eToxPred: a machine learning-based approach to estimate the toxicity of drug candidates

“DeepTox was reported to outperform deep neural networks (DNNs) [31], support vector machines (SVMs) [32], random forests (RF) [33], and elastic nets [34].”

“VS performed with high-performance computing machines renders docking programs such as AutoDock Vina [19], rDock [20], Glide [21], and FlexX [22] capable to search through millions of compounds in a matter of days or even hours.”

“DeepTox was reported to outperform deep neural networks (DNNs) [31], support vector machines (SVMs) [32], random forests (RF) [33], and elastic nets [34].”

“The training process involves the optimization of the weight vector through Gibbs sampling [38].”

“In this work, the SAscore predictor is evaluated with a 5-fold cross-validation protocol, which was empirically demonstrated to be sufficient for most applications [64].”

“Input data to machine learning models are 1024-bit Day-light fingerprints constructed for dataset compounds with Open Babel [61].”

“VS performed with high-performance computing machines renders docking programs such as AutoDock Vina [19], rDock [20], Glide [21], and FlexX [22] capable to search through millions of compounds in a matter of days or even hours.”

“Finally, the Extremely Randomized Trees, or Extra Trees (ET), algorithm [43] is used to predict the toxicity of drug candidates (Fig.”

“The presence of fluorine atoms in pharmaceuticals increases their bioavailability by modulating p K a and lipophilicity, as well as by improving their absorption and partitioning into membranes [82].”

“It is noteworthy that DBNs are first effective deep learning algorithms capable of extracting a deep hierarchical representation of the training data [41].”

“Similarly, many well established and commercially available drugs contain a piperazine ring as part of their molecular structures [77].”

“In addition, we assess the overall quality of a binary classifier with the Matthews correlation coefficient (MCC) [65] and the Receiver Operating Characteristic (ROC) analysis.”

“As a result, an estimated one-third of the top-performing drugs currently on the market contain fluorine atoms in their structure [81].”

“This architecture allows for a fast, layer-by-layer training, during which the contrast-ive divergence algorithm [40] is employed to learn a layer of features from the visible units starting from the lowest visible layer.”

“FDA-approved drugs were obtained from the DrugBank database, a widely used cheminformatics resource providing comprehensive information on known drugs and their molecular targets [51].”

“To develop a new drug requires 6 – 12 years and costs as much as $2.6 billion [1, 2].”

“Although, a breakdown of the overall capitalized costs shows that the clinical period costing $1.5 billion is economically the most critical factor, the expenditures of the pre-human phase aggregate to $1.1 billion [1].”

“Nitrogen-bearing heterocycles, piperidine and pi-perazine, are of central importance to medicinal chemistry [71].”

“In addition to natural products, we compiled a non-redundant set of mostly synthetic bioactive compounds from the Data-base of Useful Decoys, Extended (DUD-E) database [50] by selecting 17,499 active molecules against 101 pharmacologically relevant targets.”

“The KEGG-Drug resource contains drugs approved in Japan, United States, and Europe, annotated with the information on their targets, metabolizing enzymes, and molecular interactions [52].”

“CADD strategies include ligand-and structure-based drug design, lead optimization, and the comprehensive evaluation of absorption, distribution, metabolism, excretion, and toxicity (ADMET) parameters [4].”

“…example is chemTox (http://www.cyprotex.com/ insilico/physiological_modelling/chemtox) predicting key toxicity parameters, the Ames mutagenicity [27] and the median lethal dose (LD 50 ) following intravenous and oral administration, as well as the aqueous solubility. chemTox employs molecular…”

“…SAscore values are computed with an exact approach that combines the fragment-based score representing the “ historical synthetic knowledge ” with the complexity-based score penalizing the presence of ring systems, such as spiro and fused rings, multiple stereo centers, and macrocycles [62].”

“In addition, we tested two other classifiers, the Linear Discriminant Analysis (LDA) [67] and Multi-layer Perceptron (MLP) [68], however, their performance is generally not as high as those of RF and ET.”

“For instance, a blockage of the human Ether-à--go-go-Related Gene (hERG) potassium ion channels by a surprisingly diverse group of drugs can induce lethal cardiac arrhythmia [25].”

“A low metabolic stability due to cytochrome P450-mediated oxidation can be mitigated by blocking metabolically unstable hydrogen positions with fluorine atoms [84], as exemplified by drug structures shown in Fig.”

“…donepezil (Drug-Bank-ID: DB00843), a reversible acetyl cholinesterase inhibitor to treat Alzheimer ’ s disease [74], an opioid analgesic drug remifentanil (DrugBank-ID: DB00899) [75], and dipyridamole (DrugBank-ID: DB00975), a phosphodiesterase inhibitor preventing the blood clot formation [76].”

“Foremost SBDD methods include molecular docking [15], molecular dynamics [16], receptor-based VS [17], and the de novo design of active compounds [18].”

“DeepTox [30] was the best performing methods in the Tox21 Data Challenge winning the grand challenge, the nuclear receptor panel, the stress response panel, and six single assays.”

“Although the combination of pharmacophore modeling, QSAR, and VS techniques has been demonstrated to be valuable in the absence of the protein structure data [12, 13], the three-dimensional (3D) information on the target protein allows employing structure-based drug design (SBDD) [14] in CADD.”

“Foremost SBDD methods include molecular docking [15], molecular dynamics [16], receptor-based VS [17], and the de novo design of active compounds [18].”

“1C) [39].”

“VS employing features provided by pharmacophore modeling [6] and quantitative structure-activity relationship (QSAR) analysis [7] can be performed in order to identify potentially active compounds.”

“As an independent set, we employ the Traditional Chinese Medicine (TCM) Database@Taiwan, currently the largest and most comprehensive small molecule database on traditional Chinese medicine for virtual screening [55].”

“A similar method, ProTox, predicts rodent oral toxicity based on the analysis of toxic fragments present in compounds with known LD 50 values [28].”

“Other techniques to predict toxicity utilize various features such as fingerprints, physicochemical properties, and pharmaco-phore models to build predictive dose-and time-response models [29].”

“VS performed with high-performance computing machines renders docking programs such as AutoDock Vina [19], rDock [20], Glide [21], and FlexX [22] capable to search through millions of compounds in a matter of days or even hours.”

“…to treat Parkinson ’ s disease [73], donepezil (Drug-Bank-ID: DB00843), a reversible acetyl cholinesterase inhibitor to treat Alzheimer ’ s disease [74], an opioid analgesic drug remifentanil (DrugBank-ID: DB00899) [75], and dipyridamole (DrugBank-ID: DB00975), a phosphodiesterase inhibitor…”

“Ligand-based drug design (LBDD) leverages the spatial information and physicochemical features extracted from known bioactives against a given target protein to design and optimize new compounds for the same target [5].”

“Another example is cholesterol inhibitor ezetimibe (DrugBank-ID: DB00973), in which two metabolically labile sites are effectively blocked by fluorine substituents [86].”

“VS employing features provided by pharmacophore modeling [6] and quantitative structure-activity relationship (QSAR) analysis [7] can be performed in order to identify potentially active compounds.”

“Captopril, an angiotensin-converting enzyme inhibitor, was one of the first success stories of LBDD, which was considered a revo-lutionary concept in 1970s compared to conventional methods [11].”

“A recently developed method, Pred-hERG, estimates the cardiac toxicity with a set of features based on statistically significant and externally predictive QSAR models of the hERG blockage [26].”

“activities exhibited by piperazine derivatives make them attractive leads to develop new antidepressant, anticancer, anthelmintic, antibacterial, antifungal, antimalarial, and anticonvulsant therapeutics [78].”

“Re-moving the redundancy at a Tanimoto coefficient (TC) [48] of 0.8 with the SUBSET [49] program resulted in 1008 NuBBE and 81,372 UNPD molecules.”

“These results are in line with benchmarking data reported for other classifiers; for instance, e ToxPred compares favorably with different methods particularly against the AO and ED datasets [30, 70].”

“The CD set includes 350 toxic compounds with an IC 50 of 1 μ M [58].”

“NuBBE is a virtual database of natural products and derivatives from the Brazilian biodiversity [46], whereas UNPD is a general resource of natural products created primarily for virtual screening and network pharmacology [47].”

“Endocrine disrupting chemicals interfere with the normal functions of endogenous hormones causing metabolic and reproductive disorders, the dysfunction of neuronal and immune systems, and cancer growth [59].”

“T3DB houses detailed toxicity data in terms of chemical properties, molecular and cellular interactions, and medical information, for a number of pollutants, pesticides, drugs, and food toxins [54].”

“Foremost SBDD methods include molecular docking [15], molecular dynamics [16], receptor-based VS [17], and the de novo design of active compounds [18].”

“9B, are antipsychotic fluphenazine (DrugBank-ID: DB00623), antiretroviral delavirdine (DrugBank-ID: DB00705), antihistamine meclizine (DrugBank-ID: DB00737), and flibanserin (DrugBank-ID: DB04908) to treat hypoactive sexual desire disorder among pre-menopausal women [79].”

“Numerous machine learning-based techniques have been developed to reveal complex relations between chemical entities and their biological targets [35].”

“Although the combination of pharmacophore modeling, QSAR, and VS techniques has been demonstrated to be valuable in the absence of the protein structure data [12, 13], the three-dimensional (3D) information on the target protein allows employing structure-based drug design (SBDD) [14] in CADD.”

“Computer-aided drug discovery (CADD) holds a significant promise to reduce the costs and speed up the development of lead candidates at the outset of drug discovery [3].”

“Although the capabilities of the traditional LBDD to discover new classes of leads may be limited, recent advances in generating targeted virtual chemical libraries by combinatorial chemistry methods considerably extend the application of LBDD methods [8 – 10].”

“VS performed with high-performance computing machines renders docking programs such as AutoDock Vina [19], rDock [20], Glide [21], and FlexX [22] capable to search through millions of compounds in a matter of days or even hours.”

“Although the combination of pharmacophore modeling, QSAR, and VS techniques has been demonstrated to be valuable in the absence of the protein structure data [12, 13], the three-dimensional (3D) information on the target protein allows employing structure-based drug design (SBDD) [14] in CADD.”

“Among 12,612 molecules with LD 50 data provided by the SuperToxic database [60], 7392 compounds are labeled as toxic with a LD 50 of 500 mg kg − 1 .”

“…application, this technique can be incorporated into high-throughput pipelines constructing custom libraries for virtual screening, such as that based on e MolFrag [9] and e Synth [10], to eliminate from CADD those drug candidates that are potentially toxic or would be difficult to synthesize.”

“DeepTox was reported to outperform deep neural networks (DNNs) [31], support vector machines (SVMs) [32], random forests (RF) [33], and elastic nets [34].”

“Further, fluorination helps stabilize the binding of a drug to a protein pocket by creating additional favorable interactions, as it was suggested for the fluorophe-nyl ring of paroxetine (DrugBank-ID: DB00715) [83], a selective serotonin reuptake inhibitor shown in Fig.”

“1) fluorine substitution [80].”

“…application, this technique can be incorporated into high-throughput pipelines constructing custom libraries for virtual screening, such as that based on e MolFrag [9] and e Synth [10], to eliminate from CADD those drug candidates that are potentially toxic or would be difficult to synthesize.”

“9A, trihexyphenidyl (DrugBank-ID: DB00376), a muscarinic antagonist to treat Parkinson ’ s disease [73], donepezil (Drug-Bank-ID: DB00843), a reversible acetyl cholinesterase inhibitor to treat Alzheimer ’ s disease [74], an opioid analgesic drug remifentanil (DrugBank-ID: DB00899) [75], and…”

“Finally, replacing the chlorine atom with a fluorine improves safety profile and pharmacokinetic properties of prasugrel (DrugBank-ID: DB06209) compared to other thienopyridine antiplatelet drugs, ticlopidine and clopidogrel [87].”

“Indeed, a targeted fluorination of a nonsteroidal anti-inflammatory drug flurbiprofen (DrugBank-ID: DB00712) helped prolong its metabolic half-life [85].”

“The cardiotoxicity (CD) dataset contains 1571 molecules characterized with bioassay against human ether-a-go-go related gene (hERG) potassium channel. hERG channel blockade induces lethal arrhythmia causing a life-threatening symptom [57].”

“…prediction software and web-based tools such as Tox21 (https:// ntp. niehs. nih. gov/ whatw estudy/ tox21/ index. html) [327], SEA (http:// sea. bkslab. org/) [328], eToxPred (https:// www. bryli nski. org/ etoxp red-0) [329], and TargeTox (https:// github. com/ artem-lysen ko/ Targe Tox) [330].”