Introduction

Properties such as absorption, distribution, metabolism, excretion and toxicity (ADMET), are an important component of pharmaceutical drug design. It is often reported that the failure to meet requisite ADMET criteria are a common cause for the high attrition rates of drug candidates [1]. Early ADMET profiling is indeed desirable so as to mitigate the risk of attrition. Various medium and high-throughput in vitro ADMET screens have therefore been developed, that have contributed to the available experimental data. These are nonetheless quite expensive especially when thousands of compounds are involved. Furthermore, reducing animal testing has now become a priority.

With the aim of facilitating rapid and inexpensive means of ADMET profiling, various in silico tools have been developed [2]. Using databases of experimentally measured ADMET properties [3], various quantitative structure-activity/property relationship (QSAR/QSPR) models have been generated that can predict a range of ADMET properties for novel chemical entities. Other efforts have made use of ADMET predictions to evaluate drug-likeness of a compound [4, 5]. While some of the models are available as part of commercial software packages based on proprietary datasets, there has been a significant push for open source software and web services [6,7,8,9,10,11,12].

Among the popular services, ADMETLab [12] offers 53 prediction models that are calculated using a multi-task graph attention network and operates on graph-structured data. The method is able to generate customized fingerprints from the general features for a specific task. Another web tool, SwissADME [9] evaluates pharmacokinetics, drug-likeness of small molecules. The predictions are based on a combination of fragmental methods (for solubility), as well as machine-learning based binary classification methods for other ADMET properties (cytochrome-P450 inhibitor, P-glycoprotein substrate). In ADMETSar [11], models for applications in both drug discovery and environmental risk assessment are built using MACCS and Morgan fingerprints. The toxicity models used in ProTox [13] are developed based on chemical similarities between compounds with known toxic effects and the presence of toxic fragments. Other models for hepatotoxicity, cytotoxicity, mutagenicity, and carcinogenicity rely on fingerprints (MACCS/Morgan). Extended connectivity fingerprints form the basis for the prediction of 15 ADMET properties in the vNN server [10] where models are trained using variable nearest neighbourhood method. pkCSM [6], on the other hand, uses graph-based signatures to develop predictive models of central ADMET properties. Other software such as MDCKPred [14], CarcinoPred-EL [15], CapsCarcino [16] focus on a single property such as the prediction of permeability coefficient and carcinogenic compounds. Overall, the molecular representations underlying these models include various molecular and physicochemical descriptors such as fingerprints, graph signatures, and other 2D/3D indices [17, 18]. Among these, fingerprint representations which are seen as an alternative to descriptors for QSPR studies, have been quite popular given their ease of computation and predictive value.

A number of fingerprints ranging from substructure/path to feature-class/circular have been proposed many of which are used in similarity searching [19, 20]. For ADMET studies however, the fingerprints studied so far have largely been restricted to a select few. In this study, we have evaluated the predictive efficacy of 20 different fingerprints ranging from substructure and extended/functional connectivity fingerprints to various path based encodings (depth-first search, shortest path, local path environments) [21]. The fingerprint-based regression/classification models were calculated for over 50 ADMET and ADMET-related endpoints (using data collated from various literature sources) and is to our knowledge one of the most comprehensive compilations analysed. For a majority of the endpoints, the prediction results were found to be comparable with more sophisticated descriptor formulations. Although the pharmacophore fingerprints yielded consistently poor results, others such as the PUBCHEM, MACCS and ECFP/FCFP encodings were found to yield the best results for most properties. The models and related software have been bundled into a downloadable package and is released under the GNU license.

Approach

Molecular representation

In this study, we have examined 20 different fingerprints (see Table 1) that are routinely used as similarity search tools in drug discovery. The ECFP- and FCFP-class fingerprints are circular topological fingerprints, where the former focuses on the atom properties (e.g. atomic number, charge, hydrogen count), whereas in the functional connectivity FPs, the emphasis is on properties that relate to ligand binding (e.g. hydrogen donor/acceptor, polarity, aromaticity). MACCS and PUBCHEM fingerprints are substructure fingerprints that cover a wide range of features such as element counts and ring systems, atom pairing, or atom environment etc. Other fingerprints include path based fingerprints such as the depth-first search fingerprints (DFS), all-shortest path encoding (ASP), radial fingerprints (Molprint2D), topological atom pairs (AP2D) and triplets (AT2D), pharmacophore pair and triplet encodings as well as local path environments [21]. Fingerprint calculations were performed using in-house code written in Java and makes use of the Chemistry Development Kit library [22]. The software merges existing fingerprints in the library with those calculated by the software jCompoundMapper [21].

Table 1 Fingerprints used in this study to model different ADMET related properties
Full size table

Data curation

Data for different endpoints were collected from previously published articles and databases with a primary source being the Online Chemical Database (OCHEM) [3]. The molecules were subsequently cleaned and duplicates (where present) were removed. Tables 2 and 3 lists the various endpoints and associated data sources considered in this study. Brief descriptions of the endpoints and the results from previous modelling efforts are provided in Additional file 1. Since, early identification of severe toxicity is a key requirement for the safety evaluation of drug candidates, we have evaluated a number of toxicity models covering a range of endpoints such as cardiac, hepatotoxicity, endocrine, urinary tract, carcinogenicity and cytotoxicity. While a majority of the models are binary classification models, for some endpoints such the metabolic intrinsic clearance, acute oral toxicity in rats, plasma protein binding and elimination half-life, multiclass models are proposed.

Table 2 Summary of the ADMET endpoints studied
Full size table
Table 3 Summary of the ADMET and other endpoints for which fingerprint-based regression models were evaluated
Full size table

For other endpoints, regression models have been evaluated (see Table 3). These include the CACO-2 permeability which is commonly used to predict the absorption of orally administered drugs and other xenobiotics, the fraction of unbound drug in plasma, the liver microsomal clearance (typically used to predict hepatic clearance in humans), in vitro human skin permeability and the cancer potency. Models for other ADMET-related properties have also been studied. For instance, properties such as the dissociation constant (\(\text {pK}_a\)) affect solubility (\(\log\) S), permeability, distribution coefficient (\(\log\) D) and oral absorption. These in turn along with other properties such as the human serum albumin (HSA) binding impact pharmacokinetic behaviour and drug bioavailability.

Modelling

In order to build the models, the Random Forest algorithm [23] was chosen which is an ensemble learning method for both classification and regression. The algorithm makes use of bagging and feature randomness to build multiple decision trees (each trained on a random subset of data) and merges them together. The models were trained using the ranger [24] library in the statistical computing environment R [25]. The number of trees used to compute the final average predicted value was set to 500. For each endpoint, the data was split randomly into separate training (80%) and test (20%) sets. A fivefold cross-validation was used to identify the best performing model. In order to rule out any selection bias, we repeated random splitting 3 times and the results were averaged to gain an understanding of the variability. Furthermore, y-randomization tests were conducted to assess the robustness of the final model. To address the problem with unequal distribution of samples between classes, data augmentation of the minority class was carried out using the synthetic minority oversampling technique (SMOTE) [26].

For regression models, the performance was assessed using the squared regression coefficient (\(R^2\)) for the correlation between experimental and predicted values. the root mean squared error (RMSE) and the mean absolute error (MAE). For classification models, metrics that are sensitive to the class imbalance have been used. These include the balanced accuracy (BACC) given by:

$$BACC = \frac{1}{m} \sum _i^m \frac{k_i}{n_i}$$
(1)

where \(k_i\) is the number of correct predictions in class i, m is the number of classes and \(n_i\) is the number of examples in class i. In addition, other metrics such as the overall accuracy, the sensitivity (the true positive rate—TPR) and specificity (the true negative rate—TNR) and the area under the curve (AUC) are also reported (see Additional file 1).

Every model has a finite applicability domain (AD) within which its predictions can be trusted. For regression models, we quantify the prediction intervals (95%) using the quantile regression forests approach [27]. Here, a shorter prediction interval indicates the higher stability of prediction. In the case of classification, two values: confidence and credibility are associated with the predicted label based on the conformal prediction framework [28, 29]. While the confidence provides a measure of how likely a prediction is compared to all other possible classifications, the credibility measure (equal to the highest p-value of any one of the possible classifications being the true label) provides an indication of how good the training set is for classifying the given example.

Results and discussion

For the various endpoints, the relevant performance metrics associated with the best fingerprint-based models are summarized in Tables 4 (for classification models) and 5 (for regression models). The complete performance summary for the training and validations sets is listed in Additional file 1: Tables S1 and S2. For all cases, permutation tests confirmed (p-values < 0.001) that the probability that the model was obtained by chance is quite low. Overall, high classification accuracies (\(BACC > 0.80\)) are obtained for the blood brain barrier permeability, plasma protein binding, CYP450 inhibition (3A4/2C19/1A2/2C9/2C8 isoforms), human intestinal absorption, breast cancer resistance protein inhibition, p-glycoprotein inhibitor/substrate and hemolytic/respiratory toxicity. For some of the other endpoints such as the mitochondrial/urinary tract toxicity, human liver microsomal stability, metabolic intrinsic clearance, AMES mutagenecity, cytotoxicity (multiple cell lines), hERG cardiotoxicity/liability, drug induced liver injury, myelotoxicity, phospholipidosis, rhabdomyolysis, OATP1B1/OATP1B3 inhibition, BSEP and OCT2 inhibition, moderate (\(BACC = 0.71 \;\text{to}\; -0.78\)) performances were observed. Properties such as skin sensitization, acute oral toxicity, phototoxicity in humans, ototoxicity, choleostasis, hepatic steatosis, and carcinogenecity yielded somewhat average results. In the case of regression models, performances were largely on the poorer side with the exception of \(\text {pK}_a\), \(\log\) S, \(\log\) D, human serum albumin and skin penetration, \(R^2_{cv} > 0.70\).

Table 4 Performance metrics for the best performing fingerprint-based classification models
Full size table
Table 5 Performance metrics for the best performing fingerprint-based regression models
Full size table

To identify which of the fingerprints perform well on the different datasets, we plotted heatmaps (see Figs. 1 and 2) of the balanced accuracies (for classification models) and squared correlations (in the case of regression) obtained for the different endpoints. While the pharmacaphore fingerprints (2PPHAR/3PPHAR) perform poorly on all datasets, fingerprints based on substructure keys (PUBCHEM, MACCS, KR) show moderate to high accuracies for a majority of the modelled endpoints. Although the performances for regression models are somewhat less encouraging, here too the \(R^2_{cv}\) for PUBCHEM, ECFP4, and ASP fingerprints yield better models than the other fingerprints tested.

Fig. 1
figure 1

Heatmap showing the cross-validated balanced accuracies (average of 3 independent runs) achieved by different fingerprint-based models for the endpoints studied

Full size image
Fig. 2
figure 2

Heatmap showing the cross-validated correlation coefficients (average of 3 independent runs) achieved by different fingerprint-based models for the endpoints studied

Full size image

We further compared the performances achieved by the fingerprint models with those obtained for the 2D/3D descriptor based approaches. The barplots in Fig. 3 compare the accuracies achieved by the fingerprint models with values reported by the models published earlier. While results for most properties are comparable, for some endpoints such as myelotoxicity, ototoxicity, myopathy accuracies obtained using 2D/3D descriptors are only marginally better. Indeed better results are obtained for rhabdomyolysis, phospholipidosis, phototoxicity with other descriptor based models. For phototoxicity in particular, quantum chemistry-based 3D descriptors are used which can add to the time taken. It must however be pointed out that some of the better performing models take advantage of deep learning. Attempts to improve results for selected properties were carried out using support vector machines. However, the models were not always found to improve on the random forest approach.

Fig. 3
figure 3

Comparison of the accuracies achieved by the fingerprint based models in this study (“Current”) with those created using standard molecular graph based descriptors (“Original”) published in the literature. For OATP inhibtion, descriptors consist of constitutional, geometrical, electrostatic, and physicochemical indices. For phototoxicity, descriptors contain HOMO-LUMO gaps, spectral integrals, ionization potential, electron affinity and CATS descriptors. For properties such as toxic myopathy and MATE1 inhibition, the values compared are the accuracies and AUCs respectively

Full size image

For the regression models calculated for selected properties: \(\text {pK}_a\), \(\log\)S, \(\log\)D, skin penetration, human serum albumin, MDCK permeability \(\text {HD}_{{50}}\), we assessed the prediction reliability based on the prediction intervals. Plots of the prediction intervals with respect to the observed response values for the test sets (see Additional file 1: Figure S1) showed that most of the samples lie within the 95% prediction interval which indicates that the constructed prediction intervals are reliable. For classification models, we focused on excluding compounds whose labels are predicted with low confidence and credibility. Thus, different thresholds for p-values (0.5, 0.6, 0.7, 0.8, 0.9) were applied and the corresponding fraction of molecules that would be withheld from further testing was recorded. A plot of the overall error rates and the percentage reduction in compounds excluded from further processing (see Additional file 1: Figure S2) shows that for many of the endpoints modelled, the predictive performance is not significantly impacted even at cutoffs of 0.50. Such a strategy that allows for compound selection based on static thresholds for the confidence/credibility offer a way to reduce the number of compounds that typically undergo experimental testing.

Software usage

FP-ADMET is available as open access software (GNU GPL v3.0) and can be downloaded from https://gitlab.com/vishsoft/fpadmet. Use of FP-ADMET proceeds in two steps (i) fingerprint calculation followed by (ii) predicting the ADMET endpoint of interest. The software is command line driven and is governed by a shell script (runadmet.sh) that can be run as:

bash runadmet.sh -f molecule.smi -p ## -a

The input to the script is a file (molecule.smi) containing SMILES strings. The ## is a number between 1 (predict Anticommensal Effect) and 56 (predict skin penetration) and corresponds to the prediction task. The results are written to a text file where each line contains molecule name and the predicted response. The “-a” option allows for the calculation of prediction intervals (in the case of regression) and confidence (for classification). For classification, conformal prediction is used to calculate a confidence (how certain the model is that the prediction is a singleton) and a credibility. For example, predicting AMES mutagenecity (task number 4) for a series of molecules produces the following results (see Table 6). The label “inactive” for compound G00001 suggests that the compound is predicted to be non-mutagenic. A confidence value of 0.95 suggests that the classifier is quite certain that the prediction is likely to be a single label. A relatively low value of credibility (0.57) suggests that the compounds like G00001 are not sufficiently represented in the training set and that the user needs to treat the prediction with caution. In the case of regression, a 95% prediction interval (predictions at the 0.025 and 97.5 percentiles for \(pK_a\)) is calculated and provides a range for the predictions on an individual observation. Narrow prediction intervals indicate a lower uncertainty associated with the prediction.

Table 6 Example showing the property (\(pK_a\) and anticommensal effect) predictions and associated uncertainties for 3 molecules
Full size table

Conclusion

In this article, we have evaluated the performance of various molecular fingerprints for predicting a number of ADMET and ADMET-related endpoints. A total of 1500 models were analysed spanning 75 responses and 20 fingerprints. The results show that the machine learning performance using the different fingerprint encodings rival those of traditional descriptor-based methods. Future work will focus on combining different data sets in a multitask modeling approach which has been shown to yield statistically superior results compared with single-task models [12, 30]. In order to facilitate ADMET evaluation, the best performing models have been compiled into an open access software package called FPADMET that can be downloaded from https://gitlab.com/vishsoft/fpadmet.