A workflow for application of the general pharmacodynamic interaction model on high-throughput combinatorial data in order to identify, quantify and characterise drug combinations that can overcome multi-drug-resistance
Robin Michelet (1), Ana Rita Brochado (2,3), Athanasios Typas (2,4), Sebastian G. Wicha (5), Charlotte Kloft (1)
(1) Freie Universitaet Berlin, Institute of Pharmacy, Dept. of Clinical Pharmacy & Biochemistry, Berlin, Germany, (2) European Molecular Biology Laboratory, Genome Biology Unit, Heidelberg, Germany (3) Wuerzburg University, Chair of Microbiology, Wuerzburg, Germany (4) European Molecular Biology Laboratory, Structural & Computational Biology Unit, Heidelberg, Germany (5) University of Hamburg, Institute of Pharmacy, Dept. of Clinical Pharmacy, Hamburg, Germany
Background:
Antimicrobial resistance is one of the key challenges in the current global healthcare system [1]. As new antibiotics are lacking, combinations of existing drugs can help to treat multi-drug-resistant (MDR) bacterial infections. In order to detect synergistic combinations between antibiotics, human-use (non-antibiotic) drugs and other compounds (e.g. food additives), we previously combined ~3000 compound pairs and assessed their interaction in three Gram-negative species [2]. In the current work, a robust workflow to quantitatively characterise these interactions using the General PharmacoDynamic Interaction (GPDI) model [3] is presented. Using this approach, not only the magnitude but also directionality of an interaction between two or more compounds can be elucidated, possibly identifying interesting combinations for further non-/clinical development.
Methods:
A model selection & evaluation workflow was established using R (v. 3.4.4) and RStudio (v. 1.1.447). In step 1, linear, power and EMAX-type models were fitted to single concentration-effect data, after which their parameters were fixed and all possible interaction parameters for all possible model combinations were estimated (113 interaction model types). Parameter estimation was performed by ELS regression using the Nelder-Mead and BFGS algorithm. Parameter precision was assessed using the diagonal of the Fisher Information Matrix (calculated from the Hessian outputted by the last successful algorithm). For the best model combination, all parameters were estimated again and the model next evaluated. Model selection was based on the precision of parameter estimates (discarding models with parameter imprecision >50% RSE) and the AIC (penalty of 2 points per parameter). Model evaluation was performed by comparing the model to the experimental data, for which >15% deviation from observed effect, or no overlap with the 95% confidence interval of the t-distribution estimated from the data, were considered significant deviations. Per combination, the best model was then used to simulate a response surface, which was compared to the Bliss Independence surface [4], in order to visualise the interaction. The parameters of the GPDI model were then used to assess the magnitude and direction of the interaction in order to inform hypotheses about the interaction mechanism and select interesting candidates for further development.
Results:
A dataset [2] consisting of extended-dose data (8x8 checkerboard experiments, 242 drug combinations in susceptible Gram negative strains and 7 synergistic combinations in a set of 6 E.coli and K. pneumoniae MDR clinical isolates) was first analysed using the developed workflow. In general, the GPDI model described the experimental data well and identified similar synergies and antagonisms as conventional response-surface analyses suggested. Furthermore, using the estimated interaction parameters, the nature of the interactions and putative perpetrator and victim drugs could be identified. Indeed, 28.3% of the observed interactions were mono-directional synergistic, 24.1% mono-directional antagonistic, 13.6% bi-directional synergistic, 2.2% bi-directional antagonistic and 30.4% asymmetric. In the clinical isolates, strong synergies between colistin and macrolide drugs (strong decrease of macrolide EC50 in function of colistin concentration) and between colistin and loperamide (strong decrease of loperamide EC50 or increase in Emax in function of colistin concentration, depending on the strain) were characterised. Weaker synergies were quantified between doxycycline and procaine (bi-directional effect), and vanillin and spectinomycin (only for E. coli).
Conclusions:
A robust workflow was set up to apply the GPDI model to high-throughput data and select the most fitting model structure per combination. In this way, promising combination candidates could be identified and their interaction quantitatively described. This workflow can now be applied on the larger dataset consisting of 3000 combinations to identify the complete set of promising candidates. These combinations can be further investigated and pushed towards pre-clinical testing and eventual clinical application. Furthermore, clustering approaches could be applied to the generated model repository in order to group interactions according to their intensity and directionality to inform mechanistic hypothesis generation.
References:
[1] The Lancet. Antibiotic resistance: a final warning. The Lancet. 382: 1072 (2013)
[2] A. R. Brochado et al. Species-specific activity of antibacterial drug combinations. Nature. 559: 259–263 (2018).
[3] S. G. Wicha et al. A general pharmacodynamic interaction model identifies perpetrators and victims in drug interactions. Nat. Commun. 8:1-10 (2017).
[4] C.I. Bliss. The toxicity of poisons applied jointly. Ann. Appl. Biol. 26: 585-615 (1939).
