Validation of a Model Predicting Anti-infective Lung Penetration in the Epithelial Lining Fluid of Humans
Validation of a Model Predicting Anti-infective Lung Penetration in the Epithelial Lining Fluid of Humans
Linda B. S. Aulin 0 1 2 4 5 6
Pyry A. Valitalo 0 1 2 4 5 6
Matthew L. Rizk 0 1 2 4 5 6
Sandra A. G. Visser 0 1 2 4 5 6
Gauri Rao 0 1 2 4 5 6
Piet H. van der Graaf 0 1 2 4 5 6
J. G. Coen van Hasselt 0 1 2 4 5 6
0 Merck & Co. Inc. , Kenilworth, New Jersey , USA
1 Orion Corporation Orion Pharma , Kuopio , Finland
2 Leiden Academic Centre for Drug Research, Leiden University , Einsteinweg 55, 2333 CC Leiden , Netherlands
3 J. G. Coen van Hasselt
4 Certara , Canterbury , UK
5 Eshelman School of Pharmacy, University of North Carolina at Chapel Hill , Chapel Hill, North Carolina , USA
6 GSK, King of Prussia , Pennsylvania , USA
We previously developed a quantitative structurepharmacokinetic parameter relationship (QSPKR) model to predict antibiotic lung penetration of several classes of antiinfective agents (5). This model utilized a regularized elastic net regression approach to relate multiple specific chemical structural properties or descriptors to the ratio of the concentration in the ELF to the unbound plasma concentration. The model was trained based on log-transformed clinical ELF and plasma concentration data from 56 unique anti-infective compounds that were extracted from the previous publications of clinical lung penetration studies. The model was validated using a leave-one-out cross validation and by prediction of a limited set of five anti-infective compounds not used for model development. Since then several new clinical lung penetration studies have been published. The aim of this report was to perform a more extensive external validation of the published model to further evaluate its predictive value.
antibiotics; epithelial lining fluid; lung; pharmacokinetics; prediction
Community and hospital acquired bacterial pneumonias are
associated with significant mortality and morbidity (
). It is
essential to achieve sufficiently high antibiotic exposure in
the epithelial lining fluid (ELF) in order to obtain sufficient
efficacy and to prevent selection for resistant or persistent
bacterial subpopulations (
). However, antibiotic concentrations
in the ELF may be significantly different from concentration
in the plasma (
). Hence, the characterization of antibiotic
ELF concentrations is important for antibiotics aimed at
treating lung infections. Bronchoalveolar lavage (BAL) is most
commonly used to quantify antibiotic drug concentrations in
humans. However, the technique has significant limitations
including the limitation of a single sample per patient given
the invasiveness of obtaining the BAL sample (
), and the
significant variability between measurements. Approaches to
predict lung penetration of antibiotics are thus highly relevant
to support informative BAL study design and to support the
selection or prioritization of antibiotic candidates.
We searched the PubMed database and relevant
microbiology conference abstracts for clinical studies reporting
antiinfective ELF and plasma concentrations in humans between
year 2011 and 2017. We also included the five antibiotics used
for the original external validation. Antibiotics already present
in the training dataset used for the model development were
excluded. For each drug identified and included we extracted
the non-extrapolated mean AUCELF and AUCplasma values.
The AUCplasma values were converted to unbound
concentrations (fAUCplasma) using the reported protein binding values
obtained from the DrugBank database. The
AUCELFfAUCplasma ratio was collected and subsequently
logtransformed to obtain the clinically observed log ELF/
plasma penetration ratio (EPR). Using an R script (included
as supplemental material in the original model publication),
we generated the same 145 chemical descriptors used for the
developed QSPKR model using the R package Rcdk.
Subsequently, we applied the original elastic net regression
model, without any modifications, to predict the log EPR
values for each antibiotic included in the new validation data
set, which were compared with the clinically reported log
EPRs. This comparison was done by graphically assessing
the clinically reported log EPR versus the model predicted
log EPR values as well as reporting the percentage of the
predictions being outside a 3-fold change from the
observations. Additionally we assessed the models capability to
characterize drugs ability to penetrate the lungs in a
We identified nine anti-infective drugs for which EPRs could
be determined and which were not included in the original
model. Together with the five antibiotics (arbekacin,
GSK2251052, tedizolid, imipenem, peramivir) previously
included for the external validation in the original publication
the new validation dataset comprised of 14 compounds. The
studies for each of these 14 anti-infective agents are
summarized in Table I.
The model predictions for the 14 drugs were in line with the
predictive performance reported in the original publication
(Fig. 1), with a root mean squared error of 1.42. For 57%
(n=8/14) of the drugs the predicted EPRs were within a
3fold difference from the observations. The model predicted
under or over-exposure compared to plasma, i.e. EPR > 1
or EPR < 1 and log EPR > 0 or log EPR < 0 for
untransformed and log-transformed EPR respectively, correctly
for 93% (n=13/14) of the drugs. A trend towards
underprediction of the log EPR was observed, particularly for
eravacycline, temocillin, and tedizolid.
The external validation analysis of the previous developed
QSPKR lung penetration model has shown a predictive
performance in line with the reported prediction performance of
the original publication. In the current validation the number
of antibiotics outside a 3-fold prediction error for
untransformed ERP was 43% whereas in the original
publication it was 20% for the small external validation set and 39%
for the leave-one-cross validation analysis of the training
dataset. Notably, as no model development was performed
at this stage, no leave-one-cross validation analysis was
conducted and thus no comparable metric was obtained.
Three classes of anti-infective drugs were included in this
validation that were not used for the training of the original
m o d e l , i . e . t h e l e u c y l - t R N A s y n t h e t a s e i n h i b i t o r
GSK2251052, the non-beta-lactam beta-lactam inhibitors
avibactam and relebactam, and the pleuromutilin antibiotic
lefamulin. These new-in-class anti-infective agents were well
predicted, within a 2-fold difference from the observed EPRs.
Eravacycline was associated with substantial lung
penetration compared to plasma concentrations, with a clinically
observed ERP of 6.44, but was miss-classified by the model as
giving under-exposure (EPR<1), indicating that some of the
model predictions should be interpreted with caution for
compounds that are structurally related to this drug. The
model did however correctly classify all drugs associated with
under-exposure in the ELF compared to plasma. This
suggests the clinical applicability of the model as a tool to inform
when dose-adjustments may be warranted and guide in the
adjustment for clinically relevant exposure, which could
improve treatment and decrease resistance development.
We were not able to identify clear chemical structure
characteristics different from the training set that could explain the
misclassification of eravacycline or the high predictive errors
associated with eravacycline, temocillin, and tedizolid. The
three drugs are all relatively highly bound to plasma proteins
(≥ 75%), which is not currently included in the model.
However, during a post-hoc residual analysis, which included
all drugs present in the validation dataset, no correlation was
found between protein binding and residuals. Antibiotic class,
molecular descriptors, and factors relating to study design
were considered in addition to protein binding. However, no
strong correlations were found in this analysis. Worth
considering is that the ability of detecting correlations was limited
due to the small size of the dataset. In the validation dataset
eravacycline and tedizolid are the only drugs containing
fluorine, while in the training dataset all fluorine containing drugs
were primarily fluoroquinolones. Additionally, temocillin was
the only drug studied on patient with lung infections and not
healthy volunteers. Infection could affect the permeability of
the drug to the lungs, i.e. the EPR, contributing to prediction
error. In the original dataset difference in clinical observed
EPR could be seen between diseases states for some of the
drugs. No other study design aspect, such as size of study
cohort or sampling technique, could be linked to the
We expect that increasing the mechanistic aspects of this
model could improve the model predictions. Currently,
plasma protein binding is not explicitly included as a predictor in
the model but potentially its consideration could improve the
predictions. However, the maximal possible quality of this
predictive model is partially limited by the high variability of
the ELF concentrations obtained by BAL sampling. The
method is associated with inherent uncertainty that is partially
related to indirect quantification, possible contamination from
cellular release, and technical errors (
), as well as only
obtaining a single sample per individual. A more novel and
p r o m i s i n g s a m p l i n g t e c h n i q u e i s b r o n c h o s c o p i c
microsampling (BMS), used in 2 of the 14 studies (
technique is less invasive than BAL and allows for direct
repeated measurements of drug concentrations in the ELF over
Although the log EPR predictions by the QSPKR model are
still associated with a significant error, we nonetheless expect
that this QSPKR model is of relevance to determine the
expected magnitude of lung penetration on a semi-quantitative
basis, i.e. under- or over-exposure, or comparable exposure to
plasma. Specifically, the model has relevance to support
informative study design of pharmacokinetic studies (
potentially in conjunction with population pharmacokinetic models
). Future studies to predict the EPR may benefit from a
combined mechanistic PBPK approach to enable prediction
of non-steady state ELF pharmacokinetics, while supported by
a QSPKR model for estimation of the partitioning
Open Access This article is distributed under the terms of the
Creative Commons Attribution 4.0 International License
permits unrestricted use, distribution, and reproduction in any
medium, provided you give appropriate credit to the original
author(s) and the source, provide a link to the Creative
Commons license, and indicate if changes were made.
1. Anevlavis S , Bouros D. Community acquired bacterial pneumonia . Expert Opin Pharmacother . 2010 ; 11 : 361 - 74 .
2. Andersson DI , Hughes D. Microbiological effects of sublethal levels of antibiotics . Nat Rev Microbiol . 2014 ; 12 : 465 - 78 .
3. Kiem S , Schentag JJ . Interpretation of antibiotic concentration ratios measured in epithelial lining fluid . Antimicrob Agents Chemother . 2008 ; 52 : 24 - 36 .
4. Meyer KC , Raghu G , Baughman RP , Brown KK , Costabel U , Du Bois RM , et al. An official American Thoracic Society clinical practice guideline: the clinical utility of bronchoalveolar lavage cellular analysis in interstitial lung disease . Am J Respir Crit Care Med . 2012 ; 185 : 1004 - 14 .
5. Välitalo PAJ , Griffioen K , Rizk ML , Visser SAG , Danhof M , Rao G , et al. Structure-based prediction of anti-infective drug concentrations in the human lung epithelial lining fluid . Pharm Res . 2016 ; 856 - 67 .
6. Fun atsu Y , Hasegawa N , Fujiwara H , Namkoong H . Pharmacokinetics of arbekacin in bronchial epithelial lining fluid of healthy volunteers . J Infect Chemother . 2014 ; 20 : 607 - 11 .
7. Nicolau DP , Siew L , Armstrong J , Li J , Edeki T , Learoyd M , et al. Phase 1 study assessing the steady-state concentration of ceftazidime and avibactam in plasma and epithelial lining fluid following two dosing regimens . J Antimicrob Chemother . 2015 ; 70 : 2862 - 9 .
8. Riccobene TA , Pushkin R , Jandourek A , Knebel W , Khariton T. Penetration of ceftaroline into the epithelial lining fluid of healthy adult subjects . Antimicrob Agents Chemother . 2016 ; 60 : 5849 - 57 .
9. Chandorkar G , Huntington JA , Gotfried MH , Rodvold KA , Umeh O . Intrapulmonary penetration of ceftolozane / tazobactam and piperacillin / tazobactam in healthy adult subjects . J Antimicrob Chemother . 2012 ; 67 : 2463 - 9 .
10. Connors KP , Housman ST , Pope JS , Russomanno J , Salerno E , Shore E , et al. Phase I , open-label, safety and pharmacokinetic study to assess bronchopulmonary disposition of intravenous eravacycline in healthy men and women . Antimicrob Agents Chemother . 2014 ; 58 : 2113 - 8 .
11. Tenero D , Bowers G , Rodvold KA , Patel A , Kurtinecz M , Dumont E , et al. Intrapulmonary pharmacokinetics of GSK2251052 in healthy . Antimicrob Agents Chemother . 2013 ; 57 : 3334 - 9 .
12. Rhee EG , Jumes PA , Gotfried MH , Rizk ML , Liu Y , Mangin E , et al. Intrapulmonary pharmacokinetics of MK-7655 , a novel β - lactamase inhibitor, dosed in combination with imipenem / cilastatin in healthy subjects . Denver: ICAAC; 2013 . p. 7655 .
13. Zeitlinger M , Schwameis R , Burian A , Burian B , Matzneller P , Mu M , et al. Simultaneous assessment of the pharmacokinetics of a pleuromutilin, lefamulin, in plasma, soft tissues and pulmonary epithelial lining fluid . J Antimicrob Chemother . 2016 ; 1022 - 6 .
14. Horn KS , Gotfried MH , Steenbergen JN , Villano S , Tzanis E , Garrity-Ryan L , et al. Comparison of omadacycline and tigecycline pharmacodynamics in the plasma, epithelial lining fluid, and alveolar macrophages in healthy subjects . Washington DC: 113th Annu. Conf. Am. Thorac. Soc.; 2017 .
15. Food and Drug Administration . Indications and Usage Peramivir, RAPIVAB . Durham: BioCryst Pharmaceuticals Inc .; 2014 .
16. Funatsu Y , Tasaka S , Asami T , Namkoong H , Fujiwara H , Yagi K , et al. Pharmacokinetics of intravenous peramivir in the airway epithelial lining fluid of healthy volunteers . Antivir Ther . 2016 ; 21 : 621 - 5 .
17. Housman ST , Pope JS , Russomanno J , Salerno E , Shore E , Kuti JL , et al. Pulmonary disposition of tedizolid following administration of once-daily oral 200-milligram tedizolid phosphate in healthy adult volunteers . 2012 . p. 2627 - 34 .
18. Visée C , Layios N , Mistretta V , Maes N , Van Bambeke F , Frippiat F. Epithelial lining fluid penetration of temocillin administered by continuous infusion in critically ill patients with nosocomial pneumonia . Vienna: 27th ECCMID; 2017 .
19. Wenzler E , Gotfried MH , Loutit JS , Durso S , Griffith DC , Dudley MN , et al. Meropenem-RPX7009 concentrations in plasma, epithelial lining fluid, and alveolar macrophages of healthy adult subjects . Antimicrob Agents Chemother . 2015 ; 59 : 7232 - 9 .
20. Yamazaki K , Ogura S , Ishizaka A , Oh-hara T , Nishimura M. Bronchoscopic microsampling method for measuring drug concentration in epithelial lining fluid . Am J Respir Crit Care Med . 2003 ; 168 : 1304 - 7 .
21. Clewe O , Goutelle S , Conte JE Jr, Simonsson USH . A pharmacometric pulmonary model predicting the extent and rate of distribution from plasma to epithelial lining fluid and alveolar cells - using rifampicin as an example . Eur J Clin Pharmacol . 2015 ; 71 : 313 - 9 .
22. Van Hasselt JGC , Rizk ML , Lala M , Chavez-Eng C , Visser SAG , Kerbusch T , et al. Pooled population pharmacokinetic model of imipenem in plasma and the lung epithelial lining fluid . Br J Clin Pharmacol . 2016 ; 81 : 1113 - 23 .