2.8.1. Reproductive and Genetic Toxicity

2.8.1. Reproductive and Genetic Toxicity

ADMET Predictor provides a model for general reproductive toxicity (Repro_Tox) as well as two different kinds of genotoxicity endpoints: generation of chromosomal aberrations (Chrom_Aberr) in cultured mammalian cells and mutagenicity against five bacterial mutations assayed with (MUT_mxxx) and without (MUT_xxx) rat liver S9 fraction activation. The models for endocrine disruption described in the Human Toxicity section of the manual may also be relevant.

Reproductive Toxicity

Reproductive and developmental toxicity are synonymous terms. They relate to anything that disturbs the reproductive process of organisms, including adverse effects to sexual organs, behavior, ease of conception, teratogenicity and developmental toxicity to offspring before or after birth. Under the European Union’s REACH (Registration, Evaluation and Authorization of Chemicals) legislation, screening tests for reproductive and developmental toxicity (OECD TGs 421 or 422) are required for substances produced or imported in quantities between 10 and 100 tons per year. Other tests are usually required for substances produced or imported in quantities greater than or equal to 100 tons per year.

Data for our Repro_Tox model was taken from the FDA/TERIS database (Ghanooni 1997) and augmented with data obtained from the literature (Arena 2004). The database was constructed by researchers for the Department of Environmental and Occupational Health at the University of Pittsburg (Sussman 2003) and it combines subsets of information from the Teratogen Information System (TERIS) (Briggs 1990) guidelines set forth by the Food and Drug Administration (FDA) (Shepard 1992). Structures were available for 292 of the 293 compounds of which 114 were considered to be toxic. The model achieves an overall concordance of 82%.

Chromosomal aberrations

Another aspect of genotoxicity is the potential for producing chromosomal aberrations, and testing for such an effect is also used for the assessment of the genotoxic potential of chemicals and drugs that "pass" the Ames test. Just as for the Ames test, compounds are examined with and without a mammalian metabolic activation system containing liver microsomes to see if they induce structural chromosome aberrations in cultured cells. Anticipating the outcome of the chromosomal aberration test from chemical structure may save considerable time and money during drug discovery and development.

ADMET Predictor's Chrom_Aberr ANNE model predicts whether chemicals or drugs will cause chromosome aberration from their 2D structures. Test results for 1,388 compounds were collected from the literature, yielding roughly the same number of positive ("Toxic") and negative ("Nontoxic") examples. We used 1,159 compounds for the training pool and 229 compounds for external testing. Overall, the model’s prediction accuracy is about 80%.

Bacterial mutagenicity

The bacterial mutagenicity panel features a series of MUT_*** models that predict mutagenicity with or without metabolic activation of the test compound. The endpoints they model assess the ability of compounds to promote reversion in a panel of bacterial strains mutated in particular amino acid synthesis pathways (Ames 1975). The original "Ames test" relied only on Salmonella typhimurium mutants, but it has since been expanded by the addition of Escherichia coli repair-deficient strains that bear the same mutations. Such tests are now the primary screen for genotoxicity, having replaced chronic testing in rodents that took longer and cost more.

The Division of Genetics and Mutagenesis, National Institute of Health Sciences of Japan (DGM/NIHS) recently established a unique proprietary Ames database consisting of 13,729 new chemical substances that have not been previously used for developing QSAR models. These chemicals have been classified into three classes: Strongly Positive (Class A), Positive (Class B) and Negative (Class C). (Honma 2019)

The DGM/NIHS made this Ames data available to Simulations Plus and some other software companies for the purpose of building predictive QSAR models. We chose to merge Class A and Class B into a single positive class. The curated data set included 13,141 compounds. It is very imbalanced, with over 11,000 negatives and about 1,800 positives. This model exhibited 76% accuracy in prediction with a test set sensitivity of 78.0% and a test set specificity around 76%.

Dataset provided by DGM/NIHS included around 1500 compounds with experimental mutagenicity against ten different strains. These compounds were appended to our in-house collection of individual strains datasets. All the ten individual strain models were rebuilt. The models present in AMES mutagenicity models in ADMET Predictor 10.4 incorporate this newer data with larger chemical space.

The eleven in silico Ames ANNE models in ADMET Predictor classify compounds as "Positive" (mutagenic) or "Negative" (nonmutagenic). Most model tests are carried out in S. typhimurium strains designated TA98, TA100, etc. The "TA" has been omitted from the model names for brevity. Assay results on strains TA1537 and TA97 are considered interchangeable, as are results for TA102 and repair-deficient E. coli strains WP2. The former are indicated by "97+1537" in the model name and the latter by "102+wp2." A prefix of "m" indicates that the assay included metabolic activation by microsomal (S9) fractions.

Test set sensitivities for the ten models based on older data sets range from 71 to 80%, with an average of 74%. Test set specificities range from 79 to 89%, with an average of 83% (for remaining ten models).

In a regulatory context, a compound "fails" the Ames test if a positive result is obtained for any strain, and the overall reproducibility of the test is 85% or less (Hansen 2009). The models and their errors are not mutually independent, however, so classifying a compound as "Ames positive" because it returns a positive for one of the mutagenicity models is not necessarily the best way to combine their predictions.

Instead, we provide a MUT_Risk model that yields a qualitative estimate of overall mutagenicity by combining individual positive predictions. By default, each individual “Mutagenic” prediction contributes 0.6 “vote” to the score, with the exception of TA98 and TA100. Results for this pair of tests overlap mechanistically (Mortelmans 2000), and so each rule for those strains gets vote of 0.3.

    S_97:     MUT_97+1537 = Positive

    m_97:     MUT_m97+1537 = Positive AND NOT MUT_97+1537 = Positive

    S_98:     MUT_98 = Positive

    m_98:     MUT_m98 = Positive AND NOT MUT_98 = Positive

    S100:     MUT_100 = Positive

    m100:     MUT_m100 = Positive AND NOT MUT_100 = Positive

    S102:     MUT_102+wp2 = Positive

    m102:     MUT_m102+wp2 = Positive AND TOX_MUT_102+wp2 = Positive

    S535:     MUT_1535 = Positive

    m535:     MUT_m1535 = Positive AND NOT MUT_1535 = Positive

    NIHS:     MUT_NIHS = Positive

About 18% of the commercial drugs in the focused subset of the WDI have a MUT_Risk score greater than 1, which is the MUT rule threshold in the default TOX_Risk and ADMET_Risk models. Increasing that threshold to 1.2 would reduce the number of compounds in the WDI subset that violate it to 12%.

MUTx_RISK

Proceedings of an International Workshop on Genotypic Testing were recently published at which one working group assessed the sensitivity and selectivity of the bacterial strains specified in the Organization for Economic Cooperation and Development (OECD) Test Guideline TG471 and recommended several possible modifications to that guideline (Williams 2019). The important outcomes from this discussion were:

1. Of all the mutagens detected by the full TG471 strain panel, 93% were mutagenic in either TA98 or TA100

2. TA100 is derived from TA1535 and it detects more mutagens than TA1535. Testing against TA100 alone would suffice

3. TA97, TA102, and WP2uvrA could be removed from OECD TG471 with little, if any, loss of sensitivity

Based on these findings and other recent literature, we have refined our MUT_RISK model by incorporating interaction rules for the MUT_98, MUT_100 and MUT_1535 models that reflect reported correlations between the TA98, TA100 and TA1535 assay results as well as reducing the contributions made by the models based on data from TA97, TA102+wp2 and TA1537 assays.  To do this, we created 19 orthogonal rules by modifying 6 of the 11 existing rules and adding 8 more. The MUTx_Risk rules that resulted can be summarized as follows:

• The S_97+1537, S102 and NIHS rules remain the same: they are triggered whenever the prediction from the corresponding model (MUT_97+1537, MUT_102+wp2 and MUT_NIHS, respectively) is Positive.

• The m_97+1537, and m102 rules are also unchanged. They are triggered whenever the corresponding model prediction for the respective strain including S9 activation is Positive but the prediction for the assay run on the same strain but without S9 activation is Negative.

• The S_98, S100 and S535 Risk Codes are now triggered only when exactly one of the corresponding model predictions is Positive.

• The m_98 rule is triggered when MUT_m98 is Positive and MUT_98, MUT_m100 and MUT_m1535 are all Negative; the m100 and m535 rules follow the same pattern, which prevents ever counting the same activity twice when it is predicted to be mutagenic with and without S9 activation.

• The combination rule Sx01 is triggered when both MUT_98 and MUT_100 are Positive but MUT_1535 is Negative; Sx02 is triggered when MUT_98 and MUT_1535 are Positive but MUT_100 is Negative; and Sx03 is triggered when MUT_100 and MUT_1535 are Positive but MUT_98 is Negative.

• Rule Sx04 is triggered when MUT_98, MUT_100 and MUT_1535 are all Positive.

• The combination rules mx01, mx02, mx03 and mx04 are constructed analogously to m_98, though the logic involved is rather complex. Please see the MUTx_Risk.ro5 file in the Applications Directory for the details if you are interested.

The MUTx_RISK rule weights were set based on multilinear regression (MLR) analysis of the Ames mutagenicity data set compiled by Hansen et al. (Hansen 2009). The coefficients obtained were re-scaled so as to sum to 4, which yielded a Risk model with better sensitivity and specificity statistics for this data set than those shown by MUT_Risk when a decision threshold of 1.0 was used.

MUTx_Risk also out-performed MUT_Risk on our in-house hand-curated subset of the World Drug Index. Only 12% of the commercial drugs in the focused subset of the WDI have a MUTx_Risk score greater than 1.0, versus the 18% seen with the previous MUT_RISK score. The current MUT rule threshold is 1.0 in the default TOX_Risk and ADMET_Risk models. Increasing that threshold to 1.2 would reduce the number of compounds in the WDI subset that violate it to 8%.