In vitro dormancy models improve ability to predict treatment response in severe marmoset tuberculosis lesions

Abstract Tuberculosis lesions are structurally and physiologically heterogeneous, creating microenvironments that restrict antibiotic penetration and alter bacterial susceptibility. This heterogeneity remains a major impediment to shortening tuberculosis treatment, as it is difficult to predict which regimens will sterilize bacteria in these hard-to-treat lesion niches. Many in vitro systems and commonly used animal models fail to recapitulate the necrotic, caseous lesion pathology associated with treatment failure and relapse. Here, we show that resource-efficient in vitro assays engineered to mimic key lesion microenvironments, including those with lipid-rich, caseum-like matrices with controlled oxygenation and pH, generate drug-response metrics that predict lesion-level treatment responses in a non-human primate model. We stratified individual marmoset lesions by baseline 2-deoxy-2-[18F] fluoro-D-glucose ([18F] FDG) positron emission tomography/computed tomography (PET/CT) features, then quantified drug potency and interactions across three lipid-induced dormancy (LIDs) conditions under both equipotent and lesion pharmacokinetic-informed dosing schemes. LIDs-based and pharmacokinetic-informed measurements aligned more closely with treatment responses in severe lesions than conventional in vitro data and recapitulated interaction patterns consistent with known regimen performance. Integrating imaging-derived and in vitro features in multivariate, sequential computational models improved predictive accuracy over baseline imaging alone and identified a subset of informative in vitro metrics. Together, these results establish a scalable framework linking lesion-mimicking assays to lesion-specific outcomes, enabling earlier, more cost-effective prioritization of pre-clinical regimens aimed at sterilizing hard-to-treat tuberculosis lesions and shortening therapy.