Explainability Methods from Machine Learning Detect Important Drugs’ Atoms in Drug-Target

16.6 KB of attribution data from four explainable AI methods applied to graph neural network models for drug-target interaction prediction. The dataset, created by Mrinal Mahindran and last updated in April 2026, benchmarks methods on kinase and G-protein-coupled receptor targets, mapping attributed ligand atoms to 3D protein structures. Consensus attributions were highly enriched for atoms contacting the binding pocket, with up to 76% within 2 Å in kinase-inhibitor complexes.

Use Cases

• Benchmarking explainable AI attribution methods based on atom-level intersection over union scores
• Validating the biological relevance of model attributions based on mapping to 3D protein-ligand structures
• Identifying chemically meaningful ligand features for kinase and GPCR targets based on consensus attributions
• Developing more interpretable graph neural networks for drug discovery based on the provided attribution data

Strengths

• Data is validated against 3D protein-ligand structures, with consensus attributions showing up to 76% enrichment for atoms within 2 Å of the binding pocket.
• Benchmarks four distinct explainable AI attribution methods on two important target classes: kinases and G-protein-coupled receptors.
• Includes attribution mapping to experimentally important regulatory residues, such as those in the DFG motif.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is small at 16.6 KB, indicating a limited scope of analysis.

Provenance

Source

figshare

Collection Method

Attribution data generated from benchmarking four XAI methods on GNN models trained for specific biological targets.

Freshness

Last updated 2026-04-15 19:07:39; freshness should be verified.