Artificial intelligence is increasingly intersecting with antimicrobial pharmacology, offering new computational pathways to accelerate therapeutic discovery, refine dosing regimens, and address global infectious disease challenges. In a comprehensive review published in Frontiers in Bacteriology, researchers Jacob K. McPherson, Surbhi Sharma, Biswajit Mishra, Fadi Shehadeh, and Eleftherios Mylonakis from the Charles W. Duncan Jr. Department of Medicine at Houston Methodist systematically evaluated the capabilities and current limitations of machine learning and deep learning models applied to antimicrobial pharmacokinetics and pharmacodynamics. Pharmacokinetics, which tracks how drugs move through biological systems via absorption, distribution, metabolism, and excretion, represents a fundamental determinant of clinical success or therapeutic failure. Simultaneously, pharmacodynamics defines the relationship between drug concentrations and microbiological responses, prominently captured by the minimum inhibitory concentration. The investigators examined how computational architectures forecast critical pharmacokinetic and pharmacodynamic indices, assessing their predictive precision, mathematical interpretability, cross-dataset generalizability, and ultimate readiness for prospective translation into real-world antimicrobial stewardship and clinical decision-making across diverse medical environments.
The development of anti-infective therapeutics carries substantial financial exposure, even though its economic profile differs from broader pharmaceutical development. Citing macroeconomic analyses, the review authors noted that anti-infective drug development in the United States requires a mean expected capitalized cost of approximately 378.7 million dollars per agent, with a ninety-five percent confidence interval spanning 244.6 million to 556.1 million dollars, encompassing the costs of attrition and capital investments. While significantly lower than the cross-therapeutic pharmaceutical average of 879.3 million dollars per agent, this expenditure occurs in an environment where only ten percent of new molecular entities entering Phase I clinical trials ultimately secure regulatory approval. Infectious disease agents historically demonstrate higher transition success, achieving approximately twenty-five percent advancement from Phase I to commercial approval due to well-defined microbial targets and objective microbiological clearance endpoints. To mitigate high late-stage attrition, researchers increasingly evaluate structure-tissue exposure and selectivity-activity relationships alongside regulatory roadmaps from the Food and Drug Administration promoting computational in silico modeling to reduce preclinical animal testing.
To capture complex biological processes computationally, investigators have engineered multi-parameter artificial intelligence platforms capable of predicting absorption, distribution, metabolism, excretion, and toxicity profiles from chemical representations. The authors surveyed leading platforms including ADMET-AI, Deep-PK, pkCSM, ADMETlab, FP-ADMET, OptADMET, ADMETboost, HelixADMET, admetSAR, vNN-ADMET, and PKSmart, which deploy varied architectures ranging from graph neural networks and random forests to gradient boosting and multitask learning. For example, ADMET-AI evaluates benchmark datasets using multitask ensemble learning, reporting area under the receiver operating characteristic curve values near 0.84, while Deep-PK utilizes graph neural networks across curated datasets to achieve coefficient of determination values of approximately 0.72 on specific endpoints. Standardized initiatives like the Therapeutic Data Commons benchmark curated datasets across twenty-two distinct endpoints to harmonize model development, while platforms like pkCSM utilize graph-based signatures to evaluate permeability and solubility. However, the review emphasizes that platform comparisons remain inherently complex because training datasets vary substantially in size, chemical diversity, compound curation standards, and external validation rigor.
A central methodological question addressed in the review is how emerging artificial intelligence systems compare against established classical pharmacokinetic frameworks, notably nonlinear mixed-effects modeling implemented through NONMEM. Evaluating recent comparative research by Mao and colleagues, the authors highlighted investigations testing five machine learning algorithms, three deep learning frameworks comprising neural networks, gated recurrent units, and long short-term memory networks, and a neural ordinary differential equation model on simulated and empirical clinical trial datasets. Although advanced computational models demonstrated superior predictive accuracy over classical nonlinear mixed-effects models under controlled conditions, their performance exhibited marked sensitivity to clinical trial phases and underlying data distributions. The authors determined that while artificial intelligence tools represent powerful complementary mechanisms for pharmacokinetic analysis, they cannot yet supplant mechanistic mathematical frameworks in scenarios requiring reliable biological extrapolation, physiological parameter estimation, and interpretable clinical guidance, underscoring the enduring necessity of physics-based and physiologically informed modeling approaches within pharmacological research.
The divergence in predictive reliability becomes especially pronounced when examining specific pharmacokinetic endpoints, particularly the estimation of biological half-life. Whereas basic physicochemical parameters such as aqueous solubility, lipophilicity, and membrane permeability correlate closely with molecular structure and benefit from extensive in vitro experimental datasets, half-life reflects multifaceted systemic physiological interactions. It incorporates clearance mechanisms, plasma protein binding, tissue penetration, metabolic transformation, active transporter kinetics, and broad human physiological variation. Consequently, models attempting to predict biological half-life from chemical structure alone report modest and highly variable results. Deep-PK demonstrated a mean Matthews correlation coefficient of 0.32 across three-fold cross-validation, while ADMETlab 2.0 achieved a correlation coefficient of 0.478 and an area under the receiver operating characteristic curve of 0.801. Transitioning half-life prediction to continuous regression, ADMETlab 3.0 reached coefficients of determination between 0.653 and 0.748, whereas ADMET-AI and ADMETboost yielded average Spearman correlation coefficients of 0.239 and 0.396 on standardized benchmark splits.
Predicting antimicrobial pharmacodynamics introduces another layer of mathematical and biological complexity, centered on determining the minimum inhibitory concentration that governs core pharmacokinetic and pharmacodynamic exposure indices. Classical dosing efficacy relies heavily on three mainstay parameters: the duration that free unbound drug concentrations remain above the minimum inhibitory concentration, the ratio of maximum free drug concentration to that threshold, and the ratio of free drug area under the curve to the minimum inhibitory concentration. Large surveillance databases such as CARD, AMRFinderPlus, MEGARes, and ATLAS compile extensive genotype-phenotype measurements, but they frequently harbor inter-laboratory assay variations, inconsistent breakpoint definitions, and discordant data processing. The review highlights rigorous methodological solutions, noting Batisti Biffignandi and colleagues’ structured evaluation of continuous regression versus categorical classification for censored concentration thresholds, alongside log2 transformations. Similarly, the CRyPTIC Consortium examined over fifteen thousand Mycobacterium tuberculosis isolates, utilizing multiple reading modalities to incorporate measurement uncertainty into linear mixed models, uncovering 492 resistance-associated variants and 91 hypersensitivity mutations.
Beyond quantitative minimum inhibitory concentration modeling, artificial intelligence platforms increasingly classify broader antimicrobial resistance phenotypes from microbial whole-genome sequencing data. Distinguishing between quantitative susceptibility estimation and binary resistance classification, the review underscored studies demonstrating that genomic sequence features can accurately identify resistance determinants without relying on predefined mechanistic assumptions. Research by Ren and colleagues illustrated that machine learning algorithms and convolutional neural networks trained on frequency matrix chaos game representations achieved classification areas under the receiver operating characteristic curve reaching 0.96 in Escherichia coli. Parallel investigations in Acinetobacter baumannii by Gao and colleagues attained categorical agreement exceeding ninety-five percent using k-mer feature extraction, while Pataki and colleagues predicted fluoroquinolone susceptibility within two-fold and four-fold dilution margins using leave-one-country-out validation to isolate mutations in gyrA and parC genes. Nonetheless, investigators emphasize that data heterogeneity and uncaptured genetic factors—often termed missing heritability—continue to impose performance ceilings when models encounter divergent bacterial strains or variable experimental conditions.
Antimicrobial peptides represent a structurally and computationally distinct class of anti-infective agents that diverge from conventional small-molecule drug design paradigms. Rather than relying on classical molecular fingerprints or graph topologies, peptide-focused artificial intelligence models process primary amino acid sequences using protein language models, deep convolutional networks, and transformer encoders. Recent architectures highlight this technical transition: BERT-AmPEP60 leverages transformer embeddings to achieve Pearson correlation coefficients of 0.81 in Escherichia coli and 0.75 in Staphylococcus aureus, while ANIA incorporates Chaos Game Representations and Inception modules to yield correlation coefficients between 0.75 and 0.79 across multiple bacterial pathogens. Concurrently, multi-branch attention models like MBC-Attention achieve Pearson correlations of 0.775, and generative frameworks such as AMP-GAN explore de novo sequence generation. However, the authors emphasize that peptide databases, including APD, CAMP, and DRAMP, remain relatively small and heterogeneous, suffering from non-standardized synthesis methods, variable assay protocols, and ambiguous negative sample selection that complicate prospective translational validity.
The review also evaluated computational modeling across foundational physicochemical properties that govern antimicrobial distribution, namely the acid dissociation constant, aqueous solubility, and lipophilicity. Acid dissociation constants, evaluated through small-molecule models like Epik 7.0, MolGpKa, and pKaLearn or protein predictors like PyPKa and PKALM, dictate molecular ionization at infection sites such as abscesses or intracellular niches, directly controlling membrane penetration and target exposure. Aqueous solubility, essential for intravenous formulation feasibility and oral bioavailability, has evolved from descriptor-based tools like ESOL toward graph neural networks like Chemprop, contrastive frameworks like MolCLR, and transformer architectures like SolTranNet and SolvaFormer. Simultaneously, lipophilicity prediction models such as ALipSol, cx-Attentive FP, and RTlogD leverage retention time measurements and message-passing neural networks to estimate partition coefficients under physiological conditions. Because highly lipophilic molecules exhibit enhanced tissue distribution but suffer from elevated plasma protein binding that limits the pharmacologically active unbound fraction, accurate computational balancing of these physicochemical properties remains critical for achieving target exposures.
Translating artificial intelligence models into validated clinical tools requires bridging persistent gaps in algorithmic interpretability, standardization, and prospective experimental corroboration. Complex deep-learning systems often operate as opaque black boxes, providing limited biological insight into the molecular features driving predictions, although explainable techniques such as Shapley Additive Explanations, saliency maps, and applicability domain boundaries help identify key molecular descriptors and resistance determinants. Crucially, the review identified a profound scarcity of prospective in vitro and in vivo validation across the published literature, where most platforms rely strictly on retrospective internal cross-validation or historical benchmark datasets. To overcome these limitations, the authors advocate adherence to Findable, Accessible, Interoperable, and Reusable data standards, international data-sharing collaborations through initiatives like the CAMO-Net Global Data Repository, and cost-sensitive evaluation frameworks that weigh the clinical risks of false predictions. Developing transparent, experimentally corroborated computational models will ultimately determine whether artificial intelligence becomes an indispensable instrument in antimicrobial stewardship or remains an isolated computational exercise.
Subject of Research: Artificial intelligence models for antimicrobial pharmacokinetics, pharmacodynamics, and resistance prediction
Article Title: Black boxes and petri dishes: AI models for antimicrobial pharmacokinetics, drug properties, and resistance
Article References: McPherson, J. K., Sharma, S., Mishra, B., Shehadeh, F., & Mylonakis, E. (2026). Black boxes and petri dishes: AI models for antimicrobial pharmacokinetics, drug properties, and resistance. Frontiers in Bacteriology, 5, Article 1926678. https://doi.org/10.3389/fbrio.2026.1926678
Image Credits: AI Generated
DOI: 10.3389/fbrio.2026.1926678
Keywords: antimicrobial pharmacology, pharmacokinetics, pharmacodynamics, minimum inhibitory concentration, ADMET prediction, deep learning, graph neural networks, biological half-life, antimicrobial resistance, whole genome sequencing, antimicrobial peptides, explainable artificial intelligence, Therapeutic Data Commons, CAMO-Net GDR, CRyPTIC Consortium, acid dissociation constant, aqueous solubility
News Source: Blake Davidson. (November 16, 2026). AI Predicts Antimicrobial Properties but Lacks Prospective Experimental Validation Across Studies. Scienmag.



