Translating In Silico Predictions into Human Dose Feasibility in AI-Driven Antibacterial Discovery: A Conceptual Pharmacokinetic/Pharmacodynamic Framework

Abstract

Computational discovery pipelines identify antibacterial candidates primarily on potency and, increasingly, on predicted physicochemical and ADMET properties, yet many candidates are later abandoned because achievable human exposure cannot reach pharmacodynamic targets. This theoretical paper develops a conceptual framework linking in silico predictions of potency, clearance, distribution and protein binding to early judgements of human dose feasibility through pharmacokinetic/pharmacodynamic target attainment. The study is non-empirical and uses a conceptual research design. A corpus of peer-reviewed journal articles published between 2020 and 2026 was assembled across four literatures, namely antibacterial pharmacokinetics, pharmacodynamics and target attainment, translational and population dose prediction, machine-learning prediction of pharmacokinetic and ADMET properties, and computational antibacterial discovery, with a supplementary search on resistance burden, antibiotic access and laboratory capacity in sub-Saharan Africa. Every source was verified against its Crossref publisher record for authorship, journal, volume, pagination and year and screened for retractions, corrections and expressions of concern in that record and in the corresponding PubMed record; sources failing verification were excluded rather than corrected by inference. Analysis proceeded by decomposing the translational step into its constituent links, classifying each as predicted, measured or assumed, and reasoning from the resulting structure. Three findings are reported. First, because the links combine multiplicatively, prediction errors compound rather than average and each link contributes to the result in proportion to its own variance; separately, a link whose error has never been quantified cannot be propagated at all and therefore bounds what the calculation can claim, site penetration being the link most often assumed and least often quantified. Second, because feasibility is adjudicated at a threshold rather than estimated, the informativeness of in silico reasoning varies with a candidate's distance from the feasibility boundary and is lowest precisely for difficult cases. Third, feasibility is a two-sided window bounded below by target attainment and above by toxicity, so frameworks modelling only the efficacy bound systematically overestimate how many candidates are viable. The paper derives a prioritisation rule that evaluates each candidate under favourable and unfavourable assumptions, and establishes an asymmetry: in silico reasoning can reliably exclude candidates whose window is empty under optimistic assumptions but cannot reliably rank candidates whose windows are both non-empty. Six falsifiable propositions and ten reporting principles are presented. It is recommended that uncertainty be propagated rather than point-estimated, that error in currently assumed links be quantified as a priority, that predictive pharmacokinetic models be evaluated against decision accuracy rather than endpoint accuracy alone, and that for settings without routine therapeutic drug monitoring the width of the feasibility window be treated as a design criterion.