ATOM3D: Ligand Binding Affinity Data for Machine Learning

Ligand Binding Affinity (LBA) data from the ATOM3D project, provided in LMDB format. The upload includes the full dataset and pre-split versions at 60% and 30% protein sequence identity thresholds, along with corresponding train, validation, and test indices. The dataset was created by Raphael J.L. Townshend of Stanford University.

Use Cases

• Training graph neural networks for binding affinity prediction based on protein-ligand structures.
• Benchmarking machine learning models using pre-defined splits at different sequence identity thresholds.
• Developing methods for molecular property prediction based on the provided ligand binding data.

Strengths

• Includes pre-defined train/validation/test splits at two sequence identity thresholds (30% and 60%), facilitating standardized benchmarking.
• Data is provided in LMDB format, which is likely optimized for efficient data loading in machine learning workflows.
• The dataset originates from the ATOM3D project, a known source for molecular machine learning benchmarks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for specific model training needs.
• Last update date is unknown; freshness unverified.

Provenance

Source

ATOM3D project, Stanford University