Introduction

Drug-induced phospholipidosis is a lysosomal lipid storage disorder characterized by excessive intracellular accumulation of phospholipids and the formation of concentric lamellar bodies detectable by transmission electron microscopy [1,2]. The phenomenon has been observed across several tissues, including liver, lung, kidney, and nervous tissue, and remains relevant in drug development because it can complicate preclinical safety interpretation and compound prioritization [1-4]. Although PLD is often considered an adaptive and potentially reversible cellular response rather than a direct toxicological endpoint, its association with altered lipid metabolism, intracellular drug accumulation, and secondary cellular dysfunction underscores the importance of early identification of phospholipidogenic compounds in drug discovery [3,4].

The most widely accepted mechanistic framework for PLD involves cationic amphiphilic drugs, which combine lipophilicity with one or more protonatable amine groups [5-7]. These compounds can passively diffuse across cellular membranes and subsequently accumulate in acidic lysosomal compartments via ion trapping, in which protonation reduces their ability to diffuse back into the cytosol [5,6]. Lysosomal accumulation is further promoted by interactions with phospholipid-rich membranes, leading to drug-phospholipid complex formation, impaired phospholipid degradation, and inhibition or functional disruption of lysosomal phospholipases [4-6]. Previous studies have shown that PLD liability is influenced by the interplay between lipophilicity, basicity, and distribution-related properties, including volume of distribution [7,8]. However, simple rules based on calculated logarithm of the octanol-water partition coefficient (log P), the negative logarithm of the acid dissociation constant (p_K_a) or cationic amphiphilicity are not sufficient to explain all cases, because structurally diverse and non-classical CAD compounds may also induce phospholipid accumulation [8].

Several experimental approaches have therefore been developed to detect or predict PLD liability. Cell-based assays remain among the most biologically relevant methods because they directly assess intracellular phospholipid accumulation. In an early and influential study, Casartelli et al. [9] demonstrated the utility of a cell-based approach for assessing phospholipidogenic potential in pharmaceutical research and drug development. Subsequent fluorescence-based assays using probes such as NBD-PE and high-content imaging platforms further improved the quantitative assessment of intracellular lipid accumulation and enabled better distinction between PLD induction and cytotoxicity-related effects [10,11]. These methods provide valuable biological information and can support compound safety profiling; however, they are more resource-intensive than simple physicochemical or chromatographic approaches and may not always be optimal for rapid early-stage screening.

Non-cell-based methods have also been proposed to evaluate drug-phospholipid interactions as surrogate indicators of PLD risk. Fluorescence-based lipid interaction assays, including methods using Prodan and other probes, can detect drug-induced perturbation of lipid systems and have shown useful correlations with PLD-related endpoints [12]. Similarly, biochemical and physicochemical assays, combined with multivariate analysis, have been used to estimate the phospholipidosis-inducing potential [13]. These approaches support the concept that PLD is not governed by a single molecular property, but by the combined influence of amphiphilicity, ionization, membrane affinity, and intracellular distribution.

Chromatographic approaches have also been explored for early PLD screening. Jiang and Reilly proposed chromatographic methods for assessing phospholipidosis-inducing potential, showing that retention behavior in selected chromatographic systems could provide useful physicochemical information for early screening [14]. More specifically, IAM-related chromatographic approaches have already been connected to PLD. Zhao et al. developed a mixed phospholipid-functionalized monolithic column and reported that CHI IAM 7.4 values were highly correlated with drug-induced PLD potency in a set of marketed drugs. More recently, Wang et al. developed an acidic phospholipid-containing immobilized artificial membrane column designed to better mimic the negatively charged lysosomal phospholipid environment and improve prediction of drug-induced PLD potency. These studies clearly establish that IAM-based membrane-affinity measurements are relevant to PLD assessment. Therefore, the present study lies not in the first application of IAM chromatography to PLD, but in the systematic comparative evaluation of standard biomimetic chromatographic descriptors against conventional physicochemical descriptors within a unified modelling framework.

In parallel with experimental approaches, computational models have been developed for PLD prediction, including classical QSAR, rule-based methods, and machine-learning models [8,17-19]. These methods can support high-throughput compound screening and have demonstrated useful predictive performance. However, many models rely mainly on calculated descriptors, structural fingerprints, or lipophilicity/basicity rules [8,17-20]. While such descriptors are valuable, they do not directly measure membrane affinity, phospholipid interaction, or protein-binding-related distributional behaviour. This limitation is important because PLD is fundamentally a distribution-driven intracellular phenomenon involving membrane partitioning, lysosomal accumulation, and drug-phospholipid interactions [4-7,20].

Biomimetic chromatography provides experimentally derived descriptors that can complement conventional calculated physicochemical parameters. Immobilized artificial membrane chromatography quantifies the affinity of compounds for phospholipid-like environments and has been widely used as a surrogate for membrane partitioning, permeability, tissue binding, and related ADMET properties [21-24]. Human serum albumin and α1-acid glycoprotein chromatography provide additional descriptors of plasma protein binding and systemic distribution [21-24]. These chromatographic measurements do not directly reproduce the acidic lysosomal environment; rather, they provide experimentally accessible surrogates for drug-membrane and drug-protein interactions that influence compound disposition. In the context of PLD, CHI IAM is particularly relevant because membrane affinity and phospholipid interaction are central features of lysosomal phospholipid accumulation. Protein-binding descriptors may provide complementary information about distributional phenotypes, especially for basic and amphiphilic compounds, although their mechanistic role in PLD should be interpreted with caution.

Building on previous PLD screening and IAM-based studies [14-16], the present work evaluates standard biomimetic chromatographic descriptors, including the chromatographic hydrophobicity index measured by immobilized artificial membrane chromatography (CHI IAM), the logarithmic retention factor on human serum albumin chromatography (log _k_HSA), and the logarithmic retention factor on α₁-acid glycoprotein chromatography (log _k_AGP), alongside conventional physicochemical descriptors in a curated dataset of compounds with experimentally reported PLD responses. The study compares these biomimetic descriptors with log P, the logarithm of the pH-dependent distribution coefficient (log D), molecular size, hydrogen-bonding descriptors, and ionization fractions to determine whether experimentally measured membrane affinity and protein-binding descriptors provide additional predictive and mechanistic value beyond conventional lipophilicity-based QSAR. In addition, both continuous and ordinal modelling strategies are applied. Multiple linear regression is used to model PLD potency, expressed as pEC₅₀, i.e. the negative logarithm of the EC₅₀ value expressed in molar units, where EC₅₀ denotes the concentration producing half-maximal intracellular phospholipid accumulation under the assay conditions. Ordinal regression is used to classify compounds into ordered PLD severity categories, namely non-inducers, weak/moderate inducers, and strong inducers.

Accordingly, the aim of this study was to reassess and extend the role of biomimetic chromatographic descriptors in PLD liability prediction using a comparative ADMET modelling framework. By integrating experimentally derived membrane-affinity and protein-binding descriptors with continuous potency modelling and ordinal severity classification, the study seeks to clarify whether CHI IAM and related biomimetic descriptors can serve as practical, interpretable surrogates for PLD-relevant distributional behaviour. This approach is intended not to replace cell-based PLD assays, but to provide an early-stage decision-support tool to identify compounds with elevated liability for phospholipidosis and to prioritize them for further experimental evaluation.