Experimentally derived biomimetic chromatographic descriptors for drug-induced phospholipidosis liability prediction

Drug-induced phospholipidosis (PLD) is a complex intracellular liability commonly associated with the lysosomal accumulation of cationic amphiphilic drugs and remains an important concern in drug development because of its implications for safety, intracellular disposition, and compound prioritization. Although cationic amphiphilicity, membrane interactions, and lysosomal trapping are established features of PLD, experimentally accessible descriptors that capture PLD-relevant membrane affinity remain valuable for early-stage risk assessment. This study aimed to evaluate whether biomimetic chromatographic descriptors can improve the prediction and interpretation of PLD liability by providing experimentally derived information on membrane affinity and protein-binding-related distributional behavior.

A curated dataset of 65 compounds with experimentally reported PLD responses was analyzed using multiple linear regression and ordinal classification approaches. PLD potency was expressed as pEC₅₀, while ordinal models classified compounds into non-inducers, weak/moderate inducers, and strong inducers. Membrane affinity was quantified using immobilized artificial membrane chromatography, expressed as CHI IAM, while protein-related interactions were represented by human serum albumin affinity, log k _HSA, and α₁-acid glycoprotein affinity, log k_AGP. Biomimetic chromatographic descriptors were compared with conventional physicochemical parameters, including log P and log D, to assess their relative predictive and mechanistic value.

CHI IAM-containing models showed strong predictive performance for PLD potency, with external predictivity reaching Q² _ext = 0.823. CHI IAM alone also performed strongly, achieving Q² _ext = 0.821, indicating that membrane affinity captures a major component of PLD-relevant behavior. In ordinal classification models, CHI IAM alone provided robust discrimination of PLD severity, with a test accuracy of 0.769, macro F1 score of 0.778, and macro-AUC of 0.923. Protein-binding descriptors, particularly log k_AGP, showed moderate predictive value but offered limited, model-dependent improvement once membrane affinity was considered. Conventional lipophilicity descriptors did not outperform the biomimetic membrane-affinity descriptor.

Biomimetic chromatographic descriptors provide experimentally derived, mechanistically interpretable surrogates for assessing early-stage PLD liability. CHI IAM offers a practical approximation of membrane-affinity behavior relevant to lysosomal phospholipid accumulation and PLD risk. Protein-binding descriptors may provide complementary information related to distributional phenotypes, but their added value appears less consistent than that of membrane affinity. This approach extends beyond conventional lipophilicity-based QSAR by incorporating experimentally measured distribution-relevant properties and may support compound prioritization during drug discovery.