CAMIE: Context-Aware Subgraph Explanations for Multi-Task GNNs

Official implementation of Context-Aware Subgraph Explanations for Multi-Task GNNs using PyTorch.

Welcome to the official implementation codebase for CAMIE, a post-hoc explanation framework for multi-task molecular GNNs.

Unlike context-agnostic explainers that assign a single fixed explanation to a molecule, CAMIE estimates assay-specific motif importance for each molecule--assay pair. The key idea is that, in multi-task molecular prediction, the same molecule may rely on different substructures depending on the queried biological assay. CAMIE captures this by conditioning motif scoring on assay context and distilling assay-specific motif-removal responses from a frozen multi-task predictor.

---

Overview

CAMIE consists of four stages:

1. Frozen multi-task prediction backbone
   A pretrained shared GNN encodes each molecule and produces assay-specific outputs.

2. Motif decomposition and representation extraction
   Each molecule is decomposed into chemically meaningful candidate motifs using BRICS. Graph, motif, and assay-context representations are then constructed.

3. Context-aware motif scoring
   CAMIE learns a shared scorer over (graph, motif, assay) tuples using pseudo targets derived from assay-specific motif-removal responses.

4. Top-k motif explanation
   At inference time, CAMIE ranks motifs for a queried assay and returns the top-k motifs as the explanation.

---

Repository Structure

CAMIE/
├── baselines/                  # Saliency and aggregation baseline models
│   ├── common/                 # Shared utilities, aggregation and metrics
│   └── models/
│       └── gradient/           # Gradient-based saliency (SA) baseline
├── datasets/                   # Preprocessing and dataset classes
│   ├── preprocess_toxcast.py   # Maps biological context and annotations to ToxCast
│   ├── toxcast_dataset.py      # Preprocesses and splits ToxCast deepchem dataset
│   ├── toxcast_graph_dataset.py# PyG InMemoryDataset definition for ToxCast
│   └── utils.py                # Graph helper functions
├── evaluate/                   # Evaluation scripts
│   └── compute_fidelity_f1_single.py # Computes F1-Fidelity under S1/S2 masking
├── models/
│   └── gnn/                    # Backbone GNN definitions and extraction script
│       ├── gnn.py              # PyG GIN/GCN architectures
│       ├── run_toxcast.py      # GNN backbone training & evaluation
│       └── extract_toxcast_emb.py # Extracts GNN graph and node embeddings
├── subgraph/                   # Subgraph & context processing
│   ├── context/
│   │   └── hierarchical_dataset.py # Generates motif-assay tables
│   ├── motifs/
│   │   ├── extract.py          # Decomposes SMILES into motifs (BRICS/Tree)
│   │   └── build_motif_emb.py  # Builds motif embeddings via node pooling
│   └── scoring/                # Joint Scorer implementation
│       ├── scoring_dataset.py  # Prepares pseudo-targets (masking difference)
│       ├── train_mse.py        # Trains the JointMLPScorer model
│       ├── scores_mse.py       # Predicts motif joint scores
│       └── decomposition_mse.py# Hard S1/S2 partition of motifs
├── utils/
│   ├── chemutils.py            # RDKit decomposition helpers
│   └── utils.py                # Generic utilities
│
├──
└── environment.yml             # Conda environment configuration

---

Environment Setup

Create the conda environment from environment.yml:

conda env create -f environment.yml
conda activate SCAR

If your environment name is different, replace SCAR with the actual name defined in environment.yml.

---

Dataset

We use the ToxCast multi-task molecular assay benchmark.

Main statistics

• # molecules: 8,578
• # assays/tasks: 30
• # observed molecule--assay pairs: 59,131
• # motif--assay scoring instances: 413,114

Split statistics

• train pairs: 49,110

• valid pairs: 4,681

• test pairs: 5,340

• train molecules: 6,862

• valid molecules: 858

• test molecules: 858

Additional statistics

• overall positive ratio: 0.116
• pairs per assay (min / median / max): 97 / 502 / 7,934
• assays per molecule (min / median / max): 1 / 3 / 30
• avg. motifs per molecule--assay pair: 6.99
• avg. motifs per molecule: 7.61

We use Bemis--Murcko scaffold splitting with a 7:1:2 ratio and BRICS motif decomposition.

---

Baselines

We compare CAMIE against representative post-hoc explanation baselines.

Optimization-based

• eXEL-group
• eXEL-lasso

Gradient-based

• SA
• GBP
• GradCAM

Perturbation-based

• GNNExplainer
• PGExplainer

For gradient- and perturbation-based baselines, node- or edge-level scores are converted to motif-level scores using a shared aggregation protocol.

---

Execution Pipeline

Step 1. Train the CAMIE scorer

python -m subgraph.scoring.train_mse \
  --scoring_table_dir assets/scoring/scoring_dataset/motif_context_scoring_table_seed0.csv \
  --out_dir assets/scoring/joint_mlp/seed0/joint_ckpt \
  --model_type joint_mlp \
  --seed 0

Step 2. Score motif--assay tuples

python -m subgraph.scoring.scores_mse \
  --scoring_table_dir assets/scoring/scoring_dataset/motif_context_scoring_table_seed0.csv \
  --joint_ckpt_dir assets/scoring/joint_mlp/seed0/joint_ckpt \
  --out_dir assets/scoring/joint_mlp/seed0/scores \
  --model_type joint_mlp \
  --seed 0

Step 3. Decompose into top-k motifs

python -m subgraph.scoring.decomposition_mse \
  --scored_table_dir assets/scoring/joint_mlp/seed0/scores/scored_table_joint_mlp_seed0.csv \
  --score_col score_joint_mlp \
  --rule_name joint_mlp \
  --out_dir assets/scoring/joint_mlp/seed0/decomposition \
  --seed 0

Step 4. Compute Fidelity F1

python -m baselines.evaluate.compute_fidelity_f1_single \
  --decomp_csv assets/scoring/decomposition/ablation/mse/motif_decomposition_table_seed0.csv \
  --out_dir assets/baselines/fidelity_f1_single/seed0/mse \
  --model mse \
  --seed 0

---

Main Outputs

Scored motif table

assets/scoring/joint_mlp/seed0/scores/scored_table_joint_mlp_seed0.csv

Motif decomposition table

assets/scoring/decomposition/ablation/mse/motif_decomposition_table_seed0.csv

Fidelity summaries

assets/baselines/fidelity_f1_single/seed0/mse/compact_fidelity_f1_summary_mse.csv

---

Reproducibility Notes

• All main results are reported over 10 random seeds (0--9).
• Main comparisons use top-k = 2.
• Probabilities are thresholded at 0.5 for F1-based fidelity evaluation.
• Gradient- and perturbation-based baselines use the same frozen multi-task GNN backbone.

---

Citation

If you use this code or our results in your research, please cite our paper:

@inproceedings{yoon2026camie,
  title={Context-Aware Subgraph Explanations for Multi-Task GNNs},
  author={Yoon, Yejin and Lim, Jaesung and Choi, Jinhee and Jeon, Jong-June},
  booktitle={Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM '26)},
  year={2026},
  publisher={ACM},
  doi={10.1145/3799682.3840994}
}

---

Notes

• CAMIE is a post-hoc explainer and does not retrain the original multi-task predictor.
• The pseudo target is derived from the predictor's assay-specific motif-removal response, so CAMIE explanations depend on the behavior of the frozen backbone.
• Explanation resolution is limited by the predefined motif candidates produced by BRICS decomposition.