Computational drug screening struggles against drug-resistant tuberculosis, review finds

Computational tools can now search billions of molecules for candidate anti-tuberculosis drugs, yet most predicted hits never reach patients with drug-resistant disease. The authors argue that the problem is not simply weak algorithms but a structural mismatch: computational pipelines optimize for strong binding to a single molecular target, while Mycobacterium tuberculosis (Mtb) survives inside the human body through a layered set of defenses that no docking score captures. Their conclusion is that anti-tuberculosis drug discovery needs a shift from target-centric computation toward biologically grounded, regimen-aware modeling.

The stakes are high. Tuberculosis remains the leading cause of death from a single pathogen worldwide, and the World Health Organization has set a goal of reducing tuberculosis incidence and mortality by 95 percent relative to 2015 levels by 2035. Second-line treatment can stretch to two years with drugs that carry considerable toxicity, and the review notes that new anti-tuberculosis drugs have not kept pace with the evolution of resistance.

The authors identify several biological barriers that separate computational predictions from real-world efficacy. First, a target being essential for bacterial survival does not mean it is pharmacologically vulnerable: the mycolic acid–rich cell envelope and active efflux pumps severely limit intracellular drug accumulation, so strong binding affinity often fails to produce effective target engagement. Second, Mtb adapts to host stresses such as hypoxia and acidic pH by entering non-replicating, dormant states with extensive metabolic reprogramming, which makes many growth-associated targets conditionally irrelevant.

Third, tuberculosis granulomas are spatially heterogeneous, with necrotic caseum forming diffusion-limited regions where systemically effective drugs may never reach therapeutic concentrations. The review also notes that some drugs, like isoniazid, require intracellular enzymatic activation, and that tuberculosis is always treated with multi-drug regimens, so single-compound optimization does not directly translate into regimen-level success. Against this backdrop, the authors evaluate the main computational approaches in current use.

Virtual screening comes in two forms. Ligand-based screening assumes that molecules structurally or physicochemically similar to known actives will show comparable activity, using tools such as quantitative structure–activity relationship (QSAR) modeling, pharmacophore modeling, and similarity searching; newer three- and four-dimensional QSAR methods add spatial conformations and dynamic conformational changes to improve accuracy. Structure-based screening, the most widely used form, relies on three-dimensional protein structures to dock candidate compounds into targets such as the enoyl-acyl carrier protein reductase InhA, DNA gyrase, and ATP synthase, then refines hits with molecular dynamics simulations and ADMET predictions of absorption, distribution, metabolism, excretion, and toxicity.

One of the review's more pointed contributions is a classification of tuberculosis drug targets by how likely structure-based work on them is to translate. Clinically validated targets such as InhA, DNA gyrase, and ATP synthase are the most reliable class, but structure-based work there should focus on resistance-aware optimization rather than de novo discovery. Structural information alone, they conclude, is insufficient to predict druggability in Mtb.

They also caution that multi-step docking-plus-dynamics workflows can produce internally consistent metrics that reinforce one another without independent validation, supporting ligand prioritization rather than reliable prediction of biological activity. Artificial intelligence offers a different route: rather than modeling binding explicitly, machine learning models learn predictive relationships directly from chemical structure, bioactivity, drug–target interaction, and omics data. The review traces a decade of progress, from Bayesian models that identified seven of nine Mtb-active compounds reported in published studies, to refined models with positive predictive values above 77 percent, to 2022 suites of models supporting drug design, target prediction, and data visualization.

But the authors stress persistent weaknesses: dataset bias from chemically narrow or imbalanced training sets, limited interpretability of deep learning predictions, and reproducibility problems stemming from inconsistent preprocessing and incomplete reporting. Hybrid frameworks combine these tools, and the review documents how AI has shifted from an auxiliary role to a central driver of screening pipelines. In one workflow, the generative platform Reinvent4 designed novel inhibitors of fatty acid-CoA synthetase, a key enzyme in Mtb lipid metabolism, with docking and binding-energy analysis yielding four lead candidates.

Mass spectrometry and protein–protein interaction network analysis have also been combined with structure-based screening to improve compound characterization and target identification. The authors' conclusion, however, is consistent: whatever the sophistication of the pipeline, its outputs remain in silico predictions without experimental validation, and most reported studies never progress to whole-cell or in vivo testing. The review then turns to strategies aimed specifically at resistance.

Combination therapy receives particular attention because it is the clinical reality of tuberculosis treatment. Physiologically based pharmacokinetic modeling has quantified drug–drug interactions such as rifampicin-mediated reductions in bedaquiline exposure, and imaging studies of tuberculosis lesions have documented heterogeneous drug penetration into caseous necrotic regions. An AlphaFold-assisted docking study expanded the search for synergistic combinations and enabled drug repurposing in tuberculosis–diabetes comorbidity.

The authors argue that these regimen-level effects—synergy, accumulation at the site of infection, and resistance suppression—are precisely what current computational pipelines ignore. The review closes with a forward-looking agenda. Next, future models must embed multi-scale biological constraints—cell envelope permeability, efflux pump dynamics, metabolic heterogeneity, and lesion-level pharmacokinetics—rather than relying on static target-binding predictions.

Finally, because tuberculosis therapy is intrinsically combination-based, computational strategies should prioritize drug–drug interactions and regimen-level performance over the optimization of single compounds. Until those shifts occur, the authors suggest, computational screening should be regarded primarily as a hypothesis-generation and prioritization tool, its hits treated as candidates for systematic experimental validation rather than as near-clinical leads.

Subject of Research: Computational drug screening approaches and their limitations for drug-resistant tuberculosis

Article Title: Computational drug screening for drug-resistant tuberculosis: advances, challenges, and future perspectives

Article References: Hu, S., Li, X., Batool, Z., Zheng, X., Dong, J., Zhang, C., Yang, Y., & Shen, B. (2026). Computational drug screening for drug-resistant tuberculosis: advances, challenges, and future perspectives. Frontiers in Pharmacology, 17, Article 1808051. https://doi.org/10.3389/fphar.2026.1808051

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1808051

Keywords: tuberculosis, drug-resistant TB, Mycobacterium tuberculosis, virtual screening, artificial intelligence, machine learning, molecular docking, host-directed therapy, combination therapy, MDR-TB, drug discovery, pharmacokinetics