Conclusions

In this study, biomimetic chromatographic descriptors were evaluated primarily through continuous regression modelling of PLD potency, with ordinal regression retained as a complementary risk-stratification analysis. Across these analyses, membrane affinity, represented by CHI IAM, emerged as the most informative and consistent descriptor within the studied dataset. CHI IAM-based models showed strong predictive performance in the continuous pEC₅₀ analysis and provided balanced classification performance in the ordinal regression analysis, supporting the relevance of experimentally measured membrane affinity for PLD liability assessment. The inclusion of additional descriptors, such as log _k_AGP or ionization-related variables, yielded only limited, model-dependent improvements. This suggests that protein-binding and charge-related descriptors may provide complementary distributional information, but their added value appears less consistent once membrane affinity is explicitly accounted for. Conventional lipophilicity descriptors, including log P and log D, showed weaker and more variable performance, highlighting the limitation of relying solely on bulk lipophilicity to describe membrane-associated processes involved in PLD. Importantly, CHI IAM provides an experimentally accessible surrogate of membrane-affinity behaviour, which is relevant to drug-phospholipid interactions and intracellular distribution processes associated with PLD. Thus, biomimetic chromatography can complement conventional physicochemical descriptors by adding experimentally measured interaction information that is not fully captured by calculated lipophilicity parameters. Overall, the findings support the use of CHI IAM and related biomimetic chromatographic descriptors as practical, interpretable tools for early-stage PLD liability assessment and compound prioritization. The proposed approach may serve as a decision-support layer to identify compounds at elevated risk of phospholipidosis and to guide further experimental evaluation. Future studies using larger, more diverse datasets, harmonized experimental endpoints, and additional external validation will be important for further defining the applicability domain and translational value of this modelling framework.