SERVICES . USA \\MENT OF HEALTH & HOME
Published in final edited form as: Environ Sci Technol. 2025 July 15; 59(27): 13844-13856. doi:10.1021/acs.est.5c04054.
Machine Learning and Large Language Models for Modeling Complex Toxicity Pathways and Predicting Steroidogenesis
Thomas R. Lane1,*, Patricia A. Vignaux1, Joshua S. Harris1, Scott H. Snyder1, Fabio Urbina1, Sean Ekins 1,*
1Collaborations Pharmaceuticals, Inc., 840 Main Campus Drive, Lab 3510, Raleigh, North Carolina, 27606, United States of America.
Abstract
High-throughput screening and computational models have been effective in predicting chemical interactions with estrogen and androgen receptors, but similar approaches for steroidogenesis remain limited. To address this gap, we developed general steroidogenesis modulation models using data from ~1,800 chemicals screened in H295R human adrenocortical carcinoma cells. A random forest model was validated using a prospective test set of 20 compounds (14 predicted active, 6 inactive), achieving 80% accuracy with conformal prediction adjustments. In parallel, we built classification and regression models based on IC50 data from ChEMBL for key steroidogenic enzymes, including CYP17A1, CYP21A2, CYP11B1, CYP11B2, 17ß-HSD (1/2/3/5), 5a-reductase (1/2), and CYP19A1 (126-9,327 compounds per target). These models enable predictions of both general steroidogenesis inhibition and potential molecular targets. Additionally, we developed a transformer-based model (MolBART) to predict all endpoints simultaneously and validated this performance. Combined, these models may offer a rapid and scalable system for assessing chemical impacts on steroidogenesis, supporting chemical risk assessment, product stewardship, and regulatory decision-making.
Graphical Abstract
Competing interests:
Virtual Screen
Target-Specific Steroidogenesis Prediction Visualizations
Classification (Consensus)
Probability-Like Score
Built ML Target-Specific Steroidogenesis Models
0.0
1.0
CYPILA
ML Steroidogenesis Generalized Model
;- HSD(V2)
CYPRES (0)
CYP2142 (0)
CYPI82 (0)
CYP1102 0
CYPITAS (0)
CYPTTAI (0)
Forskolin (Control) Prochioraz (Control) Sulfamethazine (Na+)
3-HSD(1/2)
CYP21A2 (0)
CYPRI82 (0)
Piribedit Paronol
4-Methylumbelliferone. Dobutamine (HCI).
CYPTTA: (0)
CYPETAI (0)
Edrephonium chloride
Ritodrine (HCI)
SS-HSO(1/2)
CYP29 (0)
N-Acetyl-
Letrozole
17|-HSD(1/2/3/5) (0)
175-HSD(1/2/3/5) (0)
175-HSD(1/2/3/5) (0)
Pravastatin (Na+) (*)-Naproxen
Aldosterone
A drostenedione
Corticosterone
Cortisol
Cortison
11-Deoxycortiso
DHT
DOC
Estrone
Estradiol
170H-Progesterone
Progesterone
Testosterone
SB-HSO(1/2]
CYP19 (0)
So-reductasel (1) Sa-reductase2 [0]
Prospective Test Set: 80% Accuracy
Keywords
Steroidogenesis; endocrine disruption; machine learning; large language models; MolBART; conformal predictors
Introduction
Over the past decades, the adverse effects of endocrine disruption have been the subject of increasing scientific interest 1-4. Endocrine disruption refers to the chemical-induced alteration of normal hormone levels in the body which can have a profound impact on physiological processes 5. Such disruptions have been linked to a variety of deleterious health outcomes, including but not limited to reproductive dysfunctions, developmental abnormalities, impaired growth, heightened susceptibility to cancer, and disturbances in the immune and nervous systems 6-8. The mechanisms underlying these effects often involve the interference of endocrine-disrupting chemicals (EDC) with hormonal signaling pathways, which can lead to alterations in gene expression and cellular function that persist throughout life 5.
The growing body of evidence regarding the detrimental health effects of endocrine disruption has led to heightened regulatory and public health concerns 6-8. However, the identification and testing of EDC present significant challenges, the gold standard of assessing endocrine disruption in animals presents ethical questions as well issues around species differences and extrapolation. One of the primary obstacles is the substantial backlog of commercial chemicals that require evaluation for their potential endocrine-disrupting properties. In response to these challenges, the U.S. Environmental Protection Agency (EPA) and other regulatory agencies have emphasized the development and implementation of alternative screening methods to predict endocrine disruption more efficiently and ethically 9-16. These alternative methods range from in vitro assays to computational approaches, are designed to provide accurate, rapid, and cost-effective means of assessing the endocrine-disrupting potential of chemicals 15-19. Computational approaches also have emerged as a promising tool in this domain, offering the ability to predict endocrine
disruption from a molecule structure and are based on the growing body of publicly available data as a training set 19-31. By leveraging advanced algorithms and machine learning techniques, these ligand-based models can help to identify potential endocrine disrupting compounds early in the chemical development process, reducing the need for extensive animal testing while improving the speed and efficiency of screening efforts 32.
Our previous research has focused on the development of such predictive models for the inhibition of key targets within the endocrine system, such as the estrogen (ER), androgen (AR) and aromatase receptors 1-4. These 3 targets are critical in the regulation of sex hormone signaling, and their disruption can have significant repercussions on human health. We have now greatly extended this work to include machine learning models of upstream targets (CYP17, CYP19, 176-HSD(1/2/3/4), CYP21, CYP11B1, CYP11B2, 5a-reductase(1/2)) that are involved in steroidogenesis, the biosynthetic pathway responsible for production of steroid hormones 33-35. This process is highly conserved, in which intracellular cholesterol is transported and then converted into various steroids in the mitochondria 36. Steroidogenesis involves a complex network of enzymes that catalyze the conversion of cholesterol into biologically active steroid hormones, including glucocorticoids, mineralocorticoids, and sex steroids (Figure 1). Mutations in the genes encoding these enzymes can lead to defective steroidogenesis which has also been linked to several rare diseases 37.
Recently a high-throughput H295R steroidogenesis assay (HT-H295R) was developed with the results subsequently released for 2,060 chemical samples from a library compiled from multiple ToxCast chemical lists 38. We have now built machine learning models for general steroidogenesis inhibition using this high-throughput screening (HTS) data (representing 1,845 unique compounds) and virtually screened a drug repurposing compound library to select molecules to prospectively validate the model. We also scored these molecules from this prospective test set in our steroidogenesis enzyme activity models to predict which enzymes are likely involved in this endocrine disruption. This approach provides a more comprehensive predictive framework by not only assessing general steroidogenesis disruption, but the specific impacts of specific enzymes on steroidogenesis homeostasis. As there are likely thousands of industrial chemicals 39, 40 that likely need profiling for effects on steroidogenesis, their experimental screening may be too expensive and hence methods to prioritize testing would be a valuable alternative 15, 19.
Methods
Target selection
The main proteins responsible for the transport of cholesterol, followed by its conversion to various steroids have been extensively studied and have been covered by many reviews 33-35, 41. Our steroidogenesis target selection focuses on the steroidogenesis pathway from the human adenocarcinoma cell line (H295R), which has been shown to independently produce progestogens, corticosteroids, androgens, and estrogens following cholesterol import 42, 43.
Dataset Curation
Datasets for (CYP17, CYP19, 176-HSD(1/2/3/4), CYP21, CYP11B1, CYP11B2, 5a- reductase(1/2)) (as well as other targets that are downstream of steroidogenesis) were downloaded from ChEMBL33 44 for individual targets and EPA’s CompTox Chemicals Dashboard 45 for the H295R HTS screen 38. Data preprocessing was conducted to ensure consistency across all datasets. This process included compound neutralization, salt removal, and the standardization of SMILES representations (canonicalization). Classification of the steroidogenesis inhibition dataset was based on the “hit call” from the original publication. If an alteration (absolute value ≥1.5-fold control) was detected in any of the 10 quantified steroids by an experimental compound, either up or down, it was classified as a steroidogenesis inhibitor in our model.
For binary target-based classification models, thresholds were set at 100 nM; data points with values ≤100 nM were assigned a binary value of 1, while those above this threshold were assigned a value of 0. Ambiguities in binarization, indicated by less than 60% agreement among sources, resulted in the exclusion of the affected samples. In the case of classification tasks, binary labels were assigned prior to averaging, with a minimum consensus threshold of 60% for data retention. Any ambiguous data points that did not meet this threshold were excluded from further analysis.
For regression models, the geometric mean of IC50 values was calculated following the removal of duplicate entries and outliers. All measurement values were normalized to a consistent unit (-logM) and aggregated as a single geometric mean value after filtering for duplicates and outliers.
Computational methods
Assay Central Model Building: We developed various classification and regression models using our proprietary Assay Central software, as described in previous work 46. These models were generated with extended-connectivity fingerprints (ECFP6), commonly used for drug target applications. The algorithms applied include Deep Learning (DL), AdaBoost decision trees (ada), Bernoulli naïve bayes (bnb), Bayesian Ridge regression (br), elastic net regression (enr), k-nearest neighbors (knn/knnr), support vector machine (svc/ svr), logistic regression (lr), xgboost (xgb/xgbr) and random forest (rf/rfr). Assay Central outputs both a probability-like score or activity prediction for each chemical evaluated.
The final model selection was determined by optimizing the area under the curve (AUC) for classification models and the coefficient of determination (R2) for regression models. Prediction scores for each compound were generated, with a threshold of 0.5 applied to classify compounds as likely active.
Model Validation: Machine learning models were assessed using a nested, 5-fold cross validation methodology with five 20/80% testing/training splits (stratified for classification, random for regression). Standard cross validation metrics are reported for all classification and regression models. Statistical parameters generated for classification models included specificity, recall, precision, F1-Score, accuracy, Receiver Operating Characteristic (ROC)
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
curve area under the curve (AUC), Cohen’s Kappa 47, 48, and the Matthews correlation 49. Regression model validation statistics were also well-defined metrics for model validation and included mean absolute error (MAE), root mean squared error (RMSE) and R2.
A prospective test set was also assessed for the general steroidogenesis modulation model. A virtual screen was performed with a drug repurposing compound library (5,080 compounds) from MedChem Express (Monmouth Junction, NJ) and scored for predicted activity using the random forest model with a conformal predictor cutoff (a=0.05). Conformal predictions were calculated using a method combining class-conditional conformal predictors 50 for a prediction coverage level a such that a compound in a test set has a 1-a probability of being included in the prediction set if the label is true: P(Y test E C(X test)| Y test = y) ≤1 -a. For cross-conformal predictors 51 the dataset is divided into N folds, and a model trained with N-1 folds is calibrated on the remaining fold. This is repeated N times, using a different fold as a calibration set each time, and the union of all predictions for the N sets is used to calibrate the conformal predictor.
Overall, 150 compounds (some salt variations) were predicted to be within the domain of the rf model (predicted positive or negative). To reduce the likelihood of replicating known data or retest the active compound we also attempted to remove known androgen or estrogen receptor agonists/antagonists, duplicate active compounds, those that contained the five-ring steroid substructure and those flagged by the vendor to target cancer and likely to be cytotoxic, which left a total of 60 unique compounds.
Large Language Model Architecture and Training
Newer machine learning methods such as transfer learning and multi-task output generally use much larger datasets for related targets, and example are the large language models (LLMs) like those used in commercial products 52. Such architectures offer an alternative to machine learning methods for cheminformatics analyses. To investigate the impact of tuning-dataset size on the performance of a pretrained large LLM transformer, we fine-tuned the base Chemformer model provided by Irwin et al. 53 on datasets of progressively increasing size. This LLM is a molecular Bidirectional Auto-regressive Transformer (MolBART) model, which employs a BART architecture with both encoder and decoder layers. This structure enables the model to learn contextual molecular encodings through the encoder, while the autoregressive decoder module learns molecular structures. Following the pretraining phase, the bidirectional encoder within MolBART can be fine-tuned for downstream applications, such as predicting molecular properties (e.g., IC50 values against specific targets).
The fine-tuning procedure was conducted in PyTorch using the Lightning 54 framework, with each session running for 150 epochs. The combined dataset for all steroidogenesis targets was first split into a 70/15/15% training/validation/test set by target. All endpoint datasets were utilized for fine-tuning, enabling comprehensive adaptation of the pretrained model to various molecular prediction tasks. For a direct comparison with the performance of individual targets models, this same splitting pattern was repeated, and individual models were built for each target with a 70% training dataset and validated on a 15% leave out set.
t-SNE
t-SNE plots are a popular data visualization approach for understanding very large datasets and compress the data into a lower-dimensional space 55. 1024 ECFP6 fingerprints were generated for all compounds. The 1024-bit fingerprints were then embedded into a 2-dimensional vector using t-SNE and values were generated using the scikit-learn library in python with default hyperparameters (n_components = 2, perplexity = 30, early exaggeration = 12.0, learning rate = 200, n_iter =1000).
H295R Steroidogenesis Assay
An initial toxicity screen was performed with 60 compounds with predicted steroidogenesis modulation (filtering process described above). In short, H295R cells (ATCC CRL-2128) were grown in DMEM/F12 medium (Gibco 21331020) completed with 2.5 mM L-glutamine (Gibco A2916801), 1% ITS premix (Corning 354350), 2.5% Nu-serum IV (Corning 355504), and 100 µg/mL pen/strep. Cells were seeded at 50-60% confluency and maintained at 37°C in 5% CO2 for 48 hours before the addition of compound (100 µM and 1% DMSO final concentration). After another 48 hours, a CellTiter-Glo 2.0 Assay (Promega G9241) was performed to determine cell viability. Compounds that displayed a cell viability of >80% at 100 µM were selected for further profiling (Supplemental File 1).
Based on these toxicity results 20 compounds were selected as non-toxic were then tested for steroidogenesis modulation. The design of this H295R Steroidogenesis assay was based on the guideline OECD 456 56 with the non-GLP screening done by Charles River (Den Bosch, Netherlands). H295R cells (ATCC CRL-2128) were grown in DMEM/F12 medium with 1% ITS + Premix, 2.5% Nu-serum and 0.1% Penicillin/Streptomycin. Cells 5-10 passages from the frozen stock (passage number p9.5) were seeded into 24-well plates with 1 mL/well at 2.0-3.0×105 cells/mL and then grown for 24 hours before drugging. After 24 hours, compounds were added in duplicate at a final concentration of 100 µM and 1% DMSO. Forskolin (10 µM) was used as a positive control inducer, prochloraz (1 µM) was used as a positive control inhibitor, and DMSO was added as a solvent control. An additional two wells were treated with 1% DMSO to use as the methanol control in the cell viability assay. After 48 hours of exposure, media was removed from the cells and stored at -80℃ for analysis. Three of the DMSO-treated wells were exposed to 70% methanol, and cell viability was measured using an MTT assay, as per the OECD guidelines, and then calculated using the equation % Viability = (response in well - average response in MeOH wells)/(average response in SC wells - average response in MeOH wells) * 100%. No toxicity was detected by MTT assay at the concentrations tested. Collected cell media was then shipped on dry ice for hormonal analysis.
Hormonal Analysis for the H295R Steroidogenesis Assay
Media from H295R drug exposure assay was quantified to the OpANS sterol panel (OpANS, Durham, NC). Samples were extracted using liquid-liquid extraction using Methyl tert-butyl ether (MTBE). The analysis was performed on an Agilent 6495D triple quadrupole LC-MS/MS instrument using reversed phase chromatography in positive ion (aldosterone, androstenedione, corticosterone, cortisol, cortisone, dehydroepiandrosterone (DHEA), deoxycorticosterone (DOC), 11-deoxycortixol, dihydrotestosterone (DHT), 17a-
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
OH-progesterone, progesterone, testosterone) or negative ion (estrone and estradiol) MRM modes. Curve fitting was performed using linear (testosterone), quadratic (DHEA, estrone, and corticosterone), and power (all others) fits with 1/x2 weighting. All AQL results originated from the use of a power fit (Response Ratio = m*concentrationy, where y is near 1). Power fits are near linear fits and for the AQL samples and chromatograms were reviewed to ensure a lack of clear detector nonlinearity which is the biggest reason actual AQL results would diverge from extrapolated concentrations. For the following steroids and sensitivities: aldosterone (10 - 10,000 pg/mL), androstenedione (20 - 20,000 pg/mL), corticosterone (200 - 200,000 pg/mL), cortisol (600 - 600,000 pg/mL), cortisone (20- 200,000 pg/mL), DHEA (600-600,000 pg/mL), DOC (6-6,000 pg/mL), 11-deoxycortixol (20-20,000 pg/mL), DHT (60-60,000 pg/mL), estradiol (10 - 10,000 pg/mL), Estrone (10- 10,000 pg/mL), 17a-OH-progesterone (60 - 60,000 pg/mL), progesterone (20-20,000 pg/mL), testosterone (10 - 10,000 pg/mL). Each compound was tested in duplicate for all 14 steroids. Androstenedione, 11-deoxycortisol and deoxycorticosterone were all above the quantification level (AQL) for the DMSO controls so comparisons are based on extrapolated values. Dehydroepiandrosterone was not able to be quantified due to chromatographic interference (analyte mass channel) in all samples, suggesting matrix incompatibility. If a sample exhibited an interference in the internal standard mass channel the fold-change was not calculated and is annotated.
Results and Discussion
Firstly, we built a generalized steroidogenesis modulation model using training data from a high-throughput H295R steroidogenesis assay, in which a first pass, single-point evaluation of 2,060 chemical samples from multiple ToxCast chemical lists was undertaken that tracked the change in 13 different steroids 38. This in vitro system utilizes the H295R cell line, a commonly used surrogate for characterizing human steroidogenesis modulation. The concentration-based effects on steroidogenesis were also determined in a follow-up experiment for 524 samples, but these were not used in our models as we intended our external test validation to be done as a single point endpoint assay. For our modeling purposes, any compound that had a ‘hit call’ for at least one of the tracked steroids was defined as ‘active’. This created a balanced dataset, with 1,845 unique compounds after curation, and these models had reasonable cross validation statistics for all algorithms (Figure 2). The best algorithms, rf and svc, shows limited overall separation between classes based on the probability-like score histogram, but rf had good separation of the positive class when >0.70 (as annotated in Figure 2B) suggesting a stricter probability-like score threshold beyond 0.5 may increase the likelihood of a correct positive prediction. Based on this we used a conformal predictor with rf prediction scores to select a prospective test set.
We also compiled IC50 data from ChEMBL for nearly all the targets for steroidogenesis as exemplified in the H295R system. By including these targets into our predictive models, we offer novel insights into the early detection of endocrine disruption, facilitating more effective in silico screening and the identification of safer chemicals for human use. For these targets we have built classification and regression models using well-established machine learning models to predict the enzymatic activity of a set of compounds tested in our prospective test set for steroidogenesis modulation.
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
Based on these data, we created training datasets for both regression and classification models for individual steroidogenesis targets and a classification model for general steroidogenesis inhibition. The datasets varied in size, with an average of~650 compounds. All target-based classification training data had a threshold of 100 nM for consistency. The activity ranges typically spanned from 100 uM to less than 1 nM, representing a large range of activity (Table 1). Unfortunately, there was insufficient data for both CYP11A and 36-HSD to enable the building of machine learning models for these targets.
For all targets we built regression and classification models using several algorithms including DL (3 layers), ada, bnb, br, enr, knn, knnr, svc, svr, lr, xgb, xgbr, rfr and rf. For classification models, validation statistics are presented for standard metrics, with values closer to one indicating better performance. For regression models, the mean MAE, RMSE and R2 are shown, with all values being calculated using units of -logM. The closer the MAE/RMSE is to zero, the better, with a perfect R2 value equal to one. Cross- validation statistics for each classification and regression model are shown in Table S1 and S2, respectively. Figure 3 summarizes the distributions of cross-validation metrics for the individual steroidogenesis targets for inhibition, for both classification and regression models. In general, these models had excellent cross validation scores, with AUC scores typically above 0.85 and most datasets were relatively well balanced. Typically, for the best algorithm the R2 values were above 0.7 and MAEs were below 0.5 (-logM).
While most of these machine learning models show excellent cross validation metrics, examples of the best and worst validation scores by targets are shown in Figure 4. The 5-fold cross validation statistics for steroid 5-a-reductase 2 (Figure 4A,B) are among the best, but are also highly representative of all the other model validation statistics. Models for 176-HSD2 (Figure 4C,D) are some of the worst performing models, but overall, these metrics still suggest these models likely have good predictivity. In addition to standard metrics, this figure also shows a histogram of the probability-like score distributions for rf, which is typically one of the best performing algorithms in our tests. The “model evaluation overlay” shows the negative class in red and the positive class in blue. This shows this rf model has clear separation in their probability-like scores for both the positive and negative class prediction for steroid 5-a-reductase 2, but more limited separation for 17ß-HSD2. For regression models (Figure 4B,D) the plotted activity, with the predicted scores on the y-axis and the actual activity on the x-axis, is also displayed. We have also built models for various other targets that are related to endocrine disruption, not including AR and ER, which were shown to have similar 5-fold cross validation statistics. These are likely to be relevant to incorporate into more complex models of endocrine disruption outside of steroidogenesis and include targets such as the thyroid hormone receptor (Table S3, Figure S2). There are also other thyroid related endpoints such as the sodium iodide symporter 57 and deiodinase 58 that could be modeled similarly; however, it is likely there is less data for these. The homeostasis and proper signaling of thyroid hormone require coordination of hormone synthesis, metabolism, elimination and iodide uptake which may add complexity to models that otherwise might just be limited to the thyroid hormone receptors 59.
To test the validity of the steroidogenesis modulation model we undertook a virtual screen of a drug repurposing compound library (5,080 compounds) from MedChem Express
(Monmouth Junction, NJ) using a rf model with a conformal predictor cutoff (a=0.05) to choose compounds to use as a prospective test set. Conformal predictors have been shown to be a very effective modeling technique to add additional model confidence and accuracy 60-65, and has been implemented with other machine learning models for endocrine related targets 66. Overall, 150 compounds (some salt variations) were predicted to be within the domain of the rf model (predicted positive or negative). Additional selection criteria, described in the methods, reduced this number to 60 compounds. These remaining compounds were then subjected to toxicity testing in the H295R cell line at 100 uM. Compounds that displayed a cell viability of>80% at 100 µM, which was the typical concentration used in the initial HTS, were selected for further profiling. In total, 20 compounds (14 predicted active, 6 predicted inactive; Table S4) were ultimately tested for steroidogenesis modulation using the H295R steroidogenesis inhibition assay (Charles River).
We classified a compound as a steroidogenesis modulator based on the criteria used in the original publication 38, where an approximate ≥ 1.5-fold absolute change from the control was used to determine a “hit call”. To confirm the reliability of the assay, we also included the steroidogenesis activator and inhibitors forskolin and prochloraz, respectively. The fold-change from the DMSO controls is shown in Figure 5. Most of our compounds predicted to be active (Figure 5A) showed the modulation of one or more steroids in the H295R cell line, typically effecting 3 or more. 11 of 14 of those predicted to have activity did, but the compounds sulfamethazine, edrophonium and tyramine did not show a significant difference based on the ≥1.5 absolute change over the control. Letrozole is a known aromatase inhibitor 67, but it was not included in the original steroidogenesis model and therefore is still a valid prospective test molecule. It was also surprising that as an aromatase (CYP17A1) inhibitor it still impacted multiple steroids upstream of aromatase in the steroidogenesis pathway, including major downregulation of androstenedione, cortisol, cortisone, 11-dexycortisol and DHT and this may represent a novel finding for this study. Letrozole is a potent aromatase inhibitor (≤5 nM in MCF-7 and T-47 cells 68), but no other direct inhibition of other steroidogenesis regulating enzymes has been noted, although some upregulation and downregulation of the expression of CYP17A1 and CYP11A1/HSD3B1, respectively, have been found in rat 69. Interestingly, previous data suggests that letrozole does not affect levels of cortisol or aldosterone 70, but at the concentration tested (100uM) we saw a significant reduction of cortisol of >8-fold (aldosterone had signal interference). This finding does suggest that high doses of letrozole may have more steroidogenesis inhibition that goes beyond the intended target of aromatase, which may lead to unintended effects on other steroids. Surprisingly, two of the compounds predicted to be modulators of steroidogenesis, neostigmine and edrophonium, are both acetylcholinesterase (AChE) inhibitors 71, 72, yet only neostigmine had a ≥1.5-fold absolute change in any steroid and only with aldosterone. Whether it is related or not is unknown, but edrophonium is a much less potent inhibitor of AChE, with an IC50 in the uM range 66 as compared to 70 nM for neostigmine 65. The effects of paeonol, a chemical component commonly found in treatments used in traditional Chinese medicine, may of be of interest because it is outside the current targets known for this compound suggesting some off-target effects 73. The compound piribedil, which is marketed in Europe for Parkinson’s disease, is known to
be a dopamine (D2/D3) and adrenergic (a2) receptor antagonist 74, 75, but the effects on steroidogenesis have not previously been characterized.
Other molecules we tested include pravastatin which inhibits 3-hydroxy-3-methyl-glutaryl- coenzyme A (HMG-CoA) reductase and cholesterol metabolism and has been demonstrated to impact steroidogenesis in rat 76 but not in humans 77, 78. The non-steroidal anti- inflammatory drug naproxen is widely used globally and has been reported to demonstrate endocrine disruption in freshwater water fleas and 79 fish as well as led to steroidogenic alteration in H295R cells. Others have also shown effects on gene expression of sex hormones in rainbow trout 80. The beta-adrenoreceptor agonist dobutamine interfered with steroidogenesis in bovine adrenocortical cells 81 and this molecule stimulates progesterone production in bovine corpus luteum in vitro 82. No publications related to steroidogenesis were identified for ritodrine, neostigmine, bentiromide, N-acetyl-L-tyrosine, paeonol, edrophonium, sulfamethazine, piribedil, 4-methylumbelliferone and tyramine. The quantification for each steroid assessed (n=2), including DMSO control values, are also provided in Figure S1 and Supplemental File 2.
Based on the prediction results from the individual models, we developed a simple predictive scheme to suggest targets involved in steroidogenesis modulation. We scored the prospective test set chosen for the generalized steroidogenesis modulation model with the individual steroidogenesis target models. We have visualized an example molecule, with the molecular structure, the average regression activity prediction from all models and the consensus binary score, which is a simple majority rule decision from the 8 classification models (Figure 6). This is overlaid on the steroidogenesis pathway as described in Figure 1. Pravastatin is predicted to be an inhibitor of CYP11B2 and 5a-reductasel by these models. Follow-up assays would need to be performed to confirm the accuracy of these predictions. For completeness, the classification predictions for the active compounds in the HTS screen 38 used for steroidogenesis modulation modeling are also provided in Supplemental File 3.
Finally, we utilized a Molecular Bidirectional Auto-Regressive Transformer (MolBART) which is a model trained on 100 million SMILES strings from the ZINC-15 database, enabling it to learn the encoding of molecular structures 53. Using this pre-trained model, the encoder is employed to encode molecules and is subsequently fine-tuned to predict properties such as IC50. The MolBART model is capable of simultaneously predicting multiple targets, whether for regression or classification tasks. The MolBART model was further explored in the context of steroidogenesis, where it was fine-tuned on IC50 data for each target and indicated a similar level of performance (Figure S3, Figure S4) to the individual classification and regression models based on 5-fold cross-validation. In this case, we split the data as follows: 70% for training, 15% for validation, and 15% for testing. We built the model on the training set, tuned it using the validation set, and evaluated its performance on the test set. We then compared this overall model performance with the performance of individual models that were trained and evaluated using the same data splitting pattern (70% training, 15% test). The statistical comparisons demonstrate comparable performance (Figure S4), although the RMSE was statistically superior with the individual models. While outside the scope of this project, determining if this is a consistent phenomenon will need to be explored with additional targets. Overall, the MolBART model
shows excellent validation statistics for all targets in the test set (Figure S3). Additional targets can also be incorporated into the MolBART model, potentially enhancing the prediction accuracy for individual targets.
An additional advantage of MolBART is its flexibility. New targets can be incorporated into the model with ease, and it can also be retrained to accommodate these additions. This modular approach enables the continuous improvement and expansion of the model’s capabilities without the need for significant retraining or reengineering. For example, in this study, the model was fine-tuned on IC50 data for steroidogenesis, demonstrating its versatility in adapting to new domains. Our earlier results from studies on other toxicology targets also suggest that MolBART holds great promise in specific areas where data may be sparse 83, 84. In cases where only a small dataset is available for training, MolBART has shown the ability to deliver substantial improvements in predictive power, likely due to its deep learning architecture and its ability to capture complex, non-linear relationships in the data 84. These results indicate that MolBART could be valuable for tasks with poorly understood molecular targets with limited data, where traditional machine learning models might struggle due to the insufficient data. This work also indicated that the MolBART model has potential for enhancing predictive accuracy for targets with limited data 83, 84.
We were also interested in understanding the chemical space covered by these steroidogenesis-related models. We therefore visualized this using a t-SNE plot with ECFP6 descriptors for all the target training data combined with various other reference datasets important to different industries (Figure 7). These were compiled from the EPA CompTox dashboard and other sources 45. Firstly, we show the distribution of the compounds by target. This shows clustering by target, suggesting similar compounds were tested against the same target (Figure 7A). Overall, we also see that the HT-H295R often does not overlap with the training data for the target specific data suggesting these represent distinct chemical spaces. Looking at the overlap between these training data and various datasets we found significant overlap. Unsurprisingly, FDA approved drugs show substantial overlap with the steroidogenesis training data and there is a similar trend with over-the-counter drugs. Dyes and flavorings have about 50% overlap with the steroidogenesis training data. Scents, personal care, cleaning, industrial and home maintenance products also have substantial overlap as well. As nearly 40% of these training data have activity against one or more of these targets this suggests many of these classes of chemicals could potentially interfere with targets involved in steroidogenesis which may be worth further analysis in future.
We have built a machine learning model for general steroidogenesis modulation and validated this using a prospective test set in a H295R steroidogenesis assay. Overall, this had an 80% accuracy and identified several compounds with unknown steroidogenesis modulation and some compounds which are suggested to have additional targets involved. Our target-based machine learning models show high predictivity based on 5-fold cross validation statistics. The MolBART transformer model also performed similarly as the individual models based on external test sets and potentially has additional advantages. These machine learning models collectively cover a large chemical space enabling predictions for chemicals important to different industries.
One limitation of these machine learning models is that they only cover inhibition of the steroidogenesis targets, so it would be beneficial to expand this to activation as well. In the future we would also like to explore various graph-based neural networks that model the quantitative effects on steroidogenesis (e.g. magnitude of steroid fold-change) found in the HTS. We would also like to use both binary and regression model scores in tandem to potentially increase prediction accuracy of these models. Future prospective validation of the predictions of these models will also be key. These models can also be incorporated in our Mega Tox software to enable accessible predictions for the steroidogenesis pathway and be used alongside other models as well as read-across capabilities. It should also be of note that finding or generating additional data for the enzymes (CYP11A and 36-HSD) with missing machine learning models will also be of great interest.
In conclusion, the approaches highlighted could be used to prioritize potential compounds of concern for future testing of these chemicals and if quantitative adverse outcome pathways are available, they could help to establish potential biological plausibility, allowing for the prediction of toxicological potency and ultimately risk of exposure. Our approach may offer some benefits in screening very large numbers of molecules which would be either impossible to test in vitro or be too costly overall. The further development of such steroidogenesis machine learning models may therefore offer benefits for safety prediction of future molecules in pharmaceutical and consumer products.
Supplementary Material
Refer to Web version on PubMed Central for supplementary material.
Acknowledgements
We kindly acknowledge earlier discussions with Dr. Pat Guiney, Dr. Peter Thorne and Dr. Wendy Hillwalker on steroidogenesis. We thank Ms. Melanie Tojong for assistance with figures.
Funding Sources
This project was funded in whole or in part with federal funds from R44GM122196-02A1 from NIGMS and 1R44ES031038-01 from NIEHS.
Data and Software Availability Statement
All predictions and cross-validation statistics are available in the supplemental files. Training datasets are available upon request. These models can be made commercially available in the commercial software MegaTox.
Abbreviations
| HMG-CoA | 3-hydroxy-3-methyl-glutaryl-coenzyme A |
| AChE | Acetylcholinesterase |
| ada | AdaBoost decision trees |
| AOP | Adverse Outcome Pathway |
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
Lane et al.
| AR | androgen receptor |
| br | Bayesian Ridge regression |
| bnb | Bernoulli naïve bayes |
| MolBART | Bidirectional Auto-regressive Transformer |
| R2 | coefficient of determination |
| DL | Deep Learning |
| enr | elastic net regression |
| EDC | endocrine-disrupting chemicals |
| ER | estrogen receptor |
| ECFP6 | Extended Connectivity 6 |
| HT-H295R | high-throughput H295R steroidogenesis assay |
| HTS | high-throughput screening |
| H295R | human adenocarcinoma cell line |
| knn | k-nearest neighbors classification |
| knnr | k-nearest neighbors regression |
| LLM | Large Language Model |
| lr | logistic regression |
| MAE | Mean Absolute Error |
| rf | Random Forest classification |
| rfr | Random Forest Regression |
| AUC | receiver operator curve area under the curve |
| RMSE | Root Mean Squared Error |
| SVC | Support Vector Classification |
| svr | Support Vector Regression |
| EPA | U.S. Environmental Protection Agency |
| xgb | xgboost classification |
| xgbr | Xgboost Regression |
References
(1). Zorn KM; Foil DH; Lane TR; Hillwalker W; Feifarek DJ; Jones F; Klaren WD; Brinkman AM; Ekins S Comparing Machine Learning Models for Aromatase (P450 19A1). Environ Sci Technol 2020, 54 (23), 15546-15555. DOI: 10.1021/acs.est.0c05771. [PubMed: 33207874]
(2). Zorn KM; Foil DH; Lane TR; Hillwalker W; Feifarek DJ; Jones F; Klaren WD; Brinkman AM; Ekins S Comparison of Machine Learning Models for the Androgen Receptor. Environ Sci Technol 2020, 54 (21), 13690-13700. DOI: 10.1021/acs.est.0c03984. [PubMed: 33085465]
(3). Zorn KM; Foil DH; Lane TR; Russo DP; Hillwalker W; Feifarek DJ; Jones F; Klaren WD; Brinkman AM; Ekins S Machine Learning Models for Estrogen Receptor Bioactivity and Endocrine Disruption Prediction. Environ Sci Technol 2020, 54 (19), 12202-12213. DOI: 10.1021/acs.est.0c03982. [PubMed: 32857505]
(4). Russo DP; Zorn KM; Clark AM; Zhu H; Ekins S Comparing Multiple Machine Learning Algorithms and Metrics for Estrogen Receptor Binding Prediction. Mol Pharm 2018, 15 (10), 4361-4370. DOI: 10.1021/acs.molpharmaceut.8b00546. [PubMed: 30114914]
(5). Shanle EK; Xu W Endocrine disrupting chemicals targeting estrogen receptor signaling: identification and mechanisms of action. Chem Res Toxicol 2011, 24 (1), 6-19. DOI: 10.1021/ tx100231n. [PubMed: 21053929]
(6). Kleinstreuer NC; Ceger PC; Allen DG; Strickland J; Chang X; Hamm JT; Casey WM A Curated Database of Rodent Uterotrophic Bioactivity. Environ Health Perspect 2016, 124 (5), 556-562. DOI: 10.1289/ehp.1510183. [PubMed: 26431337]
(7). Takemura H; Sakakibara H; Yamazaki S; Shimoi K Breast cancer and flavonoids - a role in prevention. Curr Pharm Des 2013, 19 (34), 6125-6132. [PubMed: 23448447]
(8). Rodgers KM; Udesky JO; Rudel RA; Brody JG Environmental chemicals and breast cancer: An updated review of epidemiological literature informed by biological mechanisms. Environ Res 2018, 160, 152-182. DOI: 10.1016/j.envres.2017.08.045. [PubMed: 28987728]
(9). European Chemical, A .; European Food Safety Authority with the technical support of the Joint Research, C .; Andersson N; Arena M; Auteri D; Barmaz S; Grignard E; Kienzler A; Lepper P; Lostia AM; Munn S; Parra Morte JM; Pellizzato F; Tarazona J; Terron A; Van der Linden S Guidance for the identification of endocrine disruptors in the context of Regulations (EU) No 528/2012 and (EC) No 1107/2009. EFSA J 2018, 16 (6), e05311. DOI: 10.2903/j.efsa.2018.5311. [PubMed: 32625944]
(10). EPA. Endocrine Disruptor Screening Program Tier 1 Battery of Assays. https:// www.epa.gov/endocrine-disruption/endocrine-disruptor-screening-program-tier-1-battery-assays (accessed January 5, 2025).
(11). Sun H; Xia M; Austin CP; Huang R Paradigm shift in toxicity testing and modeling. AAPS J 2012, 14 (3), 473-480. DOI: 10.1208/s12248-012-9358-1. [PubMed: 22528508]
(12). Dix DJ; Houck KA; Martin MT; Richard AM; Setzer RW; Kavlock RJ The ToxCast program for prioritizing toxicity testing of environmental chemicals. Toxicol Sci 2007, 95 (1), 5-12. [PubMed: 16963515]
(13). Judson RS; Houck KA; Kavlock RJ; Knudsen TB; Martin MT; Mortensen HM; Reif DM; Rotroff DM; Shah I; Richard AM; Dix DJ In vitro screening of environmental chemicals for targeted testing prioritization: the ToxCast project. Environ Health Perspect 2010, 118 (4), 485-492. [PubMed: 20368123]
(14). Setzer RW An Introduction to EPA’s ToxCast. EPA, Ed .; 2009.
(15). Judson RS; Magpantay FM; Chickarmane V; Haskell C; Tania N; Taylor J; Xia M; Huang R; Rotroff DM; Filer DL; Houck KA; Martin MT; Sipes N; Richard AM; Mansouri K; Setzer RW; Knudsen TB; Crofton KM; Thomas RS Integrated Model of Chemical Perturbations of a Biological Pathway Using 18 In Vitro High-Throughput Screening Assays for the Estrogen Receptor. Toxicol Sci 2015, 148 (1), 137-154. DOI: 10.1093/toxsci/kfv168. [PubMed: 26272952]
(16). Browne P; Judson RS; Casey WM; Kleinstreuer NC; Thomas RS Screening Chemicals for Estrogen Receptor Bioactivity Using a Computational Model. Environ Sci Technol 2015, 49 (14), 8804-8814. DOI: 10.1021/acs.est.5b02641. [PubMed: 26066997]
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
(17). Nelms MD; Mellor CL; Enoch SJ; Judson RS; Patlewicz G; Richard AM; Madden JM; Cronin MTD; Edwards SW A mechanistic framework for integrating chemical structure and high-throughput screening results to improve toxicity predictions. Comp Toxicol 2018, 8, 1-12.
(18). Judson RS; Houck KA; Watt ED; Thomas RS On selecting a minimal set of in vitro assays to reliably determine estrogen agonist activity. Regul Toxicol Pharmacol 2017, 91, 39-49. DOI: 10.1016/j.yrtph.2017.09.022. [PubMed: 28993267]
(19). Mansouri K; Abdelaziz A; Rybacka A; Roncaglioni A; Tropsha A; Varnek A; Zakharov A; Worth A; Richard AM; Grulke CM; Trisciuzzi D; Fourches D; Horvath D; Benfenati E; Muratov E; Wedebye EB; Grisoni F; Mangiatordi GF; Incisivo GM; Hong H; Ng HW; Tetko IV; Balabin I; Kancherla J; Shen J; Burton J; Nicklaus M; Cassotti M; Nikolov NG; Nicolotti O; Andersson PL; Zang Q; Politi R; Beger RD; Todeschini R; Huang R; Farag S; Rosenberg SA; Slavov S; Hu X; Judson RS CERAPP: Collaborative Estrogen Receptor Activity Prediction Project. Environ Health Perspect 2016, 124 (7), 1023-1033. DOI: 10.1289/ehp.1510267. [PubMed: 26908244]
(20). He J; Peng T; Yang X; Liu H Development of QSAR models for predicting the binding affinity of endocrine disrupting chemicals to eight fish estrogen receptor. Ecotoxicol Environ Saf 2018, 148, 211-219. DOI: 10.1016/j.ecoenv.2017.10.023. [PubMed: 29055205]
(21). Zhao Q; Lu Y; Zhao Y; Li R; Luan F; Cordeiro MN Rational Design of Multi-Target Estrogen Receptors ERalpha and ERbeta by QSAR Approaches. Curr Drug Targets 2017, 18 (5), 576-591. DOI: 10.2174/1389450117666160401125542. [PubMed: 27033186]
(22). Sakkiah S; Selvaraj C; Gong P; Zhang C; Tong W; Hong H Development of estrogen receptor beta binding prediction model using large sets of chemicals. Oncotarget 2017, 8 (54), 92989- 93000. DOI: 10.18632/oncotarget.21723. [PubMed: 29190972]
(23). Lee S; Barron MG Structure-Based Understanding of Binding Affinity and Mode of Estrogen Receptor alpha Agonists and Antagonists. PLoS One 2017, 12 (1), e0169607. DOI: 10.1371/ journal.pone.0169607. [PubMed: 28061508]
(24). Asako Y; Uesawa Y High-Performance Prediction of Human Estrogen Receptor Agonists Based on Chemical Structures. Molecules 2017, 22 (4), 675. DOI: 10.3390/molecules22040675. [PubMed: 28441746]
(25). Wang P; Dang L; Zhu BT Use of computational modeling approaches in studying the binding interactions of compounds with human estrogen receptors. Steroids 2016, 105, 26-41. DOI: 10.1016/j.steroids.2015.11.001. [PubMed: 26639429]
(26). Ribay K; Kim MT; Wang W; Pinolini D; Zhu H Predictive Modeling of Estrogen Receptor Binding Agents Using Advanced Cheminformatics Tools and Massive Public Data. Front Environ Sci 2016, 4, 1-9. DOI: 10.3389/fenvs.2016.00012.
(27). Niu AQ; Xie LJ; Wang H; Zhu B; Wang SQ Prediction of selective estrogen receptor beta agonist using open data and machine learning approach. Drug Des Devel Ther 2016, 10, 2323-2331. DOI: 10.2147/DDDT.S110603.
(28). Niinivehmas SP; Manivannan E; Rauhamaki S; Huuskonen J; Pentikainen OT Identification of estrogen receptor alpha ligands with virtual screening techniques. J Mol Graph Model 2016, 64, 30-39. DOI: 10.1016/j.jmgm.2015.12.006. [PubMed: 26774287]
(29). Martin TM Prediction of in vitro and in vivo oestrogen receptor activity using hierarchical clustering. SAR QSAR Environ Res 2016, 27 (1), 17-30. DOI: 10.1080/1062936X.2015.1125945. [PubMed: 26784454]
(30). Huang R; Sakamuru S; Martin MT; Reif DM; Judson RS; Houck KA; Casey W; Hsieh JH; Shockley KR; Ceger P; Fostel J; Witt KL; Tong W; Rotroff DM; Zhao T; Shinn P; Simeonov A; Dix DJ; Austin CP; Kavlock RJ; Tice RR; Xia M Profiling of the Tox21 10K compound library for agonists and antagonists of the estrogen receptor alpha signaling pathway. Sci Rep 2014, 4, 5664. DOI: 10.1038/srep05664. [PubMed: 25012808]
(31). Zhang L; Sedykh A; Tripathi A; Zhu H; Afantitis A; Mouchlis VD; Melagraki G; Rusyn I; Tropsha A Identification of putative estrogen receptor-mediated endocrine disrupting chemicals using QSAR- and structure-based virtual screening approaches. Toxicol Appl Pharmacol 2013, 272 (1), 67-76. DOI: 10.1016/j.taap.2013.04.032. [PubMed: 23707773]
(32). Borrel A; Rudel RA Cheminformatics analysis of chemicals that increase estrogen and progesterone synthesis for a breast cancer hazard assessment. Sci Rep 2022, 12 (1), 20647. DOI: 10.1038/s41598-022-24889-w. [PubMed: 36450809]
Environ Sci Technol. Author manuscript; available in PMC 2025 October 01.
(33). Sanderson JT The steroid hormone biosynthesis pathway as a target for endocrine-disrupting chemicals. Toxicol Sci 2006, 94 (1), 3-21. DOI: 10.1093/toxsci/kf1051. [PubMed: 16807284]
(34). Guarnotta V; Amodei R; Frasca F; Aversa A; Giordano C Impact of Chemical Endocrine Disruptors and Hormone Modulators on the Endocrine System. Int J Mol Sci 2022, 23 (10), 5710. DOI: 10.3390/ijms23105710. [PubMed: 35628520]
(35). Storbeck KH; Schiffer L; Baranowski ES; Chortis V; Prete A; Barnard L; Gilligan LC; Taylor AE; Idkowiak J; Arlt W; Shackleton CHL Steroid Metabolome Analysis in Disorders of Adrenal Steroid Biosynthesis and Metabolism. Endocr Rev 2019, 40 (6), 1605-1625. DOI: 10.1210/ er.2018-00262. [PubMed: 31294783]
(36). Hanukoglu I; Rapoport R Routes and regulation of NADPH production in steroidogenic mitochondria. Endocr Res 1995, 21 (1-2), 231-241. DOI: 10.3109/07435809509030439. [PubMed: 7588385]
(37). Miller WL; Pandey AV; Fluck CE Disordered Electron Transfer: New Forms of Defective Steroidogenesis and Mitochondriopathy. J Clin Endocrinol Metab 2025, 110 (3), e574-e582. DOI: 10.1210/clinem/dgae815. [PubMed: 39574227]
(38). Karmaus AL; Toole CM; Filer DL; Lewis KC; Martin MT High-Throughput Screening of Chemical Effects on Steroidogenesis Using H295R Human Adrenocortical Carcinoma Cells. Toxicol Sci 2016, 150 (2), 323-332. DOI: 10.1093/toxsci/kfw002. [PubMed: 26781511]
(39). Henson MC; Chedrese PJ Endocrine disruption by cadmium, a common environmental toxicant with paradoxical effects on reproduction. Exp Biol Med (Maywood) 2004, 229 (5), 383-392. DOI: 10.1177/153537020422900506. [PubMed: 15096650]
(40). Levine SL; Webb EG; Saltmiras DA Review and analysis of the potential for glyphosate to interact with the estrogen, androgen and thyroid pathways. Pest Manag Sci 2020, 76 (9), 2886- 2906. DOI: 10.1002/ps.5983. [PubMed: 32608552]
(41). Chakraborty S; Pramanik J; Mahata B Revisiting steroidogenesis and its role in immune regulation with the advanced tools and technologies. Genes Immun 2021, 22 (3), 125-140. DOI: 10.1038/s41435-021-00139-3. [PubMed: 34127827]
(42). Breen MS; Breen M; Terasaki N; Yamazaki M; Conolly RB Computational model of steroidogenesis in human H295R cells to predict biochemical response to endocrine-active chemicals: model development for metyrapone. Environ Health Perspect 2010, 118 (2), 265-272. DOI: 10.1289/ehp.0901107. [PubMed: 20123619]
(43). Saito R; Terasaki N; Yamazaki M; Masutomi N; Tsutsui N; Okamoto M Estimation of the Mechanism of Adrenal Action of Endocrine-Disrupting Compounds Using a Computational Model of Adrenal Steroidogenesis in NCI-H295R Cells. J Toxicol 2016, 2016, 4041827. DOI: 10.1155/2016/4041827. [PubMed: 27057163]
(44). Zdrazil B Fifteen years of ChEMBL and its role in cheminformatics and drug discovery. J Cheminform 2025, 17 (1), 32. DOI: 10.1186/s13321-025-00963-z. [PubMed: 40065463]
(45). EPA. Comptox dashboard. 2024. https://comptox.epa.gov/dashboard/batch-search (accessed Febuary 12, 2024).
(46). Lane T; Russo DP; Zorn KM; Clark AM; Korotcov A; Tkachenko V; Reynolds RC; Perryman AL; Freundlich JS; Ekins S Comparing and Validating Machine Learning Models for Mycobacterium tuberculosis Drug Discovery. Mol Pharm 2018, 15 (10), 4346-4360. DOI: 10.1021/acs.molpharmaceut.8b00083. [PubMed: 29672063]
(47). Carletta J Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics 1996, 22, 249-254.
(48). Cohen J A coefficient of agreement for nominal scales. Education and Psychological Measurement 1960, 20, 37-46.
(49). Matthews BW Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim Biophys Acta 1975, 405 (2), 442-451. [PubMed: 1180967]
(50). Ding T; Angelopoulos A; Bates S; Jordan M; Tibshirani RJ Class-conditional conformal prediction with many classes. Advances in neural information processing systems 2023, 36, 64555-64576.
(51). Vovk V Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence 2015, 74, 9-28.
(52). Liu Y; Han T; Ma S; Zhang J; Yang Y; Tian J; He H; Li A; He M; Liu Z; Wu Z; Zhao L; Zhu D; Li X; Qiang N; Shen D; Liu T; Ge B Summary of ChatGPT-Related research and perspective towards the future of large language models. Meta-Radiology 2023, 1 (2), 100017. DOI: 10.1016/ j.metrad.2023.100017.
(53). Irwin R; Dimitriadis S; He J; Bjerrum EJ Chemformer: a pre-trained transformer for computational chemistry. Machine Learning: Science and Technology 2022, 3 (1), 015022.
(54). Falcon W PyTorchLightning/pytorch-lightning: 0.7.6 release (0.7.6). 2020. https://zenodo.org/ records/3828935 (accessed Feburary 08, 2024).
(55). van der Maaten L; Hinton G Visualizing Data using t-SNE. J Machine Learning Research 2008, 9, 2579-2605.
(56). Co-operation, O. f. E .; Development. Test No. 416: Two-Generation Reproduction Toxicity; OECD Publishing, 2001.
(57). Wang J; Hallinger DR; Murr AS; Buckalew AR; Lougee RR; Richard AM; Laws SC; Stoker TE High-throughput screening and chemotype-enrichment analysis of ToxCast phase II chemicals evaluated for human sodium-iodide symporter (NIS) inhibition. Environ Int 2019, 126, 377-386. DOI: 10.1016/j.envint.2019.02.024. [PubMed: 30826616]
(58). Hornung MW; Korte JJ; Olker JH; Denny JS; Knutsen C; Hartig PC; Cardon MC; Degitz SJ Screening the ToxCast Phase 1 Chemical Library for Inhibition of Deiodinase Type 1 Activity. Toxicol Sci 2018, 162 (2), 570-581. DOI: 10.1093/toxsci/kfx279. [PubMed: 29228274]
(59). Wang F; Xing J Classification of thyroid hormone receptor agonists and antagonists using statistical learning approaches. Molecular diversity 2019, 23 (1), 85-92. DOI: 10.1007/ s11030-018-9857-9. [PubMed: 30014306]
(60). Norinder U; Boyer S Binary classification of imbalanced datasets using conformal prediction. J Mol Graph Model 2017, 72, 256-265. DOI: 10.1016/j.jmgm.2017.01.008. [PubMed: 28135672]
(61). Norinder U; Boyer S Conformal Prediction Classification of a Large Data Set of Environmental Chemicals from ToxCast and Tox21 Estrogen Receptor Assays. Chem Res Toxicol 2016, 29 (6), 1003-1010. DOI: 10.1021/acs.chemrestox.6b00037. [PubMed: 27152554]
(62). Zhang J; Norinder U; Svensson F Deep Learning-Based Conformal Prediction of Toxicity. J Chem Inf Model 2021, 61 (6), 2648-2657. DOI: 10.1021/acs.jcim.1c00208. [PubMed: 34043352]
(63). Liu Y; Levis AW; Normand SL; Han L Multi-Source Conformal Inference Under Distribution Shift. Proc Mach Learn Res 2024, 235, 31344-31382. [PubMed: 39193374]
(64). Luong KD; Singh A Application of Transformers in Cheminformatics. J Chem Inf Model 2024, 64 (11), 4392-4409. DOI: 10.1021/acs.jcim.3c02070. [PubMed: 38815246]
(65). Angelopoulos AN; Bates S A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. 2021. arXiv (Machine Learning). 2107.07511v6 [cs.LG]. (accessed 2024-07-01)
(66). Sapounidou M; Norinder U; Andersson PL Predicting Endocrine Disruption Using Conformal Prediction - A Prioritization Strategy to Identify Hazardous Chemicals with Confidence. Chem Res Toxicol 2023, 36 (1), 53-65. DOI: 10.1021/acs.chemrestox.2c00267. [PubMed: 36534483]
(67). Bhatnagar AS; Hausler A; Schieweck K; Lang M; Bowman R Highly selective inhibition of estrogen biosynthesis by CGS 20267, a new non-steroidal aromatase inhibitor. J Steroid Biochem Mol Biol 1990, 37 (6), 1021-1027. DOI: 10.1016/0960-0760(90)90460-3. [PubMed: 2149502]
(68). Kijima I; Itoh T; Chen S Growth inhibition of estrogen receptor-positive and aromatase- positive human breast cancer cells in monolayer and spheroid cultures by letrozole, anastrozole, and tamoxifen. J Steroid Biochem Mol Biol 2005, 97 (4), 360-368. DOI: 10.1016/ j.jsbmb.2005.09.003. [PubMed: 16263272]
(69). Ortega I; Sokalska A; Villanueva JA; Cress AB; Wong DH; Stener-Victorin E; Stanley SD; Duleba AJ Letrozole increases ovarian growth and Cyp17a1 gene expression in the rat ovary. Fertil Steril 2013, 99 (3), 889-896. DOI: 10.1016/j.fertnstert.2012.11.006. [PubMed: 23200686]
(70). Haynes BP; Dowsett M; Miller WR; Dixon JM; Bhatnagar AS The pharmacology of letrozole. J Steroid Biochem Mol Biol 2003, 87 (1), 35-45. DOI: 10.1016/s0960-0760(03)00384-4. [PubMed: 14630089]
(71). Sadashiva CT; Narendra Sharath Chandra JN; Ponnappa KC; Veerabasappa Gowda T; Rangappa KS Synthesis and efficacy of 1-[bis(4-fluorophenyl)-methyl]piperazine derivatives for acetylcholinesterase inhibition, as a stimulant of central cholinergic neurotransmission in Alzheimer’s disease. Bioorganic & medicinal chemistry letters 2006, 16 (15), 3932-3936. DOI: 10.1016/j.bmcl.2006.05.030. [PubMed: 16735118]
(72). Komloova M; Horova A; Hrabinova M; Jun D; Dolezal M; Vinsova J; Kuca K; Musilek K Preparation, in vitro evaluation and molecular modelling of pyridinium-quinolinium/ isoquinolinium non-symmetrical bisquaternary cholinesterase inhibitors. Bioorganic & medicinal chemistry letters 2013, 23 (24), 6663-6666. DOI: 10.1016/j.bmcl.2013.10.043. [PubMed: 24220173]
(73). Zhang L; Li DC; Liu LF Paeonol: pharmacological effects and mechanisms of action. Int Immunopharmacol 2019, 72, 413-421. DOI: 10.1016/j.intimp.2019.04.033. [PubMed: 31030097]
(74). Gobert A; Di Cara B; Cistarelli L; Millan MJ Piribedil enhances frontocortical and hippocampal release of acetylcholine in freely moving rats by blockade of alpha 2A-adrenoceptors: a dialysis comparison to talipexole and quinelorane in the absence of acetylcholinesterase inhibitors. J Pharmacol Exp Ther 2003, 305 (1), 338-346. DOI: 10.1124/jpet.102.046383. [PubMed: 12649387]
(75). Millan MJ; Cussac D; Milligan G; Carr C; Audinot V; Gobert A; Lejeune F; Rivet JM; Brocco M; Duqueyroix D; Nicolas JP; Boutin JA; Newman-Tancredi A Antiparkinsonian agent piribedil displays antagonist properties at native, rat, and cloned, human alpha(2)-adrenoceptors: cellular and functional characterization. J Pharmacol Exp Ther 2001, 297 (3), 876-887. [PubMed: 11356907]
(76). Sokalska A; Stanley SD; Villanueva JA; Ortega I; Duleba AJ Comparison of effects of different statins on growth and steroidogenesis of rat ovarian theca-interstitial cells. Biol Reprod 2014, 90 (2), 44. DOI: 10.1095/biolreprod.113.114843. [PubMed: 24389875]
(77). Bohm M; Herrmann W; Wassmann S; Laufs U; Nickenig G Does statin therapy influence steroid hormone synthesis? Z Kardiol 2004, 93 (1), 43-48. DOI: 10.1007/s00392-004-1003-2. [PubMed: 14740240]
(78). Jay RH; Sturley RH; Stirling C; McGarrigle HH; Katz M; Reckless JP; Betteridge DJ Effects of pravastatin and cholestyramine on gonadal and adrenal steroid production in familial hypercholesterolaemia. Br J Clin Pharmacol 1991, 32 (4), 417-422. DOI: 10.1111/ j.1365-2125.1991.tb03924.x. [PubMed: 1958433]
(79). Kwak K; Ji K; Kho Y; Kim P; Lee J; Ryu J; Choi K Chronic toxicity and endocrine disruption of naproxen in freshwater waterfleas and fish, and steroidogenic alteration using H295R cell assay. Chemosphere 2018, 204, 156-162. DOI: 10.1016/j.chemosphere.2018.04.035. [PubMed: 29655108]
(80). Schmitz M; Beghin M; Mandiki SNM; Nott K; Gillet M; Ronkart S; Robert C; Baekelandt S; Kestemont P Environmentally-relevant mixture of pharmaceutical drugs stimulates sex-steroid hormone production and modulates the expression of candidate genes in the ovary of juvenile female rainbow trout. Aquat Toxicol 2018, 205, 89-99. DOI: 10.1016/j.aquatox.2018.10.006. [PubMed: 30347285]
(81). Lightly ER; Walker SW; Bird IM; Williams BC Subclassification of beta-adrenoceptors responsible for steroidogenesis in primary cultures of bovine adrenocortical zona fasciculata/reticularis cells. Br J Pharmacol 1990, 99 (4), 709-712. DOI: 10.1111/ j.1476-5381.1990.tb12993.x. [PubMed: 1972892]
(82). Miszkiel G; Kotwica J Mechanism of action of noradrenaline on secretion of progesterone and oxytocin by the bovine corpus luteum in vitro. Acta Vet Hung 2001, 49 (1), 39-51. DOI: 10.1556/004.49.2001.1.6. [PubMed: 11402689]
(83). Urbina F; Jones T; Harris JS; Snyder SH; Lane TR; Ekins S Predicting the Hallucinogenic Potential of Molecules Using Artificial Intelligence. ACS Chem Neurosci 2024, 15 (16), 3078- 3089. DOI: 10.1021/acschemneuro.4c00405. [PubMed: 39092989]
(84). Snyder SH; Vignaux PA; Ozalp MK; Gerlach J; Puhl AC; Lane TR; Corbett J; Urbina F; Ekins S The Goldilocks paradigm: comparing classical machine learning, large language models,
and few-shot learning for drug discovery applications. Commun Chem 2024, 7 (1), 134. DOI: 10.1038/s42004-024-01220-4. [PubMed: 38866916]
Synopsis
We have developed machine learning models to predict environmental chemicals’ impact on steroidogenesis disruption, with a validation screen achieving 80% accuracy, which may potentially aid in risk assessment and regulatory prioritization.
Cholesterol
Pregnenolone
Progesterone
11-Deoxy corticosterone
Corticosterone
Aldosterone
17a-Hydroxy pregnenolone
17a-Hydroxy progresterone
11-Deoxycortisol
Cortisol
CYPTIA
3ß-HSD(1/2)
Dehydro epiandrosterone
Androstenedione
Estrone
CYP21
CYP11B2
CYP17
CYP19
Androstenediol
Testosterone
17ß-Estradiol
CYP11B1
17ß-HSD(1/2/3/5)
5a-reductase(1/2)
Dihydro testosterone
A
| Method | AUC | F1 | Precision | Recall | Accuracy Specificity | Cohen's K MCC | |||
|---|---|---|---|---|---|---|---|---|---|
| ada | 0.69 | 0.6 | 0.67 | 0.55 | 0.65 | 0.75 | 0.3 | 0.3 | |
| bnb | 0.71 | 0.65 | 0.64 | 0.66 | 0.66 | 0.66 | 0.32 | 0.32 | |
| knn | 0.73 | 0.64 | 0.67 | 0.61 | 0.67 | 0.72 | 0.33 | 0.34 | |
| Ireg | 0.72 | 0.64 | 0.66 | 0.62 | 0.67 | 0.71 | 0.33 | 0.33 | |
| DL | 0.71 | 0.62 | 0.68 | 0.57 | 0.67 | 0.75 | 0.33 | 0.33 | |
| rf | 0.74 | 0.69 | 0.65 | 0.73 | 0.68 | 0.64 | 0.37 | 0.37 | |
| SVC | 0.75 | 0.69 | 0.66 | 0.73 | 0.69 | 0.66 | 0.39 | 0.39 | |
| xgb | 0.69 | 0.64 | 0.64 | 0.64 | 0.66 | 0.67 | 0.31 | 0.31 | |
SVC
Steroidogenesis Modulation (ECFP6-1024)
879/966
B
RF AUC=0.74, F1=0.69
1845
AUC=0.75, F1=0.69
Cross Validation ROC
Cross Validation ROC
1
1|
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
0
0.5
1
0
0.5
1
Model Evaluation Overlay
Model Evaluation Overlay
0.7
80
60
Count
Count
60
40
40
20
20
0
0.2
0.4
0.6
0.8
0
Value
0.2
0.4
0.6
Value
0.8
Model Truth Table
Model Truth Table
True Positive
False Postive
640
345
False Negative
True Negative
239
621
| True Positive 640 | False Postive 329 |
|---|---|
| False Negative | True Negative 637 |
| 239 |
Figure 2. 5-fold cross-validation metrics for classification machine learning models for the modulation of steroidogenesis in the H295R model. Training dataset information (dataset size, classification distributions) is annotated (blue and grey text boxes). (A) Performance by algorithm is given numerically, with (B) example truth tables, ROC plots and probability- like scores histogram distributions examples shown for random forest and SVC models. For the histograms, red and blue bars represent the ground truth negative and positive classes, respectively. A 0.7 probability-like score is annotated on the rf model histogram highlighting the accuracy of the positive class at this threshold. Deep Learning (DL), AdaBoost decision trees (ada), Bernoulli naïve bayes (bnb), Bayesian Ridge regression (br), elastic net regression (enr), k-nearest neighbors (knn), support vector machine (svc), logistic regression (lr), xgboost (xgb) and random forest (rf).
Figure 3. Distributions of 5-fold cross-validation (CV) metrics and training set size/balance for all (A) classification and (B) regression models built for the steroidogenesis targets (Table 1). All classification models have a unified threshold of 100 nM. Balance is the fraction of the positive class. Receiver operator curve “area under the curve” (AUC), Accuracy (ACC), Recall, Specificity (Spec), Precision (Prec), F-1 score (F1), Matthews Correlation Coefficient (MCC) and Cohen’s kappa coefficient (Cohen’s x), Mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2).
A Steroidogenesis Classification Models
B Steroidogenesis Regression Models
5-Fold CV
Dataset
2500
2.0
2000
1.0
Number of Compounds
Number of Compounds
0.8
2000
1.5
1500
1.0
Score
0.6
1500
. ..
Score
0.5
1000
0.4
1000
0.0
0.2
500
500
-0.5
0.0
MCC Precision
Balance Training Size
0
-1.0
0
ACC
F1-Score
AUC
Cohen’s K
Recall
Specificity
MAE
RMSE
R2
Training Size
A Steroid 5-a-reductase 2 (ChEMBL1856) Binary, 100nM
B Steroid 5-a-reductase 2 (ChEMBL1856) Regression
420
265/155
rf
333
4-10.4
rfr
100
Model Evaluation Overlay
Method
AUC
F1
Precision
Recall
Accuracy
Specificity
Cohen’s K MCC
Method
MAE
RMSE
R2
Cross-Validation Predictions
ada
0.95
0.94
0.94
0.94
0.92
0.9
0.84
0.84
80
adar
0.63
0.81
0.68
10
60
bnb
0.93
09
0.93
0.87
0.87
0.88
0.73
0.74
Count
40
br
0.53
0.75
0.72
9
DL
0.96
0.92
0.92
0.92
0.9
0.86
0.79
0.79
20
enr
0.62
0.8
0.69
8
knn
0.95
0.93
0.92
0.94
0.91
0.86
0.81
0.82
0
0
0.5
knnr
0.57
0.75
0.72
7
Ireg
0.96
0.94
0.93
0.94
0.92
0.88
0.83
0.83
Value
Model Truth Table
0.53
0.73
0.74
6
rf
0.97
0.93
0.94
0.92
0.92
0.9
0.82
0.82
rfr
0.97
0.94
0.94
0.94
0.92
0.9
0.84
0.84
svr
0.57
0.76
0.71
5
SVC
4
xgb
0.95
0.92
0.93
0.92
0.9
0.88
0.8
08
xgbr
0.59
0.87
0.62
4
5
6
7
8
9
10
C 17-B-HSD3 (ChEMBL4234) Binary 100nM
D
17-B-HSD3 (ChEMBL4234) Regression
rf
2.77-10
rfr
198
104/94
196
Model Evaluation Overlay
Method
AUC
F1
Precision
Recall
Accuracy Spec
pecificit
Cohen’s K
MCC
10
Method
MAE
RMSE
R2
Cross-Validation Predictions
ada
0.76
0.69
0.78
0.62
0.71
08
0.42
0.43
adar
0.62
0.81
0.51
10
bnb
0.74
0.68
0.7
0.66
0.67
0.68
0.34
0.35
Count
5
br
0.57
0.74
0.59
9
DL
0.78
0.72
0.73
0.72
0.71
0.7
0.42
0.42
8
enr
0.88
1.17
0.02
knn
7
0.72
07
0.73
0.68
0.71
0.73
0.41
0.42
0
knnr
0.62
08
0.51
6
Ireg
0.77
0.74
0.74
0.73
0.73
0.72
0.45
0.45
Q
0.5
Value
.
rf
0.78
0.76
0.76
0.77
0.75
0.72
0.49
0.49
Model Truth Table
rfr
0.61
08
0.52
5
4
SVC
0.81
0.75
0.76
0.75
0.74
0.73
0.48
0.49
svr
08
1.07
0.14
3
xgb
0.77
0.72
0.73
0.71
0.71
07
0.41
0.41
xgbr
0.78
1
0.21
2
3
4
5
6
7
8
9
| True Positive 245 | False Positive 15 |
| False Negative 20 | True Negative 140 |
| True Positive 80 | False Positive 26 |
|---|---|
| False Negative 24 | True Negative 68 |
Figure 4. 5-fold cross-validation metrics models for the inhibition (IC50) of two steroidogenesis targets, (A,B) steroid 5-a-reductase 2 and 17B-HSD2 (C,D). Training dataset information, such as dataset size, classification distributions or activity ranges are annotated below each title. Performance by algorithm is given numerically, with example truth tables, probability- like scores histogram distributions and plotted activity (predicted vs actual activity [-logM]) examples shown for random forest models. For the histograms, red and blue bars represent the ground truth negative and positive classes, respectively. Deep Learning (DL), AdaBoost decision trees (ada), Bernoulli naïve bayes (bnb), Bayesian Ridge regression (br), elastic net regression (enr), k-nearest neighbors (knn), support vector machine (svc), logistic regression (lr), xgboost (xgb) and random forest (rf).
| A Forskolin (Control) | 14.66 | 1.97 | 23.59 | 4.93 | 2.72 | 2.69 | 1.20 | 16.15 | 22.30 | 21.85 | 1.30 | 2.25 | 1.99 | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Prochloraz (Control)- | -1.14 | 63.35 | -2.94 | -8.39 | -2.72 | 52.42 | -3.55 | -4.31 | -7.86 | -1.34 | -1.65 | 12.15 | 15.32 | |
| Sulfamethazine (Na+)- | -1.06 | 1.02 | -1.01 | -1.02 | -1.26 | -1.05 | -1.18 | -1.06 | 1.10 | 1.02 | 1.05 | -1.07 | 1.00 | |
| Piribedil- | ☒ | -1.27 | -2.93 | -1.82 | -1.93 | -1.28 | 3.41 | 2.51 | -1.15 | 1.14 | -1.32 | -2.61 | -1.85 | |
| Paeonol- | 1.71 | -1.12 | 2.31 | 1.26 | -1.36 | 1.09 | -1.31 | 1.74 | 2.16 | 2.06 | -1.17 | 1.01 | -1.45 | |
| 4-Methylumbelliferone | -1.14 | -1.01 | -2.94 | -3.84 | -2.16 | -5.59 | -1.98 | -7.49 | 1.16 | 2.12 | 18.24 | 5.80 | -1.43 | |
| Dobutamine (HCI)- | 3.49 | 5.89 | 6.20 | -3.29 | -2.72 | 2.99 | 3.55 | 3.57 | 1.29 | 2.47 | 4.84 | 22.45 | -4.46 | |
| Edrophonium chloride | 1.10 | -1.01 | 1.00 | -1.02 | 1.18 | 1.00 | -1.03 | 1.20 | 1.02 | 1.08 | 1.18 | 1.26 | -1.20 | |
| Ritodrine (HCI)- | -1.14 | -1.30 | -1.19 | -1.35 | 1.01 | -1.19 | -1.48 | -1.07 | 1.34 | 1.48 | 1.38 | 1.26 | -1.54 | |
| Neostigmine (Br-)- | 1.60 | -1.11 | -1.03 | -1.11 | 1.05 | -1.00 | -1.25 | -1.03 | 1.13 | 1.08 | 1.04 | 1.42 | -1.41 | |
| Bentiromide- | ☒ | 1.08 | -1.13 | -1.09 | 1.52 | 1.01 | -1.10 | -1.01 | 1.00 | 1.15 | 1.10 | -1.02 | 1.01 | |
| Letrozole | ☒ | -4.96 | -2.94 | -8.39 | -2.72 | -2.38 | 3.26 | -1.39 | 20.44 | -2.80 | -1.14 | -1.00 | -3.69 | |
| N-Acetyl-L-tyrosine | -1.14 | 1.13 | -1.10 | -1.07 | 1.07 | -1.02 | 1.01 | -1.07 | -1.10 | -1.09 | 1.07 | -1.10 | 1.13 | |
| Tyramine | -1.07 | -1.07 | -1.44 | -1.47 | -1.02 | -1.08 | -1.19 | 1.00 | -1.02 | 1.11 | 1.02 | -1.05 | -1.24 | |
| Pravastatin (Na+)- | -1.14 | -1.38 | -2.91 | -1.40 | 1.85 | -1.50 | -1.37 | -2.87 | -1.17 | -1.02 | -2.47 | -4.25 | -1.22 | |
| (±)-Naproxen- | 1.46 | 1.09 | 1.08 | -1.08 | 1.05 | -1.14 | 2.71 | 1.09 | 1.26 | 1.21 | -1.16 | -1.16 | -1.66 | |
| B | Aldosterone | Adrostenedione | Corticosterone | Cortisol- | Cortisone | 11-Deoxycortisol- | DHT | DOC | Estrone | Estradiol- | 17OH-Progesterone | Progesterone | Testosterone | |
| Forskolin (Control). | 14.66 | 1.97 | 23.59 | 4.93 | -2.72 | 2.69 | 1.20 | 16.1522.3021.85 | 1.30 | 2.25 | 1.99 | |||
| Prochloraz (Control) | -1.14 | -63.3 | 2.94 | 8.39 | 2.72 | 52.42 | 3.55 | 4.31 | 7.86 | 1.34 | -1.65 | 12.15 | 15.3 | |
| (-)-Fucose- | 1.11 | -1.07 | -1.05 | -1.02 | 1.30 | 1.04 | -1.15 | -1.00 | 1.00 | 1.09 | 1.04 | -1.04 | 1.02 | |
| Cyclen | 1.27 | 1.02 | 1.41 | 1.26 | -1.03 | 1.04 | 1.02 | -1.12 | 1.14 | 1.17 | 1.34 | 1.01 | 1.13 | |
| Magnesium acetate- | 1.15 | 1.02 | 1.08 | -1.01 | 1.06 | -1.02 | 1.21 | 1.12 | 1.22 | 1.04 | 1.02 | -1.14 | -1.23 | |
| Nitroprusside- | 1.14 | 1.05 | -1.20 | -1.21 | -1.05 | -1.11 | -1.15 | -1.03 | -1.29 | 1.35 | 1.15 | 1.09 | 1.06 | |
| Heptaminol HCI- | -1.07 | 1.01 | -1.13 | -1.14 | 1.08 | -1.08 | -1.13 | -1.08 | -1.10 | -1.08 | 1.02 | -1.11 | 1.05 | |
| Pyruvic acid- | 1.03 | 1.11 | -1.06 | -1.10 | 1.15 | 1.03 | 1.01 | -1.08 | -1.07 | 1.09 | 1.07 | -1.00 | 1.16 | |
| Aldosterone | Adrostenedione | Corticosterone | Cortisol | Cortisone | 11-Deoxycortisol | DHT | DOC | Estrone | Estradiol | 17OH-Progesterone | Progesterone | Testosterone | ||
≥ 1.5 fold-change ± 1.49 fold-change ≤-1.5 fold-change
Classification (Consensus)
O
HO.
OH
O
OH
Pravastatin
H
Cholesterol
CYP11A
HO
0
0
0
OH
O
OH
O
0
3B-HSD(1/2)
CYP21A2 (0)
CYP11B1 (0)
HO.
OH
HO
CYP11B2 (1)
HO”
0
0
Pregnenolone
Progesterone
11-Deoxy corticosterone
CYP11B2 (1)
0
0
Corticosterone
Aldosterone
CYP17A1 (0)
CYP17A1 (0)
OH
1
OH
HO 1
O
O
o
OH
HO
5
OH
0
3B-HSD(1/2)
CYP21A2 (0)
CYP11B2 (1)
HO
0
17a-Hydroxy pregnenolone
17a-Hydroxy progresterone
0
11-Deoxycortisol
0
Cortisol
CYP17A1 (0)
CYP17A1 (0)
Regression (Consensus)
0
0
0
3B-HSD(1/2)
CYP19 (0)
CYP11B1
HỢP
HO
Dehydro epiandrosterone
Androstenedione
0
5a-reductase2-8 -… CYP11B2
Estrone
6
17B-HSD(1/2/3/5) (0)
17B-HSD(1/2/3/5) (0)
17B-HSD(1/2/3/5) (0)
5a-reductase1
CYP17A1
OH
1
OH
OH
17B-HSD5
CYP19
3B-HSD(1/2)
CYP19 (0)
HO-
0
HO
Androstenediol
Testosterone
17ß-Estradiol
17-HSD3
CYP21A2
5a-reductasel (1) 5a-reductase2 (0)
176-HSD2178-HSD1
Probability-Like Score
1
OH
-logM
0.0
1.0
Dihydrotestosterone
0
A
100
Target (Specific)
· 17BHSD1
50
· 17BHSD2
· 17BHSD3
· 17BHSD5
TSNE Y
· CYP11B1
0
· CYP11B2
· CYP17A1
· CYP19A1
-50
· CYP21A2
· Steroid 5-alpha-reductase 1
· Steroid 5-alpha-reductase 2
· Steroidogenesis HTS
-100
-100
-50
0
50
100
TSNE X
B
100
Set
50
· Primary
· Dyes
· Scents
TSNE Y
· Home Maintenance
0
. Industrial Products
· Personal Care
· Flavorings
-50
. Cleaning Products
· OTC
· FDA-Approved
-100
-100
-50
0
50
100
TSNE X
| Enzyme name | Enzyme ChEMBL ID | Regression Training Size | *Actives/Total | Activity Range (-logM) |
|---|---|---|---|---|
| CYP11A | CHEMBL2033 | N/A | N/A | N/A |
| 36-HSD1/2 | CHEMBL1958/CHEMBL3670 | N/A | N/A | N/A |
| CYP17 | CHEMBL3522 | 664 | 489/940 | 4.0-9.3 |
| CYP19 | CHEMBL1978 | 1749 | 415/1971 | 4-10.8 |
| 17B-HSD 1 | CHEMBL3181 | 528 | 253/563 | 4.1-10 |
| 17B-HSD 2 | CHEMBL2789 | 544 | 164/563 | 3.5-9.7 |
| 17B-HSD 3 | CHEMBL4234 | 196 | 104/198 | 2.8-10 |
| 176-HSD 5 | CHEMBL4681 | 623 | 159/657 | 3.7-9.3 |
| CYP21 | CHEMBL2759 | 113 | 31/127 | 5.0-9.2 |
| CYP11B1 | CHEMBL1908 | 962 | 331/1066 | 4.0-9.5 |
| CYP11B2 | CHEMBL2722 | 1254 | 851/1326 | 4.1-10 |
| 5a-reductase 1 | CHEMBL1787 | 281 | 148/309 | 4.3-9.2 |
| 5a-reductase 2 | CHEMBL1856 | 333 | 265/420 | 4-10.4 |