Maria Kalliga 1,2,3, Raju Prasad Sharma 3, Sieto Bosgra 3, Fulya Akpinar Singh 4, Terezinha de Souza 3
1 Vrije Universiteit Amsterdam (VU Amsterdam) (Amsterdam, Netherlands), 2 University of Amsterdam (UvA) (Amsterdam, Netherlands), 3 Genmab (Utrecht, Netherlands), 4 Genmab (Plainsboro, USA)
Introduction:
In antibody therapy, accurate prediction of first-in-human dose often requires accounting for target-mediated drug disposition (TMDD), for which PBPK modelling is one commonly used quantitative framework [1]. These models rely on quantitative target expression inputs in molar units. However, such data is often unavailable or incomplete across tissues. Furthermore, most available expression datasets, including large-scale proteomics resources such as PaxDb [3], provide expression values in relative abundances (parts-per-million (ppm)), or other measures of expression that are difficultly translated into absolute concentration in tissues, such as mRNA. Recent work by Sepp & Muliaditan [2] proposed an empirical sigmoidal approach to translate PaxDb ppm values into molar concentrations, with validation to date largely limited to liver.
Objectives:
The objective of this work was to test generalizability of the proposed sigmoidal conversion across different tissues and to benchmark whether alternative strategies improve accuracy using predefined accuracy thresholds (within two-fold preferred, within three-fold acceptable).
Methods:
A preprocessing pipeline was developed to harmonize protein identifiers across sources and link each protein to both its PaxDb abundance (ppm) and absolute concentration (nM). Subcellular localization was assigned using Human Protein Atlas annotations [7,8] and primary evaluation focused on proteins annotated as membrane-localized, as these are accessible to monoclonal antibodies and can drive TMDD [2]. The pipeline was applied to three datasets providing absolute concentrations: liver [4], heart [5], and cerebrospinal fluid (CSF) [6]. Generalizability beyond liver was assessed in the heart dataset, evaluating both whole-heart and region-specific estimates using corresponding ppm values. For CSF, reported concentrations (g/L) were converted to nM using protein molecular weight. The published sigmoidal conversion [2] was first applied to the liver dataset and subsequently evaluated across tissues using both the published parameters and tissue-specific refitting. Two alternative methods were benchmarked: (i) a baseline conversion using molecular-weight weighting with an assumed total tissue protein concentration, and (ii) a machine-learning (ML) method, an XGBoost regressor [10]. To assess whether additional biological descriptors improve prediction, proteins were annotated with InterPro identifiers [9] and corresponding feature flags. The model was trained on heart (region-specific), liver and CSF, using ppm, molecular weight, subcellular class, InterPro feature flags, assumed volumes, and total protein content. Model performance was summarized using median fold error and the percentage of proteins within predefined accuracy threshold of two-fold and three-fold.
Results:
In liver, using the published parameters for membrane proteins (n = 772), the sigmoidal model achieved 50% within two-fold and 72% within three-fold accuracy (~2 median fold error). In heart (whole-heart, n = 712), performance was similar, with 48% and 67% within two- and three-fold, respectively. Refitting yielded marginal improvement and minor parameter changes.
In CSF, applying liver-derived parameters resulted in poor performance (median fold error ~3.7e3), whereas refitting restored tissue-like accuracy (47% within two-fold, 68% within three-fold, median fold error ~2).
The baseline molecular-weight model showed similar median error magnitude, but reduced performance within the predefined thresholds (liver: 45% within two-fold, 65% within three-fold, median fold error 2.2; heart (region-specific): 43% within two-fold, 60% within three-fold, median fold error 2.3).
The ML model trained on heart (region-specific), liver and CSF achieved median fold errors of 1.3 in heart, 1.9 in liver, and 2.1 in CSF; overall on held-out membrane proteins, median fold error was 1.60 (n=205; 69/82% within two-/three-fold).
Conclusions:
Across liver and heart, the sigmoidal conversion yielded a consistent fold-error range for membrane proteins. Refitting the sigmoidal model was necessary to explain CSF, consistent with the substantially lower total protein content in CSF compared with solid tissues and a corresponding shift in the fitted relationship. A similar sigmoidal relationship has been reported for plasma [11], supporting the hypothesis that total protein content is one of the key drivers of cross-sample shifts.
The ML approach showed competitive performance in the present data setting, suggesting potential to improve as absolute proteomics data become available across additional organs, enabling broader training and more robust external validation. Importantly, even well-performing ppm-to-nM conversions showed target-level variability relevant for TMDD, with roughly 65-80% of targets falling within three-fold across tissues. These findings support pragmatic use of ppm-derived target estimates in PBPK/TMDD models, provided downstream dosing decisions are accompanied by sensitivity analyses exploring plausible target-expression ranges.
References:
[1] Pasquiers B et al. Review of the Existing Translational Pharmacokinetics Modeling Approaches Specific to Monoclonal Antibodies (mAbs) to Support the First-In-Human (FIH) Dose Selection. Int J Mol Sci. 2022. doi:10.3390/ijms232112754
[2] Sepp A, Muliaditan M. Application of quantitative protein mass spectrometric data in the early predictive analysis of membrane-bound target engagement by monoclonal antibodies. MAbs. 2024. doi:10.1080/19420862.2024.2324485
[3] Huang Q et al. PaxDb v6.0: reprocessed, LLM-selected, curated protein abundance data across organisms. Nucleic Acids Res. 2026. doi:10.1093/nar/gkaf1066
[4] Wegler, C., Wiśniewski, J.R., Robertsen, I., Christensen, H., Kristoffer Hertel, J., Hjelmesæth, J., Jansson-Löfmark, R., Åsberg, A., Andersson, T.B. and Artursson, P. (2022), Drug Disposition Protein Quantification in Matched Human Jejunum and Liver From Donors With Obesity. Clin Pharmacol Ther, 111: 1142-1154. https://doi.org/10.1002/cpt.2558
[5] Doll S et al. Region and cell-type resolved quantitative proteomic map of the human heart. Nat Commun. 2017. doi:10.1038/s41467-017-01747-2
[6] Zhang Y et al. A comprehensive map and functional annotation of the normal human cerebrospinal fluid proteome. J Proteomics. 2015. doi:10.1016/j.jprot.2015.01.017
[7] Thul PJ et al. A subcellular map of the human proteome. Science. 2017. doi:10.1126/science.aal3321
[8] Human Protein Atlas. proteinatlas.org
[9] Blum M et al. InterPro: the protein sequence classification resource in 2025. Nucleic Acids Res. 2025. doi:10.1093/nar/gkae1082
[10] Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proc KDD. 2016. doi:10.1145/2939672.2939785
[11] Muliaditan M, Sepp A. Application of quantitative protein mass spectrometric data in the early predictive analysis of target engagement by monoclonal antibodies. Clin Transl Sci. 2022. doi:10.1111/cts.13278
Reference: PAGE 34 (2026) Abstr 12276 [www.page-meeting.org/?abstract=12276]
Poster: Methodology - Model Evaluation