Site of Metabolism (SOM) Prediction

The foundational input for SOM prediction is the calculation of electronic and steric descriptors for each candidate atom. Key features include:

• Fukui indices: Measures of electrophilic (f-) and nucleophilic (f+) susceptibility derived from frontier molecular orbital theory.
• Atomic partial charges: Computed via methods like Gasteiger-Marsili or DFT-level electrostatic potential (ESP) charges.
• Accessible surface area: The solvent-accessible surface area of an atom, distinguishing buried sites from those sterically available to the enzyme's reactive heme center.
• Bond dissociation energies: The homolytic BDE of C-H bonds, a critical parameter for aliphatic hydroxylation reactions.

State-of-the-art models are not generic metabolism predictors; they are isoform-specific classifiers. The human liver contains multiple CYP450 enzymes with distinct substrate preferences and active site topologies.

• CYP3A4 models: Trained on substrates for the most abundant hepatic enzyme, which has a large, flexible active site.
• CYP2D6 models: Address a polymorphic enzyme with a strict requirement for a basic nitrogen atom at a specific distance from the oxidation site.
• CYP2C9 models: Specialize in substrates containing acidic moieties that interact with the enzyme's arginine recognition pocket.
• Multi-isoform models use the enzyme identity as an explicit input feature.

SOM prediction methodologies bifurcate into two fundamental paradigms:

• Ligand-based models: Rely solely on the chemical features of the substrate molecule. They use QSAR, random forests, or graph neural networks trained on experimentally determined SOM data without requiring the enzyme's 3D structure. These are faster and applicable when the enzyme structure is unknown.
• Structure-based models: Use molecular docking or molecular dynamics to place the substrate in the enzyme's active site. The reactivity distance—the geometric proximity of each atom to the heme iron's oxo-ferryl species—becomes the dominant predictive feature. These models explicitly account for steric constraints imposed by the binding pocket.

The current frontier of SOM prediction is dominated by message-passing neural networks (MPNNs) that operate directly on the molecular graph. Unlike fixed fingerprints, these models learn atom-level representations end-to-end.

• Node features: Each atom is initialized with its atomic number, hybridization, degree, and implicit valence.
• Edge features: Bond type, conjugation, and ring membership are encoded.
• Site prediction head: After message passing, each atom's latent vector is passed through a classifier to produce a probability score for being a metabolic site.
• Models like XenoSite and FAME series have demonstrated that learned representations consistently outperform hand-crafted reactivity descriptors.

Top-performing models augment learned representations with density functional theory (DFT)-derived features to capture the intrinsic chemical reactivity that governs the catalytic mechanism.

• Activation energies: The computed energy barrier for hydrogen atom abstraction by Compound I (the active oxidizing species) is a direct mechanistic predictor.
• Spin population: The unpaired electron distribution on the substrate radical intermediate influences regioselectivity.
• Semi-empirical surrogates: To avoid the computational cost of full DFT, models like SMARTCyp use pre-computed activation energies from the AM1 or PM6 levels of theory, mapped by the reaction type and carbon atom hybridization.

SOM prediction is fundamentally a ranking problem, not a binary classification. A drug molecule typically has multiple experimentally observed metabolites, and the model must reproduce the relative likelihood of each site.

• Top-k accuracy: The standard evaluation metric measures whether the experimentally observed SOM is within the model's top-1, top-2, or top-3 predicted sites.
• Probability calibration: Well-calibrated models output a probability distribution over all heavy atoms, allowing medicinal chemists to assess the relative risk of modification at each position.
• Metabolite cascade prediction: Advanced systems predict not just the primary SOM, but the subsequent secondary and tertiary sites on the first-generation metabolite.

The primary superfamily of heme-containing monooxygenases responsible for Phase I oxidative metabolism. CYP3A4, CYP2D6, and CYP2C9 isoforms alone metabolize over 50% of marketed drugs. SOM prediction models are typically isoform-specific because the active site topology and reactivity of the Compound I oxy-ferryl intermediate vary significantly between isoforms, dictating the geometric constraints for hydrogen atom abstraction.

Key atomic and molecular properties derived from density functional theory (DFT) calculations that serve as input features for SOM models:

• Bond Dissociation Energy (BDE): The energy required for homolytic C-H bond cleavage; lower values indicate higher lability
• Fukui Indices: Measures of electrophilic or nucleophilic susceptibility at each atom
• Atomic Spin Density: The unpaired electron distribution in the radical intermediate
• Electrostatic Potential Maps: Visualize charge distribution influencing substrate docking

SOM prediction fundamentally balances two competing factors:

Intrinsic Reactivity (Lability) The electronic susceptibility of an atom to oxidation, governed by local bond strengths and frontier molecular orbital energies.

Steric Accessibility The geometric constraint imposed by the enzyme's active site architecture. A highly labile site buried in the protein interior may be metabolically inert, while a less reactive but solvent-exposed position undergoes transformation. Modern models integrate both via reactivity-accessibility frameworks.

A deep learning approach to SOM prediction that uses a convolutional neural network operating on atom-centered molecular descriptors. Key innovations include:

• Topological atom fingerprints encoding circular substructural environments
• Atom-type specific networks that learn distinct reactivity patterns for sp2 carbons, sp3 carbons, and heteroatoms
• Training on a curated dataset of over 680 substrates with experimentally confirmed SOMs across nine CYP isoforms

XenoSite demonstrated that learned representations can outperform hand-crafted quantum chemical features for certain substrates.

A collection of machine learning models that predict SOMs using random forest and extreme gradient boosting algorithms. FAME models are trained on:

• Circular fingerprints (ECFP-like) describing the atomic environment
• Physicochemical atom descriptors including partial charges and polarizability
• Topological indices capturing molecular shape and branching

The FAME family includes FAME 2 (atom-based classification) and FAME 3 (site-of-metabolism ranking), with demonstrated performance across diverse chemical space.