Physicochemical Determinants of Gram-Negative Penetration and the Limits of Machine Learning in Antibacterial Discovery: A Theoretical Analysis

Abstract

Poor penetration and accumulation of compounds in Gram-negative bacteria is a principal reason that potent antibacterial leads fail, and machine-learning models are increasingly used to predict permeation and accumulation from chemical structure. This theoretical paper asks what such models can reasonably be expected to learn. 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 outer-membrane permeation and porin-mediated uptake, efflux transport and its inhibition, compound accumulation measurement and the prediction tools built upon it, and machine-learning discovery of antibacterial compounds, with a supplementary search on resistance burden and diagnostic 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 both that record and the corresponding PubMed record; sources failing verification were excluded rather than corrected by inference. Analysis proceeded by classifying each measurement in routine use according to the processes it collapses, and by theory construction from the resulting pattern. Three findings are reported. First, every label obtainable at the scale machine learning requires is a composite in which influx across the outer membrane, active efflux, partitioning between cellular compartments and target engagement are summed into one number, so the separate contributions are not identifiable and mechanistic interpretations of such models are unlicensed. Second, because the relative weight of these processes varies with organism, growth state and assay, transfer failure is governed by shift in that mixture rather than by chemical distance, which explains why accumulation rules established in one species need not hold in another whose dominant route of entry differs. Third, the inference a model licenses is fixed by its held-out design rather than by its reported accuracy. The paper develops a four-layer framework relating the quantity of interest to the available label, the confounds between them and the validation design required, distinguishes chemotype, species and condition failure as modes with distinct remedies, and derives six falsifiable propositions and eleven reporting standards. It is recommended that performance be reported separately by split design, that condition-annotated and organism-resolved measurement be prioritised over further accumulation of heterogeneous data, and that model applicability for African health systems be established in locally circulating isolates rather than assumed from reference strains.