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

🚀 Accepted at ISMB 2026

📄 Paper: ATTNSOM: learning cross-isoform attention for cytochrome P450 site-of-metabolism prediction — Bioinformatics 42(Supplement_1), 2026.

Reference implementation of ATTNSOM, an isoform-aware framework for atom-level site-of-metabolism (SoM) prediction in cytochrome P450–mediated drug metabolism.

Most SoM predictors either ignore CYP isoform identity or model isoforms independently. Human CYP isoforms have broad, partially overlapping substrate specificity, so their SoM patterns are strongly correlated (pairwise Jaccard similarity of annotated SoMs is high across all isoform pairs in the Zaretzki dataset). ATTNSOM exploits that structure explicitly.

1. Method
2. Environment Setup
3. Datasets
4. Training and Evaluation
5. Ablations
6. Inference on New Molecules
7. Cross-Isoform Similarity Analysis
8. Repository Layout
9. Citation

Given a molecular graph and a target CYP isoform, ATTNSOM predicts a metabolic likelihood for every atom. Three components (model.py):

The prediction head consumes [n'_i ‖ n_i^attn ‖ c_t] for target isoform t.

Training uses L = λ_main · L_main + λ_attn · L_attn (train.py):

• L_main — Focal loss (γ = 1) on atom-wise logits, for the severe class imbalance.
• L_attn — masked soft-label BCE that aligns the atom→isoform attention with the experimentally annotated SoMs. Probability mass 1/|P_i| is spread over the isoforms for which atom i is annotated positive, and a binary mask excludes every (molecule, isoform) pair with no experimental annotation. Pairs held out in the validation/test split are masked out of the training targets as well (apply_no_leakage_to_dataloaders), so the auxiliary supervision never leaks test annotations.

Default hyperparameters (as reported in the paper): hidden size 256, 4 message-passing layers, batch size 32, up to 50 epochs, AdamW with lr = 1e-4 and weight decay = 1e-4, Focal γ = 1, λ_attn = 0.5, 10-fold CV with 5% of each training split held out for validation. Checkpoints are selected by lowest validation loss.

1. Install Miniconda or Anaconda. Make sure your GPU driver and CUDA version are compatible with the versions pinned in envs/attnsom.yaml (PyTorch 2.4.0 / CUDA 11.8).

2. Create the environment:

   cd envs/
   conda env create -f attnsom.yaml
   conda activate attnsom
   pip install https://data.pyg.org/whl/torch-2.4.0%2Bcu118/torch_scatter-2.1.2%2Bpt24cu118-cp310-cp310-linux_x86_64.whl
   cd ..

   PyTorch ≥ 2.4 is required (torch.nn.RMSNorm).

Both benchmarks ship with the repository under cyp_dataset/.

679 substrates with atom-resolved SoM annotations for nine human CYP isoforms, giving 2,003 molecule–isoform instances. Labels are read from the PRIMARY_SOM SD property (1-based atom indices).

120 public compounds with high-confidence, atom-resolved SoMs from human hepatocyte assays (Chen et al., 2025). Only exact site annotations are used; extended-group annotations are discarded. Atom indices in this file are 0-based. The dataset carries no isoform labels — annotations are aggregate hepatocyte outcomes — so it is modelled as a single isoform.

10-fold cross-validation on the Zaretzki dataset (the setting of Tables 2 and 3):

python main.py --result_dir results/attnsom --save_ckpt --ckpt_dir checkpoint/attnsom

AZ-ExactSOM (Table 4):

python main.py --dataset_dir cyp_dataset/az --result_dir results/az

Useful flags:

Outputs written to --result_dir:

summary.json                      aggregated metrics, overall and per isoform
config.json                       model configuration (consumed by inference.py)
all_folds_predictions.{json,npz}  atom-level predictions over all folds
all_molecules_predictions.json    molecule-level predictions
folds/                            per-fold metrics and predictions

Reported metrics are Top-1/2/3 accuracy, precision, recall, F1, per-atom accuracy, molecule-level exact match, and MCC (the balanced criterion emphasised in the paper).

Every row of Table 3 is a flag combination:

All of them at once:

bash run_ablation.sh

Predictions are averaged over the fold checkpoints, matching the ensembling used for the case studies in the paper.

python inference.py \
  --ckpt_dir checkpoint/attnsom \
  --smiles "CC(C)(C)c1cc(NC(=O)c2c[nH]c3ccccc3c2=O)cc(C(C)(C)C)c1O" \
  --cyp 3A4 --save_attention --output_dir inference_results

--smiles_file accepts one SMILES per line. Outputs are predictions.json (per-atom probabilities, predicted SoMs and top-3 atoms per isoform), predictions.csv (flat, one row per atom × isoform) and, with --save_attention, the atom×isoform attention map as .npy plus a heatmap.

The example above is ivacaftor: ATTNSOM ranks the tert-butyl methyl carbons highest for CYP3A4, the position whose deuteration produced deutivacaftor.

Reproduces the SoM-pattern similarity heatmaps and dendrograms (Figures 1 and 5), plus the cophenetic correlation between them:

python analysis.py --source both --result_dir results/attnsom --output_dir figures

--source dataset uses the annotated SoMs, --source predicted uses the cross-validated predictions in --result_dir. Similarity is the mean Jaccard overlap of SoM atom sets over the molecules annotated (or predicted) for both isoforms of a pair.

main.py            cross-validated training and evaluation
model.py           ATTNSOM and the encoder/ablation variants
train.py           losses, training loop, evaluation loop
dataset.py         Zaretzki SDF and AZ-ExactSOM CSV loaders
dataset_utils.py   atom/bond featurisation, graph construction, stratified splits
metrics.py         Top-k, MCC, precision/recall/F1, exact match
inference.py       fold-ensemble prediction for arbitrary SMILES
analysis.py        cross-isoform SoM similarity (Figures 1 and 5)
run_ablation.sh    all Table 3 configurations
cyp_dataset/       Zaretzki SDFs and AZ-ExactSOM CSV
envs/attnsom.yaml  conda environment

DOI: 10.1093/bioinformatics/btag314

@article{kim2026attnsom,
  title={ATTNSOM: learning cross-isoform attention for cytochrome P450 site-of-metabolism prediction},
  author={Kim, Hajung and Lee, Eunha and Chung, Sohyun and Park, Jueon and Baek, Seungheun and Kang, Jaewoo},
  journal={Bioinformatics},
  volume={42},
  number={Supplement\_1},
  pages={btag314},
  year={2026},
  publisher={Oxford University Press}
}

Contact: kangj@korea.ac.kr

Released under the MIT License; see LICENSE.