Table 1. Overview of Methods for Predicting SOMs, Structures of Metabolites, and Interactions with Metabolizing Enzymes.
| Methods for predicting SOMs | Category | Description | Refs (Examples) |
|---|---|---|---|
| QMBO | Reactivity-based method | Derives likelihood of a metabolic reaction at a certain atom position from its hydrogen abstraction energy based on bond order, employing a DFT wave function. Considers accessibility of hydrogen atoms. | (16) |
| CypScore | Reactivity-based method | Uses AM1-based atom reactivity descriptors to estimate metabolic reactivity of a certain atom position. Six models to describe various generic CYP metabolic reactions. | (17) |
| Metaprint2D | Fingerprint-based data mining method | Derives likelihood of metabolic transformations for atoms with a defined atom environment from data mining of large biotransformation databases. Encodes atom environments using SYBYL266 atom types in combination with circular fingerprints. | (21−24) |
| ADMET Predictor – Metabolite Module | Machine learning method | Derives the likelihood of metabolic reactions to happen at specific atom positions using ANN ensembles. Classification models allow identification of substrates for five CYP isoforms. | (28) |
| ROCS | Shaped-focused method | Uses shape-focused alignment of molecules to known CYP substrates in order to derive a potential geometric orientation to the catalytic heme iron. Atom positions in the proximity of the heme iron are considered potential SOMs. | 37 |
| Classic docking tools (AutoDock, FlexX, GLIDE, GOLD, etc.) | Protein–ligand docking-based method | Evaluate orientation of the ligand to the enzyme catalytic center in order to identify potential SOMs. Atom positions in the proximity of the heme iron are considered potential SOMs. | (51, 52) |
| MetaSite | Combined approach | Uses protein structural information, GRID-derived MIFs of protein and ligand, as well as molecular orbital calculations to estimate the likelihood of a metabolic reaction at a certain atom position. | (40, 54, 55) |
| Combined pharmacophore, homology modeling and quantum chemical approach | Combined approach | Combines a pharmacophore-based approach, homology modeling and molecular orbital calculations to pinpoint potential SOMs. | (60) |
| SMARTCyp | Combined approach | Utilizes a set of precalculated DFT activation energies in combination with topological accessibility descriptors for prognosis of potential SOMs (CYP3A4 and 2D6). | (61, 62) |
| StarDrop | Combined approach | Combines quantum chemical analysis and a ligand-based model of CYP substrates to highlight potential SOMs. Takes into account calculated logP values. | (64) |
| RS-Predictor | Combined approach | Utilizes a set of 148 topological and 392 quantum chemical atom-specific descriptors in combination with a SVM-like ranking and a multiple instance learning method to identify potential SOMs. | (15) |
| Machine learning-based multidescriptor approach | Combined approach | Takes into account quantum chemical, SASA and pharmacophoric descriptors using a random forest/ensemble decision tree approach to identify potential SOMs. | (65) |
| Machine learning-based multidescriptor approach | Combined approach | Employs electrostatic, inductive, energetic, topological, steric, and distance properties in combination with a SVM to predict potential SOMs of endogenous substrates. | (67) |
| Combined quantum chemical/docking/MD approach | Combined approach | Combination of quantum chemical methods with docking to account for reactivity, followed by MD simulations to predict potential SOMs. | (69) |
| MLite | Combined approach | Combines quantum chemistry-derived reactivity estimation with docking. | (73) |
| Ensemble-based/MD-supported docking | Protein–ligand docking-based method | Accounts for protein target flexibility with conformational ensembles of proteins generated using MD simulations and related techniques. | (79−82) |
| IDSite | Combined approach | Combination of an induced fit docking approach (GLIDE, PLOP) with a reactivity model (Jaguar). | (84) |
| Methods for predicting xenobiotic metabolites | Category | Description | References (Examples) |
|---|---|---|---|
| MetabolExpert | Expert system | Uses knowledge database of rules to predict the structures of likely metabolites. Predicts pathways in animals, plants, or through photodegradation. | (88) |
| META | Expert system | Uses a large dictionary of biotransformations to predict the structure of likely metabolites. Analyzes metabolite stability. Predicts pathways in mammals, through aerobic and anaerobic biodegradation. | (89) |
| Meteor | Expert system | Employs a collection of knowledge-based biotransformation rules defined using a dedicated structure representation language to derive the structure of likely metabolites. Considers calculated logP values for predictions. | (90) |
| University of Minnesota Pathway Prediction System (UM-PPS) | Expert system | Utilizes biotransformation rules to predict the structure of likely metabolites. Specific to microbial catabolic metabolism. | (91) |
| SyGMa | Expert system | Predicts structures of likely metabolites based on rules derived from the Accelrys Metabolite Database and assigns probability scores for each metabolite. | (92) |
| TIMES | Expert system | Employs a biotransformation library and a heuristic algorithm to generate metabolic maps. | (93) |
| JChem Metabolizer module | Expert system | Enumerates all possible metabolites of a given compound. Supports species-specific predictions of likely metabolites. | (94) |
| Metaprint2D-React | Fingerprint-based data mining approach | Predicts structures of likely metabolites based on the MetaPrint2D data mining approach. | (21) |
| Machine learning-based multidescriptor approach | Combined approach | See description of this software in Methods of Predicting SOMs. | (67) |
| MetaSite | Combined approach | See description of this software in Methods of Predicting SOMs. | (40, 54, 55) |
| Methods for predicting CYP binding affinity/inhibition by xenobiotics | Category | Description | References (Examples) |
|---|---|---|---|
| Linear interaction energy (LIE) | MD simulation | semi-empirical method for calculating free energy of binding for a ligand using ensemble averaged nonbonded interaction energies. | (156−159) |
| Free energy perturbation (FEP)/thermodynamic integration (TI) | MD simulation | Simulations of unphysical states along a thermodynamic cycle connecting the bound and unbound ligand form for calculating the free energy of binding. | (131, 161−163) |
| Decision tree, k-nearest neighbor, ANN, random forest, SVM, etc. | QSAR and machine learning method: Classification model | Classify compounds for enzyme inhibition. Allow conclusions to be drawn on isoform specificity. | (202−206, 209) |
| isoCyp | QSAR and machine learning method: Classification model | Classifies compounds for CYP3A4, 2D6, and 2C9 inhibition. | (207, 208) |
| PCA, PLS, multiple linear regression, etc. | QSAR and machine learning method: Regression model | Predict enzyme inhibition rates. | (210, 212−216) |
| CoMFA, GRID/GOLPE | 3D-QSAR method: Classification model | Classify compounds for enzyme inhibition. Allow conclusions to be drawn on isoform specificity. | (222, 224) |
| CoMFA, GRID/GOLPE | 3D-QSAR method: Regression model | Predict enzyme inhibition rates. Allow derivation of 3D properties crucial for bioactivity. | (224−229, 231, 232) |
| Pharmacophore models | Pharmacophore-based method | Predict quantitative and qualitative enzyme inhibition. Allow conclusions to be drawn on isoform specificity. | (213, 233, 236, 238−240) |
| Protein–ligand docking | Protein–ligand docking-based method | Predict binding mode and binding affinity. Allow conclusions to be drawn on isoform specificity. | (159, 241, 242) |
| Combined pharmacohore ensemble/SVM approach | Combined approach | Uses an ensemble of pharmacophores to account for protein flexibility. | (243) |
| Proteochemometric analysis supported by GRIND and further physicochemical descriptors | Combined approach | Considers protein sequences of 14 CYPs as well as GRIND and further descriptors for substrates. | (244) |
| Combined machine learning, protein modeling, and docking approach | Combined approach | Uses simulated annealing to render the conformational space of the target protein and docking scores as attributes for subsequent ANN model generation. | (246) |
| VirtualToxLab | Combined approach | Uses flexible docking in combination with a multidimensional QSAR approach to predict ligand interaction with 16 antitargets, including CYP450 1A2, 2A13, 2C9, 2D6, and 3A4. | (247) |
| Methods for predicting CYP induction by xenobiotics | Category | Description | Examples (References) |
|---|---|---|---|
| PLS using VolSurf descriptors | QSAR approach | Uses atPLS and VolSurf descriptors for the development of a QSAR model for PXR/AhR interaction. | PXR: (251); AhR: (251) |
| SVM, k-nearest neighbor, probabilistic neural network | Machine learning approach | Uses various machine learning methods to predict human PXR interaction. | PXR: (252) |
| Pharmacophore modeling | Pharmacophore-based method | Uses pharmacophores to predict PXR/CAR interaction. | PXR: (253−256); CAR: (262) |
| Protein–ligand docking | Protein–ligand docking-based method | Uses protein–ligand docking to predict PXR/AhR/CAR activation, rationalize SARs and gain insight on likely molecular interaction modes. | PXR: (256−258, 263); AhR: 259, (263); CAR: (263) |
| GRID/GOLPE | 3D-QSAR method | Uses MIF-derived 3D-QSAR models to predict CAR interaction. | CAR: (262) |
| VirtualToxLab | Combined approach | Uses flexible docking in combination with a multidimensional QSAR approach to predict ligand interaction with AhR and 15 other targets (see above). | AhR: (247) |