Cross-Species Molecular Docking Method to Support Predictions of Species Susceptibility to Chemical Effects

1. Introduction

An essential aspect of environmental risk assessment for chemicals involves the estimation of toxicity across the diversity of species within an environment. Standard toxicity tests in environmental risk assessment that rely on a limited set of representative species might not be adequately protective for whole ecosystems (1). Given the amount of new and existing anthropogenic chemicals, the application of historical, whole-animal toxicity testing strategies to every possible species-chemical combination is impractical (2). Biological read-across aims to address these data gaps by extrapolating toxicity knowledge across species based on their relatedness (3–5). However, methods capable of generating lines of evidence toward chemical susceptibility with greater taxonomic resolution hold promise for potential benefits for ecosystem or species-specific evaluations (6, 7). For instance, when evaluating chemical risk with respect to threatened or endangered species or specific ecosystems of concern. The ability to make predictions of species susceptibility at high-taxonomic resolution is enabled by technological advancements that have facilitated increasingly comprehensive cross-species comparisons at the molecular and biochemical levels (8–14). By harnessing the wealth of existing biological and chemical data, computational New Approach Methodologies (NAMs), which are by definition non-animal based, are contributing to the establishment of the Next Generation Risk Assessment (NGRA) paradigm (15, 16).

Numerous computational tools have already been developed to generate and support predictions of cross-species susceptibility to chemical toxicity (8, 11, 17, 18). However, combining application of these tools can synergistically strengthen a weight-of-evidence approach (19). In this regard, this paper introduces a computational approach that utilizes advanced protein structure prediction techniques in combination with molecular docking simulations for generating an additional line of evidence to support predictions of species susceptibility to chemical effects (Fig. 1).

Fig. 1.

This cross-species molecular docking method makes use of several existing tools for molecular modeling and docking to support predictions of species susceptibility. (A) The protein structures for query species were first generated using the Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS v7.0) tool through the Iterative-Threading ASSEmbly Refinement (I-TASSER) platform. (B) After receptor and ligand preparation, molecular docking simulations were performed using AutoDock Vina. (C) Using the docking simulation results, the protein-ligand interaction fingerprint (PLIF) similarity, ligand root-mean squared deviation (RMSD), and binding pocket shape similarity were calculated for each query structure based on comparisons with an experimentally derived reference structure. The docking score was calculated by default from the docking simulation. (D) These binding metrics were interpreted using a k-nearest neighbors (kNN) algorithm to classify the binding modes for each of the query species to make predictions on species susceptibility.

Chemical toxicity is often characterized using the Adverse Outcome Pathway (AOP) framework, which contextualizes mechanistic and/or predictive relationships between chemical–biological interactions leading to adverse outcomes relevant to risk assessment (20). Within the AOP framework, the action of a chemical on a biological system begins with a molecular initiating event (MIE), which is often described by direct chemical binding to a protein target. Consequently, information on the conservation of a protein target across species can generate and support predictions of species susceptibility to chemical toxicity (12, 21–23). This is the theoretical basis of the United States Environmental Protection Agency’s (US EPA) Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS) tool (8), which uses protein sequence and structure information to estimate protein conservation across species.

With its latest expansion (version 7.0), the SeqAPASS tool can now generate protein structures across species using protein sequence information to further support predictions of chemical susceptibility based on structural conservation (24). This advancement was enabled through the incorporation of the Iterative-Threading ASSEmbly Refinement (I-TASSER) platform (25) and template modeling (TM)-align (26) algorithm. Protein structure prediction methods like I-TASSER and Alphafold (27) have allowed for the generation of quality protein structures for virtually any protein across any species where relevant sequence information is available. The application of these predicted structures within SeqAPASS for predicting cross-species susceptibility prompted a deeper investigation of these protein structures within a functional context using molecular docking.

Molecular docking is a conventional technique used for the virtual screening of chemicals in drug discovery (28, 29). This is achieved by either sifting an extensive chemical space via scoring simulated binding modes (30, 31) or through “reverse/inverse” docking methods to identify likely molecular targets of a chemical (32). In this study, similar principles and underlying assumptions of a traditional molecular docking virtual-screening approach were applied. However, instead of screening thousands of chemicals across a single target, we screened a single chemical against various protein orthologs (i.e., the same protein target from different species).

A well-known limitation of traditional molecular docking methods is that docking scores generally do not correlate well with measured binding affinities (33). To overcome this, we evaluated resulting binding modes using a combination of four metrics: docking score (calculated by the docking program in kcal/mol), ligand root-mean squared deviation (RMSD), binding pocket shape similarity (calculated as PPS-scores), and Protein-Ligand Interaction Fingerprint (PLIF) similarity (calculated as Tanimoto coefficients). Apart from the docking score, each of these metrics was calculated through direct comparisons with pre-existing, experimentally derived reference structures. The interpretation of these binding metrics was facilitated using the k-nearest neighbors (kNN) algorithm, a non-parametric supervised learning classifier, to assign susceptibility calls to each species.

This molecular-docking based method was demonstrated using chemical interaction with the androgen receptor (AR) as a case study. The AR is a thoroughly studied nuclear receptor with a well-resolved ligand-binding domain (LBD) that is known to be critical for mediating the toxic effects for various environmentally relevant, endocrine disrupting chemicals (34). In fact, in support of the US EPA’s Endocrine Disruptor Screening Program (https://www.regulations.gov/document/EPA-HQ-OPP-2021–0756-0002), a worldwide consortium was created called the Collaborative Modeling Project for Androgen Receptor Activity or CoMPARA, underscoring the importance of this molecular target in environmental and human toxicology (35). Moreover, the focus on AR facilitated a comparative analysis between the results of this study and previous work that focused on cross-species extrapolation using AR (24, 34).

Two chemicals that are known to target AR were used: 5α-dihydrotestosterone (DHT), an endogenous ligand of AR (36) and 5-(2-fluoro-4-hydroxyphenyl)-1-methyl-1H-pyrrole-2-carbonitrile (which we refer to as FHPMPC, here), a synthetic selective androgen receptor modulator (SARM) (37). From an ecotoxicology perspective, neither DHT, an endogenous, naturally occuring androgen hormone, nor FHPMPC, an experimental research chemical designed for the treatment of osteoporosis, are currently known chemicals of concern (37, 38). Nevertheless, both chemicals were useful in demonstrating the effectiveness of this cross-species molecular docking method in predicting species susceptibility to chemical effects. Each has an experimentally derived AR-bound crystal structure available in the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB), and their relative structural distinctiveness makes them valuable for comparison purposes. Certainly, this method could be applied to chemicals with better characterized toxicity, provided relevant crystal structures are available.

2. Methods and Results

2.1. Generation of initial species susceptibility calls and protein structures using SeqAPASS.

The SeqAPASS (v7.0; https://seqapass.epa.gov/seqapass/) tool was queried using human AR (NCBI Protein Accession No. AAI32976.1) and evaluated at Levels 1 – 3, previously (24). Level 2 evaluations were completed using the AR LBD (cd_07073), and individual amino acid residues determined as critical for AR function were evaluated within Level 3. As part of the Level 4 evaluation within SeqAPASS, 268 AR structural models (each representing a unique species) were generated from primary amino acid sequences that passed the initial prioritization step (see SI Appendix, seqapass_summary.xlsx for a summary of SeqAPASS results at each level of evaluation). Most of these 268 species were birds (Aves; 73.9%) followed by bony fish (Actinopteri; 14.6%) and mammals (Mammalia; 6.72%). Each model was generated using the I-TASSER algorithm (25).

Of the 268 structural models generated, 184 were identified as low-quality structures based on SeqAPASS Level 4 criteria (24) and were not evaluated for susceptibility by SeqAPASS. Despite this, all 268 models were used in the molecular docking simulation of both DHT and FHPMPC since docking was performed with respect to the LBD and less ordered regions of the proteins could be trimmed. When accounting only for the LBD, the protein RMSD values calculated with respect to the PDB ID:2AMA reference structure ranged from 0.268 – 2.346 Å with one outlier value of 8.154 Å. This outlier was derived from the androgen receptor beta sequence (NCBI Protein Accession No. NP_001117657.1) of rainbow trout (Oncorhynchus mykiss), which represents a distinct isoform (39). The greater relative similarity of the LBDs supported the use of all 268 structures for demonstrating this method, with the rainbow trout structure retained to support method verification (i.e., structural outliers should yield negative results).

2.2. Preparation of protein and ligand structures for docking.

Comparing binding modes across species posed challenges distinct from conventional virtual screening methods. As opposed to docking many chemicals against a single protein target, our study involved docking a single chemical to various protein orthologs. Many of these proteins exhibited unique sequences, leading to variations in the numbering of conserved amino acid residues across species. Since these structures were computationally generated, there were often extraneous disordered regions, and the centroid coordinates of the structures differed. Collectively, these differences made calculating meaningful PLIF comparisons and ligand RMSD values impracticable. To address these challenges, a custom Python script was used to perform both multiple sequence alignments and structural alignments across each query protein.

First, the PDB files for each protein were converted to FASTA format and aligned using the MUSCLE algorithm (40), which was implemented using the latest version release of the MUSCLE executable v5.1 (https://github.com/rcedgar/muscle/releases/tag/5.1.0). Using the multiple-sequence alignment results, the residue position numbers were harmonized across all 268 species structures for all aligned residues. Further, each query sequence was trimmed to retain the residues aligned within the range of the reference protein sequence encompassing the LBD, inclusive of an extension of 10 residues at each terminus. These trimmed structures were then structurally aligned with superimposition to the reference structure using the default “align” function available through Open-Source PyMOL (v4.6.0; Schrodinger, LLC). Following the alignments, water molecules were removed, polar hydrogens were added, and Kollman charges were added to each structure using a Python script available through AutoDock Tools (41).

2.3. Flexible docking protocol.

All docking simulations were performed using AutoDock Vina v1.2.5 (42). Limited flexible receptor binding was performed across all query species structures to help compensate for minor structural inaccuracies resulting from the I-TASSER predictions, which could interfere with ligand recognition and pose prediction (43, 44). Residues were selected as flexible based on a distance-threshold criterion. First, the binding pocket of the reference structure was defined using the coordinates of the bound ligand following a technique parallel to that within the Protein-Ligand Interaction Profiler tool (45). Essentially, a Euclidean distance-based cutoff is calculated based on the size of the bound ligand (i.e., the maximum distance of a ligand atom to the ligand centroid), and protein residues with at least one heavy atom within that distance are counted as part of the binding pocket (or “site”, depending on the protein). Then, Euclidean distances were computed between all atoms of the identical binding site residues within the superimposed query and reference structures. A residue within the query structure was designated as “flexible” in the docking simulation if it had an interatomic distance exceeding 3 Å (roughly twice the length of the average C-C bond) between any two identical residue atoms. This distance calculation was repeated for each query structure.

AutoDock Vina requires that the user specifies the dimensions of the simulation search space, called the “grid box”. The dimensions of the grid box, in this case, were determined based on the size of the ligand. Each dimension was set to four times the maximum distance between any ligand atom and the ligand’s centroid (i.e., $x=y=z=4*ligand“radius”$). The coordinates of the center of the grid box were the same as the ligand centroid coordinates of the reference structure. The Vina RaDii Optimized (Vinardo) scoring function was used in place of the default Vina scoring function due to apparent improvements in scoring power (46). For every docking simulation, the top 3 scoring binding modes (i.e., the orientation of the ligand relative to the receptor when bound) were saved. AutoDock Vina, by default, produces two separate PDBQT files for these binding modes—one containing the coordinates of the bound ligand and another containing the resulting coordinates of any flexible residues from the simulation. To generate PDB-formatted models (i.e., the three-dimensional structure of the total protein domain with the bound ligand) for each resulting binding mode, these PDBQT files were combined with the corresponding structural file containing the coordinates of all the rigid receptor residues.

2.4. Four binding metrics were used to characterize the docking results.

Each protein-ligand model was evaluated based on four quantitative metrics: the docking score (in units of kcal/mol), ligand RMSD, binding pocket shape similarity (calculated as PPS-scores), and PLIF similarity (calculated as Tanimoto coefficients). Docking scores were computed during docking simulations using the Vinardo scoring function. Ligand RMSD values were computed for each protein-ligand model relative to the reference structure (SI Appendix, Fig. S1). The PPS-align tool (https://seq2fun.dcmb.med.umich.edu//PPS-align/) performs non-order structural alignments of two protein binding pockets using an enhanced-greedy based, iterative heuristic search algorithm to score (i.e., PPS-score) the shape similarities between the pockets (SI Appendix, Fig. S2). Additionally, PLIFs were generated by making use of the Protein-Ligand Interaction Profiler (PLIP) tool (https://plip-tool.biotec.tu-dresden.de/plip-web/plip/index) (45) in combination with a custom function for calculating van der Waals interactions based on previous methods (47).

A PLIF is a bit vector representation of the interactions between protein residues and bound ligands in terms of various, pre-defined interaction types (48). In this case, 7 interaction types were considered: hydrogen bonds, hydrophobic contacts, pi-stacking, pi-cation interactions, salt bridges, halogen bonds, and van der Waals contacts. In a PLIF, a “1” represents the presence of an interaction and a “0” represents the absence of an interaction (SI Appendix, Fig. S3). Similarities between the PLIFs of the query docking conformations and the reference were calculated using the Tanimoto coefficient, which is a similarity measure ranging between 0 and 1, with 1 denoting identical fingerprints.

The reference structures that were used to calculate the ligand RMSD, PPS-scores, and PLIF Tanimoto coefficients were PDB:2AMA for DHT (https://doi.org/10.2210/pdb2AMA/pdb) and PDB:5VO4 for FHPMPC (https://doi.org/10.2210/pdb5VO4/pdb). The structures of the ligands and the major intermolecular interactions they display based on their reference structures (with waters removed) are shown in Figure 2.

Fig. 2.

Two-dimensional diagrams of (A) 5α-dihydrotestosterone (DHT) and (B) 5-(2-fluoro-4-hydroxyphenyl)-1-methyl-1H-pyrrole-2-carbonitrile (FHPMPC). Rotatable bonds are highlighted in pink. Three-dimensional diagrams of (C) DHT and (D) FHPMPC displaying residue interactions according to their respective crystal structures (with waters removed)—PDB:2AMA for DHT and PDB:5VO4 for FHPMPC. The interactions were determined using a protein-ligand interaction profiler (PLIP) analysis. The dashed orange lines represent hydrophobic interactions, the dashed magenta lines represent pi-stacking, and the solid light blue lines represent hydrogen bonding.

2.5. Defining susceptibility using an ensemble docking training dataset.

An experimentally derived crystal structure for a protein-ligand complex provides strong evidence that the species from which that protein was derived is likely susceptible to a chemical effect exerted by the bound ligand. Therefore, experimental structures allow for the estimation of a “biologically active” binding mode with respect to the protein-ligand pair. If a biologically active binding mode can be predicted, then predictions of chemical “susceptibility” can be made, which is at the heart of predictive toxicology (i.e., the AOP framework). Using the four binding metrics described earlier, the binding simulation results were classified into either likely “biologically active” modes or “non-biologically active” modes. If a resulting binding mode for a protein was classified as “biologically active,” then the species from which that protein structure was derived from was given a susceptibility prediction of “yes.”

We used an ensemble docking approach to generate a set of models that represented both biologically active and inactive binding modes. Ensemble docking refers to the generation of an “ensemble” of receptor conformations that are used in docking simulations (49). In this case, an ensemble of 47 human AR structures was generated by using an advanced search within the RCSB PDB (www.rcsb.org/search/advanced) that restricted the search criteria to structures with a refinement resolution ≤ 2.5Å, having no mutations, and were “strictly” (as defined by PDB) structurally similar to the AR-DHT reference structure, PDB:2AMA (see SI Appendix for more detail on the search criteria). Both test chemicals (DHT and FHPMPC) were used in flexible-receptor molecular docking simulations (as described earlier) against each of the human AR structures in the ensemble. The top five scoring binding modes were saved and ligand RMSD values were calculated with respect to the reference structures for each chemical. The binding modes were filtered to include only those that were most similar to and least similar to the reference pose based on the calculated RMSD values. The “best” result had the lowest RMSD value whereas the “worst” result had the highest RMSD value. To help reduce bias, docking results with a ligand RMSD ≥ 10 Å were considered as outliers (Figure 3; a visual explanation of the ensemble docking approach).

Fig. 3.

(A) An ensemble of 47 human androgen receptors (ARs) was used in docking simulations for both 5α-dihydrotestosterone (DHT) and 5-(2-fluoro-4-hydroxyphenyl)-1-methyl-1H-pyrrole-2-carbonitrile (FHPMPC). Using the results from the docking of DHT to PDB ID:3RLJ as an example, the “best” resulting pose is shown in (B) with an RMSD = 0.579 Å and the “worst” resulting pose is shown in (C) with an RMSD = 6.813 Å. The distribution of the ligand RMSD values across all 47 structures for the DHT binding simulations is shown as a histogram in (D). The range of the best poses is shown by the two red dotted lines, and the range of the worst poses is shown by the two blue dotted lines. Any binding result with a RMSD value ≥ 10 Å was considered as an outlier and is marked with an asterisk (*). The rationale for using this ensemble approach to define susceptibility is that the best poses likely represent biologically active binding modes, and the worst poses likely represent non-biologically active binding modes. If the binding mode is considered biologically active, this translates to a prediction of chemical susceptibility.

The “best” resulting binding modes from the ensemble docking simulations were used to define a biologically active mode whereas the “worst” resulting binding modes were considered as non-biologically active modes. In this way, these best and worst binding modes were used as positive and negative controls, respectively, to develop the susceptibility classifier model used to generate species susceptibility predictions for all 268 query species.

2.6. Species susceptibility predictions were made using a k-nearest neighbors classifier.

The defined control sets from the ensemble-docking results were used to make susceptibility calls based on cluster analysis. Specifically, the four binding metrics described earlier were first standardized across the ensemble dataset such that the mean values for each metric equaled 0 and the standard deviation equaled 1. Using this standardized dataset an unsupervised k-means clustering analysis was performed (SI Appendix Fig. S4) and visualized in two dimensions using a principal component analysis (Fig. 4 A and B). The cluster containing the best self-docking pose (which is the result of docking DHT to the apo-2AMA structure or FHPMPC to the apo-5VO4 structure) was determined as the “susceptible” control group whereas all other clusters represented the “not likely susceptible” controls (Fig. 4 A and B).

Fig. 4.

A principal component analysis (PCA) was performed to help visualize the ensemble docking results of the k-means clustering analysis for (A) DHT and (B) FHPMPC. The points circled in black represent the self-docking result with the lowest ligand RMSD value for each respective chemical. Correlation matrices for the four binding metrics are shown for (C) DHT and (D) FHPMPC. The results indicate that the DHT ensemble docking analysis generally led to greater correlated values across each of the metrics when compared to those of FHPMPC, which likely contributed to the less distinct clusters for the FHPMPC results.

Furthermore, the four binding metrics had correlation values with a range of 0.68 to 0.82 for the DHT ensemble docking results whereas the range was 0.31 to 0.74 for the FHPMPC results (Fig. 4 C and D). The lowest correlation values for FHPMPC related to docking score, which suggested that the inclusion of the additional metrics, in this case, likely improved discrimination power. The lower docking score correlation was not surprising given that scoring power for molecular docking programs is typically considered poor (50). Regardless, the correlations between each of the binding metrics demonstrates harmonization in characterizing the binding modes.

A k-nearest neighbor (kNN) algorithm was used to develop a classifier model based on the k-means clusters of the ensemble docking results on a per-chemical basis. Thus, the features of the classifier models consisted of the four binding metrics derived from each docking simulation, and the targets of the classifier models were the clusters defined using the k-means clustering analysis. The optimal number of neighbors (k) for each model was determined using five-fold cross validation. The model accuracy was calculated across n values where n equaled half of the total number of samples in the ensemble dataset (n = 47). The value of k with the greatest accuracy was selected (SI Appendix, Fig. S4). These clustering analyses were performed using the Scikit-learn Python module (51). The cross-validated kNN model was then used to classify the standardized docking results (i.e., where the mean = 0 and the standard deviation = 1 for each binding metric) from the 268 query species. If a species model was classified into the same cluster as the “susceptible” controls, then that species was predicted as being likely susceptible to the effects of the bound chemical. This resulted in 188 (70.1%) and 207 (77.2%) of the 268 query species as susceptible for DHT and FHPMPC, respectively.

2.7. Docking susceptibility predictions were conservative compared to other methods.

The species susceptibility predictions made by this molecular docking approach were compared to in vitro AR binding or activation assay results. This in vitro data was compiled previously through a systematic literature review process (34). No data for FHPMPC was available. However, there was data for 7 species that were also evaluated via docking for DHT. One of those species was rainbow trout (O. mykiss), which was tested using a different isoform than what was used in the docking evaluation preventing meaningful comparisons for this species. Of the 6 species remaining, there were 3 species that shared a susceptibility prediction with the docking results (i.e., there was a 50% overlap; SI Appendix, Table S1).

Comparisons were also made with respect to previous SeqAPASS predictions. Of the species found to be susceptible according to the docking analysis, 87.5% of those were also assessed as susceptible across all four levels of SeqAPASS evaluation, coincidentally, for both DHT and FHPMPC (Fig. 5). This comparison was made with respect to 48 of the 84 species that met the Level 4 quality criteria and had a “yes” susceptibility call across all four levels of evaluation. When accounting for all 84 of the species, the agreement between the docking susceptibility calls and the SeqAPASS Level 4 susceptibility calls was 82.14% and 89.29% for DHT and FHPMPC, respectively (SI Appendix, seqapass_summary.xlsx). Overall, the docking method resulted in a lower number of predicted susceptible species than from these other two NAMs indicating that it gave more conservative predictions.

Fig. 5.

The results of the k-nearest neighbors (kNN) analysis are shown using a principal component analysis (PCA) for (A) DHT and (B) FHPMPC for each binding mode generated through the AutoDock Vina molecular docking simulation across the 268 species structures for AR. The circled points represent the training set (i.e., ensemble) data used for the kNN model construction. The docking results resembling a biologically active mode are represented by the “0” (or red) cluster for DHT and the “1” (or blue) cluster for FHPMPC. Binding modes that resemble a biologically active mode are considered to represent “susceptibility” to chemical effects for the respective species from which the docking result was obtained. The pie charts summarize the percentage of the 268 species that were predicted as susceptible according to the docking simulation for (C) DHT and (D) FHPMPC. The agreement in the docking predictions with SeqAPASS predictions (from Level 1 – 4 evaluations) are also shown. Coincidentally, for both DHT and FHPMPC, 87.5% of the species predicted as susceptible according to the docking method were also predicted as susceptible by SeqAPASS.

3. Discussion

We have introduced an innovative cross-species molecular docking method that generated an additional line of evidence as part of a weight of evidence approach to support predictions of species susceptibility to two chemicals, DHT and FHPMPC, across 268 species. The goal of this molecular docking approach was to generate and subset the protein-ligand binding models that are most likely to represent biologically active binding modes to support species susceptibility predictions. The major underlying assumption here was that the binding mode for a species with known susceptibility to a chemical will be similar across other susceptible species with respect to the four binding metrics evaluated in this approach. This method leverages both computationally and experimentally derived protein structural data to help move in silico predictions into a biologically functional framework. This method adds to the repertoire of computational NAMs available to support NGRA and opens many opportunities to improve the use of molecular modeling in environmental toxicology, at large. Moreover, we have made the code required to implement this method publicly available on GitHub (https://github.com/pschumann4/cross_species_docking.git).

The androgen receptor was chosen as a case study to demonstrate this cross-species docking method due to its importance in environmental toxicology (35), having been well characterized both structurally and functionally (52), and for the ability to compare the species susceptibility predictions to those from previous work (24). We chose DHT and FHPMPC as the chemicals of interest because they are known to target AR and have experimentally derived crystal structures available to be used as references.

The species susceptibility predictions were compared to those from both empirical, in vitro assay data and from a separate computational prediction tool – SeqAPASS. The comparisons that could be made with existing in vitro data was limited and should be taken with care. Especially considering that the data was derived from distinct assay systems with variable metrics from various research groups. Based on the in vitro and SeqAPASS comparisons, the docking method made more conservative estimates of susceptibility. In future studies, it would be useful to make additional comparisons to in vivo data to further support validation of the use of this method in environmental risk assessment.

While this method is applicable to a wide array of protein-chemical combinations where experimental crystal structures exist for the chemical of interest, there are several limitations. First, this method is dependent on AutoDock Vina to make accurate predictions of binding poses. As with any docking program, there is a tendency towards the generation of both false-positives and false-negatives (33). We attempted to overcome some of the limitations of traditional docking metrics, such as the docking score and ligand RMSD values, by incorporating additional binding metrics. The method also relies on computationally predicted protein structures where minor inaccuracies in modeling binding sites can undermine pose prediction (53). Limited flexible docking was employed to help overcome this but still led to some unexpected susceptibility predictions (e.g., Danio rerio). Encouragingly, a recent study has shown that in a virtual screening approach using docking, predicted structures gave similar hit rates as crystal structures (54). However, the reliance on structural prediction tools may limit what chemical-protein systems can be evaluated by this method. For instance, multi-domain models would require additional predictions of domain placements, which currently needs improvement (55). Additionally, ligands engaged in covalent protein interactions, like organophosphates and acetylcholine receptors (56), are challenging to screen using molecular docking as they would require a modified docking approach (57, 58). Moreover, our method relies on existing structures for use as references, and developing reliable in silico models for this purpose is a work-in-progress.

Protein structure prediction methods like I-TASSER and AlphaFold are fundamentally limited by available protein sequence information. In this study, of the 268 species selected based on protein structure quality criteria established as part of SeqAPASS Level 4, about 74% of these species were birds. The value of large-scale species sequencing efforts is demonstrated by this taxonomic disproportion, which is a result of projects such as Bird10k (https://b10k.genomics.cn/) and the Avian Phylogenetics Project (http://avian.genomics.cn/en/index.html). Of the species determined as susceptible based on the docking results, 75.5% and 71.5% of those species were also birds for DHT and FHPMPC, respectively (SI Appendix, docking_susceptibility.xlsx). Including greater taxonomic diversity in the docking procedure would ultimately improve the ability to assess for taxonomic bias.

The binding of a chemical to a protein is an important descriptor of the molecular initiating event of an AOP and is therefore essential for understanding the molecular basis of toxicity. However, predictive toxicology is a multidimensional challenge, as toxicity is influenced by many factors, such as dose, exposure, and toxicokinetics. For instance, while species can share susceptibility to a chemical effect, their sensitivity to that effect can vary (59, 7). These differences in species sensitivity can be explained, in part, by differences in binding affinity to the target protein of the chemical (7). Therefore, the evaluation of chemical-protein interaction offers a valuable line of evidence within a weight-of-evidence approach for understanding toxicity across species.

As a potential next step, the reliability of the docked chemical-protein models assigned as “susceptible” through this method can be more thoroughly scrutinized using molecular dynamics simulations (MDS). With careful design and interpretation of MDS experiments using the “susceptible” models generated through this docking approach as inputs, it may be possible to develop a computational means of predicting species sensitivity based on more rigorous predictions of binding affinity. In turn, these sensitivity predictions would support allow for the ranking of species for supporting chemical prioritization efforts in environmental risk assessment (7, 59).

Abbreviations

NAMs

New Approach Methodologies

NGRA

Next Generation Risk Assessment

AOP

Adverse Outcome Pathway

SeqAPASS

Sequence Alignment to Predict Across Species Susceptibility

I-TASSER

Iterative-Threading ASSEmbly Refinement

RMSD

root-mean squared deviation

PLIF

protein-ligand interaction fingerprint

PLIP

Protein-Ligand Interaction Profiler

kNN

k-nearest neighbors

AR

androgen receptor

LBD

ligand binding domain

DHT

5α-dihydrotestosterone

FHPMPC

5-(2-fluoro-4-hydroxyphenyl)-1-methyl-1H-pyrrole-2-carbonitrile