Congenital malformations are defined as the malformations of organs or body parts during development in utero, mainly attributed to hereditary, maternal, external environmental or some unknown factors. Major congenital malformations (MCMs) are defined as structural abnormalities with surgical, medical, functional, and or cosmetic importance (Tomson et al., 2019). It is not only an irreparable blow to the family, but also a burden on society. As a potential teratogen, a number of drugs can pass through the placental barrier and affect the growth and development of the fetus in utero. It is commonly accepted that the most sensitive period to teratogens is during active organogenesis, which is from three to 8 weeks after fertilization. Some organs, such as the brain, will continue to be very active developmentally after active organogenesis and may still be affected by teratogens (Ornoy, 2009).

Antiseizure medications (ASMs) are used to control various types of convulsive disorders or as mood stabilizers. About 0.3–0.7% of pregnant women are diagnosed with epilepsy (Viinikainen et al., 2006), most of which receive ASMs monotherapy (74%) (Meador et al., 2018). Although the incidence of MCMs in offspring of women with epilepsy (WWE) receiving ASMs monotherapy during pregnancy is very low (4%), the rate is still higher than that of MCMs in the general population (Campbell et al., 2014). In clinical practice, the lowest effective dose is usually recommended for seizure control for the safety of both mother and child. Furthermore, WWE should avoid unexpected pregnancies because ASMs will take a period of time to metabolize to a safe level for embryo. Therefore, the teratogenic risk evaluation of ASMs is crucial for the clinical medications.

During the last decade, data from large prospective epilepsy and pregnancy registries have been used to compare the incidence and the risk of MCMs in offspring exposed to different ASMs. When exposed to different antiseizure medications (ASMs), the incidence of teratogenesis exhibits significant difference. The incidence of MCMs in offspring exposed to valproate and phenobarbital is 5–11% and 6–7%, respectively, while it is only 2–3% for lamotrigine and levetiracetam (Brodie, 2006; Hernandez-Diaz et al., 2012; Tomson and Battino, 2012; Bromley et al., 2017; Tomson et al., 2018). Fetal exposure to ASMs is associated with an increased risk of MCMs. Although there is a variation in the rate, the evidence in pregnancy registries implies that valproate is associated with the highest risk of MCMs among ASMs. The post-marketing surveillance data in US Food and Drug Administration (FDA) Spontaneous Adverse Events Reporting System (FAERS) also indicates the high prevalence of congenital malformations induced by valproate. (Supplementary Table S1). In addition, carbamazepine, phenobarbital, phenytoin, and topiramate are also associated with the increased risk of MCMs in offspring, while lamotrigine and levetiracetam are of relatively low teratogenic risk (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Campbell et al., 2014; Tomson et al., 2018). In preclinical studies, zebrafish embryo model and the animal models have also been established to evaluate the developmental toxicity, and teratogenic activity of the drugs (Murayama et al., 2018; Chetot et al., 2020; Jarque et al., 2020; Nguyen et al., 2021; Simeon et al., 2021). In addition, a few in silico models have been proposed to predict the reproductive toxicity of the chemical compounds (Basant et al., 2016; Jiang et al., 2019), but none of them is developed for the purpose of teratogenic risk prediction.

Although large scale epilepsy pregnancy registries played an important role in evaluating the teratogenic risk of ASMs, the teratogenic risks of a large proportion of ASMs, especially the newly approved ones, are still unknown because of the limited post-marketing surveillance data. To fill this gap, we endeavoured to construct a model to predict the potential teratogenic risk of the drugs based on their chemical structural information. The detailed workflow was depicted in Figure 1.

To facilitate the prevention and early detection of drug-induced teratogenesis in pregnancy, we generated a curated dataset and developed a SAR model to predict the potential teratogenic risk of a drug. The model was subsequently applied to evaluating the teratogenic risk of ASMs.

Although the incidence of drug-induced malformations accounts for only 1% or less of the cases of birth defects with known causes (De Santis et al., 2001), the teratogenicity of drugs has always been a widespread concern in clinical medication. Since the first case report indicated that ASMs were associated with congenital malformations in the 1960s (Janz, 1964; Meadow, 1968), a number of efforts had been made for assessing the teratogenic risk of ASMs, so as to help pregnant women avoid the harm of drugs to the offspring while treating epilepsy (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Campbell et al., 2014; Thomas et al., 2017; Tomson et al., 2018; Tomson et al., 2019). Case-controlled studies and animal studies are currently reliable approaches to evaluate the teratogenic risk of ASMs. The teratogenicity of commonly used ASMs including valproate, carbamazepine, and lamotrigine had been systematically investigated by reviewing the prevalence of major congenital malformations in epilepsy pregnancy registries (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Campbell et al., 2014; Tomson et al., 2018; Tomson et al., 2019). However, for a large proportion of ASMs, especially the newly approved ones, the lack of evidence is a big hurdle to effectively estimate the teratogenic risk. To fill this gap, we proposed a SAR model to detect the drugs with high teratogenic risk and applied it to assess the teratogenic risk of ASMs. The model was constructed by using SVM algorithm with the chemical structural descriptors of the drugs as features.

As a result, the SVM model performed the best when compared to the state-of-art machine learning approaches (Figure 2). The average MCC, F1 score and recall rate for classifying the teratogenic risk of drugs in the curated dataset were 0.312, 0.762, and 0.851, respectively. When validated by the FAERS dataset, the SVM model detected 136 of 196 drugs with teratogenic risk (recall rate = 0.694). The distribution of the predicted probability of the teratogenic risk of the drugs were shown in Supplementary Figure S1. For the subsets of top ten and top 25 drugs ranked by their RORs, the recall rates achieved by the SVM model were 0.700 and 0.800, respectively, indicating that the model was more sensitive to detecting the high teratogenic risk drugs (Figure 3B). Interestingly, the model seems to be sensitive to detecting the teratogenic risk drugs in two therapeutic categories, namely antiinfectives for systemic use and antineoplastic and immunomodulating agents, for which the recall rates were 91.7% and 92.3% (Figure 3A).

For the risk prediction of ASMs, valproate, topiramate, phenobarbital and phenytoin were predicted to be high teratogenic risk, which were consistent with drug labeling information, and reported prevalence in pregnancy registries (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Campbell et al., 2014; Tomson et al., 2018; Tomson et al., 2019). The FDA-approved drug labeling of these four drugs included clear statement of the teratogenic risk of use during pregnancy. Valproate was warned in the BOXED WARNING labeling section that it can cause major congenital malformations, e.g., neural tube defects, and should avoid treating women with epilepsy during pregnancy. The teratogenic severity of topiramate, phenobarbital and phenytoin was also indicated in WARNINGS AND PRECAUSIONS labeling section that infants exposed to these drugs in utero had an increased risk for congenital malformations and other developmental outcomes. Especially for topiramate and phenobarbital, case-controlled studies and the data in pregnancy registries indicated that the use of these two drugs might increase the risk of fetal abnormalities. Furthermore, relatively high prevalence of major congenital malformations of valproate, topiramate, phenobarbital, and phenytoin is reported in the pregnancy registries (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Campbell et al., 2014; Thomas et al., 2017; Tomson et al., 2018; Tomson et al., 2019). In addition, five commonly used ASMs including lamotrigine, levetiracetam, gabapentin, carbamazepine, and oxcarbazepine were predicted to be low teratogenic risk, which was also consistent with the low reported prevalence of major congenital malformations in pregnancy registries (Tomson et al., 2011; Hernandez-Diaz et al., 2012; Thomas et al., 2017; Tomson et al., 2018; Tomson et al., 2019).

The chemical structures of drugs containing the toxicophores (or structural alerts (SAs)) was considered as one of the important factors of drug-induced teratogenesis (Sankar and Lerner, 2008). For instance, the aromatic ring in the chemical structure of a drug, e.g., lamotrigine, may produce a highly reactive epoxide or arene oxide in the metabolism through cytochrome P-450 family of enzymes and generate reactive metabolites, which may contribute to teratogenesis (Sankar and Lerner, 2008). To further investigate the structural features of ASMs and help better understand the mechanisms of teratogenicity, the SAs associated with known toxic property were extracted from the structures of 39 ASMs. We uploaded the SDFs of the ASMs into the Online Chemical Modeling Environment (OCHEM, https://ochem.eu/) (Sushko et al., 2011) and used a web server named ToxAlerts (Sushko et al., 2012) for mapping the toxicological SAs. Consequently, a total of 23 SAs that might be associated with teratogenicity were extracted, which belonged to four endpoints, namely genotoxic carcinogenicity and mutagenicity, idiosyncratic adverse reaction, potential electrophilic agents and endocrine disruption (Table 4). The chemical structure of the common antiepileptic drug valproate contained the carboxylic acids, which might lead to the formation of reactive metabolites, and then induce the dose-dependent idiosyncratic adverse reaction (Kalgutkar and Soglia, 2005). Another drug phenytoin contained a toxicological alert ethane-1,1-diyldibenzene, which was reported to be of potential estrogenic and androgenic activities (Nendza et al., 2016). In addition, ezogabine was predicted to be a high teratogenic risk drug by the SVM model with the highest predicted probability (0.858), for which the chemical structure contained three SAs. One of the SA was anilines, which might form the reactive metabolites by bioactivation and was linked to the idiosyncratic adverse reaction (Kalgutkar and Soglia, 2005). The other two SAs were aromatic amines and derivatives of urethane, which were reported to be associated with genotoxic carcinogenicity and mutagenicity (Ashby and Tennant, 1988).

At present, three third-generation antiepileptic drugs, namely perampanel, rufinamide, and brivaracetam were used in clinics and there is still insufficient data to support the risk assessment of teratogenicity. The predicted probability of teratogenicity was 0.80, 0.75, and 0.50 for perampanel, rufinamide and brivaracetam, respectively, indicating a high risk of perampanel and rufinamide and a moderate risk of brivaracetam. In the FDA drug labeling, the evidence provided by animal studies showed that these three ASMs had developmental toxicity in pregnant rats or rabbits at clinically relevant doses or at plasma exposures greater than clinical exposures, including visceral abnormalities, skeletal abnormalities, decreased fetal weight and embryo-fetal death for perampanel exposure at dose of 1, 3 or 10 mg/kg/day, skeletal abnormalities, decreased fetal weight and embryo-fetal death for rufinamide exposure at dose of 20, 100, and 300 mg/kg/day, and embryo-fetal death and decreased fetal weight only for highest dose brivaracetam exposure. One case report reported 2 cases of minor congenital malformations in the offspring exposed to brivaracetam during pregnancy (Paolini et al., 2020). As a result, our model can be used as a tool to alert new drugs that still lack clinical evidence for teratogenicity.

Some caveats of this study were worth further discussing. It is still a big challenge to develop a reproducible procedure to assess the teratogenic risk for the drugs. A set of attributes, such as severity, expectedness and causality, should be considered in annotating the teratogenic risk for a drug. In this study, we annotated the teratogenic risk for the drugs based on the FDA-approved drug labeling by considering the fact that drug labeling is a compilation of up-to-date drug safety information and a relatively accurate and stable data source. Consequently, we annotated 67 drugs with teratogenic risk (positive samples) and 45 drugs with low teratogenic risk (negative samples). The average precision achieved by our model in 1,000 iterations of 5-fold cross-validation was 0.690. When validating the model by using the FAERS data, we can only determine the drugs with teratogenic risk when the teratogenicity is reported. As a result, we only tested the recall rate of our model. Note that the annotation was based on the US FDA-approved drug labeling. The classification of the drugs might be different when using the drug labeling of other countries. The teratogenic risk of the drugs used in the modeling is determined by the FDA drug labeling. In the drug list, all of the 26 drugs in the therapeutic category of antineoplastic and immunomodulating agents (L) were assigned to the high teratogenic risk group (positive samples). It may be one of the main reasons that 92.3% of drugs were predicted to have teratogenic risk when using the FAERS dataset to validate the predictive model. In comparison, the number of drugs in the therapeutic category N (Nervous system) in the high teratogenic risk group (positive samples) and the low teratogenic risk group (negative samples) is relatively balanced (7 positive and 10 negative samples). As a result, 53.8% of the drugs in the FAERS dataset were predicted to be teratogenic. When more balanced data can be obtained with the update of the FDA drug labeling, the predictive performance of our model can be further improved. Our proposed model only based on the chemical structures of the drugs. Although it is simple and easy to use, the limited information restricted the predictive performance of the model. Considering that drug-induced malformations are usually dose-dependent and affected by many factors. Additional information, such as the cell-based in vitro data, the data from animal studies as well as the pharmacogenomics data, involved in the modeling procedure may improve the model performance. Furthermore, a comprehensive model is also needed.

In this study, we proposed a SAR model for predicting the potential teratogenic risk of the ASMs based on the chemical structures of the drugs. The model was sensitive to detecting the high teratogenic risk drugs. Our findings can be helpful for the prevention and early detection of drug-induced teratogenesis.