OpenProteinSet: 16+ Million Precomputed Protein Alignments and Structures

More than 16 million precomputed multiple sequence alignments, structural template hits, and AlphaFold2 structure predictions are included in this corpus. It was released by the OpenFold team to reproduce and extend the training data used for AlphaFold2. The dataset was used to train the OpenFold model from scratch to parity with AlphaFold2.

Use Cases

• Training protein structure prediction models based on the provided multiple sequence alignments.
• Benchmarking new algorithms against precomputed AlphaFold2 structure predictions.
• Analyzing protein sequence-structure relationships using the large-scale alignment corpus.

Strengths

• Contains more than 16 million precomputed multiple sequence alignments.
• Includes structural template hits from the Protein Data Bank and AlphaFold2 predictions.
• Successfully used to train the OpenFold model to parity with AlphaFold2.

Limitations

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

Provenance

Source

OpenFold team (Ahdritz et al., NeurIPS 2023 Datasets and Benchmarks).

Collection Method

Reproduces and extends the kind of training data used for AlphaFold2.

Freshness

Last updated 2026-05-27 11:41:41; freshness should be verified.