{"paper_id":"2b1728f2-4035-4d9e-a18e-57d985ba70a9","body_text":"Distinct mechanisms drive post-antibiotic \nTuberculosis relapse post-cure versus post-\ntreatment-failure \nAuthors: Christian T. Michael1, Maral Budak1, Philana Ling Lin2, Denise Kirschner1  \n1Kirschner lab, Department of Microbiology & Immunology, University of Michigan - Michigan Medicine, Ann \nArbor, MI, USA  \n2Department of Pediatrics, UPMC Children’s Hospital of Pittsburgh, University of Pittsburgh School of \nMedicine, Pittsburgh, PA, USA \nAbstract \nTuberculosis (TB) remains a global health concern, as Mycobacterium tuberculosis (Mtb) infects a quarter of the \nworld's population. Though many TB patients sterilize infection with treatment  regimens including the current \nstandard, incomplete sterilization leads to post -treatment relapse and development of drug resistance. Two \nmechanisms have been hypothesized as driving relapse: persistence, where treatment kills all replicating Mtb, \nand relapse follows once non-replicating Mtb return to a replicative niche; and threshold, where replicating Mtb \nremain alive, yet  below detectable levels. Relapse is often detected through a combination of clinical and \nbacteriological testing, often clinically described as recurrence of TB <2 years after a \"cure\" diagnosis, while \nmany experimental studies examine relapse ~2-months after treatment completion. Our capacity to untangle \nthese considerations and identify mechanisms driving relapse in vivo are limited. Here, we examine the impact \nof both threshold and persistence mechanisms on relapse  post-treatment completion and post-cure diagnosis \nusing our computational model capturing whole-host Mtb infection dynamics. Simulations show that erroneous \nTB-negative diagnosis post-treatment (false cure) rates are regimen-specific, specifically, the historic standard \nHRZE is more likely to result in false cure than the contemporary regimens RMZE or BPaL. We also identify how \nthreshold-driven or persistence-driven relapse correlates with both pre-treatment bacterial burden and diagnostic \ntests used at treatment completion. Simulations show that post -cure relapse is almost exclusively persistence \ndriven, while threshold-driven relapse is most common without a \"cured\" inclusion criterion . Thus, for patients \nwith negative bacteriological diagnostic results at treatment completion, subsequent relapse may best be \npersonalized by targeting non-replicating Mtb. \nImportance \nIncomplete treatment of TB leads to antibiotic resistance and risks relapse, which may occur years later. \nUnderstanding relapse is methodologically challenging but may allow some cases to shorten the 4-month-long \nrecommended treatment timeframe. Predictors of relapse are not well-defined given variability in technical \ndefinitions of relapse and nuances of study designs. Here, we simulate both clinical and experimental relapse \nstudies, including multiple diagnostic tests and relapse definitions, using biologically-based computation. We find \nthat two hypothesized types of relapses, each potentially requiring a different treatment strategy, are \nsimultaneously at play. Simulations suggest that relapse after \"cure\" diagnosis (most clinical studies) is caused \nby reactivation of non -replicating bacteria hidden from treatment within necrotic granuloma tissue, whereas \nexperiments that classify any live Mtb post -treatment as relapse  are more likely to report relapse caused by \nincomplete sterilization of replicating Mtb. Predictions depend on  the currently unclear relationship between \nbacterial burden and clinical symptoms.  \n1. Introduction \nPulmonary tuberculosis (TB) remains a dire concern for countries across the globe. A staggering one-quarter of \nthe human population is infected with Mycobacterium tuberculosis (Mtb)1, with recently rising rates of infection \ndocumented within the United States 2. Approximately 90% of those infected individuals harbor clinically \nasymptomatic Latent Tuberculosis Infection (LTBI), while the remainder progress to clinically active TB disease3–\n5. Accurate diagnosis and successful treatment of Mtb infection is required to eradicate TB disease worldwide, \nas incomplete treatment and adherence issues have led to a rise in drug-resistant Mtb strains6. The WHO- and \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nCDC-recommended treatments for drug -susceptible active TB are between 4-9 months7,8, and the WHO -\nrecommended treatment for LTBI takes 3-6 months9. Human studies suggest that standard treatment results in \napproximately a 5% relapse  rate (i.e. recurrent TB disease after treatment, see Table 1), a number that increases \nto 20% for patients undergoing short-course treatment10. Still, that more than half of TB patients are successfully \ntreated within 2-3 months11 presents a critical balancing act for public health agencies and clinicians alike. On \none hand, shortening treatment has the potential to reduce the risk of drug-induced side effects, financial and \nhuman public health resources. On the other, incomplete treatment risks  the possibility of  relapse and \ndevelopment of resistance. Confounding this, we lack reliable predictors of post-treatment relapse.  \nTable 1. Definitions of terms used to describe TB assessment and state. We have broadened these definitions to reconcile both \nclinical terminology and experimental terminology, including distinctions where they may not align. This is partially based on Table 1 from \n11. Many of these entries have no good tests or biomarkers support. We use host and patient interchangeably, although we use patient \nto encourage clinical consideration and host as species-agnostic, also applying to animal models. *These definitions are not standard \nterminology but are related to sensitivity and specificity of diagnostic tests; we use these terms because the specific meaning of positive \n(and thus sensitivity/specificity) depends on whether one analyzes cure positivity or Mtb-infection positivity.  \nSterilization  \n(or clearance) \nAn actual host state wherein all Mtb are dead or nonviable. \nDiagnostic test  \n(or diagnostic) \nA test that seeks to classify disease state at a specific time, typically assessing for active TB, LTBI, or cure. \nThe WHO-recommended diagnostic varies based on the age and history of the patient12. Many diagnostics \nalso include tests for drug resistance13. \nCure and Non-cure \n(or Negative diagnosis and \nPositive diagnosis) \nPotential results of diagnostics test that seek to classify a patient as sterile and/or disease-free after a \npreviously Mtb-positive result (not to be confused with sterilization ). Results are occasionally presented \nalongside the general type of assessment used to establish cure, e.g. \"microbiological cure\". We use cure \nterminology after prior infection is established.  \nControlled infection A non-sterile infection state wherein the total Mtb population within a host or granuloma is prevented from \nexpanding.  \nSubclinical infection \n(or subclinical disease/TB) \nA non-sterile infection wherein the patient does not exhibit overt TB disease symptoms yet may shed viable \nMtb. \nLatent TB infection \n(or LTBI) \nA non-sterile infection state assumed to be controlled and non-contagious, typically diagnosed by immune \nor radiological methods4. \nTreatment Completion A milestone within a host's regimen wherein the patient has finished the prescribed regimen, independent \nof cure status. \nTreatment Success A disease state wherein a patient is cure-positive at treatment completion. \nTreatment Failure A disease state wherein a patient is cure-negative at treatment completion. \nRecurrence An instance of diagnosable Mtb infection that had previously been assessed as cured. Recurrence most \noften describes presentation of active TB disease after cure ( recurrent TB, most common ), but may \ntheoretically describe recurrent LTBI (i.e., a situation where a previously-cured patient later is determined \nto have LTBI.) \nRelapse An instance of recurrence after cure (most often and intended11) or treatment completion (less common, \noften due to experimental limitations under the assumption that cure is likely14). Relapse can be defined \nas after either recurrent symptomatic TB (clinical definition) or recurrent LTBI (possible in experiments if \nonly microbiological tests are used and animals do not have human-like symptoms). \nReactivation An instance of active TB after a state of stable LTBI has progressed and infection is no longer controlled15. \nReinfection An instance of Mtb infection resulting from a separate exogenous inoculum after a prior Mtb infection, \ntypically after the prior infection is declared cured. Epidemiological factors can make this difficult to \ndistinguish from relapse, requiring molecular markers from both episodes16.  \nFalse Cure* A cure diagnosis while the patient is not sterilized. \nFalse Positive* A positive or active TB diagnostic test result while the patient is actually sterilized, e.g. a positive tuberculin \nskin test on a Mtb-free BCG-vaccinated individual17. \nRecent studies have revived a deep interest in the importance of understanding relapse for TB drug \ntreatment11,18. A key complicating factor is the presence  of granulomas†. These complex tumor-like structures \nare hallmarks of Mtb infection, and they provide infection-site-specific pharmacokinetic and pharmacodynamic \nbarriers to treatment. Mtb trapped within necrotic cores found in most granulomas (i.e. caseum) maintains a non-\nreplicating metabolic state (see Supplemental Material S1)23, potentially delaying relapse for long tim es. Our \ngoal is to re visit the contexts under which relapse occurs and use comparisons between experimental animal \nmodels, human datasets and computational modeling to elaborate mechanisms driving relapse and their \nrelationships to clinically-detectable TB treatment outcomes. \n1.1 Relapse \n \n†Note that we refer to lung granulomas unless otherwise specified, though granulomas commonly form in thoracic \nlymph nodes19, likely impacting adaptive immune priming20,21; granulomas may also form during extrapulmonary \nTB—e.g., liver22. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nClinically, active TB is typically associated with signs (cavitary lesions on chest x-rays, fever, etc.) and symptoms \n(cough, fatigue) of TB , often with microbiologic confirmation of Mtb  infection. During treatment, signs and \nsymptoms often improve (improved chest x-ray, sputum smear and culture are Mtb-negative, resolution of cough \nand weight loss).  Clinical r elapse refers to having new signs and symptoms of TB after resolution during \ntreatment when exogenous Mtb re -infection has been ruled out 11. Table 1 contains detailed terminology for \ndescribing TB disease state in the context of treatment. \nA recent review11 theorized two mutually-compatible mechanisms (referred to there as concepts) that may lead \nto relapse: persistence and threshold. Persistence suggests that TB treatment leaves a reservoir of non-\nreplicating Mtb alive in caseum, transitioning the host into a state of LTBI; and/or a Threshold where TB treatment \nleaves a few replicating Mtb alive that are below a level of detection (LOD) until treatment stops and the bacterial \nniche repopulates. These mechanisms are challenging to separate experimentally and clinically11, so there is no \ndefinitive answer as to which drives relapse.  \nIf indeed both mechanisms are at play,  it is not clear how persistence - and threshold-driven relapse differ in \nterms of optimal diagnosis or treatment. Relapse is a phenomenological  term, typically reported using a \ncomposite of one or more measures including positive bacterial smears or cultures, rapid-test molecular analysis, \nradiographic exams, or clinical TB symptom presentation11,18,24–26. Each test for TB cure has advantages and \ndisadvantages (Table 2 and 11), although none can categorically predict relapse. Using cynomolgus macaques, \na non-human primate (NHP) model of human-like LTBI27, it was found that measuring Bronchoalveolar Lavage \nFluid (BALF) for CFU reduction over time was insufficient for assessing effective antibiotic treatment, particularly \nduring short-term treatment28.  \nTable 2: Common methods for assessing M. tuberculosis infection and/or TB disease in vivo.  \nDiagnostic Advantages Drawbacks CFU Level of Detection (LOD)  References \nSolid or liquid culture of \nBAL \nGold standard for positive \ndiagnosis when repeated \nCan take weeks to months, \nrequires 2+ tests. Poor \npredictor of relapse during \nearly infection \n1 CFU within BALF or sputum \n(ideally) \n11,29,30 \nWHO-approved rapid \ndiagnostics (e.g., Xpert \nultra) \nFast, accurate Expensive, can bear false \npositive after previous \ninfection \n16 CFU/mL (BALF or sputum) 31,32 \nAcid-fast sputum or BALF \nsmear \nFast, inexpensive Yields false positives after \nprevious infection, false-\npositive rate may depend \non treatment regimen \n~5,000 CFU/mL (BALF or sputum) 11,33,34 \nWhole-lung or tissue \nhomogenate \nHighly accurate, detects \ninfections during active TB \nor LTBI \nAnimal model only \n(Performed post-necropsy) \n~5CFU/sample 18(NHP); 35,36(Mouse) \nRadiographic tests (e.g., \nchecking for dissemination \nusing a 18F-\nfluorodeoxyglucose radio-\ntagged PET/CT probe) \nNon-invasive, spatially \nresolves individual \ngranuloma disseminations, \ncorrelates with Mtb burden, \nlongitudinal data available \nMeasures inflammation \ntypically associated with \nbacterial burden but can be \nhost response, which may \npersist post-cure; unknown \nprecise Mtb LOD \n- Not a direct measure of CFU, but \nFDG avidity from PET/CT is \nproportional to CFU37. \n \n(Returns a measure of inflammation \nin host) \n18,37 (NHP); 38(Human) \nImmune assays Some, like tuberculin skin \ntests (TSTs) or interferon \ngamma release assays \n(IGRAs), only measure Mtb \ninfection. Standard \ndiagnostic tests include \nTSTs and QuantiFERON-\nTB Gold In-Tube (an \nIGRA).  \nUnclear as to how long \ninflammation lasts after \nMtb are cleared, not \nalways able to distinguish \nactive TB from LTBI \nLow quality evidence for \nthe ability to detect LTBI \n- Not based on CFU \n \n(Returns a positive/negative based \non sufficient presence/absence of \nTB-associated immune markers) \n9,17,39,40 \nRecording clinical TB \nindicators (cough, weight \nloss, etc.) \nNoninvasive. Motivates the \npatient / care team to test \nfor TB that may be \notherwise missed \nInsufficient to diagnose TB \nby themselves \nUnclear connection to CFU \n \n(Gives clinical picture not linked to \ninfection measures) \n8 \n \n1.2 Experimental Models of relapse \nMultiple in vivo animal and human studies are employed  to study TB outcomes41 including relapse, notably \nincluding NHP18,27,42 and mouse43–46 models. Relapse mouse models (RMMs) can be readily treated with human-\nequivalent doses (i.e., corresponding average plasma concentrations 47,48) of first and second -line antibiotics. \nHowever, most mice do not form necrotic granulomas  with the exception of C3HeB/FeJ mice 35,46,49, and mice \nhave disease states incongruent from those in humans (e.g., no true latent infection). NHPs most closely model \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nhuman TB, but require expensive resources and expertise50. Animal models have been given (intentionally) weak \nantibiotic regimens to allow for reliable and experimentally -controllable relapse 18, though this  experimental \nformulation of relapse may be subsequent to treatment completion rather than cure (Figure 1 ). Moreover, \nassessing cure in animal models is not always straightforward —repeated sampling of CFU in RMMs is not \nconsistently possible, and such limitations may require testing for relapse post treatment-completion rather than \npost-cure14. Contrasting this, clinical trials are the gold standard for analyzing treatment regimens, but can take \nyears to complete and cost an average of $40 M51. Testing drug regimens at high risk of TB relapse in humans \nwould be unethical and relapse events in clinical trial datasets are rare and are reported as composites of multiple \ndetection methods18,24–26. All clinical relapse studies we examined considered relapse to be post -cure, though \nrelapse was occasionally reported as grouped with treatment failure as \"unfavorable outcomes\" . Additionally, \ndiagnostic tests used to identify relapse in individual patients was not always reported.  \n \nFigure 1: Classifications of TB recurrence and relapse. This flowchart shows the connection of all host states and tests listed in \nTable 2. Note that this shows three key confounding factors frequently addressed in relapse research. First an overload in discussing \ntreatment success or failure: in clinical trials, conclusions about an actual sterilizing population (third row) must be based on cure \nreports (fourth row), and \"treatment success/failure\" is used interchangeably to describe rows 3-5. Second is that clinical studies define \nrelapse to occur after cure, (bottom row, solid arrows) though some studies implicitly include relap se as active TB after treatment \ncompletion and/or failure if diagnostic tests are not run upon treatment-completion (bottom row, dashed arrows). Third is the complexity \nof having interchangeable diagnostic tests at initial diagnosis, treatment completion, and during relapse follow-up. This classification \nassumes no reinfection (Table 1). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n \nA confounding factor when studying relapse is that there is no directly confirmed connection between the state \nof a host’s lung infection and their clinical symptoms. Even if one knew the precise number of bacteria and their \nmetabolic state at every timepoint within lung s, we would not know whether culturable or symptomatic relapse \nwill result. It is possible for asymptomatic hosts to shed viable Mtb  (subclinical disease), observed in ~5% of \nnonhuman primates (NHPs), (referred to as percolators52–54) as have been described now in humans 55. Model \nhosts have also been observed as having levels of inflammation markers such as interferon gamma (IFN -γ), \nwhich cannot directly predict host outcomes56, highlighting this complication. Previous studies found 10-25% of \nLTBI patients exhibited neutrophilic whole -blood transcriptomic signatures consistent with active TB 4,57. \nConsistent with this, clinical studies have shown multiple inflammatory markers in patients with LTBI17. Together, \nthis suggests  some LTBI cases involve an ongoing immune siege  to contain metabolically-active Mtb from \nescaping granulomas, while other cases involve surveillance over non-replicating, caseum-trapped Mtb.  \nThere is conflicting terminology between experimental and clinical studies of relapse (Table 1), and it is unclear \nwhether the persistence or threshold mechanisms equivalently underpin distinct notions of relapse. Challenges \nto experimentally reproducing relapse in vivo (e.g., species differing in symptom presentation)  necessitate the \nuse of alternative measures than diagnostics used for humans. For instance, CFU enumeration in mouse lung \nhomogenates or NHP granulomas is used as one proxy for clinical relapse (Table 2), even though the detection \nof small or caseum-bound Mtb populations in such samples may resemble reactivation from latency or recurrent \nLTBI rather than symptomatic clinical relapse. \nTo examine mechanisms underpinning various measures of relapse, we extend our virtual host model HostSim \n(see Methods and Supplemental Material S2) to include both persistence and threshold mechanisms as well \nas virtual diagnostic  tests that mimic in vivo diagnostics. We use HostSim to reproduce both clinical and \nexperimental relapse studies. Our simulations suggest that treatment failure leaves a reservoir of replicating \nbacteria, leading to rapid threshold-driven relapse (Figure 1). On the other hand, post-cure relapse is primarily \npersistence-based, i.e., driven by a continuous release of non-replicating bacteria derived from caseum, typically \ntaking much longer between treatment completion and relapse. Simulation results depend on levels of detection \nof diagnostic tests, many of which are only estimates due to a current lack of understanding of how lung CFU \nrelates to sputum/BALF CFU. \n2. Results  \nTo predict drivers of relapse, we use our recently -calibrated and validated whole-host TB model, HostSim58–60. \nBriefly, HostSim is a hybrid agent-based, ordinary differential equation model that represents interaction between \nhost immune cell populations in lungs , uninfected lymph nodes , and blood with Mtb subpopulations within \nmultiple lung granulomas of a virtual host. HostSim allows us to simulate temporal trajectories of various T-cell \nand macrophage population sizes, as well as replicating and non-replicating Mtb within each granuloma of each \nhost (see Methods) and their total in lungs . Recently, we used HostSim to compare multiple measurements of \ndrug efficacy of various first- and second-line antibiotic regimens58.  \nFor all analyses in this paper, we create a single virtual cohort of 500 Mtb-infected hosts, all inoculated with a \nsingle Mtb at 𝑡𝑖𝑚𝑒 = 0. We allow each virtual patient to develop a mature Mtb infection until 300 days p.i. then \ntake a snapshot of  its simulation configuration. From those snapshots, we can use one virtual cohort to study \nmultiple what-if scenarios wherein we treat the entire cohort with one antibiotic regimen, halt treatment, then \nobserve regimen-specific host and cohort outcomes.  \nTo compare our virtual predictions to both animal and human datasets, we develop a set of virtual TB diagnostics \nthat mimic those currently used (Table 2; Methods). Three key virtual diagnostics are Virtual Mtb plate, Virtual \nPET/CT, and Virtual clinic score (Table 3). Given these, we simulate relapse given choices of regimen, diagnostic \ntests, and follow-up times (Figure 1). For indicated virtual relapse studies, we exclude virtual hosts that are cure-\nnegative at treatment completion. For all virtual relapse stud ies, we define relapse as a non -cure result at a \nfollow-up time Δ𝑡 after treatment completion.  We classify each virtual relapse as persistence-based if, at time of \ntreatment completion, there are 0 replicating CFU within the virtual host and threshold-based otherwise.  \n \nTable 3: Virtual diagnosis overview. Each virtual diagnostic test has free parameters adjustable to mimic one or more \nreal diagnostics. See Methods for details. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nDiagnostic Test \nname \nBrief description Free parameters in their calculation  \nVirtual Mtb plate Evaluating for presence of Mtb derived from a patient sample . \nPatient sample can be derived from a lung homogenate culture, \nsmear, etc. Descriptions of how we configure Virtual Mtb plate  to \nmimic various tests are in Methods.  \nBinary - returns 1 (not cured) or 0 (cured).  \n- LOD  \n- Ratio of CFU in sample to CFU in lung \n(or granuloma) \n- Whether non -replicating Mtb may be \npresent in virtual sample \n- Whether or not recently -dead Mtb may \nbe counted towards CFU levels  \nVirtual PET/CT Examining numbers of metabolically -active immune cells in \ngranulomas allows for the prediction of 18F -fluorodeoxyglucose \n(FDG) avidity (output of PET/CT scans 61). Sufficiently increasing \nFDG avidity indicates a positive diagnosis.  \nBinary - returns 1 (Positive Increase) or 0 (No change). \n- LOD \n- Fold-increase of FDG -avidity required \nfor a positive diagnosis \n- Predicted relative contribution of \nmetabolically-active immune cells toward \nFDG avidity (Methods). \nVirtual clinic score Implemented as a combination of the above  diagnostics. We \nassume that if a virtual host's CFU increase is uncontrolled or if a \ngranuloma disseminates, they will have symptoms and receive a \nclinical score.  \nBinary - Returns 1 (Either or both of the above are both positive, \nlikely symptomatic), or 0 (Neither of the above are positive, likely \nasymptomatic). \n- All of the parameters of Virtual Mtb Plate \nand Virtual PET/CT \n \n2.1 Rates of false cure differ between regimens  \nAn essential first step to  model any notion of relapse is to capture false cure, i.e. when a non -sterile host is \ndiagnosed as cured at treatment completion (Table 1 and Figure 1). For this analysis, we compare true virtual \nhost sterilization against Virtual Mtb plate results, which mimic experimental lung and granuloma homogenate \ncultures capable of detecting non-replicating bacteria (see Methods and Table 2). We assume there is a LOD of \n50 CFU when analyzing the  whole-host scale and 10 CFU at the granuloma scale (Methods and 58,62). As \npreviously58, we also delineate between a host that is initially-high-CFU (>total 10,000 CFU pre-treatment) and \ninitially-low-CFU (<total 10,000 CFU pre-treatment), and a granuloma that is initially-high-CFU (>1,000 CFU pre-\ntreatment) and initially-low-CFU (<1,000 CFU pre-treatment). It should be noted that high CFU hosts may also \nhave some low CFU granulomas but must have at least one high-CFU granuloma. \nWhen examining 2 -month treatment regimens (group (i) in Methods), w e observe a regimen-dependent \nfrequency of false cure at treatment completion at both host and granuloma scales (Figure 2Error! Reference \nsource not found.). Our initial observations are not surprising in that most short-course mono-treatments for \ninitially-low-CFU hosts results in large rates of false cure and have little impact on initially-high-CFU hosts. Next, \nwe observe that large granuloma-scale false cure rates amplify into even larger rates at the host scale, as \npreviously predicted63. Rates of false cure are increased in initially-high-CFU hosts when compared to initially-\nlow-CFU hosts, particularly after multi-drug regimens, suggesting higher pre -treatment CFU counts are more \nlikely to lead to post-cure relapse. This notably includes HRZE—a first-line standard of care for treating drug -\nsusceptible TB8. As HostSim is calibrated to recreate treatment of drug -susceptible TB58, this suggests that \ntreating high CFU-burden patients with only 8 weeks of HRZE will have a higher risk of false cure (42% sterile \nvs 76% cured) as compared with regimens including Moxifloxacin or Bedaquiline (e.g., BPaL with 74% sterile, \n80% cured and RMZE with 80% sterile and 96% cured). This is consistent with the ~40% relapse rates found for \nshort-course HRZE treatment in a 2016-2018 clinical study11,64–66.  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n \nFigure 2. Cure versus sterilization of virtual hosts and granulomas after short -course antibiotic regimens. \nThe virtual cohort containing 500 virtual hosts (shown in top row) and a total of 6500 primary granulomas (shown in \nbottom row) are each administered a drug regimen for 8 weeks (regimens group (i), see Methods). For each regimen, \nthe percentages for sterilization (black boxes) and lack of detectable infection (green boxes) after 8 weeks of \ntreatment are shown. The left column shows treatment results f or each of these virtual hosts and granulomas, \nwhereas the center  (right) columns show results for hosts and granulomas that were low -CFU (high-CFU) pre-\ntreatment. We notice that higher-CFU granulomas and hosts are more liable to have subclinical infection (i.e., non-\nsterile but with CFU < 10 (granuloma) or 50 (host)). \n \nWe examine the spatial locations of surviving Mtb within false-cured hosts and their non-sterilized granulomas. \nThe location of CFU within granulomas of false -cured hosts depend s on their pre -treatment CFU levels, \nindependent of drug regimen. In initially-low-CFU false-cured hosts (middle column of Figure 2), almost all \nremaining CFU are trapped within caseum. By contrast, in  initially-high-CFU false-cured hosts (right column of \nFigure 2), remaining Mtb are split spatially between trapped in macrophages and caseum. After HRZE treatment, \n7/500 virtual hosts are false-cured, and 6 of those 7 are initially -high-CFU. Within the false-cured initially-low-\nCFU host, nearly all CFU were caseum-bound and non-replicating. One of the six false-cured initially-high-CFU \nhosts held Mtb intracellularly (inside macrophages), while the others held Mtb within caseum. Pooling HRZE-\ntreated granulomas from all hosts, we observe t hirteen virtual granulomas had CFU below LOD —two initially-\nlow-CFU and eleven initially-high-CFU. In those initially-low-CFU granulomas, all undetectable Mtb are trapped \nwithin caseum. Of the eleven high-CFU granulomas, eight harbor Mtb solely within caseum while the other three \nretain Mtb within macrophages at treatment completion. After treating the same virtual cohort with a second-line \nregimen, BPaL44,67, only one initially-high-CFU host holds live CFU below LOD at treatment completion, with all \nbacteria held entirely within the caseum of an initially-low-CFU granuloma at treatment completion. Overall, these \nresults suggest that when using tests that can detect Mtb  within caseum, (i) when clinical relapse occur s in \ninitially-low-CFU hosts, it will likely be persistence -based; and (ii) of relapsing hosts, higher CFU levels pre-\ntreatment are associated with greater chances of  threshold-based relapse. Note that these results interrogate \nthe relative likelihood of types of relapse, should it occur, rather than predicting relapse frequency itself events; \nthe NHP study showed that pre-treatment radiographic features did not predict relapse18.  \nTogether, these results imply that infection persists in virtual hosts after treatment completion in the form of both \nfalse cures and treatment failures  (Table 1). Moreover, we have shown that in false -cure cases, Mtb is more \nlikely to be trapped within caseum if the host had low CFU levels prior to treatment , and if a false -cure finds \nreplicating Mtb, such a host had high-CFU prior to treatment. This suggests that if we group relapsing hosts by \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nwhether their relapse was persistence-driven versus threshold-driven, those cohorts likely had distinct infection \nstates prior to treatment. \n2.2 Cure status impacts mechanisms driving virtual relapse in Virtual PET/CT   \nOur current version of HostSim captures both persistence and threshold mechanisms that we expect to \ncontribute to post-treatment relapse. First, persistence is a fine-grained representation of non-replicating bacteria \nstochastically transported out of caseum over time, likely by neutrophils68. Such events allow for the possibility \nof granulomas that harbor CFU within caseum to experience persistence-driven relapse and/or to seed a new \ngranuloma68. We calibrate rates of caseum-transport to capture Mtb reactivation of virtual hosts with LTBI during \nanti-TNF treatment such as etanercept or infliximab69,70 (Methods). Secondly, the threshold mechanism is at play \nwhen small numbers of live intracellular Mtb can replicate after the end of antibiotic treatment. The threshold \nmechanism is already present in our model in that intracellular growth of Mtb after treatment is allowable. \nTypically, to validate a computational model of, say, relapse, we would generate several virtual hosts and recreate \nrelapse using the two proposed mechanisms above. However, our previous result tells us that the host state \nprior to treatment must be carefully considered as it likely affect s presentation of relapse. Towards defining an \nappropriate validation dataset, we examine infection state s of Mtb hosts from a recent experimental relapse \nstudy that employed the use of  simian immunodeficiency virus (SIV)18. There, NHPs were treated with short -\ncourse antibiotic regimens for active-TB only monkeys (8 weeks of isoniazid and rifampicin, i.e. HR). After a one-\nmonth rest, they were then infected with SIV to induce relapse5. Relapse was primarily measured via observation \nof newly seeded granulomas using radiolabeled PET/CT scans; the study reported that 8/12 NHPs relapsed. To \nreplicate this experiment using HostSim, we represent HIV-1 co-infection as a linear decline of CD4+ T-cells \ncalibrated to T-cell blood concentrations reported from the 1990s71,72 measured prior to any treatment studies \n(see Supplemental Material S3). We assume HIV-1 and SIV virtual T-cell depletion behaves similarly73–76,‡. We \nfind that many studies report relapse only as a recurrence of active TB disease , possibly with bacteriological \nfollow-up10,36,77–79, though detection of non-symptomatic infection after treatment is sometimes  reported or \ndiscussed as relapse18,25 (See bottom row of Figure 1).  \nSome relapse studies only admit subjects with specific pre-treatment disease states (e.g., active TB), so we \nsimilarly predict disease states for virtual hosts. We classify virtual hosts as having active TB if continued bacterial \ngrowth is sufficient to cause active disease, which we characterize by a species-specific fold-increase in CFU \nbetween treatment completion and follow -up (Methods). In this way, we adjust our classification framework to \ntreat virtual hosts as analogous to humans or NHPs for purposes of disease state prediction. To ensure results \nare robust to different characterizations of virtual active TB, we also consider virtual hosts with granulomas that \nindividually maintain CFU>1000 for over a month as having active TB.  \nVirtual disease state classification allows us to examine relapse rates within four subcohorts of our original 𝑛 =\n500 virtual hosts, defined by human-like or NHP -like disease state criteria and active/LTBI inclusion criteria, \n(Table 4). First, we simulate 2 months of HR treatment, at which point we use virtual diagnostics to assess for \ncure (Virtual Clinic score, see Table 3 and Methods), labeling each host as cured or treatment failure at treatment \ncompletion. After 4 weeks of drug rest, we simulate virtual SIV infection for a further eight weeks and then assess \neach virtual host in the full cohort for relapse using Virtual PET/CT (Table 3 and Methods) to mimic PET/CT -\nbased relapse NHP observations18 (Table 4). We find that the reported percentage of relapsing virtual hosts \nheavily depends on treatment success or failure (Figure 3). If we only search for relapse in hosts with active TB \nand reported treatment failure, relapse is nearly guaranteed (>90%). However, approximately 30% of cured hosts \nrelapse within 8 weeks (Figure 3, green curves). The majority of post-cure relapse appears as persistence-driven \nrelapse, mostly exhibiting  slowly-increasing levels of intracellular bacteria , consequent to persistence. These \nrelapses tend to have a low, stochastically -oscillating level of Mtb  (as bacteria move between niches  which is \ntoo fine-grained for experimental detection ). Consistent with our previous result , more than half of simulated \nrelapse subsequent to treatment failure results from  incomplete sterilization of macrophages (Figure 3 , red \ncurves), indicating threshold -based relapse. Some of these relapses resul t in rapid bacterial regrowth in  1-2-\nweek periods that restore CFU to pre -treatment levels. Others exhibit either slow or no regrowth of Mtb \npopulations.  \n \n‡ Note that (1)  HIV-1 and SIV are not identical - e.g., SIV rarely causing AIDS-like symptoms in natural hosts73;  \nand (2) there is evidence that SIV-induced relapse is not identical to CD4+ T cell depletion-induced relapse74. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nTable 4. Predicted radiographic relapse rates by subcohort design. Full virtual cohort of 500 are analyzed for relapse using virtual PET/CT testing  \n(see Methods). Relapse rates depend on which virtual hosts are analyzed in a subcohort of interest, analogous to study eligib ility criteria based on \nexperimental design. In all rows, we assume active hosts are part of the cohort.  Each row shows the relapse rates of virtual hosts by subcohort along \nwith details of how many of those virtual hosts relapse post-TC, post-TF, or post-cure; and whether those relapses are persistence or threshold-based. \nBolded cells indicate simulated results that are the most directly comparable to those from a recent NHP relapse experiment 18. SC: Subcohort. TC: \nTreatment completion. C: Cured. TF: Treatment failed. P: Persisten t relapse. T: Threshold relapse. \nAre LTBI hosts  \nincluded? \nDoes {CFU > \n1000} also \nmean active? \n# hosts that \nreached TC \n(full SC) \n[% relapsed  \nat 8 weeks] \nProportion of \nP vs T relapse  \n(full SC) \n%P | %T \n# C at TC in SC \n[% relapsed  \nat 8 weeks] \nProportion of \nP vs T relapse  \n(post-cure) \n%P | %T \n# TF in SC \n[% relapsed  \nat 8 weeks] \nProportion of \nP vs T relapse  \n(post-TF) \n%P | %T \nNHP-like classification of active/LTBI state \n❌ \n ❌ 100 [50%] 16% P | 84% T  68/100 [29%] 40% P | 60% T 32/100 [93%] 0% P | 100% T \n❌ \n ✅ 109 [51%] 20% P | 80% T 75/109 [32%] 46% P | 54% T 34/109 [94%]  0% P | 100% T \nHuman-like classification of active/LTBI state \n❌ \n ❌ 47 [89%] 5% P | 95% T 15/47 [80%] 17% P | 83% T 32/47 [94%] 0% P | 100% T \n❌ \n ✅ 63 [82%] 13% P | 87% T 29/63 [69%] 35% P | 65% T 34/63 [94%] 0% P | 100% T \nPooling LTBI + active TB (independent of classification -type) \n✅ -\t 483 [17%] 34% P | 66% T 447/483 [11%] 56% P | 44% T 36/483 [88%] 0% P | 100% T \n \nFigure 3. Relapse of virtual hosts after short course of HR and HIV-1 co-infection as in 18. Virtual host CFU levels over \ntime for whole-lung total CFU (A), replicating (intracellular + replicating extracellular) Mtb only (B), and non-replicating Mtb only \n(C). Each line corresponds to a trajectory from one host that was categorized as having active TB prior to receiving HR then \nrelapsing after the end of treatment. Key time points in days post-infection are shown, indicating the start of treatment, the end \ntime of treatment and Virtual clinic score diagnostic (Dx1), the onset of HIV, and the time of relapse follow-up diagnostic Virtual \nPET/CT (Dx2). Each curve is colored based on whether the host was assessed as cured at Dx1 (green) or treatment -failed \n(red) at Dx1. The Virtual Clinic score assumes that non-replicating bacteria are not detectable, and the LOD of 50 is shown for \nreplicating CFU.  \nThese results confirm expectations set by the previous section  when considering Virtual PET/CT. Post-cure \nhosts that relapse are generally low CFU prior to treatment and had primarily persistence-driven relapse, while \nrelapse subsequent to treatment failure appears to be threshold-driven.  \n2.3 Cure status of cohort impacts relapse rates in multiple study designs \nWe next examine whether our previous observations are preserved across differently-structured relapse studies. \nFor this, we recreate various in vivo relapse studies, using virtual diagnostics and subcohorts analogous to those \nreported in vivo. In Table 5, we simulate relapse rates analogous to several studies that (1) report either relapse \nrates or percent non-cure after standard of care using at least one non-composite diagnostic measure; (2) had \nvaried cohort inclusion criteria and species (including humans, NHPs, and RMMs); and (3) sufficiently indicate \nwhether reported relapse was post -cure or post -treatment-completion. Studies mimicked in Table 5 analyze \nrelapse following many more regimens and timepoints than those recreated here. HostSim captures the trend \nthat longer treatment periods are less likely to relapse. In all cases, relapse rates are far higher when including \ntreatment failure in the cohort assessed for relapse.  \nTable 5: Relapse rates from virtual relapse studies that mimic in vivo studies. Columns describe several criteria used to configure \nvirtual relapse studies to match an analogous study from literature, with particular emphasis on the selection of the virtual host subcohort \nfor relapse analysis. The full virtual cohort of 𝑛 = 500 hosts are administered antibiotics from Regimen set (ii) (Methods for duration and \ndosage) and then are simulated without-treatment for a follow -up time before relapse assessment. SC: Subcohort. HC: Homogenate \nCulture. HS: Host Scale. GS: Granuloma scale. LOD: Level of Detection.  \nA – We assume that the same diagnostic test is used for both cure and relapse assessment (Methods).  \nB – Ansatz.  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nC – This LOD is the level of detection of CFU for granulomas in HostSim, though these reference values come from the CFU/mL in \nsputum.  \nD – This did not measure “relapse,” but rather was “still culture positive during continuing treatment” – but was still discussed as potential \nfor relapse in literature11.  \nE – This is a fold-increase of FDG avidity from the time of treatment end, as PET/CT does not directly measure CFU.  \nF – This is a simplified percentage of 12/99 hosts with recurrent disease post-treatment; actual relationship between PET/CT inflammation \nand Mtb sterilization is discussed in the original study 77.  \nG – We also include virtual hosts with LTBI, whereas the analogous study includes patients with recurrent active TB after maintaining a \nlong, stable cure state.  \nH – Actual regimens unknown; we assume that these followed the standard of care.  \nI – This work did not measure relapse with CFU; this percentage reflects what percentage of NHPs had detectable lung CFU at time of \nassessment out of all NHPs whose CFU could be counted (See original study 18 Fig 3C).  \nJ - These ranges came from experiments using different RMM species, each repeated twice. \nVirtual Relapse Study Analogous \nin vivo study Criteria determining subcohort for relapse analysis  Virtual experiment Outcome \nIs SC \nrestricted to \ncure at \ntreatment \ncompletion? \nReplicated \nin vivo \ndiagnostic \ntest A \n𝐋𝐎𝐃 in # \nCFU &  \ntest \nscale \nDoes test \ndetect \nnon-\nreplicating \nMtb? \nDoes test \ndetect \nrecently-\ndead \nMtb? \nSpecies-\nparameter \nused for \nclassifying \nactive TB \nSC \nSize \nTreatment \nregimen \n \nFollow-up \ntime post \ntreatment-\ncompletion \n(𝚫𝒕) \nVirtual  \nRelapse % \nRelapse % Species Study \n \n❌ \nSputum  \nculture \nconversion \n100\nB,C\n, \nHS \n❌ \n ❌ Human 63 \n \n2HRZE + \n4HR \n14 days D 88% 60% \nHuman 64–66 28 days  D 66% 40% \n60 days  D 22% 10% \n✅ \nPET/CT \ninflammation  \nrecurrence \n1.2xE, \nGS \n ❌\nB\n \n ❌ Human 51 2HRZE + \n4HR 6 months 43% 12% F Human 77 \n✅\n G\n Rapid test 16C, HS \n ❌ \n ✅ Human 378 2HRZE  + \n4HR H 2 years 16% 2.2% Human 80G \n❌ \nGranuloma  \nHomogenate \nCulture \n10, GS \n ✅ \n ❌ NHP 109 \nHR with \ncomorbid \nSIV \n8 weeks 100% 72%I Macaque 18 \n❌ \nLung \nHomogenate \nCulture \n10, HS \n ✅ \n ❌ Mouse 104 2RMZ + \n2RM \n3 months 46% 84% Outbred \nSwiss \nMouse \n36 4 months 50% 42% \n❌ \nLung \nHomogenate \nculture \n10, HS \n ✅ \n ❌ Mouse 105 2HRZ+ \n1HR \n3 months 74% 87% BALB/c \nMouse \n45 4 months 75% 5% \n❌ \nLung \nHomogenate \nCulture \n10, HS \n ✅ \n ❌ Mouse 104 2HRZE+ \n2HR 3 months 53% 13-27% \n[0-7%] J \nC3HeB/FeJ \n[BALB/c] \nMouse \n46 \n❌ \nLung \nHomogenate \nCulture \n10, HS \n ✅ \n ❌ Mouse 104 2RMZE+ \n1RM 3 months 41% 20-60% \n[0-20%] J \nC3HeB/FeJ \n[BALB/c] \nMouse \n46 \n❌ \nLung \nHomogenate \nCulture \n10, HS \n ✅ \n ❌ Mouse 104 1BPaMZ+1\nBPaM 3 months 48% 7%  BALB/c \nMouse \n44 \n \nLastly, we want to quantify whether reported persistence or threshold-based relapse is affected by sensitivity of \ndiagnostic test used. For this, we examine relapse 1 year after completing a 2-month short-course treatment of \nHRZE. We assess cure of virtual hosts both at time of treatment completion and one year later by using Virtual \nMtb plate (Methods), varying the LOD of both initial and follow-up tests from 0 to 100 CFU. We also vary whether \nor not these diagnostics c an detect non-replicating CFU. Figure 4 shows  that extremely sensitive tests are \nunlikely to classify patients as cured, especially if they detect non -replicating bacteria. As tests become less \nsensitive, more hosts diagnose  as cured, and cured hosts are unlikely to relapse.  More than half of p ost-cure \nrelapses are persistence-based. All relapses are more frequently threshold-based if the diagnostic tests do not \ndetect non-replicating bacteria. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n \nFigure 4: Persistence versus threshold relapse rates by test LOD. (Top row) Each plot shows, for different \nassumed LODs for Virtual Mtb plates, percentage of active TB hosts and relapse events as well as whether \nthose relapse events are persistence-based or threshold-based. (Second and third rows) Portion of the active \nTB cohort that were cured (middle row) or failed treatment (bottom row), as well as the respective percentages \nof those cohorts that relapsed, and the percentage of those relapses that are persistence-based. Analyses \nare repeated assuming the Virtual Mtb plate can (left) or cannot (right) detect non-replicating bacteria for \ndiagnosis of cure or relapse. \n \n3. Discussion \nSafely shortening antibiotic treatment for TB is a critical step towards eradicating the world's leading cause of \ndeath by infectious disease. Aside from development of resistance,  the main threat of shortening treatment \nregimen administration is relapse, where treated infections later recur. There are likely two modes of relapse: \npersistence-based (slow, rooted in the transport  of non-replicating Mtb out of caseum) and threshold -based \n(fast, rooted in incomplete sterilization of intracellular -Mtb)11. Given the considerable experimental challenges \ninvestigating relapse, we explore these phenomena using our whole -host computational model  grounded in \nhuman and primate datasets.  \nUsing HostSim, we compared rates of infection sterilization versus diagnostic cure status after virtual short-\ncourse antibiotic treatment regimens , focusing on antibiotics present in standard -of-care treatments such as \nHRZE8,81 (Methods). Simulations suggest that several regimens, including HRZE and most mono-treatments, \nare far more likely to result in false cure, where non -sterile infection falls below LOD at the time of treatment \ncompletion (Figure 2 ). To investigate outcomes at later times, we develop and calibrate a new  persistence \nmechanism in HostSim that allows non-replicating Mtb to be transported from caseum , allowing for new \ngranuloma reseeding and TB relapse. With it, simulations reproduce a recent short -course treatment relapse \nstudy in NHPs18.  \nWe find that reported relapse rates are considerably impacted by study-specific relapse definition—specifically, \nwhether it is based  on (1) cure status upon treatment completion (not standardized between clinical and \nexperimental studies) and (2) whether LTBI patients are admitted to relapse studies (Table 4). This remains true \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nwhen recreating several other relapse studies in humans, mice, and NHPs. When analyzing datasets as closely \nas possible to their in vivo counterparts, we recapitulate qualitative trends of relapse (e.g., lower relapse rates \nfollow longer treatment regimens). However, we also find a paucity of calibration-relevant measurements in the \nliterature, such as LOD of many diagnostic tests used. Our simulated relapse rates are higher than those reported \nin literature (Table 5). For human and NHP studies, this is reasonable since virtual diagnostic LODs assume that \n1 CFU in the lungs corresponds to 1 0 CFU/mL in the BAL, amplifying sensitivity. For RMMs, our relapse rates \nare closer to those observed in caseum-forming C3HeB/FeJ mice than non-caseum-forming BALB/c mice. \nFor research purposes, r ecurrent LTBI is seldom discussed in relapse studies because  (i) clinical relapse \ndefinitions do not typically include recurrent LTBI, and (ii) animal models such as RMMs often use detectable \nMtb post-treatment as a proxy for relapse. These proxies may implicitly classify as relapse small amounts of Mtb \nthat more closely resemble recurrent LTBI or reactivation than clinical relapse  (Table 1). Further, c linically \nstudying recurrent LTBI would be challenging because, by definition, it is asymptomatic in humans. The absence \nof such datasets leaves little guidance for what LOD thresholds to use for more detailed simulations. Moreover, \nregimen-specific risk of recurrent LTBI may be a silent contributor toward drug resistance. \nFor clinical consideration, a TB patient may be asymptomatic at treatment completion regardless of cure status. \nHowever, simulations suggest cure status at treatment completion may predict the mechanism of subsequent \nrelapse, should it arise . Specifically, patients relapsing  post-microbiologic-cure are more likely to harbor Mtb \nreservoirs within caseum rather than residual replicating populations , even if the diagnostic test cannot detect \ncaseum-trapped Mtb (Figure 4). This may inform personalized treatment strategies targeting non-replicating Mtb \nsubpopulations.  \nWe have made several important simplifications to reduce model complexity. We do not model drug resistance, \nthough drug resistance is believed to correlate with relapse rates 10. Lymph nodes are known to be important \nreservoirs of Mtb during TB disease and relapse, so we will combine our recent model of lymph node infection \ninto HostSim in future work18,19,21. Moreover, we calibrate our persistence-like mechanism using reactivation rates \nsubsequent to TNFα depletion (Methods), which may ignore subtleties of TB disease during immunosuppression, \nevident in how TNFα-induced reactivation of TB still often test skin-test negative, and may have other qualitatively \ndistinct immune factors11. Finally, the potential for diagnostic tests to yield false positive results follows from test-\nspecific causes that are beyond the scope of this work. For example, simulating false -positive results from \nimmune assays that likely require detailed represent ations of markers used in that assay, whereas predicting \nfalse positive results from culture -based tests would depend on detailed representation of BALF sampling or \nlaboratory contamination.   \nThere are gaps in literature precluding any comprehensive model of relapse. Our simulation framework (as many \nothers) is modular, allowing us to refine individual components of HostSim as necessary. However, our virtual \ndiagnostics rely on a coarse-grain representation—e.g., predicting symptoms via  𝑅, the fold -increase of CFU \nover 200 days, or assuming direct proportionality between CFU in BALF, sputum and lung tissue. We find such \nabstractions necessary as biological studies relating host/infection state appear noisy, and it is not \nmechanistically clear why some hosts with active TB are not culture -positive while some subclinical hosts are  \nculture-negative4,82. Once these factors are elucidated experimentally, they may be incorporated into  a more \ncomprehensive symptom and diagnostic model of TB to predict relapse . With sufficient mechanism, we could \ndevelop HostSim into a digital twin, a personalized model that uses individual patient datasets to predict the \nlikelihood and mechanism of a specific patient's relapse. A digital twin predicting relapse mechanism for a given \npatient could provide decision support for which Mtb subpopulations should be priority targets for the treatment \nregimen. Similar models have been made for a variety of complex diseases83–86.  \nComputational models serve as repositories that can synthesize large amount s of biological knowledge and \nclinical data. In this case, we find that relapse, often reported as a single phenomenon, is a combination of at \nleast two major presentations. Our  predictions suggest that after treatment failure, relapse is likely to be \nthreshold-driven, appearing rapidly and reproducibly. On the other hand, persistence -driven relapses appear \nmore rarely overall, yet more often within false-cured patients. One observation that is confirmed by our results \nsuggests that if a patient has LTBI (culture -negative TB) at the start of treatment, treatment regimen choice \nshould target non-replicating bacteria towards complete sterilization. \n4. Methods \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nRecently, we developed a detailed and complex in silico  whole-host model of pulmonary Mtb infection and \ntreatment, HostSim58–60,87 (Section 4.1 and Appendix A ). Moreover, HostSim contains a whole -host \npharmacokinetics (PK) and pharmacodynamics (PD) model, which we used to perform virtual pre -clinical trials \nto characterize bactericidal activity of various antibiotic regimens58. \nFor all analyses, we generate and calibrated a set of 𝑛 = 500 virtual hosts, each with 13 primary granulomas. \nThese virtual hosts are calibrate and validated as previously described58. After calibration, our untreated virtual \ncohort has 89.2% virtual hosts with LTBI, 0.8% sterilizing, and 10% virtual hosts with active TB disease  at 300 \ndays post-infection. \n4.1 Overview of HostSim \nBriefly, HostSim is a multi-scale hybrid computational model of pulmonary TB, including multiple lung granuloma \nagents within a single virtual host lung that is  coordinated with both a blood compartment and an uninfected \nvirtual lymph node system58,59. As pulmonary TB is typically contained to these 3 compartments, we refer to this \nas a “whole host” for TB. Each virtual granuloma is defined by a set of ordinary differential equations (ODEs) \nthat describe interactions between host immune cells (CD4+ T cells, CD8+ T cells, and macrophages) \ncommunicating, polarizing, and differentiating in response to cytokine signals (IFN-γ, TNFα, IL-4, IL-10, and IL-\n12). These immune cells respond to three distinct Mtb subpopulations : intracellular (within macrophages) , \nextracellular within the granuloma , and non -replicating (trapped within caseum ). As Mtb are killed, antigen \naccumulates and traffics to the lung draining lymph node system, which is represented as another set of ODEs. \nThe lymph node ODE tracks priming of Mtb-specific CD4+ and CD8+ immune cells which migrate to the host's \nblood, which is the third and final set of ODEs in HostSim. CD4+ and CD8+ Mtb-specific and nonspecific T cells \nmay then be recruited to the virtual lung granulomas based on the granuloma state. This model also has \nstochastic elements such as granuloma dissemination (i.e., seeding of new granulomas) and the persistence -\nbased mechanism of Mtb transport from caseum to macrophages. \nEach ODE term and inter-physiological compartmental transition term describe key dynamics of pulmonary TB \n(both untreated infection progression and the pharmacokinetics/pharmacodynamics (PK/PD) governing \ntreatment efficacy). The terms in HostSim include (i) metabolic differences between replicating Mtb (internalized \nby macrophages or not) and non -replicating Mtb; (ii) dynamic priming of T cells and the (de)activation of \nmonocyte-derived macrophages; (iii) the dynamic accumulation of caseum based on macrophage necrosis; (iv) \nantibiotic penetration into caseum; and (v) synergistic/antagonistic drug-drug impacts on PD.  \n4.2 Classifying Virtual Host State \nTo systematically examine case -specific patient outcomes with or without treatment ( Table 1 ), we need to \ndetermine how those outcomes are measured.  \n4.2.1 Classifying virtual host LTBI versus active TB by species \nEstimating which virtual hosts suffer active disease is a nuanced challenge , as there is no currently known \nmechanism connecting CFU (lung or BALF) to symptoms. As an estimate, if CFU levels reach a species-specific \nthreshold for fold-increase over when we would have expected hosts with LTBI hosts to stabilize  (~200 days), \nthen we label the hosts as having active TB. W e also predict hosts with large-CFU levels (>10,000 for over 30 \nday timeframe) also experience active TB, as we have in previous work58. See Supplemental Material S4 for \nmore details. \n4.2.2 Virtual Diagnostic tests \nClinically or experimentally, disease state is assessed using one or more diagnostics (Table 2). To assess virtual \nhost outcomes, we mimic those detection methods in silico using virtual diagnostics ( Table 3). These are the \nbuilding-blocks of testing for more complex outcomes like relapse or reactivation. \nA virtual diagnostic test takes a virtual host at a given time and assesses their infection state. All host-scale tests \nassume that pulmonary TB infection states can be precisely measured by looking at the lung state, but this may \nbe expanded in future studies to include more detailed regarding the role of lymph node infection21. \nVirtual Mtb Plate \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nThis virtual test recreates various culture, smear, or NAAT-based tests (Table 6) to assess a host as either cured \nor not cured at a given time. Virtual Mtb plates sum together CFU counts of relevant Mtb bacilli and compare \nthem against a LOD. The test returns not cured  if the sum  of Mtb exceed s the LOD, and it returns  cured \notherwise. The exact definition for which Mtb are relevant is configured by checking a number of biological \nassumptions; in principle, we can calibrate these further once more biology is known.  \nTable 6: Choices and assumptions that we make to configure virtual Mtb plates into various clinical \ndiagnostics in silico. The configuration-based questions are: (1) Does the diagnostic look at the entire host (H) \nor a specific granuloma tissue (G)? (2) What is the LOD of the test?  (3) Are non-replicating or caseum-bound \nbacteria detectable with this test? (4) Are dead Mtb able to be detected in this test? If so, how long can bacilli be \ndetected for after they have been killed? †Assume direct proportionality between lung-tissue CFU counts and \nsputum/BAL CFU counts. ‡This assumes that the nucleic acid being tested decays rapidly. \nReplicated test Q1 (Scale) Q2 (What is the \nLOD?) \nQ3 (Are non-\nreplicating  Mtb \nrelevant?) \nQ4 (Are recently-\ndead Mtb relevant?) \nSputum or BAL culture  Host 50 CFU† No† No \nAcid-fast Mtb smear Host 5000 CFU No Yes, 10 days† \nGranuloma tissue homogenate Mtb \nculture \nGranuloma 5 CFU Yes No \nNAAT-based rapid test Host 16 CFU No† No†,‡ \nLung homogenate Mtb culture Host 10 CFU Yes No \n \nVirtual PET/CT \nCombined positron emission tomography (PET) and computed tomography (CT) is a radiographic measure of \ninflammation ( Supplemental Material S1 ). As previously 87, w e measure virtual FDG avidity (measured by \nPET/CT) as a weighted sum of metabolically-active immune cells, including activated macrophages and T cells \nwithin each granuloma. In summary, FDG avidity is calculated as \nVirtual\tFDG\tavidity = 𝑤!MR + 𝑤\"MI + 𝑤#MA + 𝑤$T0 + 𝑤%TE + 𝑤&TEM\t\nwhere ⟨𝑤!⟩ = ⟨0,5,6,2,4,3⟩ are the relative weights of resting, infected, and activated macrophages; and primed, \neffector, and effector memory cell populations within a granuloma. Resting macrophages have a contribution of \n0 as we assume they are at the same level of background activity as the uninvolved lung tissue.  \nThe Virtual PET/CT test assesses hosts as relapsed if: \n1. Any new granulomas disseminate (Supplemental Material S5).  \n2. There is > 20% increase of virtual FDG avidity between time of treatment completion an d time of \nassessment for any granuloma within the host.  \nVirtual clinic score \nVirtual clinic score is a combination of the above tests. A host is determined to be cured if none of the following \nreturn not cured: \n1. A Virtual Mtb plate with LOD = 16 CFU, assuming that non-replicating bacteria cannot be detected (i.e., \nVirtual rapid test, Table 6). \n2. If any new granulomas have formed within the host since the last virtual diagnostic, even below LOD —\nwe assume that this is indicative of symptom recurrence as is assumed in 18.  \n4.3 Persistence mechanism - stochastic transport of Mtb from caseum \nTo simulate both persistence- and threshold-driven relapse, we include persistence and threshold mechanisms \nin HostSim. Simulated persistence allows an activated immune system  (such as neutrophils) to transport Mtb \nfrom caseum, transitioning them to an intracellular state as suggested by our previous work68. We assume that \nthe persistence mechanism  is involved in both relapse and reactivation (Table 1). LTBI hosts may go years \nwithout reactivation—only 10% of subclinical infection hosts with no comorbidities reactivate at any point in their \nlife4,15—so we calibrate our reactivation mechanism for studying relapse model in the context of reactivation \ninduced by TNFα depletion. Details are in Supplemental Material S2. \n4.4 Simulated Antibiotic Regimens \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\nRelapse has been studied in the context of shortening standard antibiotic regimens, which typically include a \ncombination of Isoniazid (INH; H), Rifampicin (RIF; R), Ethambutol (ETH; E), and Pyrazinamide (PZA; Z), called \nHRZE8. The standard treatment is 6 -9 months of HRZE, or 2 months of HRZE and 4 months of HR 88. Several \nstudies have tested for sputum culture conversion after only 2 months of antibiotic treatment 89. Second-line \nregimens for RIF-resistant TB include a combination of Bedaquiline (BDQ; B), Pretomanid (PTM; Pa), Linezolid \n(LZD; L), and (sometimes) Moxifloxacin (MXF; M), called BPaL(M) 6,25. \nWe simulated two families of antibiotic regimens, summarized in Table 7. To determine variability in the level of \nnon-detectable bacteria across multiple regimens, we reproduce a family of regimens including H, R, Z, E, B, \nPa, L, and M tested in our previous work58 (group i). We also simulate scenarios wherein virtual hosts are treated \nwith multi-phase regimens (group ii), such as the WHO standard of care (Regimen 2HRZE4HR) that discontinues \nuse of PZA and EMB after two months. After virtual treatments end, we continue simulations for without treatment \nto examine relapse events, though regimens may be assessed at more than one follow-up time (Table 5). \nTable 7: Regimens from previous studies recreated using HostSim. This table is partly based on Table 6 from \nour previous work58. Columns labeled by antibiotic name (e.g., INH) indicate the human-equivalent dosage of that \nantibiotic in human -equivalent mg/kg, administered daily. Each regimen in set (i) lasts 6 months (180 days). \nRegimens in set (ii) are split into phases, with each phase duration and the drugs administered during that phase \nindicated by regimen name —e.g., 2HR1R indicates 2 months of INH+RIF followed by 1 month of RIF \nmonotreatment. *Antibiotic halted after initial phase. †Culture positivity was reported during treatment, so we \npredicted relapse at days -post-treatment-start for these specific regimens and times. ‡Specific timing of virtual \nHRZE and virtual SIV are given in the Results text, and are adapted from an NHP model18. §Dose adjusted to the \nhuman equivalent standard dose. \nRegimen Name INH RIF PZA EMB BDQ PTM LZD MXF \nRegimen set (i) - Marmoset studies from a previous in silico / NHP study90 (set reproduced from our previous HostSim study58). \nRMZE - 10 25 20 - - - 7 \nBPa - - - - 20 20 - - \nBPaL - - - - 20 20 90 - \nBL - - - - 20 - 90 - \nRM - 10 - - - - - 7 \nHRZE 5 10 25 20 - - - - \nPaL - - - - - 20 90 - \nBedaquiline - - - - 20 - - - \nPretomanid - - - - - 20 - - \nRZ - 10 25 - - - - - \nHZ 5 - 25 - - - - - \nMoxifloxacin - - - - - - - 7 \nPyrazinamide - - 25 - - - - - \nRifampicin - 10 - - - - - - \nIsoniazid 5 - - - - - - - \nRegimen set (ii) - Multi-phase regimens used in standard of care and/or experimental relapse studies.  \n2HR + SIV‡ 6 10 - - - - - - \n2HRZE4HR 6 10 25* 20* - - - - \n2RMZ2RM§ (36) - 10 25* - - - - 7 \n2HRZE2HR§ (46) 5 10 25* 20* - - - - \n2RMZE1RM§ (46) - 10 25* 20* - - - 7 \n2HRZ1HR§ (45) 6 10 25* - - - - - \n1BPaMZ1BPaM§ (44) - - 25* - 20 20 - 7 \nAcknowledgements \nC.T.M. was supported by the Molecular Mechanisms in Microbial Pathogenesis Training Program (T32 \nAI007528). This work was by National Institutes of Health Grant R01 AI50684 (D.K.) and the Center for Data -\nDriven Drug Development and Treatment Assessment (DATA; D.K. and M.B.).  We thank Paul Wolberg for \ncomputational assistance and support. \nReferences \n1. Global Tuberculosis Report 2024. (World Health Organiza4on, Geneva, 2024). \n2. CDC. Provisional 2024 Tuberculosis Data, United States. Tuberculosis Data hFps://www.cdc.gov/tb-\ndata/2024-provisional/index.html (2025). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n3. Cadena, A. M., Fortune, S. M. & Flynn, J. L. Heterogeneity in tuberculosis. Nat. Rev. Immunol. 17, 691–702 \n(2017). \n4. Lin, P . L. & Flynn, J. L. The End of the Binary Era: Revisi4ng the Spectrum of Tuberculosis. J Immunol 201, \n2541–2548 (2018). \n5. Lin, P . L. et al. PET CT Iden4fies Reac4va4on Risk in Cynomolgus Macaques with Latent M. tuberculosis. PLoS \nPathog 12, e1005739 (2016). \n6. Dheda, K. et al. Global control of tuberculosis: from extensively drug-resistant to untreatable tuberculosis. \nLancet Respir. Med.  2, 321–338 (2014). \n7. Saukkonen, J. J. et al. Updates on the Treatment of Drug-Suscep4ble and Drug-Resistant Tuberculosis: An \nOfficial ATS/CDC/ERS/IDSA Clinical Prac4ce Guideline. Am. J. Respir. Crit. Care Med.  211, 15–33 (2025). \n8. WHO Consolidated Guidelines on Tuberculosis. Module 4: Treatment - Drug-Resistant Tuberculosis \nTreatment, 2022 Update. (World Health Organiza4on, Geneva, 2022). \n9. Latent Tuberculosis InfecMon: Updated and Consolidated Guidelines for ProgrammaMc Management . (World \nHealth Organiza4on, Global TB Programme, Geneva, Switzerland, 2018). \n10. Colangeli, R. et al. Bacterial Factors That Predict Relapse afer Tuberculosis Therapy. N. Engl. J. Med.  379, \n823–833 (2018). \n11. Goleg, D. et al. Can we predict tuberculosis cure? What tools are available? Eur. Respir. J. 52, (2018). \n12. WHO Consolidated Guidelines on Tuberculosis. Module 3: Diagnosis - Rapid DiagnosMcs for Tuberculosis \nDetecMon, 2021 Update. (World Health Organiza4on, Geneva, 2021). \n13. Brady, M. F., Coronel, J., Gilman, R. H. & Moore, D. A. The MODS method for diagnosis of tuberculosis and \nmul4drug resistant tuberculosis. J. Vis. Exp. JoVE 845 (2008) doi:10.3791/845. \n14. Berg, A. et al. Model-Based Meta-Analysis of Relapsing Mouse Model Studies from the Cri4cal Path to \nTuberculosis Drug Regimens Ini4a4ve Database. AnMmicrob. Agents Chemother. 66, e01793-21 (2022). \n15. Flynn, J. L. & Chan, J. Tuberculosis: latency and reac4va4on. Infect Immun 69, 4195–201 (2001). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n16. Guerra-Assunção, J. A. et al. Recurrence due to Relapse or Reinfec4on With Mycobacterium tuberculosis: A \nWhole-Genome Sequencing Approach in a Large, Popula4on-Based Cohort With a High HIV Infec4on \nPrevalence and Ac4ve Follow-up. J. Infect. Dis. 211, 1154–1163 (2015). \n17. Selvavinayagam, S. T. et al. Clinical laboratory analytes and platelet-associated parameters as surrogate \nmarkers of subclinical inflamma4on in latent tuberculosis infec4on. Front. Immunol. 16, (2025). \n18. Maiello, P . et al. Characterizing PET CT paFerns and bacterial dissemina4on features of tuberculosis relapse \nin the macaque model. Infect. Immun. 93, e00177-25 (2025). \n19. Ganchua, S. K. C., White, A. G., Klein, E. C. & Flynn, J. L. Lymph nodes-The neglected baFlefield in \ntuberculosis. PLoS Pathog 16, e1008632 (2020). \n20. Lin, P . L. et al. Steriliza4on of granulomas is common in ac4ve and latent tuberculosis despite within-host \nvariability in bacterial killing. Nat Med  20, 75–9 (2014). \n21. Krupinsky, K. C., Michael, C. T., Nanda, P ., Magla, J. T. & Kirschner, D. Dis4nguishing mul4ple roles of T cell \nand macrophage involvement in determining lymph node fates during Mycobacterium tuberculosis \ninfec4on. PLOS Comput. Biol. 21, e1013033 (2025). \n22. Choudhury, A. et al. Hepa4c tuberculosis: myriad of hues. GERMS  11, 310–313 (2021). \n23. Ramakrishnan, L. Revisi4ng the role of the granuloma in tuberculosis. Nat Rev Immunol 12, 352–66 (2012). \n24. Benator, D. et al. Rifapen4ne and isoniazid once a week versus rifampicin and isoniazid twice a week for \ntreatment of drug-suscep4ble pulmonary tuberculosis in HIV-nega4ve pa4ents: a randomised clinical trial. \nLancet 360, 528–34 (2002). \n25. Nyang’wa, B.-T. et al. A 24-Week, All-Oral Regimen for Rifampin-Resistant Tuberculosis. N. Engl. J. Med.  \n387, 2331–2343 (2022). \n26. Conradie, F. et al. Bedaquiline-Pretomanid-Linezolid Regimens for Drug-Resistant Tuberculosis. N Engl J \nMed  387, 810–823 (2022). \n27. Flynn, J. L. et al. Non-human primates: a model for tuberculosis research. Tuberculosis 83, 116–118 (2003). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n28. Lin, P . L. et al. Metronidazole prevents reac4va4on of latent Mycobacterium tuberculosis infec4on in \nmacaques. Proc Natl Acad Sci U A 109, 14188–93 (2012). \n29. Johnson, J. L. et al. Shortening treatment in adults with noncavitary tuberculosis and 2-month culture \nconversion. Am J Respir Crit Care Med  180, 558–63 (2009). \n30. Lu, P . et al. Time to sputum culture conversion and treatment outcome of pa4ents with mul4drug-resistant \ntuberculosis: a prospec4ve cohort study from urban China. Eur. Respir. J. 49, (2017). \n31. Dorman, S. E. et al. Xpert MTB/RIF Ultra for detec4on of Mycobacterium tuberculosis and rifampicin \nresistance: a prospec4ve mul4centre diagnos4c accuracy study. Lancet Infect. Dis. 18, 76–84 (2018). \n32. Steingart, K. R. et al. Xpert® MTB/RIF assay for pulmonary tuberculosis and rifampicin resistance in adults - \nSteingart, KR - 2014 | Cochrane Library. \nhFps://www.cochranelibrary.com/cdsr/doi/10.1002/14651858.CD009593.pub3/informa4on. \n33. 9.3 Sputum smear and culture | TB Knowledge Sharing. hFps://tbksp.who.int/en/node/1970. \n34. Chen, W.-C., Chang, C.-C. & Lin, Y . E. Pulmonary Tuberculosis Diagnosis Using an Intelligent Microscopy \nScanner and Image Recogni4on Model for Improved Acid-Fast Bacilli Detec4on in Smears. Microorganisms  \n12, 1734 (2024). \n35. Kramnik, I. & Beamer, G. Mouse models of human TB pathology: roles in the analysis of necrosis and the \ndevelopment of host-directed therapies. Semin. Immunopathol. 38, 221–237 (2016). \n36. Ibrahim, M., Truffot-Pernot, C., Andries, K., Jarlier, V. & Veziris, N. Sterilizing Ac4vity of R207910 (TMC207)-\ncontaining Regimens in the Murine Model of Tuberculosis. Am. J. Respir. Crit. Care Med.  180, 553–557 \n(2009). \n37. Maiello, P . et al. Rhesus Macaques Are More Suscep4ble to Progressive Tuberculosis than Cynomolgus \nMacaques: a Quan4ta4ve Comparison. Infect Immun 86, (2018). \n38. Chen, R. Y . et al. PET/CT imaging correlates with treatment outcome in pa4ents with mul4drug-resistant \ntuberculosis. Sci. Transl. Med.  6, 265ra166-265ra166 (2014). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n39. Goleg, D. et al. Accuracy of Immunodiagnos4c Tests for Ac4ve Tuberculosis Using Single and Combined \nResults: A Mul4center TBNET-Study. PLOS ONE 3, e3417 (2008). \n40. Zhang, H. et al. Reversion of Quan4FERON-TB Gold In-Tube test in individuals with and without \nprophylac4c treatment for latent tuberculosis infec4on: A systema4c review and meta-analysis. J. Infect. \n77, 276–282 (2018). \n41. Zhan, L., Tang, J., Sun, M. & Qin, C. Animal Models for Tuberculosis in Transla4onal and Precision Medicine. \nFront. Microbiol.  8, (2017). \n42. Capuano, S. V. et al. Experimental Mycobacterium tuberculosis infec4on of cynomolgus macaques closely \nresembles the various manifesta4ons of human M. tuberculosis infec4on. Infect Immun 71, 5831–44 \n(2003). \n43. Xu, J. et al. Contribu4on of Pretomanid to Novel Regimens Containing Bedaquiline with either Linezolid or \nMoxifloxacin and Pyrazinamide in Murine Models of Tuberculosis. AnMmicrob Agents Chemother 63, \n(2019). \n44. Li, S.-Y.  et al. Bactericidal and Sterilizing Ac4vity of a Novel Regimen with Bedaquiline, Pretomanid, \nMoxifloxacin, and Pyrazinamide in a Murine Model of Tuberculosis. AnMmicrob. Agents Chemother. 61, \n10.1128/aac.00913-17 (2017). \n45. Tasneen, R. et al. Contribu4on of Oxazolidinones to the Efficacy of Novel Regimens Containing Bedaquiline \nand Pretomanid in a Mouse Model of Tuberculosis. AnMmicrob Agents Chemother 60, 270–7 (2016). \n46. Li, S. Y . et al. Evalua4on of moxifloxacin-containing regimens in pathologically dis4nct murine tuberculosis \nmodels. AnMmicrob Agents Chemother 59, 4026–30 (2015). \n47. U.S. Department of Health and Human Services, Food and Drug Administra4on & Center for Drug \nEvalua4on and Research. Guidance for Industry: Es4ma4ng the Maximum Safe Star4ng Dose in Ini4al \nClinical Trials for Therapeu4cs in Adult Healthy Volunteers. (2005). \n48. Morden4, J. Man versus Beast: Pharmacokine4c Scaling in Mammals. J. Pharm. Sci. 75, 1028–1040 (1986). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n49. Lanoix, J.-P ., Lenaerts, A. J. & Nuermberger, E. L. Heterogeneous disease progression and treatment \nresponse in a C3HeB/FeJ mouse model of tuberculosis. Dis. Model. Mech.  8, 603–610 (2015). \n50. Scanga, C. A. & Flynn, J. L. Modeling tuberculosis in nonhuman primates. Cold Spring Harb. Perspect. Med.  \n4, a018564 (2014). \n51. Examina4on of Clinical Trial Costs and Barriers for Drug Development. ASPE \nhFp://aspe.hhs.gov/reports/examina4on-clinical-trial-costs-barriers-drug-development-0 (2014). \n52. Lin, P . L. et al. Quan4ta4ve comparison of ac4ve and latent tuberculosis in the cynomolgus macaque \nmodel. Infect Immun 77, 4631–42 (2009). \n53. Coleman, M. T. et al. Early Changes by (18)Fluorodeoxyglucose positron emission tomography coregistered \nwith computed tomography predict outcome afer Mycobacterium tuberculosis infec4on in cynomolgus \nmacaques. Infect Immun 82, 2400–4 (2014). \n54. Lin, P . L. & Flynn, J. L. Understanding latent tuberculosis: a moving target. J Immunol 185, 15–22 (2010). \n55. Kendall, E. A., Shrestha, S. & Dowdy, D. W. The Epidemiological Importance of Subclinical Tuberculosis. A \nCri4cal Reappraisal. Am. J. Respir. Crit. Care Med.  203, 168–174 (2021). \n56. Wigginton, J. E. & Kirschner, D. A model to predict cell-mediated immune regulatory mechanisms during \nhuman infec4on with Mycobacterium tuberculosis. J Immunol 166, 1951–67 (2001). \n57. Berry, M. P . R. et al. An interferon-inducible neutrophil-driven blood transcrip4onal signature in human \ntuberculosis. Nature 466, 973–977 (2010). \n58. Michael, C. T. et al. Rankings of tuberculosis an4bio4c treatment regimens are sensi4ve to spa4al scale, \ndetec4on limit, and ini4al host bacterial burden. J. Theor. Biol. 611, 112176 (2025). \n59. Joslyn, L. R., Linderman, J. J. & Kirschner, D. E. A virtual host model of Mycobacterium tuberculosis \ninfec4on iden4fies early immune events as predic4ve of infec4on outcomes. J. Theor. Biol. 539, 111042 \n(2022). \n60. Joslyn, L., Linderman, J. J. & Kirschner, D. Concomitant immunity is mediated by effector memory T cells in \ntuberculosis. (2020). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n61. Lin, P . L. et al. Radiologic responses in cynomolgous macaques for assessing tuberculosis chemotherapy \nregimens. AnMmicrob Agents Chemother hFps://doi.org/10.1128/AAC.00277-13 (2013) \ndoi:10.1128/AAC.00277-13. \n62. Budak, M. et al. GEODE: an in silico tool that translates in vitro to in vivo predic4ons of tuberculosis \nan4bio4c combina4on efficacy. Front. Pharmacol. 16, (2025). \n63. Pienaar, E., Dartois, V., Linderman, J. J. & Kirschner, D. E. In silico evalua4on and explora4on of an4bio4c \ntuberculosis treatment regimens. BMC Syst Biol  9, 79 (2015). \n64. Fox, K. A. et al. Platelets Regulate Pulmonary Inflamma4on and Tissue Destruc4on in Tuberculosis. Am. J. \nRespir. Crit. Care Med.  198, 245–255 (2018). \n65. Resolución Ministerial N.° 715-2013-MINSA. hFps://www.gob.pe/ins4tucion/minsa/normas-\nlegales/198713-715-2013-minsa. \n66. Proaño, A. et al. Dynamics of Cough Frequency in Adults Undergoing Treatment for Pulmonary \nTuberculosis. Clin. Infect. Dis. 64, 1174–1181 (2017). \n67. Conradie, F. et al. Treatment of Highly Drug-Resistant Pulmonary Tuberculosis. N Engl J Med  382, 893–902 \n(2020). \n68. Hult, C., Magla, J. T., Gideon, H. P ., Linderman, J. J. & Kirschner, D. E. Neutrophil Dynamics Affect \nMycobacterium tuberculosis Granuloma Outcomes and Dissemina4on. Front Immunol 12, 712457 (2021). \n69. Fallahi-Sichani, M., Schaller, M. A., Kirschner, D. E., Kunkel, S. L. & Linderman, J. J. Iden4fica4on of key \nprocesses that control tumor necrosis factor availability in a tuberculosis granuloma. PLoS Comput Biol 6, \ne1000778 (2010). \n70. Fallahi-Sichani, M., Flynn, J. L., Linderman, J. J. & Kirschner, D. E. Differen4al risk of tuberculosis reac4va4on \namong an4-TNF therapies is due to drug binding kine4cs and permeability. J Immunol 188, 3169–78 \n(2012). \n71. Pennisi, E. & Cohen, J. Eradica4ng HIV From a Pa4ent: Not Just a Dream? Science 272, 1884–1884 (1996). \n72. Kirschner, D. & Webb, G. F. Immunotherapy of HIV-1 Infec4on. (1997). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n73. Raehtz, K., Pandrea, I. & Apetrei, C. The well-tempered SIV infec4on: Pathogenesis of SIV infec4on in \nnatural hosts in the wild, with emphasis on virus transmission and early events post-infec4on that may \ncontribute to protec4on from disease progression. Infect. Genet. Evol. 46, 308–323 (2016). \n74. Diedrich, C. R. et al. SIV and Mycobacterium tuberculosis synergy within the granuloma accelerates the \nreac4va4on paFern of latent tuberculosis. PLOS Pathog. 16, e1008413 (2020). \n75. Diedrich, C. R. et al. Reac4va4on of latent tuberculosis in cynomolgus macaques infected with SIV is \nassociated with early peripheral T cell deple4on and not virus load. PLoS One 5, e9611 (2010). \n76. Burdo, T. H., Walker, J. & Williams, K. C. Macrophage Polariza4on in AIDS: Dynamic Interface between An4-\nViral and An4-Inflammatory Macrophages during Acute and Chronic Infec4on. J. Clin. Cell. Immunol. 6, 333 \n(2015). \n77. Malherbe, S. T. et al. Persis4ng positron emission tomography lesion ac4vity and Mycobacterium \ntuberculosis mRNA afer tuberculosis cure. Nat. Med.  22, 1094–1100 (2016). \n78. Narayanan, P . R. SHORTENING SHORT COURSE CHEMOTHERAPY: A RANDOMISED CLINICAL TRIAL FOR \nTREATMENT OF SMEAR POSITIVE PULMONARY TUBERCULOSIS WITH REGIMENS USING OFLOXACIN IN THE \nINTENSIVE PHASE. Ind. J Tub. 49, 27–38 (2002). \n79. Carvalho, A. C. C. et al. Pre-Treatment Neutrophil Count as a Predictor of An4tuberculosis Therapy \nOutcomes: A Mul4center Prospec4ve Cohort Study. Front. Immunol. 12, (2021). \n80. Jafari, C., Olaru, I. D., Daduna, F., Lange, C. & Kalsdorf, B. Rapid Diagnosis of Recurrent Paucibacillary \nTuberculosis. Pathog. Immun. 7, 189–202 (2022). \n81. Budak, M. et al. Op4mizing tuberculosis treatment efficacy: Comparing the standard regimen with \nMoxifloxacin-containing regimens. PLoS Comput. Biol. 19, e1010823 (2023). \n82. Blankley, S. et al. The Transcrip4onal Signature of Ac4ve Tuberculosis Reflects Symptom Status in Extra-\nPulmonary and Pulmonary Tuberculosis. PLOS ONE 11, e0162220 (2016). \n83. Fonseca, L. L., BöFcher, L., Mehrad, B. & Laubenbacher, R. C. Op4mal control of agent-based models via \nsurrogate modeling. PLOS Comput. Biol. 21, e1012138 (2025). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint \n\n84. Knapp, A. C., Cruz, D. A., Mehrad, B. & Laubenbacher, R. C. Personalizing computa4onal models to \nconstruct medical digital twins. bioRxiv 2024.05.31.596692 (2024) doi:10.1101/2024.05.31.596692. \n85. An, G. & Cockrell, C. Drug Development Digital Twins for Drug Discovery, Tes4ng and Repurposing: A \nSchema for Requirements and Development. Front. Syst. Biol. 2, 928387 (2022). \n86. Nanda, P ., Budak, M., Michael, C. T., Krupinsky, K. & Kirschner, D. E. Development and Analysis of Mul4scale \nModels for Tuberculosis: From Molecules to Popula4ons. in Development and Analysis of MulMscale \nModels for Tuberculosis: From Molecules to PopulaMons  (2023). \n87. Michael, C. T., Almohri, S. A., Linderman, J. J. & Kirschner, D. E. A framework for mul4-scale interven4on \nmodeling: virtual cohorts, virtual clinical trials, and model-to-model comparisons. Front. Syst. Biol. 3, \n(2024). \n88. Nahid, P . et al. Execu4ve Summary: Official American Thoracic Society/Centers for Disease Control and \nPreven4on/Infec4ous Diseases Society of America Clinical Prac4ce Guidelines: Treatment of Drug-\nSuscep4ble Tuberculosis. Clin Infect Dis 63, 853–67 (2016). \n89. Lanoix, J.-P ., Chaisson, R. E. & Nuermberger, E. L. Shortening Tuberculosis Treatment With \nFluoroquinolones: Lost in Transla4on? Clin. Infect. Dis. 62, 484–490 (2016). \n90. Budak, M. et al. A systema4c efficacy analysis of tuberculosis treatment with BPaL-containing regimens \nusing a mul4scale modeling approach. CPT Pharmacomet. Syst. Pharmacol. 13, 673–685 (2024). \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted January 5, 2026. ; https://doi.org/10.64898/2026.01.04.697520doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}