In silico and in vitro studies of a number PILs as new antibacterials against MDR clinical isolate Acinetobacter baumannii

QSAR analysis of a set of previously synthesized phosphonium ionic liquids (PILs) tested against Gram‐negative multidrug‐resistant clinical isolate Acinetobacter baumannii was done using the Online Chemical Modeling Environment (OCHEM). To overcome the problem of overfitting due to descriptor selection, fivefold cross‐validation with variable selection in each step of the model development was applied. The predictive ability of the classification models was tested by cross‐validation, giving balanced accuracies (BA) of 76%–82%. The validation of the models using an external test set proved that the models can be used to predict the activity of newly designed compounds with a reasonable accuracy within the applicability domain (BA = 83%–89%). The models were applied to screen a virtual chemical library with expected activity of compounds against MDR Acinetobacter baumannii. The eighteen most promising compounds were identified, synthesized, and tested. Biological testing of compounds was performed using the disk diffusion method in Mueller‐Hinton agar. All tested molecules demonstrated high anti‐A. baumannii activity and different toxicity levels. The developed classification SAR models are freely available online at http://ochem.eu/article/113921 and could be used by scientists for design of new more effective antibiotics.


| 625
TRUSH eT al. different kinds of hospital-acquired infections caused by opportunistic microbe A. baumannii have increased in recent years. It is known that A. baumannii isolates are able to form biofilms on the surfaces of surgical instruments, medical devices, implants, etc (Asif, Alvi, & Rehman, 2018). It should also be noted that Acinetobacter strains were widely detected in acute wound infections of patients in military conflicts in Iraq, Afghanistan, Ukraine, and other countries (Khan et al., 2018;O'Toole et al., 2019).
Acinetobacter baumannii strains are capable of colonizing skin, mucosal, and epithelial surfaces, and most often affect the respiratory, circulatory, and urinogenital systems (Doi, Murray, & Peleg, 2015;Eliopoulos, Maragakis, & Perl, 2008). Senile age, alcoholism, heavy smoking, diabetes mellitus, severe injuries, extensive burns, maligned neoplasms, broad surgical interventions, radiation, hormonal and cytostatic therapy, newborns pathology, immunodeficiency syndrome, chronic obstructive pulmonary disease, and renal disease are the risk factors of A. baumannii infections (Visca, Seifert, & Towner, 2011). The mortality rate from the diseases of the blood system caused by MDR A. baumannii is 49% and from resulting central nervous system damage is 70%. According to the literature, there are cases of A. baumannii infection as a result of procedures such as catheterization, lumbar puncture, myelography and ventriculography, intrathecal infusions or spinal anesthesia, and gastrointestinal endoscopy (Doi et al., 2015;Eliopoulos et al., 2008;O'Toole et al., 2019).
Acinetobacter spp. are usually highly resistant to a number of commonly used antibiotics, including ampicillin, carbenicillin, cefoxitin, gentamicin, and chloramphenicol (Seifert, Baginski, Schulze, & Pulverer, 1993;Vila et al., 1993). The multidrug resistance of these bacteria is mediated by all of the major resistance mechanisms that are known to occur in bacteria. These include the modification of target sites (the methylation of ribosomal RNA), enzymatic inactivation (β-lactamases, aminoglycosidases, tetraciclinases), and the slowing of drug influx and direct active efflux (Dijkshoorn, Nemec, & Seifert, 2007). Thus, extensive research and the development of new antibacterials for the prevention or treatment of MDR Acinetobacter infection are highly relevant and necessary.
Phosphonium liquids/salts (PILs) are a promising class of pharmacologically active organophosphorus compounds. They are used for imaging and diagnosing tumors, and also as anticancer (Madar et al., 2007), antioxidant (Sheu, Nauduri, & Anders, 2006), antiviral, and antiparasitic agents (Spivak et al., 2014). PILs are also used extensively to fight bacterial and fungal infections (Trush et al., 2018). and additionally as biocides and antifouling agents (Joseph et al., 2016). One of the advantages of phosphonium compounds is the high thermal and chemical stability they present due to the absence of acidic protons in their structure (Fraser & MacFarlane, 2009). Moreover, phosphonium liquids, in general, are less toxic and more environmentally friendly compared to, for example, ammonium salts (Kumar & Malhotra, 2009).
The main goal of this study was to investigate the anti-A. baumanii activity of a number of PILs based on structure-activity relationship studies (Dearden, 2016) derived using computational models developed by the On-line Chemical Database and Modeling Environment (OCHEM) (Tetko, Maran, & Tropsha, 2017 (Cheng et al., 2010;Singh et al., 2015;Sun et al., 2011). The compounds had MIC values ranging from 0.137 to 900 µM. All compounds in this set were unique. The collected data are freely accessible under the CC-BY license on the OCHEM (Tetko et al., 2017).

Modeling Environment
The OCHEM platform was used to develop public and freely accessible models for predicting the antibacterial activity of compounds. The resulting OCHEM models estimate standard mean errors for each prediction and allow users to decide whether the provided predictions are sufficiently accurate for their studies.

Descriptors
OCHEM supports multiple software packages for calculation of diverse molecular descriptors. For this study, we used electrotopological (E-state) (Hall & Kier, 1995) descriptors and AlogPS, (Tetko & Tanchuk, 2002) which were frequently the top-performing descriptors in our previous studies. AlogPS estimates lipophilicity and solubility of chemical compounds, while E-state descriptors describe their electronic and topological characteristics.

Descriptor preprocessing
The unsupervised filtering of descriptors was used. Descriptors with fewer than two unique variables or with a coefficient of variance <0.01 were excluded. Moreover, descriptors with a pairwise non-parametric Pearson's correlation coefficient R > 0.95 were grouped. Additionally, the Unsupervised Forward Selection (UFS) method (Whitley, Ford, & Livingstone, 2000) was used to select a representative non-redundant set of descriptors for model development.

Model validation
Two validation protocols were used. First of all, the initial set of 263 compounds was split by chance into training (210 compounds) and test (53) sets. Fivefold cross-validation (CV) with variable selection in each step of the analysis was used to estimate the accuracy of the models for the training set (Tetko et al., 2006). The dataset of 210 compounds was divided into 5 subsets of equal size. Out of each 5 subsets, a single subset was retained for validation, while the remaining sets were used for training. For each subdivision, we first selected descriptors using the respective training set, developed the model, and then applied it to predict the excluded molecules from the respective validation set. Then, the statistical coefficients over the combined five validation sets were computed. The prediction performance of the final model developed with 210 compounds was also tested using the test set of 53 compounds (see Table 1).

Estimation of prediction accuracy
The OCHEM estimates the applicability domain (AD) and the accuracy for each prediction, which are calibrated using CV (Sushko et al., 2010). A probability of incorrect classification of molecules based on the standard deviation and average prediction of a model class (PROB-STD), which provided the best accuracy for classification models as described elsewhere (Sushko et al., 2010), was used as a distance to model. The model AD corresponded to PROB-STD covering 90% of molecules within the training set.
A detailed description of used machine-learning methods, all selected descriptors, and validation procedures can be found in the Supplementary Materials and the OCHEM manual.

| Anti-A. baumannii activity
The antibacterial activity of the studied PILs was estimated against Gram-negative MDR clinical isolate A. baumannii obtained from the Museum of Microbial Culture Collection of the P.L. Shupyk National Medical Academy of Postgraduate Education.
Antibacterial properties were determined by disk diffusion method in Mueller-Hinton agar (Bauer, Kirby, Sherris, & Turck, 1966). A final inoculum concentration of 1 × 10 5 colony-forming unit (CFU) per ml was established using a 0.5 McFarland turbidity standard, and a subsequent dilution of 0.02 ml of the tested compounds was applied to standard paper disks (6 mm) which were placed on agar plate. All compounds were dissolved in 0.1% dimethyl sulfoxide solution (DMSO). The compound content on a disk was 0.3 µM. The studied bacterial culture was not sensitive to DMSO solvent, which was used as a negative control in the experiment. The known antibiotics ampicillin, oxacillin, and ceftriaxone were applied as positive control using the same experimental protocol.
The plates were incubated for 24 hr at 37°C. The tests were repeated three times. The antibacterial activity of the tested compounds was identified by measuring zone diameter of the growth inhibition, which indicates the degree of susceptibility or resistance of A. baumannii isolate against the test compounds. Inhibition zones were measured in millimeters. The compounds with growth inhibition zones >15 mm were selected as active.

| Acute Daphnia magna toxicity testing
In order to exclude the observed activity of compounds due to non-specific toxicity, we evaluated the acute toxicity test of the investigated compounds to Daphnia magna. LD 50 was estimated according to the procedures set out in the Organization for Economic Co-operation and Development Guideline 202 (OECD, 2014 (6-24 hr old) were used for the controls and for the geometric series of concentrations of each test compound. The mortality of the neonates was observed after 48 hr. The LD 50 values in mg/L with their 95% confidence intervals (CI) were determined using Statistica 7 program. All experiments were run in triplicate. The sensitivity of D. magna to the reference toxicant potassium dichromate (K 2 Cr 2 O 7 ) was determined as well.

| Models
The activity data were collected from different sources and thus could have some variations due to differences across laboratory protocols. In order to decrease this variation, we decided to develop classification models. All compounds were assigned to active (147 compounds with MIC ≤ 50 µM) and inactive (116, MIC > 50 µM) classes. The classification models were developed using the protocols described in Section 2.2. All models had similar performance in terms of sensitivity, specificity, and balanced accuracy (BA) as summarized in Table 1 (see also Figure S1). The BAs for the training sets were in the range of 76%-82% (Table 1). The compounds in the test sets were predicted with similar accuracies: BA = 83%-89%. The statistical parameters for both training and test sets have large standard mean errors due to small dataset sizes. Despite some variations, the model performances for both these sets were not significantly different.
A consensus model was built as a simple average of the three models. Despite its statistical parameters being similar to those of individual models (see Table 1), the consensus model was expected to be have a smaller variance. Moreover, the predictions of the consensus model were used to estimate its applicability domain (Sushko et al., 2010).

| Evaluation activity of new compounds
A virtual set of drug-like compounds was generated based on available synthetic blocks and reactions. It included 41 compounds with different substitution patterns (Table S2). Fourteen compounds predicted as inactive as well as three compounds predicted outside of the AD of the model were excluded. The remaining 24 compounds were evaluated for synthetic feasibility by organic chemists, as well as for the absence of potential toxicity (mutagenicity, tumorigenicity, irritation, and reproductive effectiveness) using the DataWarrior software (Datawarrior, 2019). As a result of this analysis, 18 compounds were selected for synthesis and testing (Table 1 and Table S1). The biological testing (Table 2) confirmed their biological activity.
Particular attention was paid to 13 C NMR data. These spectra were remarkable as signals of a 1,3-oxazole ring appear as doublets due to the spin-spin coupling with the phosphorus atom. Thus, C 5 and C 2 atoms resonate at 156.4-174.2 ppm as doublets with coupling constants of 18.9-29.4 Hz. The signal of the C 4 is a doublet at 115.3-128.5 ppm with spin-spin coupling constants of 122.5-146.1 Hz.
The signal of phosphorus nuclei in 31 P NMR spectra of compounds is observed at 9.80-13.21 ppm.
The detailed description of the chemical synthesis of selected compounds (Table S1) was provided in the Supplementary Materials.

| Biology
The results of biological activity are summarized in Table 2.
All synthesized PILs exhibited high activity against MDR clinical isolate A. baumannii with diameters of inhibition zones ranging from 22 to 32 mm (Table 2), while ampicillin, oxacillin, and ceftriaxone antibiotics were inactive. The acute toxicity of PILs ranged from 0.063 to 9.36 mg/L. According to the D.R. Passino classification, six (33%) compounds were classified as moderately toxic (1-10 mg/L), 11 (61%) were highly toxic (0.1-1 mg/L), and only one compound was extremely toxic (0.01-0.1 mg/L). However, despite the analyzed compounds having different toxicity classes, all of them exhibited strong in vitro activity against A. baumannii. Thus, their inhibitory activity was specific and was not connected to their toxicity.
We investigated antibacterial activity of new PILs. Firstly, predictive in silico models based on different machine-learning techniques were built using the OCHEM platform. The created classification models demonstrated good stability, robustness, and predictive power. Eighteen new compounds were synthesized, and their activity against A. baumanii MDR isolate was evaluated. All PILs showed high activity against MDR clinical isolate A. baumannii. According to the acute toxicity results, all tested PILs belonged to moderately and highly toxic classes, and only one substance was classified as extremely toxic. Therefore, the activity of the analyzed structures against A. baumannii was not related to their toxicity and these structures need to be further investigated as promising antibacterial agents. The developed models (http://ochem.eu/artic le/113921) can be used by scientific community to search for new antibiotics against MDR A. baumannii strains. Toxicity classification and LD 50 range: "+"-practically harmless (100-1,000 mg/L); "++"-slightly toxic (10-100 mg/L); "+++"-moderately toxic (1-10 mg/L); "++++"-highly toxic (0.1-1mg/L); "+++++"-extremely toxic (0.01-0.1mg/L).

CONFLICT OF INTEREST
IVT is CEO of BIGCHEM GmbH, which licenses the OCHEM software. Other authors declare no conflict of interest.

DATA AVAILABILITY STATEMENT
All data used in this article are publicly available under CC-BY license at http://ochem.eu/artic le/113921.