Endocrine disruptors (EDs) are exogenous compounds that affect the endocrine system of humans and other vertebrates. Endocrine activity of environmental or foreign chemicals has the potential to cause numerous adverse outcomes, including disrupting the physiologic function of endogenous hormones and altering homeostasis. The known EDs include polychlorinated biphenyls (PCBs), the synthetic estrogen diethylstilbestrol (DES), dichlorodiphenyltrichloroethane (DDT) and other pesticides. For example, DES was approved the Food and Drug Administration (FDA) for treatment of menopausal symptoms, gonorrheal vaginitis, atrophic vaginitis, postpartum lactation suppression, and prostate cancer [1
]. DES was shown to disrupt the endocrine system causing vaginal tumors in girls and women and other adverse medical complications [3
] and thus was withdrawn from the market by the FDA. Concern about EDs has invigorated intense discussion and debate over the past two decades in the scientific community [4
] and promoted the legislation for regulation of environmental chemicals mandated by the Environmental Protection Agency (EPA) and development of the Endocrine Disruptor Screening Program (EDSP) to screen potential EDs in the environment [6
EDs can disrupt the endocrine system through different mechanisms [7
]. One of the well-known mechanisms is mediated by the hormone receptors such estrogen receptor (ER) and androgen receptor (AR), in which EDs exhibit their estrogenic and androgenic effects through binding to the ER and AR in the target cells [13
]. Therefore, a huge amount of estrogenic and androgenic activity data of structurally diverse chemicals have been generated and organized in sophisticated databases such as the FDA’s Endocrine Disruptors Knowledge Base (EDKB) [17
] and Estrogenic Activity Database (EADB) [18
]. These databases have been used for the development of a diverse set of quantitative structure-activity relationship (QSAR) models for predicting estrogenic and androgenic activity and to assist evaluation of endocrine disruption potential of environmental chemicals [19
ED binding to hormone receptors in target cells is the key mechanism to display endocrine disruption. However, the affinity of binding to ER is not the sole criterion to determine EDs’ potential to disrupt the endocrine system. For example, EDs cannot bind to ER or AR in the target cells if they cannot pass the cell membrane. Therefore, in vitro
ER and AR binding data of chemicals may not reflect well their in vivo
endocrine activity, even for chemicals with high in vitro
binding affinity. To accurately estimate the endocrine disruption potential of environmental chemicals, it is necessary to have both their binding activities to hormone receptors and to competing serum proteins such as alpha-fetoprotein (AFP) [30
] and human sex hormone-binding globulin (SHBG) [32
There are different transporter proteins in serum, including albumin, globulin, fibrinogen, and others. The transporter proteins can transport hormones, vitamins and other chemicals within and between cells and organs. SHBG is one of the major transporter proteins that bind to hormones and other chemicals in human serum [33
]. AFP is a major transport protein in rat and was first discovered approximately 60 years ago [34
]. It is a serum biomarker of Down’s syndrome and neural tube defects in the clinical practice and alters the growth of fetal and cancer cells [35
]. Entrance of AFP into cells through receptor-mediated endocytosis was observed in fetal cells of different species including rat [37
], mouse [38
], human [39
] chicken [40
] and baboon [41
]. Elevated AFP level was observed in maternal circulation through transplacental passage from the fetal circulation and amniotic fluid by the placental or allantois [42
]. This protein competes with ER to bind estrogens in the blood and thus inhibits EDs access to the target cells [46
]. It has been found that diverse chemicals bind AFP [30
A huge amount of in vitro
binding assays data have been generated for the targets such as ER and AR involved in the endocrine system. However, available in vivo
bioactivity data related to endocrine disruption potential are relatively less than the in vitro
data. Moreover, most of the in vivo
data are obtained using rats uterotrophic assays [17
]. To better assess endocrine disruption potential of environmental chemicals, we measured rat AFP binding affinity for 125 chemicals with diverse structures using a competitive binding assay according to the methods published in our previous study [30
]. Our rat AFP binding data represent the largest such data set to date. Compared with the experimental data on the hormone receptors such as ER and AR, there are fewer chemicals with experimental AFP binding data, hindering the risk assessment of environmental chemicals in terms of endocrine disruption potential. Therefore, for an enhanced risk assessment it was necessary to obtain AFP binding data for those environmental chemicals lacking AFP binding data. To this end, we developed an in silico
model for prediction of AFP binding activity of environmental chemicals using our previously reported data [30
]. The performance of the model was internally evaluated through cross validations and permutation tests. It was also validated externally using the AFP binding activity data curated from the literature. We demonstrated that the model has suitable predictive power and is expected to better assist endocrine disruption assessment of environmental chemicals.
AFP is a protein in the plasma that binds to estrogens with high affinity. It can sequester EDs in the plasma and thereby reduces the concentration of EDs that can enter into the target cells. Thus, AFP can protect EDs in maternal circulation. Hence, AFP binding activity of chemicals is important information for assessment of endocrine disruption potential. If a chemical does not bind to AFP but binds to hormone receptors such as AR and ER, it can bypass AFP protection and has the potential to disrupt the endocrine system. In contrast, if a chemical binds to AFP, AFP could protect against endocrine disruption even if it has the potential to bind AR or ER. However, a very limited number of chemicals have been experimentally assayed for their AFP binding activity. Thus, we previously measured AFP binding activity for 125 structurally diverse chemicals using the competitive assay developed from rat amniotic fluid [30
]. The number of chemicals with AFP binding activity data is still much smaller than the chemicals having ER and AR binding activity, hampering comprehensive assessment of endocrine disruption potential for environmental chemicals. Therefore, in this study, we developed and extensively validated AFP binding activity prediction models using the data published in the literature including our in-house data set. Our model showed a reasonable predictive power and robustness and could be expected to help assess endocrine disruption potential of environmental chemicals.
The DF prediction model was constructed using rat AFP binding data. It could be used for prediction of rat AFP binding activity for the environmental chemicals that have no experimental data. However, the limitation of current model should be noticed when applying the model in applications of human risk assessment of environmental chemicals because the human AFP is not completely homologous to the rat AFP.
Prediction confidence analysis showed that the DF models predicted AFP binding activity very accurately for some chemicals but not so well for other chemicals. The higher the prediction confidence, more likely the prediction is accurate as demonstrated in Figure 4
. Therefore, we suggest that the AFP binding activity prediction (binder or non-binder) should be combined with the prediction confidence to better apply the DF model in assessment of endocrine disruption potential of environmental chemicals.
Though AFP was identified long time ago and has been extensively studied, no three-dimensional structure (3D) of AFP or complexes of AFP bound to ligands has been determined by X-ray crystallization. The structural features of this protein, especially in its ligand binding domain, were understood based only on the experimental binding activity data. Therefore, a homology model of rat AFP was constructed and the ligand binding interactions of this protein were elucidated using molecular docking and molecular dynamics simulations in our previous study [31
]. The computationally constructed 3D structure of rat AFP and the in silico
elucidated ligand binding interactions are expected to help the estimated AFP binding activity of environmental chemicals. Our previous study identified two different binding pockets in rat AFP, consistent with the two putative estrogen binding sites in AFP [81
]. The ligand binding interactions of rat AFP contribute from residues Glu206, Glu209, Gly210, Leu213, Lys236, His260, Try306 and His310 in the first binding site and from residues Leu233, Gln239 and Glu312 in the second binding site [31
]. Most of these amino acids have charged or have polar residues. Thus, hydrophilic and electrostatic interactions are important for a chemical to bind to AFP. Furthermore, the binding pockets were found to be different in size and shape. In this study, 16 Mold2
descriptors (Table 2
) were identified as the informative descriptors to the DF prediction models. Therefore, these molecular descriptors represent the important structural features that are determinant to AFP binding activity of chemicals. The 16 Mold2
descriptors are the structural features of the chemicals interacting with AFP related to molecular shape, electronegativity, and polarizability of chemicals indicating molecular shape, hydrophilic and electrostatic interaction capability. These molecular characteristics are used to differentiate AFP binders from non-binders. The informative descriptors identified in this study confirmed the reliability of our previously constructed 3D structure of rat AFP and the elucidated ligand binding interactions.
Recently EPA considered utilization of high throughput screening assays and computational models in the endocrine disruptor screening program [82
]. EPA led CERAPP project to develop QSAR models for prediction of estrogenic activity and the models were used for prioritize environmental chemicals for Tier-2 testing [83
]. With binding data of transporter proteins obtained from experiments or in silico
predictions, it is speculated that better priority setting the environmental chemicals for testing would be yielded.
The DF prediction models showed lower prediction accuracy than the DF model we previously developed for prediction of ER binding activity [20
]. The less predictive power of the AFP binding activity prediction models may be partially due to the relatively small sample size. We expected more accurate DF prediction models would be constructed when AFP binding activity is experimentally measured for more chemicals that can be used as training samples. Another speculation on the cause of the relatively low prediction accuracy is the multiple binding sites in AFP. The 125 chemicals bind AFP in different interaction regions. The first ligand binding site in rat AFP lies in the region of amino acids 419–433 and the second ligand binding site consists of amino acids 450–464. The chemicals that displayed rat AFP binding activity in our previous study are structurally diverse [30
]. The existence of two distinct ligand binding sites in AFP indicates that prediction of binding activity of a chemical depends on the AFP site where the chemical binds [84
]. Therefore, we assume separate prediction models should be developed, each for one of the two ligand binding sites, to improve the performance of AFP binding activity prediction model. Our previous study demonstrated competitive modeling based on molecular docking may perform better than the DF modeling for AFP binding prediction. Lack of knowledge on the binding sites for chemicals and the limited number of experimental binding data available is a major impediment in the development of such separate prediction models. Our results indicated that simple predictive models such as the DF models in this study sometimes yield inaccurate predictions, especially when the system in modeling is not simple. Even though a moderate prediction power has been shown for the AFP binding activity prediction DF model, caution is warranted in application of the DF model in assessment of endocrine disruption potential of environment chemicals, especially when a prediction has a low prediction confidence. Nonetheless, the rat AFP binding activity predictions of high confidence from the DF models should be useful for assistance in estimation of rat AFP binding activity of environmental chemicals.