Introduction

Drug-induced liver injury (DILI) remains a leading cause of drug attrition and post-marketing withdrawal, with cholestatic and mixed phenotypes frequently linked to disruption of bile acid homeostasis [1]. The bile salt export pump (BSEP; ABCB11), an ATP-dependent transporter located at the canalicular membrane of hepatocytes, represents the rate-limiting step in bile acid secretion. Inhibition of BSEP can result in intracellular accumulation of bile acids, triggering hepatocellular stress, inflammation, and ultimately cholestasis [2]. Clinical and genetic evidence, including progressive familial intrahepatic cholestasis type 2 (PFIC2), supports a causal role of impaired BSEP function in liver injury [1,2].

In pharmaceutical development, BSEP inhibition is therefore considered a critical mechanistic liability [1,2]. However, translation of in vitro BSEP inhibition data into clinical risk remains challenging [1,3]. Standard vesicle-based assays provide estimates of inhibitory potency (e.g. IC50), but their predictive value is limited by several factors, including uncertainty in unbound concentrations, the absence of metabolic processes, and limited representation of hepatocellular drug accumulation [1,4]. Importantly, clinical DILI risk is not governed solely by inhibitory potency, but by the interplay between potency and effective exposure at the site of interaction [1,3,4]. As a result, contemporary frameworks emphasize exposure-adjusted metrics (e.g. Css/IC50: steady-state systemic exposure divided by the in vitro inhibitory concentration) yet estimating relevant intracellular and membrane-proximal concentrations remains a major unresolved challenge [1,2,4]. Notably, BSEP inhibition alone is not a universal predictor of clinical DILI, as hepatotoxic risk is modulated by additional factors, including intracellular accumulation, metabolism, and mitochondrial liability [3].

Several computational models have been developed to predict BSEP inhibition, including quantitative structure-activity relationship (QSAR) models, machine learning approaches, and structure-based methods. QSAR and classification models based on molecular descriptors and physicochemical properties have demonstrated moderate-to-good predictive performance for BSEP inhibition [5-12]. More recently, machine learning and deep learning approaches, including support vector machines and graph-based neural networks, have further improved prediction accuracy by capturing nonlinear structure-activity relationships [9,12]. Their broader predictive utility has historically been constrained by limited structural information, although recent structural studies have begun to clarify the molecular basis of bile salt extrusion and small-molecule inhibition in human BSEP [13].

Despite these advances, existing models primarily rely on calculated molecular descriptors or structural features and often do not explicitly incorporate experimentally derived distribution parameters such as membrane partitioning and plasma protein binding [4,14]. This is important because BSEP inhibition depends not only on intrinsic inhibitory potency, but also on the effective local concentration of the compound at the canalicular membrane and the unbound fraction available for transporter interaction [1,2,4]. Conventional physicochemical descriptors, such as log P and log D, provide useful approximations of lipophilicity and distribution behaviour, but they do not fully capture microenvironmental processes such as phospholipid affinity, protein binding, and ionization-dependent interactions [4,14]. Consequently, experimentally derived descriptors that reflect these properties may improve both the mechanistic interpretation and prediction of BSEP inhibition [1,4,14].

Biomimetic chromatography provides a practical and experimentally grounded approach to address this gap [14]. Immobilized artificial membrane (IAM) chromatography enables the quantification of phospholipid affinity, serving as a proxy for membrane partitioning and potential enrichment near membrane-bound targets such as BSEP [14,15,17]. Complementarily, human serum albumin (HSA) affinity chromatography provides insight into plasma protein binding, a key determinant of unbound systemic exposure and hepatic drug availability [14,15]. Together, these two orthogonal descriptors capture fundamental aspects of drug distribution: membrane affinity (IAM) and free fraction modulation (HSA) [14-16].

Although biomimetic chromatography has been successfully applied in pharmacokinetic modelling, including the prediction of volume of distribution and tissue binding, its application to transporter-mediated toxicity, particularly BSEP inhibition, remains comparatively underexplored relative to QSAR and machine learning approaches [14,16]. Given that effective inhibition depends on both intrinsic potency and local exposure at the membrane interface, experimentally derived IAM and HSA retention parameters may provide mechanistically relevant predictors that complement traditional in vitro assays [1,4,14].

While the ultimate concern in drug discovery is clinical drug-induced liver injury (DILI), the present study focuses on BSEP inhibition as a mechanistically defined and experimentally tractable endpoint [1-3]. BSEP inhibition is a well-established mechanistic contributor to cholestatic DILI and is supported by clinical, genetic, and regulatory evidence as a key liability in drug development [1,2]. In contrast, clinical DILI is a multifactorial and heterogeneous outcome influenced by numerous factors, including metabolism, immune response, mitochondrial toxicity, exposure, dose, and patient-specific variability, thereby complicating direct modelling and reducing mechanistic interpretability [1,3,4]. Consequently, mechanistic intermediate endpoints such as BSEP inhibition provide a more controlled and interpretable framework for investigating structure-activity and distribution-toxicity relationships [1,4]. Direct modelling of clinical DILI would be valuable as a later-stage objective but would require larger curated datasets and additional descriptors that reflect complementary toxicity mechanisms [3,4]. The present approach is therefore positioned as a mechanistically informed screening layer that complements, rather than replaces, integrated DILI risk assessment strategies [1,3,4].

In this study, we propose an experimentally grounded, chromatography-informed modelling framework to predict BSEP inhibition using IAM and HSA descriptors. By integrating experimentally measured phospholipid and protein affinity with statistical modelling approaches, we aim to (i) improve prediction of BSEP inhibitory potency, (ii) capture microenvironmental exposure effects not represented by conventional physicochemical parameters, and (iii) provide an interpretable and scalable tool for early compound prioritization. This approach is positioned as a decision-support layer within an integrated DILI risk assessment, supporting a more informed selection of compounds for confirmatory transporter- and hepatocyte-based assays.