Transitioning from Transcriptomics to Proteomics: Enhancing Mechanistic Accuracy in PBPK Modeling via Absolute Protein Abundances

ABSTRACT

Reliable physiologically based pharmacokinetic (PBPK) modeling depends on tissue‐specific expression profiles that reflect protein activities governing drug disposition. In PK‐Sim, existing expression databases rely on transcriptomics data. However, mRNA levels often exhibit limited correlation with protein abundance, frequently necessitating empirical expression modification to align bottom‐up simulations with clinical observations. To address this limitation, we developed ProteinDB as a proteomics‐based expression database for PK‐Sim, using proteomics data primarily from PaxDb v6.0. Raw proteomics data were mapped to gene identifiers and standardized into absolute concentrations (μmol/L tissue) before integration into PK‐Sim. Cross‐platform comparisons were performed for hepatic protein abundance across PBPK platforms, while cross‐omics comparisons were made of relative tissue distributions with transcriptomics‐based PK‐Sim databases. The performance of ProteinDB in PBPK modeling was evaluated using the probe substrates midazolam, digoxin, rifampicin, and tizanidine, with associated drug–drug interactions. Cross‐platform comparisons showed strong agreement for most hepatic enzymes and transporters, while revealing divergences for proteins with greater inter‐individual variability, lower abundance, or limited evidence base. Cross‐omics analyses demonstrated tissue‐dependent discrepancies between transcript‐ and protein‐based expression patterns, with higher consistency observed for kidney and small intestine, particularly with the RT‐PCR and Bgee databases. For PBPK modeling, ProteinDB showed consistently comparable or superior predictive performance for systemic exposure and other clinical endpoints compared with transcriptomics‐based baseline and empirically modified library profiles. By providing a direct physiological basis for system parameterization, ProteinDB offers a robust alternative to current transcriptomics PK‐Sim databases and reduces the reliance on empirical expression modification, thus improving the reliability of prospective PBPK modeling.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    The current PK‐Sim platform primarily relies on transcriptomics‐derived tissue expression databases, despite the growing availability of large‐scale absolute proteomics datasets that may better represent functional protein abundance.
• What question did this study address?

  • ○
    This study developed a proteomics‐based whole‐body expression database (ProteinDB), contextualized its role in the current PBPK landscape, and evaluated whether it can improve PBPK model performance compared with existing transcriptomics‐based expression databases.
• What does this study add to our knowledge?

  • ○
    Absolute ADME protein abundances in ProteinDB, derived from large‐scale proteomics, showed overall agreement with curated PBPK platform databases. Transcriptomics‐based databases exhibited systematic bias and compressed tissue expression ranges compared with proteomics‐derived profiles.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    Proteomics‐based expression databases such as ProteinDB may provide a more mechanistically relevant foundation for modeling in PK‐Sim, supporting prospective PBPK prediction, clinical extrapolation, and drug‐drug interaction assessment.

1. Introduction

Quantitative characterization and prediction of human pharmacokinetics (PK) are central objectives in drug metabolism and disposition research. A mechanistic understanding of absorption, distribution, metabolism, and excretion (ADME) processes, particularly those mediated by drug‐metabolizing enzymes, transporters, and target protein binders, combined with the physicochemical properties of drug molecules and human system‐specific information, is essential for translating in vitro findings to clinical outcomes [1]. In this context, physiologically based pharmacokinetic (PBPK) modeling has emerged as an established framework for integrating drug‐specific properties with system‐level physiology to quantitatively characterize and predict human PK, assess drug–drug interaction (DDI) risk, and inform clinical study design [2, 3].

PBPK models describe drug disposition through explicit representation of anatomical compartments and physiological processes within the virtual subject [4]. In bottom‐up PBPK modeling, in vitro kinetic parameters characterizing drug metabolism, transport, or target binding are scaled to the organ level using organ‐specific protein abundances [5, 6]. Therefore, uncertainty or systematic differences in expression data can propagate into pathway attribution, organ‐specific contribution, and the predicted magnitude of enzyme‐ or transporter‐mediated DDIs [7].

Major PBPK platforms all include tissue expression databases for ADME proteins, with differences in data source, transparency, and implementation. Simcyp (Certara UK Ltd., Sheffield, UK) uses a curated absolute abundance database that is periodically updated and verified [8]. Similarly, GastroPlus (Simulations Plus Inc., Lancaster, CA, USA) implements an internal protein expression database that provides predefined tissue abundance values, with a specific focus on the gastrointestinal tract [9]. In contrast, the PK‐Sim platform (Open Systems Pharmacology suite, http://www.open‐systems‐pharmacology.org) has historically relied on user‐defined protein abundances, with predefined values available for selected cytochrome P450 (CYP) enzymes, and transcriptomics‐based relative expression data to map tissue‐level abundances [7].

Currently, four transcriptomics‐based databases are available in PK‐Sim, including microarray datasets (ArrayExpress), reverse transcription polymerase chain reaction (RT‐PCR) data reported in the literature, expressed sequence tag (EST) data from UniGene [7], and more recently bulk RNA‐seq datasets from Bgee [10]. Although transcriptomic datasets offer broad tissue coverage and accessibility, mRNA levels do not necessarily quantitatively reflect absolute protein concentrations due to post‐transcriptional regulation [11, 12], variability in translational efficiency [13, 14], and differences in protein turnover [15, 16]. In addition, defining reference concentrations in transcriptomics‐based PK‐Sim frameworks often requires manual curation, subjective selection of reference values, and scaling assumptions from heterogeneous literature sources, which can introduce artificial variability and limited reproducibility.

Advances in large‐scale quantitative proteomics have enabled broad and direct quantification of protein abundances across human tissues, creating new opportunities to improve the physiological basis of PBPK system parameters [17, 18]. Compared with transcriptomics‐derived inputs, proteomics data offer a more direct representation of the functional protein molecules that govern drug metabolism, transport, and target interaction. However, the integration of proteomics‐derived abundance data into PBPK software remains limited and uneven, and differences in protein abundance values across available databases, as well as differences between transcript‐ and protein‐based tissue distribution patterns, have not been systematically characterized.

To address these challenges, we developed ProteinDB, a curated proteomics‐based expression database for direct integration into PK‐Sim. ProteinDB provides harmonized absolute organ‐level protein abundances derived from large‐scale quantitative proteomics datasets and aligned with the current PK‐Sim framework. To contextualize this expression database within the broader PBPK landscape, we compared liver‐specific protein abundances across documented expression datasets from PK‐Sim, Simcyp, and GastroPlus, and evaluated correlations of organ‐specific distribution patterns between ProteinDB and the existing transcriptomics‐based databases in PK‐Sim. Finally, we evaluated the performance of ProteinDB in PBPK modeling using established probe substrate models and associated DDI scenarios in PK‐Sim. Through these analyses, we examined whether the incorporation of proteomics‐derived absolute abundance data could improve consistency, transparency, and predictive robustness in bottom‐up PBPK modeling.

2. ProteinDB Construction and Evaluation

2.1. Database Construction

Construction of ProteinDB incorporated quantitative proteomics data from PaxDb v6.0 [19] for standard solid organs and plasma, together with a specialized erythrocyte dataset from Bryk et al. [20]. To comply with the architectural constraints of the PK‐Sim platform, which organizes protein expression data by unique gene identifiers, a gene‐centered mapping strategy was employed. Raw proteomics data were first standardized by removing formatting artifacts from associated gene symbols (e.g., species annotation such as “_HUMAN”). To resolve protein groups formed by ambiguous peptide‐to‐protein mappings (e.g., highly conserved paralogs such as the tubulin family) in the erythrocyte dataset, only the primary leading gene was retained from delimited lists of gene symbols. Cleaned gene symbols were then mapped to Ensembl gene identifiers (ENSG) using the Ensembl BiomaRt human gene dataset (Ensembl Release 115) [21], and associated annotations including Entrez Gene IDs and official full names were retrieved. When direct mapping by the cleaned gene symbols failed, available protein‐level identifiers, including Ensembl protein IDs or UniProt IDs, were used to rescue the corresponding gene. The resulting mappings were then checked against the original gene annotations, and entries were retained when discrepancies could be attributed to recognized synonyms. Entries that could not be assigned to a valid gene symbol or ENSG ID were excluded. To prevent multiple records for the same gene within a given tissue, arising from isoforms, redundant identifiers, or overlapping peptide assignments, a maximum‐abundance rule was applied, such that only the entry with the highest reported abundance was retained for each gene in each tissue‐specific dataset. The duplicated records for each organ, along with their corresponding mean and sum values, are provided in Data S1, and the complete list of duplicated genes is available in the Supporting Information. Protein abundance values from both sources were converted to absolute tissue concentrations (μmol/L) using a unified scaling framework (See Supporting Information). The overall ProteinDB construction, integration, and evaluation workflow is summarized in Figure 1.

FIGURE 1.

Workflow for ProteinDB development, integration, and evaluation within PK‐Sim.

Finally, the compiled database was implemented as a relational SQLite .expressionDB file, following the schema described by Cordes and Rapp for developing expression databases for PK‐Sim [10]. The ProteinDB database file and associated code are openly available in the GitHub repository at https://github.com/translatPK‐KUL/ProteinDB.

2.2. Cross‐PBPK Platform Comparison of Hepatic Protein Abundance Values

To contextualize ProteinDB within the current PBPK modeling landscape, hepatic ADME protein abundances were compared with those implemented in mainstream platforms, including PK‐Sim v12, Simcyp V22, and GastroPlus X.2. The comparison focused on the ADME proteins available in the platform databases, including CYP enzymes, UDP‐glucuronosyltransferases (UGTs), additional metabolic enzymes, and clinically relevant transporters (the complete protein list is provided in the Supporting Information). Simcyp abundances were taken from the White American population. GastroPlus abundances were extracted for a virtual population representing 30‐year‐old healthy American males, and PK‐Sim CYP abundances were obtained from platform default reference values derived from the Rodrigues dataset [22]. To enable direct cross‐platform comparison, all abundance values were harmonized to absolute concentrations (μmol/L tissue) using the scaling conventions implemented in the respective platforms. Comparative analyses were performed separately for CYPs, UGTs, other metabolic enzymes, and transporters.

2.3. Cross‐Omics Comparison of Liver‐Normalized Abundance Between ProteinDB and Transcriptomics PK‐Sim Expression Databases

To examine how ProteinDB tissue expression patterns relate to the transcriptomics‐derived profiles in PK‐Sim, ProteinDB was compared with the four existing PK‐Sim expression databases, EST, Array, RT‐PCR, and Bgee [7, 10]. Organ‐specific expression patterns in the small intestine, kidney, and brain were evaluated in relation to the liver, due to its central role in drug metabolism and transport and its extensive characterization in PBPK modeling. The analyzed protein set comprised the union of proteins represented across three mainstream PBPK platform databases (Supporting Information), reflecting the most PBPK‐relevant ADME proteins. Tissue abundance data were extracted from ProteinDB, GENEDB_human Version 2.0.0, and GENEDB_human_TPM_ONLY_Bgee Version 3.0.1 (https://github.com/Open‐Systems‐Pharmacology/Gene‐Expression‐Databases/releases). To remain consistent with the mapping and data‐handling logic implemented in PK‐Sim, entries annotated only as “INTESTINE” were treated as “SMALL INTESTINE” when no finer subdivision was available, and the arithmetic mean was used when multiple records were reported.

2.4. ProteinDB Evaluation in Single‐Compound PBPK Modeling

Midazolam, digoxin, rifampicin, and tizanidine were selected as probe compounds to evaluate ProteinDB based on their specific disposition mechanisms, including CYP3A4‐, CYP1A2‐, UGT1A4‐ and arylacetamide deacetylase (AADAC)‐mediated metabolism, P‐glycoprotein (P‐gp)‐ and organic anion‐transporting polypeptide 1B1 (OATP1B1)‐mediated transport, gamma‐aminobutyric acid type A receptor subunit gamma 2 (GABRG2)‐ and ATPase Na⁺/K⁺ transporting subunit alpha 2 (ATP1A2)‐mediated target‐mediated drug disposition, and dynamic multi‐pathway modulation. PBPK models for those compounds were reconstructed in PK‐Sim using their published structural frameworks from evaluation reports in the PK‐Sim library (https://github.com/Open‐Systems‐Pharmacology/OSP‐PBPK‐Model‐Library) and the original DDI network study [23]. To isolate the impact of the underlying protein expression sources, model redevelopment was strictly limited to updating the expression‐level inputs. Specifically, the comparative simulations evaluated distinct expression profiles. The baseline profile employed the original transcriptomics expression data without any manual modification. The library profile used transcriptomics‐based expression data with empirical adjustments, including a 3.57‐fold increase in intestinal P‐gp expression and restriction of UGT1A4 expression to liver. The ProteinDB profile incorporated absolute proteomics abundance data using ProteinDB. The tizanidine model was selected to carry out purely bottom‐up simulations incorporating all the originally reported physicochemical properties and in vitro ADME kinetics to test the ProteinDB performance without any parameter optimization. For the purely bottom‐up tizanidine simulation, empirically optimized formulation and clearance parameters from the original library model were replaced with literature‐informed inputs. The dissolution process was represented using a Lint80 formulation, corresponding to 80% dissolution within 30 min [24], and CYP1A2‐mediated hepatic clearance was assigned the reported value of 17 mL/min/kg [25]. No additional parameter optimization was performed.

All relevant active processes described in the library model verification reports were implemented accordingly, including metabolism, transport, protein binding, inhibition, and induction where applicable. After establishing the models with those three expression profiles, a standardized parameter identification procedure was applied, consisting of an initial Monte‐Carlo optimization followed by a Levenberg–Marquardt optimization (both with a maximum of 10,000 iterations). For each compound, the optimization space remained entirely constant across all expression profiles according to the respective PK‐Sim library model verification report and the original PBPK modeling study [23]. This standardization included the use of identical training data, estimable parameters, initial values, optimization boundaries, partition coefficient and cellular permeability calculation methods across all simulations; where the optimization setting was not available, the software default was applied (Table S1). The inhibition and induction parameters of rifampicin were sourced from the verified library model. To be consistent with the original optimization strategy, interaction parameters grounded in the literature were kept fixed, and only parameters lacking experimental data were optimized during the single‐compound model development (Table S2).

2.5. ProteinDB Evaluation in DDI Modeling

Rifampicin‐midazolam and rifampicin‐digoxin co‐administration simulations were performed to evaluate ProteinDB in clinically relevant DDI scenarios. For each DDI scenario, the model structures and interaction mechanisms described in the original study were strictly retained [23]. This included the implementation of relevant perpetrator effects of rifampicin, such as competitive inhibition and induction, together with the victim elimination and transport processes governing midazolam and digoxin disposition. To isolate the impact of the protein expression, all simulation settings remained entirely identical across the three distinct expression profiles. The interaction parameters were strictly preserved and were not further optimized following the integration into DDI models (Table S2).

2.6. Statistics

Cross‐platform comparisons of absolute hepatic protein abundances were evaluated using Pearson correlations calculated on log₂‐transformed abundances, Spearman rank correlations, fold agreement (within 2‐fold and 5‐fold), and geometric mean fold error (GMFE). Leave‐one‐out and leave‐two‐out sensitivity analyses were performed to assess the influence of highly discordant proteins on class‐level GMFE. For cross‐omics relative tissue distribution analyses, tissue abundances were normalized to liver within each database, and similarity between databases was assessed using Pearson correlations calculated on log₂‐transformed values, Spearman rank correlations, together with 2‐fold and 5‐fold agreement. Statistical significance was defined as ns for p ≥ 0.05, * for p < 0.05, ** for p < 0.01, and *** for p < 0.001. PBPK model performance was evaluated using mean relative deviation (MRD) for concentration‐time profiles and other observational endpoints, and GMFE for predicted‐to‐observed AUC and Cmax ratios.

2.7. Software

Raw proteomics data processing, database construction, expression analysis, plotting, and statistical analyses were performed in R (version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria) and RStudio (version 2025.05.1+513; Posit Software, Boston, MA, USA). PBPK model development and simulations were performed using PK‐Sim (Open Systems Pharmacology Suite, version 12.1). A workflow diagram was created with http://BioRender.com.

3. ProteinDB Content and Comparative Evaluation

3.1. Overview of ProteinDB

The constructed ProteinDB database provides gene‐centered protein abundance data across 20 human tissues corresponding to PK‐Sim compartments, thereby enabling broad whole‐body representation for PBPK modeling. Across all tissue datasets, an initial pool of 18,561 globally unique raw gene symbol‐based entries was processed. During the mapping phase, 44 unique entries (0.24%) were excluded due to unresolvable ENSG IDs based on the raw gene symbols or available protein identifiers. The remaining 18,517 valid gene symbols converged onto 18,377 unique ENSG loci, indicating the presence of 140 synonymous gene identifiers. To satisfy the strict 1:1 gene‐centered structure required by PK‐Sim, a pragmatic maximum‐abundance rule was applied to resolve multiple records mapping to the same ENSG IDs, successfully deduplicating the redundant protein entries from 129 (0.70%) ENSG IDs.

The complete list of these 129 duplicated genes, along with the raw duplicated records and their corresponding arithmetic and PaxDb‐weighted mean and sum values for each organ, is provided in the Supporting Information and Data S1. Cross‐referencing these 129 duplicated genes with the extended ADME gene list (298 genes) by the PharmaADME consortium revealed that only three genes—HAGH (in blood cells only), SOD2 (in 18 organs except for blood cells and fat), and SLC29A1 (ENT1, in blood cells only)—were subject to deduplication [26]. A detailed sensitivity analysis of the duplicated records is provided in the Supporting Information.

Ultimately, this yielded a consolidated database of 18,377 globally unique gene‐centered protein entries across tissues. All core ADME‐related proteins across mainstream PBPK platforms (see Supporting Information) were successfully retained without attrition during this deduplication process. The final dataset was implemented as a relational SQLite file comprising 19 interconnected tables, storing information for gene symbols, Entrez Gene IDs, official full names, protein abundances, and other supporting data. All data, including study IDs (ProteomeXchange accession with author and year), assay types (quantification method), detection frequency (total number of constituent studies and number of studies in which the protein was detected), and weighted coefficient of variation (weighted CV%), are provided in Data S2.

The final protein abundance values were stored as absolute concentrations (μmol/L tissue), allowing direct integration into PK‐Sim models without additional scaling. Tissue names were mapped to PK‐Sim anatomical compartments, including assignment to predefined containers such as liver sub‐compartments, intestinal segments, and blood compartments. The ENSG‐centered design ensures unique gene representation and consistent linkage of expression values across tissues.

3.2. Cross‐Platform Comparison of Hepatic Protein Abundances

Cross‐platform comparisons of hepatic ADME protein abundances revealed strong protein‐class dependencies. CYP enzymes demonstrated the highest overall consistency across ProteinDB, PK‐Sim v12, Simcyp V22, and GastroPlus X.2. GastroPlus achieved the closest quantitative CYP agreement with 80% of proteins within a 2‐fold boundary and a GMFE of 1.85 (Table 1). PK‐Sim and Simcyp also showed strong CYP concordance with 5‐fold agreements exceeding 85% and GMFEs below 2.8. However, the correlations for CYPs remained only moderate across all platforms (Pearson r = 0.57–0.79; Spearman ρ = 0.65–0.72).

TABLE 1.

Cross‐platform comparison of hepatic protein abundances based on the shared protein intersection within each protein class in each database pair.

	N overlap	Pearson r (log2)	Spearman ρ	% within 2‐fold	% within 5‐fold	GMFE	Minimum leave‐one‐out GMFE (excluded protein)*	Minimum leave‐two‐out GMFE (excluded proteins)*
PK‐Sim v12
CYP	11	0.57	0.70	63.64	90.91	2.42	1.84 (CYP3A5)	1.68 (CYP3A5; CYP2B6)
Simcyp V22
CYP	14	0.79	0.72	28.57	85.71	2.71	2.34 (CYP2C18)	2.14 (CYP2C18; CYP2C19)
UGT	11	0.87	0.95	72.73	81.82	2.63	1.64 (UGT1A10)	1.44 (UGT1A10; UGT1A3)
Other enzymes	8	0.67	0.81	50.00	87.50	2.37	1.80 (SULT1A2)	1.62 (SULT1A2; FMO3)
Transporters	16	0.57	0.58	50.00	81.25	2.93	2.24 (BCRP)	1.99 (BCRP; MRP6)
GastroPlus X.2
CYP	10	0.68	0.65	80.00	100.00	1.85	1.69 (CYP2D6)	1.55 (CYP2D6; CYP2C8)
UGT	12	0.83	0.82	41.67	58.33	5.08	3.89 (UGT1A7)	2.94 (UGT1A7; UGT1A5)
Other enzymes	13	0.61	0.61	46.15	76.92	3.29	2.56 (GSTP1)	1.98 (GSTP1; AOX1)
Transporters	11	0.84	0.77	36.36	81.82	2.97	2.20 (BCRP)	2.00 (BCRP; MRP3)

^*The lowest GMFE achieved by excluding one and two protein pairs between the databases.

UGT enzymes displayed a contrasting pattern of strong correlation but diverse quantitative variability. UGTs in Simcyp exhibited robust rank‐order correlation (Spearman ρ = 0.95) and high 2‐fold agreement (73%). Conversely, UGTs in GastroPlus yielded the highest GMFE (5.08) and the lowest 5‐fold agreement (58%) despite maintaining good correlations. Other metabolic enzymes and transporters generally showed intermediate similarity, achieving roughly 40% to 50% agreement at the 2‐fold threshold and approximately 80% at the 5‐fold threshold across Simcyp and GastroPlus.

Sensitivity analyses demonstrated that a small number of highly discordant proteins drove much of the quantitative divergence across all platforms (Table 1; Figure 2). Specifically, CYP3A5 was 36‐fold higher in ProteinDB compared to PK‐Sim. BCRP abundance was roughly 170‐fold lower in ProteinDB than in Simcyp but over 50‐fold higher than in GastroPlus. Low‐abundance UGTs such as UGT1A7 and UGT1A5 were markedly lower in GastroPlus, whereas UGT1A10 was roughly 300‐fold lower in Simcyp. Excluding just one or two extreme outliers consistently reduced the GMFE values to 2.0 or below across nearly all comparisons. To understand the origin of these discrepancies, the intra‐PaxDb inter‐study variability was examined for the 17 proteins collectively identified by the leave‐two‐out analysis (Table S3). For 16 of these 17 proteins, the maximum cross‐platform ratio fell within the inter‐study variability range already observed across PaxDb constituent datasets, indicating that the discrepancies largely reflect the inherent variability of quantitative proteomics data and the divergent source databases underlying each PBPK platform rather than systematic errors in the ProteinDB curation pipeline. The sole exception was UGT1A10, for which the 293‐fold cross‐platform discrepancy with Simcyp V22 exceeded the intra‐PaxDb max/min ratio of 105‐fold, suggesting that the Simcyp‐embedded value for this enzyme may originate from a substantively different source dataset or quantification framework.

FIGURE 2.

Hepatic abundances of CYPs (A), UGTs (B), other enzymes (C), and transporters (D) in ProteinDB, GastroPlus X.2, PK‐Sim v12, and Simcyp V22.

3.3. Cross‐Omics Comparison of Relative Tissue Distribution Patterns

Cross‐omics comparison of liver‐normalized tissue expression identified systematic differences between transcriptomics databases and ProteinDB (Figure 3). Most notably, transcriptomics databases exhibited narrower dynamic ranges than ProteinDB. This compression was accompanied by systematic shifts in relative abundance, with transcriptomics databases consistently assigning substantially higher brain‐to‐liver expression ratios. EST and Array also assigned higher small intestine‐ and kidney‐to‐liver ratios than ProteinDB. In contrast, RT‐PCR and Bgee showed broader distributions and closer correspondence with ProteinDB, particularly for small intestine and kidney.

FIGURE 3.

Pairwise comparison of liver‐normalized relative protein expression patterns between ProteinDB and transcriptomics databases in the small intestine, kidney, and brain. (A) Scatter plots comparing the liver‐normalized relative expression values from ProteinDB with the values from the Array, EST, RT‐PCR, and Bgee databases. Each point represents a shared protein and is colored by tissue. The solid diagonal line indicates identity, and the dotted and dashed lines indicate 2‐fold and 5‐fold deviations. (B) Heat maps summarizing Pearson and Spearman rank correlations and fold‐based agreement between ProteinDB and each transcriptomics database across tissues. Fold‐based agreement is presented as the percentage of overlapping proteins within 2‐fold or 5‐fold difference. Numbers in parentheses indicate the number of overlapping proteins.

Despite these systematic offsets, correlation analyses indicated that relative expression patterns were partially preserved (Figure 3). Bgee and RT‐PCR showed the most consistent concordance with ProteinDB across tissues (Pearson r = 0.53–0.71; Spearman ρ = 0.56–0.72). Array showed more modest but still significant correlations. In contrast, EST displayed pronounced tissue‐dependent inconsistency, including negligible and non‐significant correlations in small intestine (Pearson r = 0.07; Spearman ρ = 0.07), despite retaining moderate correlations in kidney and brain.

Fold‐based agreement further highlighted limited quantitative concordance across databases. Even in the better‐performing comparisons, particularly for Bgee and RT‐PCR in small intestine and kidney, only a minority of ADME‐related proteins fell within 2‐fold agreement (23%–37%), although agreement improved to 46%–60% within 5‐fold. Across all databases, brain consistently showed the weakest quantitative agreement, with 2‐fold concordance frequently below 10%, whereas kidney and small intestine generally showed higher agreement.

3.4. PBPK Modeling for Midazolam, Digoxin, Rifampicin, Tizanidine and Associated DDI Scenarios

Across the four single‐compound PBPK models, implementation of ProteinDB preserved overall predictive performance relative to available baseline profiles and empirically modified library profiles (Figure S1; Table 2). For midazolam, systemic exposure predictions remained similar across all three expression settings, with concentration‐time profile MRD values and global AUC and Cmax ratio GMFE values remaining comparable. In contrast, the baseline expression model yielded MRD values of 3.22 and 4.61 for urinary excretion and glucuronidation extent, respectively, whereas both the empirically modified library profile and the ProteinDB profile substantially reduced these values. For digoxin, all three expression settings maintained good prediction of systemic exposure, with only minor differences in circulating concentration MRD and global AUC and Cmax ratio GMFE values. Predictions of urinary excretion were likewise highly similar across the baseline, library, and ProteinDB profiles. For rifampicin, the most mechanistically complex single‐compound model, all three expression settings demonstrated similar predictive performance for systemic exposure, urinary excretion, and enzyme induction. For tizanidine, the concentration‐time profile MRD and global AUC and Cmax ratio GMFE values were virtually identical between the ProteinDB and library expression profiles, indicating that the proteomics‐based parameterization achieves equivalent predictive accuracy to the transcriptomics baseline. In the purely bottom‐up tizanidine model without parameter optimization, predictive performance deteriorated under both expression settings, as expected in the absence of compound‐specific calibration. Nevertheless, the ProteinDB expression profile consistently yielded superior overall predictive performance relative to the library profile under these conditions (Figure S1; Table 2).

TABLE 2.

Predictive performance of single‐compound and DDI models across different expression profiles. Evaluated profiles included the baseline expression profile (original transcriptomics‐based profile), the library expression profile (transcriptomics‐based profile with empirical adjustments for intestinal P‐gp scaling and hepatic UGT1A4 restriction), and the ProteinDB expression profile (absolute proteomics‐based profile).

Compound	Expression profile	Mean plasma/blood/serum concentration MRD	Global AUC ratio GMFE	Global Cmax ratio (po) GMFE	Mean urinary excretion MRD	Mean glucuronidation MRD
Midazolam	Baseline	1.82	1.46	1.56	3.22	4.61
	Library	1.72	1.30	1.51	2.21	1.10
	ProteinDB	1.75	1.35	1.52	2.31	1.17
Digoxin	Baseline	1.67	1.35	1.28	1.20	/
	Library	1.67	1.35	1.29	1.19	/
	ProteinDB	1.71	1.36	1.31	1.22	/

Compound	Expression profile	Mean plasma/blood/serum concentration MRD	Global AUC ratio GMFE	Global Cmax ratio GMFE	Mean urinary/biliary excretion MRD	Mean enzyme induction fold (P‐gp/CYP3A4) MRD
Rifampicin	Baseline	1.89	1.49	1.48	1.63	1.67
	Library	1.90	1.49	1.48	1.63	1.68
	ProteinDB	1.90	1.48	1.41	1.69	1.72
Rifampicin–Midazolam DDI	Baseline	2.07	1.73	2.03	/	/
	Library	2.63	1.90	1.98	/	/
	ProteinDB	2.15	1.66	1.76	/	/
Rifampicin–Digoxin DDI	Baseline	2.02	1.69	1.85	1.43	/
	Library	1.58	1.34	1.42	1.01	/
	ProteinDB	1.58	1.30	1.46	1.01	/
Tizanidine	Library a	1.82	1.34	1.41	/	/
	ProteinDB	1.83	1.34	1.40	/	/
Tizanidine (without parameter optimization)	Library a	4.86	4.72	5.48	/	/
	ProteinDB	3.08	3.11	2.83	/	/

^a For tizanidine, the library and baseline expression profiles are identical because no P‐gp or UGT1A4 expression adjustment was applied in this model.

The DDI simulations provided a more stringent evaluation of the different expression settings. In the rifampicin‐midazolam DDI model, the baseline expression model yielded better overall prediction than the empirically modified library profile for midazolam, with lower systemic concentration MRD and lower AUC and Cmax ratio GMFE values (Figure 4; Table 2). Relative to the modified library profile, ProteinDB improved DDI prediction substantially, reducing concentration‐time profile MRD from 2.63 to 2.15 and lowering the AUC ratio GMFE from 1.90 to 1.66. Compared with the baseline expression model, ProteinDB still improved AUC and Cmax ratio predictions, although the circulating concentration MRD remained slightly higher. In the rifampicin‐digoxin DDI model, the empirically modified library profile and ProteinDB showed similar performance for plasma exposure and urinary excretion, and both outperformed the baseline expression (Figure 4; Figure S2; Table 2).

FIGURE 4.

Goodness‐of‐fit plots for midazolam and digoxin systemic exposure in the rifampicin‐midazolam and rifampicin‐digoxin DDI scenarios. The solid diagonal line indicates identity, and the dotted and dashed lines indicate 1.25‐fold and 2‐fold deviations. Evaluated profiles included the baseline expression profile (original transcriptomics‐based profile), the library expression profile (transcriptomics‐based profile with empirical adjustments for intestinal P‐gp scaling and hepatic UGT1A4 restriction), and the ProteinDB expression profile (absolute proteomics‐based profile).

4. Discussion

A core challenge in PBPK modeling is that the system parameters should closely approximate the biological processes that govern drug ADME and systemic PK. The current PK‐Sim workflows rely on tissue transcriptomics to distribute active‐process capacity across the body, although protein levels are more directly related to drug disposition and target engagement [27]. Relying on mRNA levels instead of absolute protein abundances also creates a structural uncertainty in bottom‐up PBPK modeling, in which acceptable model fits may still be obtained through compensatory parameter adjustments even when the physiological representation is not fully appropriate. To address this source of uncertainty, ProteinDB was developed as a proteomics‐based expression resource directly compatible with PK‐Sim.

ProteinDB was designed as a practical SQLite database that stores absolute protein abundances in units of μmol/L tissue for direct incorporation in PK‐Sim. By providing a standardized system‐level expression resource, it eliminates the need for reference concentration compilation and reduces unit conversion and scaling inconsistencies. During database integration, deduplication was required to produce a single gene‐level abundance value for each organ or compartment because multiple source protein records occasionally mapped to the same gene. We used the maximum‐abundance value as a pragmatic rule intended to retain the strongest supported abundance signal while avoiding potential double‐counting of redundant records. Because the biological origin of these multiple mappings does not always permit a uniquely correct aggregation rule, we evaluated the sensitivity of affected entries to alternative mean‐ and sum‐based aggregation approaches, including both arithmetic and PaxDb‐weighted calculations (Data S1).

Cross‐referencing against the extended PharmaADME gene list [26] identified only three affected genes, HAGH, SOD2, and SLC29A1, and sensitivity analyses indicated that resolved concentrations were generally robust to the deduplication strategy. Nevertheless, attention is warranted when working specifically with these genes. These overlaps primarily arise from the lack of a strict one‐to‐one mapping between ENSG IDs and the STRING database‐derived identifiers used in PaxDb [19], and from redundant accessions due to ambiguities in peptide‐to‐protein mapping in the erythrocyte dataset [28]. A maximum‐abundance approach was chosen over summation and averaging to prevent inflation from shared peptides [28, 29] and dilution by trace variants [30], respectively, thereby preserving the predominant abundance [31, 32].

Although proteomics provides a more direct basis for PBPK system parameters, the established ProteinDB remains less comprehensive than some transcriptomics databases regarding coverage and stratification. ProteinDB currently includes abundance data for 18,377 genes across PK‐Sim compartments, compared with 13,749 in Array, 21,969 in EST, 290 in RT‐PCR, and 57,427 genes in Bgee [7, 10]. In addition, because ProteinDB relies on PaxDb‐reported average abundances, it does not include age, disease states, or inter‐individual variability, whereas some existing transcriptomics databases provide these stratifications. This is a notable limitation, given that recent proteomics studies have reported very large inter‐individual variability, particularly in the intestine, where some proteins vary by more than 100‐fold [33]. Consequently, ProteinDB is intended as a physiologically improved baseline rather than a complete solution for biological variability. Importantly, within the PK‐Sim software architecture, individual expression profiles are designed as deterministic building blocks; once a specific source dataset or stratification is selected, the expression values remain fixed at the arithmetic mean values. Protein expression distribution is only introduced subsequently at the population building block level, defined by user‐input statistical distributions. Nonetheless, the study‐level variability (weighted CV%) provided in the database can serve as a valuable conservative starting point for estimating these population‐level protein expression distributions in prospective modeling (Data S2), though dedicated inter‐individual variability measurements with sufficient sample sizes would yield more representative population distribution estimates.

Cross‐platform comparison further contextualized the role of ProteinDB in the PBPK modeling landscape. Agreement between ProteinDB and existing PBPK platforms was clearly protein‐class dependent. CYP abundances were the most consistent across platforms, possibly reflecting their mature evidence base [34, 35, 36]. UGTs, other enzymes, and transporters showed greater divergence (Table 1), which could be due to their more limited supporting evidence, where inter‐study variability may be more vulnerable to methodological differences in peptide choice, digestion efficiency, membrane enrichment, subtype annotation, and assay sensitivity [37, 38, 39]. Part of the observed divergence may also reflect biological variability, such as the bimodal, genotype‐driven expression of CYP3A5 [33, 40]. However, for two low‐abundance hepatic proteins, UGT1A10 and BCRP, the more than 100‐fold inter‐platform differences (Figure 2) far exceeded clinically observed inter‐individual variability [41, 42]. This may reflect limitations in the current proteomics evidence base, where heterogeneous quantification workflows and limited assay sensitivity may not consistently capture the abundances of low‐level proteins [43, 44, 45, 46].

The sensitivity analyses further clarified that much of the divergence between ProteinDB and other platforms was driven by a few highly discordant proteins (Table 1), while overall similarity for most core ADME proteins was maintained. However, the minimal leave‐two‐out GMFE value for shared UGTs between ProteinDB and GastroPlus remained about 3, indicating persistent class‐wide quantitative differences. Moreover, substantial discrepancies in CYP2D6, CYP2C18, and BCRP consistently persisted between ProteinDB and other platforms (Figure 2), warranting careful consideration for “sensitive” compounds. Notably, for almost all the most discordant proteins from the leave‐out sensitivity, the maximum cross‐platform ratio fell within the inter‐study range observed across PaxDb constituent datasets (Table S3), reflecting the inherent variability of proteomics data. Across all evaluated expression databases, large cross‐platform abundance differences for proteins such as CYP3A5, CYP2C18, UGT1A7, UGT1A10, AOX1, GSTP1, SULT1A2, NAT2, and BCRP may materially alter pathway contributions in PBPK models, particularly for compounds involving parallel pathways, polymorphism, or drug interactions. These findings highlight current expression databases as an important source of PBPK model uncertainty, because expression profiles influence the physiological baseline underlying mechanistic model structures [10, 47].

The cross‐omics comparison of liver‐normalized expression patterns between PK‐Sim sources and ProteinDB was used to assess relative tissue distributions of ADME‐related proteins, reflecting organ‐specific activities and contributions relative to the liver. A unified set of metabolic enzymes and transporters was analyzed to capture overall trends, and the implications for PBPK modeling remain pathway‐dependent. Transcriptomics databases preserved broad inter‐tissue patterns but showed systematic differences from proteomics in dynamic range and relative abundance (Figure 3). The generally narrow distributions of liver‐normalized expression levels in the transcriptomics databases, particularly EST and Array, indicate reduced separation between hepatic and extrahepatic expression levels for ADME‐related proteins. Notably, EST‐derived intestinal ratios showed negligible correlation with proteomics, suggesting limited suitability for scaling intestinal first‐pass metabolism and active efflux. Fold‐level agreement was moderate for small intestine and kidney but consistently poor for brain, indicating incomplete capture of protein‐level differences. These discrepancies highlight the inherent limitations of relying on mRNA as a surrogate for scaling protein abundance across the body, due to highly tissue‐specific differences in translation efficiency, post‐transcriptional regulation, and protein turnover [48, 49].

The differences in relative expression are particularly relevant for tissue contribution in PBPK modeling. The substantial divergence in small intestine and kidney underscores the risk of misestimating intestinal absorption, first‐pass metabolism, and urinary clearance when relying solely on transcriptomics proxies. Given the key role of the small intestine in transporter‐mediated absorption and first‐pass metabolism, overestimation of liver‐normalized intestinal expression in transcriptomics databases may lead to misprediction of oral bioavailability or intestinal DDI magnitude. Recent quantitative proteomics studies in human intestine support protein‐based data for in silico modeling, showing improved agreement with functional activity compared with transcript‐derived estimates [27, 50]. Overall, those findings reveal systematic and tissue‐dependent differences between transcriptomics‐ and proteomics‐derived expression data, supporting the use of direct proteomics expressions to improve the reliability of organ‐specific scaling and pathway attribution in PBPK models.

The PBPK assessments provided a key functional evaluation of ProteinDB's performance relative to the transcriptomics baseline and library expression profiles. A standardized parameter identification procedure was carried out to ensure comparability while allowing the scaling of compound‐dependent parameters, particularly in vitro kinetics, to match clinical observations, representing conventional PBPK modeling practice. Across the four single‐compound models, ProteinDB maintained overall predictive performance. For midazolam, differences between expression settings primarily emerged in secondary elimination pathways rather than systemic exposure. The improved prediction of glucuronidation and urinary excretion with ProteinDB and modified library profiles, relative to the baseline profile, suggests that proteomics‐derived abundances can better approximate pathway‐specific processes than transcriptomics‐based expression. In contrast, digoxin showed limited sensitivity to the choice of expression profiles. Although its disposition relies heavily on P‐gp, the parameter optimization settings inherited from the library model allow adjustment of passive permeability across organs to compensate for P‐gp expression discrepancies. Consequently, similar predictive performance was achieved across all expression settings, though potentially at the expense of mechanistic validity. Rifampicin similarly showed comparable performance across expression settings, as its highly parameterized model structure allows coordinated adjustments across metabolic, transport, and induction processes during parameter optimization. This flexibility enables divergent mechanistic structures derived from distinct expression profiles to yield similar pharmacokinetic behavior, highlighting structural non‐identifiability in complex PBPK models, where agreement with observed PK data alone does not guarantee physiological consistency. For the CYP1A2 probe substrate tizanidine, ProteinDB and the library profile achieved comparable predictive performance after model calibration. In the purely bottom‐up simulations without parameter optimization, however, ProteinDB consistently outperformed the library profile, suggesting that absolute proteomics‐derived abundances provide a more physiologically accurate starting point for system parameterization. DDI simulations served as a more stringent evaluation by probing system consistency under perturbation. Unlike single‐compound models, DDI outcomes depend on the interplay between tissue‐specific perpetrator exposure and the expression of enzymes and transporters [23], where compensatory offsetting arising from model flexibility is often limited. Compared to the good predictive performance of the baseline profile in the digoxin single‐compound model, the deteriorated performance in the rifampicin‐digoxin DDI (Table 2) indicates that expression inaccuracies become more apparent under multi‐process perturbation. Conversely, reasonable DDI predictions for midazolam with the baseline profile, despite known pathway misrepresentation, underscore that good PK predictions can occasionally arise from offsetting errors rather than mechanistic accuracy. Overall, ProteinDB showed comparable or superior performance across single‐compound and DDI scenarios, suggesting a more physiologically consistent whole‐body expression framework for PBPK modeling.

Collectively, these results highlight a fundamental limitation of relying on transcriptomic proxies for system parameterization and underscore the necessity of a physiologically grounded proteomics database such as ProteinDB. Because transcriptomic levels do not consistently reflect absolute protein abundance, compensatory optimization of other model parameters is often required to match observed pharmacokinetic data [7, 51]. Although such optimization can produce acceptable predictions for a specific study, it may mask errors in pathway capacity and organ‐specific contributions, weakening mechanistic interpretability [7]. By providing organ‐specific absolute protein abundances, ProteinDB reduces reliance on empirical adjustments and establishes a more self‐consistent whole‐body baseline, thereby improving mechanistic coherence and the reliability of PBPK prediction and clinical extrapolation.

In conclusion, the present findings support ProteinDB as a practical, proteomics‐based alternative to transcriptomics‐derived expression databases in PK‐Sim and represent a step toward more physiologically grounded in vitro‐in vivo extrapolation (IVIVE). Nevertheless, further evaluation across more compounds, multi‐pathway situations, genetic polymorphism cases, and DDI scenarios will be essential to establish its generalizability and to achieve a level of confidence comparable to the extensively validated transcriptomics‐based PK‐Sim framework. Future large‐scale and standardized proteomics studies will further strengthen such proteomics‐based databases through improved source transparency, incorporation of inter‐individual variability, and inclusion of disease‐ and demographic‐specific expression information [52]. Continued progress in these areas will be important for increasing regulatory confidence and establishing a broader proteomics‐based foundation for PBPK modeling with the Open Systems Pharmacology suite.

Funding

Chen Ning is funded by the China Scholarship Council (CSC) from the Ministry of Education of P.R. China.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: psp470316‐sup‐0001‐DataS1.xlsx.

Data S2: psp470316‐sup‐0002‐DataS2.xlsx.

Figure S1: Goodness‐of‐fit plots for the plasma, blood, or serum concentration of the midazolam, digoxin, rifampicin, and tizanidine models. The solid diagonal line indicates identity, and the dotted and dashed lines indicate 1.25‐fold and 2‐fold deviations. Evaluated profiles included the baseline expression profile (original transcriptomics‐based profile, if available), the library expression profile (transcriptomics‐based profile with empirical adjustments for intestinal P‐gp scaling and hepatic UGT1A4 restriction), and the ProteinDB expression profile (absolute proteomics‐based profile).

Figure S2: Goodness‐of‐fit plots for urinary excretion, glucuronidation, or enzyme induction fold in the single‐compound models of midazolam, digoxin, and rifampicin. The solid diagonal line indicates identity, and the dotted and dashed lines indicate 1.25‐fold and 2‐fold deviations. Evaluated profiles included the baseline expression profile (original transcriptomics‐based profile), the library expression profile (transcriptomics‐based profile with empirical adjustments for intestinal P‐gp scaling and hepatic UGT1A4 restriction), and the ProteinDB expression profile (absolute proteomics‐based profile).

Figure S3: Goodness‐of‐fit plots for urinary excretion of digoxin in the rifampicin‐digoxin model. The solid diagonal line indicates identity, and the dotted and dashed lines indicate 1.25‐fold and 2‐fold deviations. Evaluated profiles included the baseline expression profile (original transcriptomics‐based profile), the library expression profile (transcriptomics‐based profile with empirical adjustments for intestinal P‐gp scaling and hepatic UGT1A4 restriction), and the ProteinDB expression profile (absolute proteomics‐based profile).

Table S1: Parameter optimization settings for midazolam, digoxin, rifampicin, and tizanidine.

Table S2: Inhibition and induction settings for rifampicin.

Table S3: Source‐level variability and cross‐platform comparison of PaxDb‐derived hepatic ADME protein concentrations.

Data S3: psp470316‐sup‐0004‐DataS4.r.

Data Availability Statement

The authors declare that all the data supporting the findings of this study are contained within the paper. The ProteinDB database files and associated code are openly available in the GitHub repository at https://github.com/translatPK‐KUL/ProteinDB.