ATTNSOM: Learning Cross-Isoform Attention for Cytochrome P450 Site-of-Metabolism

TL;DR

ATTNSOM addresses the challenge of predicting CYP-mediated site-of-metabolism by explicitly modeling cross-isoform patterns. It integrates a shared graph encoder for intrinsic reactivity, FiLM-based molecule-conditioned atom features, and a cross-isoform attention mechanism that aligns atom-level predictions with CYP isoform relationships. The approach achieves state-of-the-art MCC across isoforms and strong Top-$k$ performance on Zaretzki and AZ-ExactSOM datasets, while offering interpretable cross-isoform attention patterns and generalization to unseen compounds. This cross-isoform framework enhances discrimination of true metabolic sites and supports practical drug design by improving identification and prioritization of SOMs.

Abstract

Identifying metabolic sites where cytochrome P450 enzymes metabolize small-molecule drugs is essential for drug discovery. Although existing computational approaches have been proposed for site-of-metabolism prediction, they typically ignore cytochrome P450 isoform identity or model isoforms independently, thereby failing to fully capture inherent cross-isoform metabolic patterns. In addition, prior evaluations often rely on top-k metrics, where false positive atoms may be included among the top predictions, underscoring the need for complementary metrics that more directly assess binary atom-level discrimination under severe class imbalance. We propose ATTNSOM, an atom-level site-of-metabolism prediction framework that integrates intrinsic molecular reactivity with cross-isoform relationships. The model combines a shared graph encoder, molecule-conditioned atom representations, and a cross-attention mechanism to capture correlated metabolic patterns across cytochrome P450 isoforms. The model is evaluated on two benchmark datasets annotated with site-of-metabolism labels at atom resolution. Across these benchmarks, the model achieves consistently strong top-k performance across multiple cytochrome P450 isoforms. Relative to ablated variants, the model yields higher Matthews correlation coefficient, indicating improved discrimination of true metabolic sites. These results support the importance of explicitly modeling cross-isoform relationships for site-of-metabolism prediction. The code and datasets are available at https://github.com/dmis-lab/ATTNSOM.

Figures (5)

• Figure 1: Hierarchical clustering and heatmap of pairwise SOM pattern similarity across CYP isoforms in the Zaretzki dataset. The dendrogram illustrates the relational structure among isoforms based on shared metabolic patterns.
• Figure 2: Overview of the ATTNSOM architecture for atom-level SOM prediction across CYP isoforms. The framework combines a shared graph-based molecular encoder with molecule-conditioned atom representations and a cross-attention module that models inter-isoform metabolic relationships. By integrating atom-level features with relational information across CYP isoforms, ATTNSOM captures both shared and distinct metabolic patterns.
• Figure 3: SOM predictions for ivacaftor illustrating atom-level localization performance across different methods. Experimentally annotated SOM atoms are shown in the left panel (GT), followed by predictions from ATTNSOM, XenoSite, SMARTCyp, and FAME 3. ATTNSOM accurately localizes all experimentally validated SOM positions, demonstrating precise atom-level discrimination and strong generalization.
• Figure 4: Attention maps and SOM predictions for a single molecule across five CYP isoforms. ATTNSOM captures shared metabolic patterns by distributing attention for atoms 20 and 21 across related CYPs, while producing sharp isoform-specific attention for CYP2D6 at atom 2. Conservative co-prediction of atoms 20/21 for CYP2D6 reflects a trade-off of cross-isoform relational modeling.
• Figure 5: Hierarchical clustering and heatmap of pairwise SOM pattern similarity predicted by ATTNSOM across CYP isoforms.