B-XAIC

B-XAIC is a benchmark dataset for evaluating Explainable Artificial Intelligence (XAI) methods for Graph Neural Networks (GNNs) in the molecular domain.

The dataset consists of exactly 50,000 real-world molecules from ChEMBL, covering seven binary molecular structure-detection tasks. Each molecule is associated with ground-truth explanations identifying relevant atoms and bonds responsible for the corresponding labels.

The benchmark enables systematic evaluation of molecular explanations at both node and edge levels.

Dataset Contents

The dataset consists of two main components:

• CSV file: Contains molecular structures (SMILES), ChEMBL identifiers, binary labels, molecular descriptors, and five predefined data splits.
• explanations.sdf: Contains ground-truth molecular explanations, including atom- and bond-level annotations.

Tasks

The dataset covers seven molecular structure-detection tasks:

• Boron atom detection (B).
• Phosphorus atom detection (P).
• Halogen detection (X).
• Indole ring detection.
• PAINS substructure detection.
• Ring count detection.
• Maximum ring size detection.

Data Splits

Five predefined data splits (split_0–split_4) are provided. Each split defines training, validation, and test subsets, supporting reproducible model development and evaluation.

Uses

Intended Use

B-XAIC is intended for research and benchmarking in explainable molecular machine learning, particularly:

• Benchmarking explainability methods for molecular GNNs.
• Evaluating node- and edge-level explanations.
• Assessing explanation faithfulness.
• Developing interpretable molecular machine learning models.
• Research in explainable AI and cheminformatics.

Usage

Hugging Face Datasets

Load the dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("mproszewska/B-XAIC")

print(dataset)

PyTorch Geometric

For loading and processing the dataset in PyTorch graph format, see the official implementation:

https://github.com/mproszewska/B-XAIC

from bxaic import XAIMolecularDataset, TASKS

# Available molecular tasks
print(TASKS) 
# ['B', 'P', 'X', 'indole', 'PAINS', 'rings-count', 'rings-max']

# Load the dataset for a specific task
dataset = XAIMolecularDataset("./data", "PAINS")
print(dataset[0])
# Data(x=[33, 11], edge_index=[2, 68], edge_attr=[68, 1], y=[1], expl_node_mask=[33], expl_edge_mask=[68])

Citation

If you use B-XAIC in your research, please cite the original paper: https://arxiv.org/abs/2505.22252