Explainable AI Attribution for Drug-Target Interactions with GNNs

A benchmark dataset for four explainable AI attribution methods applied to graph neural network models predicting drug-target interactions. The data was created by Mrinal Mahindran and last updated in April 2026, focusing on kinase and G-protein-coupled receptor targets. It includes validation metrics like atom-level intersection over union and mappings to 3D protein-ligand structures.

Use Cases

• Benchmarking explainable AI methods for graph neural networks based on the described attribution techniques.
• Validating the biological relevance of model attributions by mapping atoms to 3D protein-ligand structures as mentioned in the description.
• Identifying chemically meaningful ligand features in drug-target interactions based on the described consensus attribution analysis.
• Developing more interpretable GNNs for kinase and GPCR targets, as referenced in the study's stated goal.

Strengths

• Includes validation against 3D protein-ligand structures, with up to 76% of consensus-attributed atoms found within 2 Å of the binding pocket for kinase-inhibitor complexes.
• Benchmarks four distinct explainable AI attribution methods on GNN models for two specific target classes (kinases and GPCRs).
• Data is available in a standard CSV format with a file size of 6.0 MB, facilitating accessibility.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

figshare

Collection Method

Likely generated from computational experiments benchmarking XAI methods on trained GNN models.

Freshness

Last updated 2026-04-15 19:07:36; freshness should be verified.