Explainability Methods for Drug-Target Interaction Predictions from GNNs

Mrinal Mahindran's dataset, last updated April 15, 2026, benchmarks four explainable AI attribution methods on graph neural network models for drug-target interaction prediction. The data likely contains atom-level attribution results for kinase and G-protein-coupled receptor targets, validated against 3D protein-ligand structures. The dataset is 63.5 KB in size and is stored in CSV format.

Use Cases

• Benchmarking explainable AI methods based on atom-level intersection over union (IoU) metrics.
• Validating the biological relevance of model attributions based on mapping to 3D protein-ligand structures.
• Identifying chemically meaningful ligand features in kinase-inhibitor complexes based on attributed atom contacts.
• Developing more interpretable graph neural networks for drug discovery based on consensus attribution results.

Strengths

• Data is validated against 3D protein-ligand structures, with attributed atoms showing up to 76% enrichment within 2 Å of the binding pocket.
• Benchmarks four distinct explainable AI attribution methods on specific biological targets (kinases and GPCRs).
• Attributed atoms were frequently found contacting experimentally important regulatory residues like 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.
• Consistency across the four explainable AI methods was described as modest.

Provenance

Source

Mrinal Mahindran via figshare.

Collection Method

Likely contains results from benchmarking explainable AI methods on trained graph neural network models.

Freshness

Last updated 2026-04-15 19:07:38; freshness should be verified.