;=\u0026thinsp;0.5, sensitivity\u0026thinsp;=\u0026thinsp;0, specificity\u0026thinsp;=\u0026thinsp;1). To address this, all subsequent models were trained on an oversampled dataset, where minority MIC categories were boosted to at least 50% of the size of the most represented category, mitigating the skew and enabling more informative learning. All MIC values across drugs are nested and predicted in the same model. Pairwise correlations between MIC values across drugs highlighted KAN and AMI (both injectables) and MXF and LEV (both fluoroquinolones) cross-resistance (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Mono-resistance was rare, with isoniazid being the highest (5%).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eModel Performance\u003c/h3\u003e\n\u003cp\u003eAcross the full drug panel, substantial binary imbalance was observed between resistant and susceptible isolates (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This was most pronounced in BDQ and DLM, where fewer than 2% of samples were resistant. However, this binary skew does not always align with inter-MIC class imbalance. For example, AMI displayed highly uneven MIC category distributions but retained a relatively balanced binary resistance profile. This highlights an added layer of complexity in the dataset, categorical imbalance within MIC bins may not correspond directly to resistance thresholds, complicating classification.\u003c/p\u003e\u003cp\u003eModel comparisons underscore how these imbalances impact performance (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Binary-supported CNNs outperformed conventional CNNs in double dilution accuracy across 8 of 12 drugs, including low-resistance drugs such as BDQ and CFZ, where standard CNNs failed to generalise (0.05 and 0.01 accuracy, respectively). Binary support consistently rescued CNN performance in these cases, suggesting that label-aware sampling plays a critical role in mitigating skew. In contrast, XGBoost had greater stable performance with or without binary support. Gains from binary-aware sampling were modest, such as in BDQ where AUC improved from 0.51 to 0.53. Notably, XGBoost achieved higher AUC than CNNs in 7 of 12 drugs, pointing to its stronger class separation ability. However, in terms of double dilution, i.e., accurate MIC bin prediction, CNNs often had the edge when balanced.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModel performance across binary-supported CNN, conventional CNN, and XGBoost, assessed using Doubling Dilution Accuracy and AUC-OVR (one-vs-rest strategy).\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c5\" namest=\"c2\"\u003e\u003cp\u003eDouble dilution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c9\" namest=\"c6\"\u003e\u003cp\u003eAUC-OVR\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBinary supported CNN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCNN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eXGB\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eBinary supported XGB\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eBinary supported CNN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eCNN\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eXGB\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eBinary supported XGB\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAMI\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBDQ\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.53\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCFZ\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.53\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDLM\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.46\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.55\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eEMB\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.81\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.76\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eETH\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.62\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eINH\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.75\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eKAN\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLEV\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.60\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLZD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.58\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMXF\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.81\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.73\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRIF\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.77\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.82\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.71\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"9\"\u003eAMI: Amikacin, BDQ: Bedaquiline, CFZ: Clofazimine, DLM: Delamanid, EMB: Ethambutol, ETH: Ethionamide, INH: Isoniazid, KAN: Kanamycin, LEV: Levofloxacin, LZD: Linezolid, MXF: Moxifloxacin, RIF: Rifampicin, CNN: Convolutional Neural Network, XGB: Extreme Gradient Boost\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTaken together, these results show that binary-supported CNNs are highly effective when class balance is enforced, offering fine-grained MIC prediction. Yet they remain fragile under imbalance. In contrast, XGBoost is robust under skewed distributions, but may lack the resolution for precise MIC classification. This trade-off between discrete classification and probabilistic ranking is central to model selection in computational DST.\u003c/p\u003e\u003cp\u003eFour drugs are presented (AMI, BDQ, RIF, EMB), and capture a diverse range of resistance profiles and modelling complexities. AMI and BDQ exhibit pronounced MIC-level imbalance, with BDQ also showing extreme binary skew due to its low resistance prevalence. RIF, by contrast, offers a relatively balanced dataset, while EMB sits between the two extremes with moderate imbalance. This selection enables a more comprehensive evaluation of model behaviour under different data distribution scenarios. Both binary-supported CNN and XGBoost models produce similar proportions of correct and incorrect predictions across all four drugs (\u003cb\u003eSupplementary Fig.\u0026nbsp;2\u003c/b\u003e). In highly imbalanced cases such as BDQ, both models misclassify large portions of the data, often redirecting predictions to more populated categories (\u003cb\u003eSupplementary Fig.\u0026nbsp;3\u003c/b\u003e). Interestingly, the nature of these misclassifications differs by model, with CNN and XGBoost tending to err in different directions. Notably, XGBoost fails to predict certain low-frequency MIC bins (e.g., 0.5 and 1), effectively ignoring them during classification. A similar pattern is observed for AMI, another imbalanced case. In contrast, RIF and EMB, which have more even distributions, show prediction errors primarily in minor classes, but all MIC bins remain represented in the model outputs.\u003c/p\u003e\n\u003ch3\u003eSNP feature importance in rifampicin resistance\u003c/h3\u003e\n\u003cp\u003eTo illustrate the interpretability of the binary-supported CNN model, feature ablation was applied to assess the relative contribution of individual SNPs and auxiliary features to RIF resistance prediction (\u003cb\u003eSupplementary Table\u0026nbsp;1\u003c/b\u003e). RIF was selected because its resistance mechanisms are well characterised. Ranked by absolute integrated gradients (IG\u003csub\u003eabs\u003c/sub\u003e), which reflect a feature's influence on the model output, the most dominant signal was the binary drug resistance label predicted by TB-Profiler, vastly exceeding the contributions of individual SNPs. This highlights the strength of combining curated rule-based predictions with learned representations. Among the top 20 genomic variants, the majority fall within the \u003cem\u003erpoA, rpoB\u003c/em\u003e, and \u003cem\u003erpoC\u003c/em\u003e genes, consistent with established mechanisms of rifampicin resistance. However, several variants, including \u003cem\u003erpoC\u003c/em\u003e Leu1245Arg and \u003cem\u003erpoA\u003c/em\u003e Ile3Val, while not currently annotated by WHO, exhibited strong positive IG direction and high ΔAUC, indicating both directional and predictive importance.\u003c/p\u003e\u003cp\u003eInterestingly, canonical rifampicin resistance mutations such as \u003cem\u003erpoB\u003c/em\u003e Ser450Leu were not among the top-ranked features in this model, suggesting they may contribute more to classification certainty rather than to MIC shifts in this regression framework, in this case their influence may have been covered in the binary drug resistance input. Several mutations, including \u003cem\u003erpoB\u003c/em\u003e Val970Met and \u003cem\u003eRv2752c\u003c/em\u003e Gly28Ser, displayed negative IG direction, implying an inverse contribution to resistance probability in the modelled output range. This directional signal reflects the model\u0026rsquo;s interpretation of these variants as associated with lower MIC values, despite their presence in resistant isolates. In contrast, ΔAUC captures the impact of ablation on predictive performance across the test set, revealing variants such as \u003cem\u003erpoA\u003c/em\u003e c.-141C\u0026thinsp;\u0026gt;\u0026thinsp;G and \u003cem\u003erpoC\u003c/em\u003e Val1039Gly as enhancing the discriminative capacity of the model.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThe CRyPTIC dataset offers a valuable foundation for advancing computational drug susceptibility testing in \u003cem\u003eMtb\u003c/em\u003e, but it also highlights limitations and significant challenges. In addition to the binary imbalance between susceptible and resistant isolates, there is substantial skew across the MIC spectrum, with certain MIC categories, particularly for newer and repurposed drugs such as BDQ and DLM, being sparsely represented. This imbalance is most pronounced in cryptic resistance cases, where canonical resistance mutations are absent and phenotypic resistance remains rare. Such distributional bias complicates both training and evaluation, especially for deep learning architectures that depend on balanced examples to generalise effectively.\u003c/p\u003e\u003cp\u003eOur work reframes prediction from a binary classification task to a more granular MIC regression problem, building on binary tools while addressing the imbalance inherent to cryptic subsets. To improve model stability and accuracy under these conditions, we evaluated strategies such as ensemble learning, sample reweighting, and feature dropout. CNNs demonstrated strong performance for fine-grained MIC estimation when guided by these mitigation strategies, benefiting from their capacity to detect complex, non-linear genomic patterns. XGBoost, while less sensitive to imbalance, delivered competitive results with fewer architectural adjustments, suggesting complementary strengths between the two approaches.\u003c/p\u003e\u003cp\u003eThe analysis of feature attribution uncovered several mutations with negative integrated gradient (IG) direction, indicating that their presence systematically lowers the predicted MIC. Notably, mutations such as \u003cem\u003eRv2752c\u003c/em\u003e Gly28Ser and \u003cem\u003erpoC\u003c/em\u003e c.339T\u0026thinsp;\u0026gt;\u0026thinsp;C, despite being enriched in resistant isolates and previously catalogued as drug-resistance-associated, appear to shift model predictions toward susceptibility. This discrepancy suggests these variants may play a role in low drug resistance or tolerance rather than conferring full resistance. Such mutations may permit effective treatment through intensified dosing of first-line agents like rifampicin, potentially reducing reliance on newer drugs with greater toxicity profiles. This raises important clinical implications: better stratification of resistant subtypes could optimise regimen design, enhance adherence, and slow the emergence of resistance to next-generation therapeutics.\u003c/p\u003e\u003cp\u003eOur results highlight two key advances. First, MIC-based modelling can meaningfully extend binary prediction tools, offering richer phenotypic resolution for clinical decision-making. Second, careful handling of data imbalance, particularly in cryptic subsets, can stabilise training and preserve performance in deep learning models. Beyond improving predictive accuracy, integrating interpretable machine learning into this workflow enhances transparency and opens avenues for hypothesis generation, guiding experimental validation of both known and novel variants. Lastly, to fundamentally cope with the data imbalance problem, more clinical samples or lab-generated samples with various levels of MIC need to be generated across different lineages. More generally, increased MIC characterisation combined with WGS will lead to improvements in TB clinical decision-making, thereby improving patient outcomes.\u003c/p\u003e"},{"header":"METHODOLOGY","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eDataset\u003c/h2\u003e\u003cp\u003eThe CRyPTIC dataset\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e encompasses minimum inhibitory concentration (MIC) phenotypes for 12,289 \u003cem\u003eMtb\u003c/em\u003e complex isolates from 23 different countries, spanning lineages L1 to L4, and L6. A total of 6,814 isolates show resistance to at least one drug. Among these, 2,129 (31.2%) isolates met the clinical criteria for being categorised as rifampicin-resistant (RR), multidrug-resistant (MDR), pre-extensively drug-resistant (pre-XDR), or extensively drug-resistant (XDR). The MIC data was gathered using a 96-well broth microdilution plate, covering a spectrum of anti-TB drugs including three primary ones (isoniazid, rifampicin, ethambutol), alongside rifabutin, two fluoroquinolones (moxifloxacin, levofloxacin), two injectables (amikacin, kanamycin), ethionamide, and four newer or repurposed medications (bedaquiline, linezolid, clofazimine, delamanid). The edge case MIC values that signify a boundary were replaced by the halving or doubling the lower or upper edge case MIC values to follow double dilution as shown in \u003cb\u003eSupplementary Table\u0026nbsp;2\u003c/b\u003e. All other edge cases equal or smaller than this value is merged into the stated MIC edge value. The specific cut-offs used are the same as The CRyPTIC consortium publication in 2022 \u003csup\u003e14\u003c/sup\u003e, according to the ECOFF values\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eData Preprocessing\u003c/h3\u003e\n\u003cp\u003eIn this study, to reduce computational load and memory usage, we adopted a streamlined representation of genetic variation by constructing drug-specific resistance vectors for each isolate. For each antibiotic, only SNPs located within gene regions with well-established links to resistance were included. These genes were as follows: \u003cem\u003einhA\u003c/em\u003e, \u003cem\u003ekatG\u003c/em\u003e, \u003cem\u003ekasA\u003c/em\u003e, \u003cem\u003eoxyR-ahpC\u003c/em\u003e, and \u003cem\u003efabG1\u003c/em\u003e for isoniazid; \u003cem\u003eethA\u003c/em\u003e and \u003cem\u003einhA\u003c/em\u003e for ethionamide; \u003cem\u003eembA\u003c/em\u003e, \u003cem\u003eembB\u003c/em\u003e, and \u003cem\u003eembC\u003c/em\u003e for ethambutol; \u003cem\u003erpoA\u003c/em\u003e, \u003cem\u003erpoB\u003c/em\u003e, and \u003cem\u003erpoC\u003c/em\u003e for rifampicin; \u003cem\u003egyrA\u003c/em\u003e and \u003cem\u003egyrB\u003c/em\u003e for levofloxacin and moxifloxacin; \u003cem\u003errs\u003c/em\u003e and \u003cem\u003eeis\u003c/em\u003e for kanamycin and amikacin; \u003cem\u003eRv0678\u003c/em\u003e, \u003cem\u003eatpE\u003c/em\u003e, and \u003cem\u003epepQ\u003c/em\u003e for bedaquiline; \u003cem\u003eRv0678\u003c/em\u003e and \u003cem\u003epepQ\u003c/em\u003e for clofazimine; \u003cem\u003efgd1\u003c/em\u003e, \u003cem\u003efbiA\u003c/em\u003e, \u003cem\u003efbiB\u003c/em\u003e, \u003cem\u003efbiC\u003c/em\u003e, \u003cem\u003efbiD\u003c/em\u003e, and \u003cem\u003eddn\u003c/em\u003e for delamanid; and \u003cem\u003errl\u003c/em\u003e and \u003cem\u003erplC\u003c/em\u003e for linezolid. Each resistance vector was composed of binary indicators representing the presence or absence of SNPs in these loci, derived from TB-Profiler variant calls. This targeted encoding strategy reduced the input dimensionality per model while preserving biological relevance to the resistance phenotype being predicted. To ensure compatibility with batch processing and model input requirements, sequence padding and normalisation were applied where necessary. All datasets were stratified into training, validation, and testing subsets to allow for consistent and rigorous performance evaluation.\u003c/p\u003e\n\u003ch3\u003eCNN model\u003c/h3\u003e\n\u003cp\u003eThe convolutional neural network (CNN) architecture (\u003cb\u003eSupplementary Table\u0026nbsp;3\u003c/b\u003e) used for MIC prediction comprised an initial feature extraction block with 64 filters and kernel size of 25, followed by two stacked convolutional modules with dynamic filter scaling. Each convolutional layer was followed by batch normalisation, Rectified Linear Unit (ReLU) activation, and dropout regularisation at a rate of 0.4. Downstream processing was performed by two dense layers, each with 128 neurons and dropout at a rate of 0.7. A prediction head merged latent features with an additional scalar input, followed by a sequence of fully connected layers and non-linear activations to yield multi-class predictions. Kaiming normal initialisation was applied to all linear layers. Training proceeded for up to 500 epochs. A grid search was performed to optimise model hyperparameters. Oversampling and category weighting were incorporated into a custom cross-entropy loss function to address class imbalance. The Adam optimiser was used to update network weights, with a learning rate fixed at 1x10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e and weight decay (L2 regularisation) set at 1x10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e. For the binary supported CNN architecture, the binary predictions for drug resistance from TB-profiler was fed into the final layer of the CNN. The CNN was developed and implemented using Pytorch 1.12.1 in Python (v3.9.15) with CUDA 11.3. Model training was performed on a Tesla v100-pcie-32gb graphics processing unit (GPU).\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eBenchmarking\u003c/h2\u003e\u003cp\u003eTo benchmark the results, the binary supported CNN model was compared against a normal CNN of the same architecture, as well as XGBoost. Two performance evaluation metrics were calculated, namely the Doubling Dilution Accuracy and the AUC-OVR. Doubling Dilution Accuracy reflects how closely predicted MICs align with actual laboratory-measured values on a two-fold dilution scale, while AUC-OVR quantifies overall classification performance across all MIC levels\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eFeature importance\u003c/h2\u003e\u003cp\u003eTo assess the contribution of individual features to drug resistance prediction, we applied an integrated attribution framework combining model ablation with gradient-based interpretation. Three complementary metrics were computed: delta AUC (ΔAUC), integrated gradients direction (IG\u003csub\u003edir\u003c/sub\u003e), and integrated gradients magnitude (IG\u003csub\u003eabs\u003c/sub\u003e). ΔAUC quantifies the drop in predictive performance when a feature is replaced with baseline values (default: 0). For each SNP or categorical input, we masked the feature across a test batch and measured the resulting change in area under the ROC curve, averaged over multiple batches. A larger positive ΔAUC indicates greater loss of predictive power upon feature ablation.\u003c/p\u003e\u003cp\u003eIn parallel, integrated gradients were computed with respect to each input feature. IG\u003csub\u003edir\u003c/sub\u003e captures the signed attribution, indicating whether a feature increases or decreases the model's prediction for the target class, while IG\u003csub\u003eabs\u003c/sub\u003e reflects the overall magnitude of influence. Together, these metrics provide complementary insights into model behaviour. ΔAUC assesses contribution to class discrimination, IG\u003csub\u003edir\u003c/sub\u003e reveals the direction of predictive influence, and IG\u003csub\u003eabs\u003c/sub\u003e ranks the salience of features regardless of direction.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eETHICS APPROVAL AND CONSENT TO \u0026nbsp;PARTICIPATE\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not involve the collection of new human or animal data. All analyses were conducted on de-identified, publicly available data from the CRyPTIC project (https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3001721), an open-access resource. The CRyPTIC dataset is curated in accordance with ethical standards, including removal of personal identifiers, and is made available for research under the program\u0026rsquo;s data-sharing policies. As no direct patient contact or intervention was performed by the authors, additional institutional ethical approval was not required. This study was conducted in accordance with the ethical principles set out in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONSENT TO PARTICIPATE DECLARATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONSENT FOR PUBLICATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors have read and approved the final manuscript. They consent to the publication of this work and confirm that the content is original and has not been published or submitted for publication elsewhere.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAVAILABILITY OF DATA AND MATERIAL\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot data collection was done. All data used is from online database TB portal. All data is available upon request from TB portal: https://tbportals.niaid.nih.gov/ The code and model can be found in the author\u0026rsquo;s Github: https://github.com/linfeng-wang/TBpt\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot data collection was done. All data used is from online database TB portal. CRyPTIC data used in the study can be found in: \u0026nbsp;\u003c/p\u003e\n\u003cp\u003ehttps://ftp.ebi.ac.uk/pub/databases/cryptic/release_june2022/reuse/\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe code and model can be found in the author\u0026rsquo;s Github: https://github.com/linfeng-wang/tb_dr_MIC3\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOMPETING INTERESTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflicts of interest.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRIBUTIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJEP and LW conceived and directed the project. LW developed the models under the supervision of SC, TGC and JEP. LW wrote the first draft of the manuscript. All authors commented and edited various versions of the draft manuscript and approved the final manuscript. LW, TGC and JEP compiled the final manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;FUNDING\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLW is funded by a BBSRC LIDO studentship (Ref. BB/T008709/1). TGC and SC are funded by the UKRI (BBSRC BB/X018156/1; MRC MR/R020973/1, MRC MR/X005895/1; EPSRC EP/Y018842/1). The funders had no role in the study design, data collection and analysis, the decision to publish, or preparation of the manuscript. The authors declare no conflicts of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGlobal Tuberculosis Report. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/teams/global-programme-on-tuberculosis-and-lung-health/tb-reports/global-tuberculosis-report-2024\u003c/span\u003e\u003cspan address=\"https://www.who.int/teams/global-programme-on-tuberculosis-and-lung-health/tb-reports/global-tuberculosis-report-2024\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVanino E, et al. Update of drug-resistant tuberculosis treatment guidelines: A turning point. Int J Infect Dis. 2023;130:S12\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWHO consolidated guidelines on tuberculosis. Module 4: treatment - drug-resistant tuberculosis treatment, 2022 update. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/publications/i/item/9789240063129\u003c/span\u003e\u003cspan address=\"https://www.who.int/publications/i/item/9789240063129\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePhelan JE, et al. Integrating informatics tools and portable sequencing technology for rapid detection of resistance to anti-tuberculous drugs. Genome Med. 2019;11:1\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePopov S, Kuzmin A, Sabgayda T, Vedenina N. Minimum inhibitory concentrations (MIC) determination of TB drugs and broad-spectrum antibiotics in M.tuberculosis with M/X/TDR. Eur Respir J 46, (2015).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeelder W, et al. Using deep learning to identify recent positive selection in malaria parasite sequence data. Malar J. 2021;20:270.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLibiseller-Egger J, et al. TB-ML\u0026mdash;a framework for comparing machine learning approaches to predict drug resistance of Mycobacterium tuberculosis. Bioinform Adv. 2023;3:vbad040.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang L, Campino S, Phelan J, Clark TG. Mixed infections in genotypic drug-resistant Mycobacterium tuberculosis. Sci Rep. 2023;13:17100.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKrizhevsky A, Sutskever I, Hinton GE. ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems. Volume 25. Curran Associates, Inc.; 2012.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGreen AG et al. A convolutional neural network highlights mutations relevant to antimicrobial resistance in Mycobacterium tuberculosis. \u003cem\u003ebioRxiv\u003c/em\u003e 2021.12.06.471431 (2021) \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1101/2021.12.06.471431\u003c/span\u003e\u003cspan address=\"10.1101/2021.12.06.471431\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen T, Guestrin C, XGBoost:. A Scalable Tree Boosting System. in \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e 785\u0026ndash;794Association for Computing Machinery, New York, NY, USA, (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eConsortium T, Cr. A data compendium associating the genomes of 12,289 Mycobacterium tuberculosis isolates with quantitative resistance phenotypes to 13 antibiotics. PLoS Biol. 2022;20:e3001721.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCRyPTIC Consortium. Epidemiological cut-off values for a 96-well broth microdilution plate for high-throughput research antibiotic susceptibility testing of M. tuberculosis. Eur Respir J. 2022;60:2200239.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eConsortium T, Cr. Quantitative drug susceptibility testing for Mycobacterium tuberculosis using unassembled sequencing data and machine learning. PLoS Comput Biol. 2024;20:e1012260.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-applied-sciences","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Applied Sciences](https://link.springer.com/journal/42452)","snPcode":"42452","submissionUrl":"https://submission.springernature.com/new-submission/42452/3","title":"Discover Applied Sciences","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7621453/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7621453/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTuberculosis (TB), caused by \u003cem\u003eMycobacterium tuberculosis\u003c/em\u003e, remains a major global health concern, with case numbers increasing since 2021. In 2024, over 10\u0026nbsp;million new cases were reported, along with 1.1\u0026nbsp;million deaths\u003csup\u003e1\u003c/sup\u003e. At the same time, the widespread adoption of whole genome sequencing (WGS) has made it possible to predict drug resistance from genetic data, allowing for faster diagnostic workflows. Several tools have been developed to classify isolates as \"resistant\" or \"susceptible\". However, drug resistance is not always binary and often exists along a continuum. Cryptic resistance is one example, where phenotypic resistance occurs in the absence of known resistance mutations. The CRyPTIC dataset, which includes over 12,000 isolates and MIC (minimum inhibitory concentration) measurements across 12 anti-TB drugs, provides a valuable resource for moving beyond binary classification. MIC values quantify the lowest concentration of a drug needed to inhibit bacterial growth, offering a more detailed picture of drug susceptibility. In this study, we build models that extend binary classification to directly predict MIC levels from genomic features. To handle the severe class imbalance in completeness across MIC distributions, especially for newer or second-line drugs, we use oversampling and label-aware training techniques. We compare two modelling approaches: XGBoost, which is well-suited to structured data, and convolutional neural networks (CNNs), which can capture spatial and hierarchical relationships within genomic inputs. XGBoost demonstrated more consistent performance in the presence of imbalance, while CNNs achieved higher resolution when the MIC classes were more evenly distributed. Feature importance analysis revealed that some variants previously thought to cause resistance were linked to lower MIC values, suggesting they may only contribute to low drug resistance instead, where the isolate can be killed with higher doage of the same drug. These insights open the door to more tailored treatment strategies, including the use of higher doses of first-line drugs, which could reduce toxicity, improve patient adherence, and slow the emergence of resistance to newer therapies.\u003c/p\u003e","manuscriptTitle":"Expanding Tuberculosis Drug Resistance Prediction beyond binary: Deep Learning for Minimum Inhibitory Concentration prediction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-10 05:23:26","doi":"10.21203/rs.3.rs-7621453/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-11-12T18:24:11+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-12T07:43:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"276853881746005402766953227360524445735","date":"2025-11-11T16:50:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-11T09:41:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-10T23:54:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"211567176199379839774832436199472166539","date":"2025-11-10T05:40:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"251643187103750010233318256923887913","date":"2025-11-10T05:25:10+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-09T09:50:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"158923388709808759337743904325994155054","date":"2025-11-09T09:25:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-09T08:36:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"45521826654128595028745836250243631499","date":"2025-11-09T07:52:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"62313520836180695864673307726763893925","date":"2025-11-09T07:41:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"223135876325315019016189762715262136518","date":"2025-11-09T07:26:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"222767606599491191130998040495927552878","date":"2025-11-09T07:17:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"58444099208319859878748798884877425254","date":"2025-11-09T06:37:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"320038663139219479564996935269433396612","date":"2025-11-03T03:09:04+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"110619210986950130510930344147856019102","date":"2025-10-31T19:15:39+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"264394407813680149749666691098825869816","date":"2025-10-30T07:05:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"16646097501555619070540051176050322365","date":"2025-10-29T14:36:53+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-28T05:59:12+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-26T04:41:57+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-17T09:58:40+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-17T09:58:32+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Applied Sciences","date":"2025-09-15T13:55:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-applied-sciences","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Applied Sciences](https://link.springer.com/journal/42452)","snPcode":"42452","submissionUrl":"https://submission.springernature.com/new-submission/42452/3","title":"Discover Applied Sciences","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a3ce8da5-0b10-40ba-b08a-1a92968c32f2","owner":[],"postedDate":"November 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-24T12:38:49+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-10 05:23:26","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7621453","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7621453","identity":"rs-7621453","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}