Unveiling the interspecies correlation and sensitivity factor analysis of rat and mouse acute oral toxicity of antimicrobial agents: first QSTR and QTTR Modeling report

Abstract

This study aims to identify toxic potential and environmental hazardousness of antimicrobials. In this regard, the available experimental toxicity data with rat and mouse acute oral toxicity have been gathered from ChemID Plus database (n = 202) and subjected to data curation. Upon the data curation 51 and 68 compounds were left for the rat and mouse respectively for the modeling. The quantitative structure toxicity relationship (QSTR) and interspecies correlation analysis by quantitative toxicity-toxicity relationship (QTTR) modeling was approached in this study. The models were developed from 2D descriptors under OECD guidelines by using multiple linear regressions (MLR) with genetic algorithm (GA) for feature selection as a chemometric tool. The developed models were robust (Q²_LOO = 0.600–0.679) and predictive enough (Q²F_n = 0.626–0.958, CCC_Ext = 0.840–0.893). The leverage approach of applicability domain (ad) analysis assures the model’s reliability. The antimicrobials without experimental toxicity values were classified as high, moderate and low toxic based on prediction and ad. The occurrence of the same classification from QSTR and QTTR models revealed the reliability of QTTR models.Finally, the applied “sensitivity factor analysis” typifies the sensitivity of chemicals toward each species. Overall, the first report will be helpful in the toxicity assessment of upcoming antimicrobials in rodents.

Graphical Abstract

Graphical Abstract.

Introduction

As the populations increasing, the demand of agrochemicals and pharmaceuticals are also increasing. The agrochemicals are used for food demand and pharmaceuticals for health management. Antimicrobials play a vital role among all the pharmaceuticals and it is considered as lifesaving drug.¹ The antimicrobial agents are used in a broad range, from personal care products to waste material treatments.

We benefit greatly from the use of antimicrobial agents, but their use has led to potentially toxic levels of antimicrobials found in the environment.^2,3 Literature reports reveal the alteration of biological and physiological responses, endocrine disruption, and different organ systems toxicity of antimicrobials.^4,5 Several antimicrobials were used as magic bullets during the COVID-19, led to immunosuppression, hypersensitivity and alteration in gut microbiota. These health issues have been experienced in the era of the COVID-19 pandemic and it made situation worse due to the extensive application of antimicrobials.^6–9 These pessimistic effects indicate the need for ecotoxicity assessment of antimicrobials. The toxicity assessments of antimicrobial agents are seldom done during their antimicrobial evaluation. But in the current scenario, environmental toxicity is a great concern for society. The regulatory agencies of developed nations (USA and EU) are seriously working on the parameters of human health and environmental safety assessment of chemicals.

The acute oral toxicity estimation is essential because it indicates food, occupational or accidental domestic poisoning. The experimental toxicity estimation proves to be costly when time, money, workforce, and number of animal sacrifices (animal ethical issue) are considered. There has been a massive demand of different regulations for the reduction of animal testing for over the years. Article 13 (point 1) of REACH clearly states, “In particular for human toxicity, the information shall be generated whenever possible by means other than vertebrate animal tests, through the use of alternative methods, for example, in vitro methods or qualitative or quantitative structure-activity relationship models or from information from structurally related substances (grouping or read-across)“.¹⁰

These, researchers are prompted to use quantitative structure toxicity relationships (QSTRs), species sensitivity differences and interspecies correlations for the toxicity estimation of the new and existing chemicals. Species Sensitivity Distributions (SSDs) are widely used in ecological risk assessments to set protective chemical levels for various species. However, SSDs face uncertainties in data quality, quantity, and model fitting, with intraspecific variability contributing to significant uncertainty. Toxicity values for a single species and chemical can vary over 100-fold, even under controlled conditions. While SSDs are used globally to establish ecological screening levels and water quality standards, broader application to numerous environmental contaminants is limited by a lack of diverse species-specific toxicity data.^11–13 The interspecies toxicity study plays a significant role in toxicity analysis and has already been acknowledged by different regulatory bodies to get the toxicity information of endangered species. Literature reports suggest many advantages of interspecies toxicity: reducing the reliance on expensive testing animals, easy interpretation of the mechanism of toxicity, fulfilling the concept of “the three Rs (replacement, reduction, and refinement)”, extrapolation of toxicity of one species to another, filling the data gap, etc.^14,15

Till the late 2020s, two techniques of interspecies toxicity study were being used

• a) The model of one species is applied to another species to predict the toxicity, and

• b) I-QTTR: The experimental value of surrogate species is used as one of the independent variables.

Our previous reports on acute rodent toxicity demonstrated the correlation between mouse and rat oral toxicity. The robust local QSTR and QTTR models were reported.¹⁶ In this, “The prediction comparison method of interspecies” is also helpful for interspecies toxicity analysis.¹⁴ In continuation, we are proposing the “sensitivity factor analysis” by making the ratio of toxicity value of two species as a response. “Sensitivity factor analysis” will exemplify the sensitivity of the chemicals towards each species apart from the correlation obtained from the QTTR models.^17–22

Rat and mouse have served as the preferred biomedical research species due to their anatomical, physiological, and genetic similarity to humans. OECD guideline (423) also suggested that these rodents can be preferred for the analysis.²³ Approximately 95% of the total genes are similar for all these three species, i.e. rat, mouse, and humans.²⁴ In this regard, we have selected rat and mouse for the toxicity assessment of antimicrobials.

Many in silico toxicity studies were reported in the literature based on rat and mouse acute oral toxicity studies of various chemicals.^16,25–28 To the best of our knowledge, no reports are available in the literature for the in silico toxicity study of antimicrobials on rat and mouse.^29–34 In these contexts, the objectives of this research work were

1. To develop a series of robust, reliable and predictive QSTR and QTTR models under OECD principles for the validation of regulatory (Q)SAR for the purpose of predicting and interspecies toxicity correlation of different antimicrobials in rodents (rat and mouse).

2. To explore the application of developed models to the antimicrobials without experimental toxicity value for filling the toxicity data gaps.

3. Introduction of “sensitivity factor analysis” was the final objective of this research work.

Materials and methods

Dataset and descriptors

We have selected dataset of antimicrobial agents containing 202 compounds for rat and mouse acute oral toxicity for the present work. The toxicity data in the form of LD₅₀ (Median lethal dose: measurement of short-term poisoning potential (acute toxicity) of chemicals which indicates 50% killing of animals for a given period) for antimicrobial agents have been gathered from ChemIDplus database.³⁵ The structures of the collected dataset were downloaded from Pubchem³⁶ and Chemspider³⁷ database and were saved in .mol format and finally checked carefully manually for their correctness. The structures were confirmed by their CAS number. Further, the data was curated as per regulatory guidelines by omitting salts, duplicate, and multiple toxicity values for the same compound, the least value, i.e. most toxic values, was considered for further analysis (worst case scenario).³⁸ The final dataset contains 51 compounds for rat oral LD₅₀ and 68 compounds for mouse oral LD₅₀ (Table S1). The experimental rat and mouse oral toxicity value unit was converted from mg/kg to mmol/kg and further converted into a negative logarithmic scale suitable for QSTR modeling and expressed as pLD₅₀. A total 1,444 1D and 2D molecular descriptors were calculated in PaDEL descriptor software (V 2.21).³⁹ The structure were minimized in Chemdraw Ultra (V10) before the descriptor calculation. After the descriptors calculation, the pre-treatment of descriptors was performed to remove the constant, near-constant, non-informative and highly correlated values. Finally, 450 descriptors for rat acute oral LD₅₀ and 540 descriptors for mouse acute oral LD₅₀ were subjected to further analysis.⁴⁰

Model development

In the recent era, externally validated QSTR models are accepted especially for regulatory purposes. In this regard, the original datasets were divided into training and test set compounds to judge the developed model’s predictive power, and external validation in the present work. The dataset division was done by using structure-based, response-based, and random splitting techniques.^41–43 The approximate ratio of 70:30 was applied for training and test set selection, respectively. For rat oral toxicity number of compounds in the training and test set were 36 compounds and 15 compounds, respectively (Table S1). Similarly, 48 compounds were in the training set for mouse oral toxicity and 20 compounds in the test sets (Table S1).

After the division of the dataset, model development was done by the application of multiple linear regressions (MLR) using ordinary least squares (OLS) and genetic algorithm (GA) for feature selection under OECD principles for validation, for regulatory purposes of quantitative structure–activity relationship models.⁴⁴ In QSARINS, the model development process is done in two steps, i.e. first subset selection up to optimal variable subset and then genetic algorithm among the subsets for the final model. Several models were developed by using genetic algorithm (Genetic iteration = 10,000, mutation rate = 50% and other default settings) for each toxicity endpoint. The best model was initially selected for each splitting based on internal fitting criteria and then externally validated by applying models to respective test set compounds in each species. From dataset splitting to model development, all the process was done by using QSARINS software running under the Windows operating system.⁴⁰

Statistical parameters for validation of models

Several statistical parameters were used to verify the internal stability of the developed QSTR models. The coefficient of determination R² was used as a measure of the goodness-of-fit,⁴⁵ a modification of R²(R²_a) has been proposed due to bias nature of R² with further addition of variables. The internal robustness of the models was verified by the cross-validation coefficient Q²_LOO (leave-one-out). To reduce the data co-linearity QUIK rule was applied.⁴⁶ Additionally, Y randomization, Y-scrambling analysis [2000 iteration] was done to evident that the models were justifiable, not by chance. The application of the developed models to respective test set compounds assured the predictive ability of models. Further, different external validation parameters such as r²m_avg,⁴⁷ Q²F_1, Q²F₂, and Q²F₃,⁴⁸ CCC_ext,^49,50 and the mean absolute error of prediction (MAE_ext)⁵¹ were used to evaluate the model’s performance. Additionally, the root mean squared errors (RMSE) prediction accuracy in the training (RMSE_tr) and the prediction (RMSE_ext) sets were also calculated.⁵¹

Applicability domain

The third and most important OECD guideline (OECD guideline 3) of QSAR is applicability domain (ad) analysis which assures the model’s reliability. The applicability domain of a QSTR model is a theoretical spatial region defined by the specific molecular descriptors of the model and the studied response. On the other hand, it can be defined completely by the nature of the chemicals in the training set but also by the values of the specific model used in the model itself. Considering a QSTR model, just robust, significant, and validated is not sufficient, here, the compounds present in the training set and their chemical space play a very crucial role in the validation of the model. In the present work, we calculated the applicability domain by leverage approach. The leverage method is based on the calculation of the hat matrix: the diagonal value of this matrix (the leverage values: h^*) is used to verify the presence of structural outlier, i.e. those compounds with h greater than the cut-off values of h^*. The h^* value was here calculated as 3p/n, where p is the number of the model variables plus one, and n is the number of training compounds.³⁸

Results and discussion

Rat and mouse acute oral toxicity

In the beginning of model development, the original dataset of rat acute oral LD₅₀ and mouse acute oral LD₅₀ were divided into training and test set compounds by using three splitting techniques as mentioned in section 2.2. With the application of the genetic algorithm number of models were developed by using training set compounds for each splitting in both endpoints, i.e. rat and mouse acute oral LD₅₀. Among the population of models in each splitting initially, the top 50 models were selected based on the cross-validation correlation (Q²_LOO) value. Finally, the best models were selected for each splitting based on the value of different internal and external validation parameters and variable reflection in models. The developed GA-MLR models were found to be statistically robust and interpretable mechanistically. The application of the QUIK rule maintained the co-linearity reduction. The variables that appeared in models were reported in descending order of their contribution reflected in the standardized coefficient value (Table 1), and the correlation matrix of variables for each model was reported in supporting information (Table S2). A significant value was observed for different parameters, such as the adjusted coefficient of variation (R²_a: 64.3%–72.3%) and leave-one-out predicted variance (Q²_LOO: 60.0%–66.6%) in each model. When the models (1–6) were applied to the respective test set for the prediction of test set compounds, the predictive R² value for the test set was statistically significant (Q²F_n: 62.6%–95.8%) (Table 1). A very small value of RMSE for both internal and external also indicated the less prediction errors and goodness of developed.

Table 1.

Values of different internal and external validation parameters of developed QSTR models indicating the robustness of the models.

Rat acute oral LD50
Sl No.	Splitting	Training/Test/ unknown	Equations	R2	R2adj	Q2Loo	Q2Fn	R2mavg	CCCExt	MAETr	MAEExt	RMSEtr	RMSEext
1.	Model 1 Response Based	36/15/139		0.716	0.690	0.621	0.669–0.877	0.674	0.856	0.269	0.170	0.342	0.225
2.	Model 2 Structure-Based	36/15/139		0.747	0.720	0.666	0.691–0.958	0.713	0.853	0.288	0.108	0.337	0.137
3.	Model 3 Random	36/15/139		0.747	0.723	0.653	0.717–0.908	0.682	0.871	0.267	0.163	0.325	0.196
Mouse acute oral LD 50
4.	Model 4 Response Based	48/20/117		0.674	0.643	0.600	0.626–0.769	0.645	0.840	0.276	0.203	0.302	0.255
5.	Model 5 Structure-Based	48/20/117		0.688	0.659	0.600	0.703–0.712	0.631	0.856	0.237	0.222	0.279	0.269
6.	Model 6 Random	48/20/117		0.694	0.665	0.600	0.733–0.791	0.668	0.865	0.250	0.187	0.285	0.235

QSTR models. The scattered plots of the developed models were reported in Fig. 1, which represents the experimental and predicted toxicity values.

Fig. 1.

Scattered plot of developed models indicating experimental vs predicted activity (A = Model1, B = Model2, C = Model3, D = Model4, E = Model5, F = Model6.

In order to minimize the prediction error for the QSAR models, the Arithmetic Residuals in K-groups Analysis (ARKA) descriptors were employed to identify potential outliers.⁵² The ARKA analysis indicated that no outliers were detected in the mouse model. However, in the rat model, compounds 3 and 6 were identified as outliers in models 2 and 3. Initially, these outliers were excluded, and the models were redeveloped using both the original data splitting as reported in the manuscript and an alternative splitting method. Upon further analysis, two consistent outliers were identified in both models 2 and 3. These outliers were removed, and the models were redeveloped again using the same and alternative splitting approaches. The results showed no significant changes in the statistical parameters compared to those reported in the Table S7. Therefore, it was concluded that the inclusion or exclusion of these compounds does not significantly affect the model quality, confirming that the reported models are robust and applicable.

Mechanistic interpretation of descriptors

We tried to address the OECD principle 5 for QSAR model validation for regulatory application, i.e. “a mechanistic interpretation, if possible” (OECD Guideline 5). The mechanistic interpretation aims to explain the relationship between the structural features of the compounds and the activity/toxicity. The interpretation of the variables was briefly described in this section.

Rat acute oral LD₅₀

Variable MDEC-23 (Molecular distance edge between all secondary and tertiary carbons) with the positive contribution and highest standard coefficient value was most important for Rat oral LD₅₀ models (Model 1–3). The code MDEC-23 indicated the Molecular (M) distance (D) edge (E) between all secondary (2) and tertiary (3) carbons (C). The rat oral toxicity of antimicrobials was directly correlated with the increasing value of this variable. The minimum distance between secondary and tertiary carbon atoms increases the value of this variable and the toxicity simultaneously (such as Rifampin, Ketoconazole, Praziquantel, Clotrimazole, and Miconazole). The structure analysis of these compounds showed that the secondary and tertiary carbon atoms are directly attached to each other.

maxsssCH (Maximum atom-type E-State: >CH-) is one of the atom type Electrotopological descriptors that positively contribute to rat toxicity. Electrotopological state indices are calculated for each atom in a molecule and encode the information about the topological environment of the atom and electronic interactions due to all other atoms.

Variable ATSC3m (CenteredBroto-Moreau autocorrelation—lag 3/weighted by mass) is an autocorrelation descriptor in which the autocorrelations are calculated by Broto-Moreau algorithm. The number 3 indicated the lag3, i.e. the topological distance between the pair of electron and m stand for the relative mass. A negative contribution suggestsan inverse correlation with the activity that an increase in ATSC3m value willdecrease the rat toxicityand vice versa. The structure analysis indicates the presence of two hydrophilic functionality with the same mass in 3 chain lengths increases the numerical value of this variable and decreases the rat oral toxicity. This functionality was observed in compounds (Chloramphenicol, Neomycin, and Lomefloxacin).

maxHBint7 (Maximum E-State descriptors of strength for potential Hydrogen Bonds of path length 7) belongs to atom type electrotopological descriptors and has a negative contribution to rat toxicity. The structure analysis indicated that the presence of a hydrophilic group or atom (-OH, O) in 7 order length increases the value of this variable which was observed in compounds (Lomefloxacin, Levofloxacin, and Ofloxacin).

Mouse oral LD₅₀

Variable GATS3c (Geary autocorrelation—lag 3 / weighted by charges) is the autocorrelation variable in which the Geary algorithm calculates the autocorrelations. The lag 3 and weighted by charges indicated that the presence of the same charged atom or group in 3 topological distances enhances the value of this variable. The structure analysis of the compounds with the higher and lower numerical value of this variable indicated that the presence of the same charged functional group or atom at 3 topological distances in any compound increases the value of this variable. On the other hand absence of this functionality decreases the value of this variable. The compounds (Proguanil, Chloroquine, Albendazole, Piperaquine and Nitazoxanide) with high value of this variable showed high toxicity towards mouse oral toxicity.

Variable SdsN (Sum of atom-type E-State: =N-) with positive contribution influences the toxicity of the antimicrobials in the mouse. The type of nitrogen influences the value of this variable. This functionality was observed in Rifabutin, Levamisole, Rifampin, and Pyrimethamine.

In variable nX (Number of halogen atoms (F, Cl, Br, I, At, Uus) the code nX indicated the number of the halogen atom. A Positive contribution of nX stated the presence of halogens would impart toxicity to the mouse. The investigation of structures revealed that the presence of halogen atoms (Cl, F) increases the toxicity of the compounds. The compounds (Chloramphenicol, Diloxanidefuroate, and Fluconazole) with chlorine and/or fluorine atom showed high toxicity compared to the compounds (Cycloserine, Undecylenicacid,and Benzoic acid) without any halogen atoms.

Variable MDEO-12 (Molecular distance edge between all primary and secondary oxygens) is a molecular distance edge vector type of variable. The coding MDE indicates the molecular distance edges, and O stands for the oxygen, while the number 1 and 2 codify the primary and secondary oxygen types. Hence the code suggests the molecular distance edges between all the primary and secondary oxygen present in molecules influence the value of this variable and thus the activity/toxicity of the compounds. It has a positive contribution, and the least influential variable appeared in model4.Therefore, the distance between primary and secondary oxygen atoms also influences the antimicrobial’s toxicity towards the mouse oral toxicity.

With positive contribution MDEC-33 (Molecular distance edge between all tertiary carbons) was the least influential variable for models (4 and 6). Increasing chain length among all tertiary carbon atoms increases the value of this variable, and the increasing value of this variable increases the toxicity of the compounds as observed in compounds (Clarithromycin, Erythromycin, Spiramycin, and Paromomycin).

Applicability domain analysis of the developed models

In this work, we have calculated the applicability domain by leverage approach as mentioned earlier in section 2.4. The observation of the ad analysis of the developed models revealed that approximately 90%–99% of the compounds were inside the ad for model (1–6), indicating the high prediction reliability. The william’s plots of developed models were given in Fig. 2(1), showing the outside ad compounds in the model (1–6). The detail of ad analysis is reported in Table S1.

Fig. 2.

William’s (1), and Insubria (2) plot of developed models indicating outside ad compounds (a = Model1, B = Model2, C = Model3, D = Model4, E = Model5, F = Model6).

Estimation of predictions of non-tested antimicrobials for rodent toxicity from the developed models: Data-gap filling

The developed rat acute oral LD₅₀ models (models 1–3) were applied to 139 antimicrobial agents and mouse acute oral LD₅₀ models (Models 4–6) to 117 antimicrobial agents to predict the toxicity value and for the classification of the unknown compounds (Having no experimental toxicity values).

Based on prediction and ad analysis, the antimicrobial agents were classified as high (LD₅₀ < 100), moderate (LD₅₀ = 100–2,000), and low toxic (LD₅₀ > 2,000) according to USEPA classification⁴⁷ in a consensus manner (Fig. 3). The outside ad antimicrobials were not considered for the classification. The detail of prediction, ad and classification was reported in supporting information (Table S1). The Insubria plots of developed models were given in Fig. 2(2), representing the outside ad compounds.

Fig. 3.

Classification of unknown antimicrobials (high (LD50 < 100), moderate (LD50 = 100–2000) and low toxic (LD50 > 2000)).

The toxicity classification of unknown antimicrobials based on predictions from developed models revealed that none of the antimicrobials was highly toxic for rat and mouse. Among the unknown antimicrobials, 76 common compounds were observed for both endpoints, in which 36 compounds were the same classified for both the species (Table S3).

Quantitative toxicity-toxicity relationship (QTTR): Rat and mouse oral LD₅₀

The idea of toxicity-toxicity relationship study comes from the genetic linkage among animal species. Genetically rat, mouse and humans have around thirty thousand genes, of which approximately 95% are shared by all three species already reported in the introduction section.¹⁵ Our previous study also suggests the same.¹³ High toxicity correlation (R = 0.780) was observed between rat and mouse oral experimental toxicity value of common antimicrobials (n = 57), which also encouraged us for the toxicity-toxicity relationship analysis (Table S4).

QTTR Modeling for rat and mouse acute oral LD₅₀

To get a toxicity-toxicity correlation between rat and mouse oral acute toxicity initially, we have selected the same antimicrobials having the experimental toxicity data for both, i.e. rat and mouse. A total of 57 antimicrobials were selected for the QTTR modeling. The toxicity value of surrogate species was taken as one of the independent variables for QTTR model development. The model development and selection were made in the same way as QSTR modeling (Model 1–6). The selected QTTR models showed significant value for different internal and external validation parameters (Table 2). The three-descriptor combination of nHBint7, SIC0 along with pLD_50(M) for rat oral QTTR models (model 7 and 8) and three descriptor combination of nHBint7, hmin along with pLD_50(M) for rat oral QTTR model 9 (Table 2). Similarly, three descriptor combinations of SHBint10, SHsOH along with pLD_50(R) for mouse oral QTTR models (model 10–12) were observed (Table 2). The scattered plots of the developed QTTR models were given in supporting information (Fig. S1). The ad analysis of the developed models revealed that the more than 90% of antimicrobials were inside the ad for rat oral QTTR models (models 7–9), and more than 95% of the antimicrobials were inside the ad for mouse oral QTTR models (models 10–12) which clearly indicates the developed QTTR models were satisfactory reliable. The outside ad compounds can be seen in William’s plots of developed models (Fig. S2).

Table 2.

Values of different internal and external validation parameters of developed QTTR models indicating the robustness of the models.

Rat acute oral LD50
Sl No.	Splitting	Training/Test/ unknown	Equations	R2	R2adj	Q2Loo	Q2Fn	R2mavg	CCCExt	MAETr	MAEExt	RMSEtr	RMSEext
1.	Model 7 Response Based	40/17/27		0.756	0.736	0.668	0.692–0.839	0.650	0.850	0.269	0.195	0.317	0.257
2.	Model 8 Structure-Based	40/17/27		0.746	0.725	0.660	0.766–0.888	0.729	0.893	0.267	0.181	0.325	0.216
3.	Model 9 Random	42/15/27		0.754	0.735	0.669	0.721–0.880	0.689	0.877	0.260	0.179	0.316	0.221
Mouse acute oral LD 50
4.	Model 10 Response Based	41/16/5		0.744	0.723	0.679	0.741–0.844	0.658	0.865	0.253	0.196	0.327	0.256
5.	Model 11 Structure-Based	41/16/5		0.739	0.718	0.678	0.754–0.859	0.661	0.868	0.258	0.174	0.330	0.242
6.	Model 12 Random	41/16/5		0.739	0.718	0.664	0.739–0.781	0.619	0.840	0.246	0.219	0.318	0.291

Note: pLD _50m and pLD _50r  = −log(mmol/kg)

Finally, the developed QTTR models for rat oral (Model 7–9) were applied to antimicrobials (n = 27), having no experimental value for rat oral toxicity but having an experimental toxicity value for mouse oral toxicity; similarly, mouse oral QTTR models (Model 10–12) were applied to antimicrobials (n = 5) having no experimental toxicity value for mouse oral toxicity but having experimental toxicity value of rat oral toxicity to get prediction and to classify them. The classification of the unknown antimicrobials was done similarly as that of QSTR models (Fig. 4). The detailed classification and ad analysis were reported in supporting information (Table S5 and Fig. S3 (Insubria plots)).

Fig. 4.

Classification of unknown antimicrobials (high (LD50 < 100), moderate (LD50 = 100–2000) and low toxic (LD50 > 2000)).

The prediction and classification obtained from the QTTR models (Models 7–9 for rat oral and 10–12 for mouse oral) of these unknown antimicrobials (n = 27 for rat oral and n = 5 for mouse oral) were compared with the prediction and classification obtained from QSTR models (Models 1–3 for rat acute oral LD₅₀ and 4–6 for mouse acute oral LD₅₀). A very interesting result was observed; among the 21 antimicrobials for rat acute oral LD₅₀, 15 antimicrobials (>70% antimicrobials) were observed to be the same classified to that classification based on a prediction by QSTR models (Model1–3) (Table S5). Similarly, the prediction and classification comparison of 4 unknown antimicrobials for mouse oral LD₅₀ was done in the same fashion as that of rat acute oral LD₅₀ (Table S5). Among the 4 antimicrobials, 1 antimicrobial was observed to be the same classified that classification based on a prediction by QSTR models (Model4–6) (Table S5). This observation suggests significant applicability of the QTTR models for toxicity-toxicity analysis between rat and mouse.

Introduction of sensitivity factor analysis

Finally, we have introduced the sensitivity factor analysis that helps to predict sensitivity of the particular antimicrobial towards a specific species, i.e. rat or mouse. The sensitivity factor calculation was done in a similar fashion to that of the selectivity factor calculation. In the case of the selectivity factor, a higher value of the selectivity factor indicates the particular compound is selective toward the denominator target, while the lower value of the selectivity factor indicates the specific compound is selective toward the numerator target. The following formula was applied for the calculation of the sensitivity factor:

(1)

Equation (1) indicates that the higher value of the sensitivity factor of a particular antimicrobial is more sensitive to the mouse. On the other hand, lower value of the sensitivity factor of a specific antimicrobial is more sensitive toward the rat.

To calculate the sensitivity factor, we have selected the whole antimicrobials (Known and Unknown) except outside the ad (Consensus ad) and the antimicrobials expelled out during the model development due to their influential property in model development. A total of 133 antimicrobials have been gathered for sensitivity factor analysis. The analysis of sensitivity factor calculation revealed that the 73 (54.89%) compounds were more sensitive toward mouse with SF value of more than 1, while 60 (45.11%) compounds were more sensitive to the rat (Table S6). Further, the SF analysis was done for the major classes of antimicrobials, and a detailed analysis was given in supporting information (Table S6).

Comparison with the published models

Numerous QSAR models have been developed to predict rat and mouse acute oral toxicity across various chemical classes, but none specifically address antimicrobials (Table S8).^{16,27,53–56} Antimicrobials are structurally diverse and generally larger, which makes them more likely to fall outside the applicability domain of models designed for other chemicals. Of the reported QSAR models, only four have focused on interspecies toxicity,^16,53–55 with just two offering models for surrogate species, and three demonstrating practical applications.^16,53 Additionally, these models do not account for species-specific chemical sensitivity, further limiting their applicability to antimicrobials.^16,54,55

In our analysis, we applied the rat acute oral toxicity model available in QSARINS software to a set of antimicrobial compounds and found that more than 75% were outside the model’s applicability domain (Fig. S6). This suggests that existing QSAR models are not well-suited for predicting antimicrobial toxicity, emphasizing the need for dedicated models that account for the unique properties of antimicrobials to improve predictive accuracy and reliability.

Overview and conclusion

In the present work, a series of QSTR and QTTR models were developed for rat and mouse acute oral toxicity of antimicrobial agents in accordance with OECD principles for validation for regulatory purposes of quantitative structure–activity relationship models. Different statistical matrices assured the robustness and reliability of developed models (QSTR and QTTR models). The mechanistic interpretation of the descriptors has been briefly described, which suggests the distance between secondary and tertiary carbon atoms, electrotopological state, and hydrophilic-hydrophobic functionality influence the toxicity of the antimicrobials towards rat oral toxicity while the number of halogen atoms (F, Cl), presence of (–N=) type of nitrogen atoms, the distance between primary and secondary oxygen atom influence the toxicity of the antimicrobials towards mouse oral toxicity.

Applications of the models have been commented on, and the antimicrobials not having experimental toxicity value have been classified as high (LD₅₀ < 100), moderate (LD₅₀ = 100–2,000) and low toxic (LD₅₀ > 2,000) toxic based on the prediction in a consensus manner for each species, i.e. rat and mouse.

Interesting results were obtained upon the comparison of classification of antimicrobials between rat and mouse; 47% of the common antimicrobials have been classified the same. Further, the classification of antimicrobials has been compared based on the prediction from QSTR and QTTR for both species. More than 70% antimicrobials classified same in the case of rat oral, and 25% antimicrobials classified same in the case of mouse oral.

Finally, “the sensitivity factor analysis” was introduced that helps to evaluate the sensitivity of the compounds towards specific species. The observations revealed that ≈55% of the antimicrobials were more sensitive to mouse. Further, the antiviral, azoles, quinolones, tetracycline etc classes of antimicrobial were more sensitive to mouse.

The overall results obtained from this research work will be helpful for the society and scientific community in the toxicity assessment of upcoming antimicrobial agents.

Funding

Financial assistance from the Science & Engineering Research Board (SERB) DST, Govt. of India, New Delhi (File No. EMR/2017/004497) is gratefully acknowledged by Dr. Partha Pratim Roy.

Conflict of interest statement: There are no conflicts to declare.