ProteinGym: Benchmark for Protein Fitness Prediction Models

ProteinGym is a benchmark suite for protein fitness prediction and design models. It comprises a curated collection of over 200 high-throughput experimental assays covering approximately 3 million mutated sequences, along with expert clinical annotations on pathogenicity for mutants in over 3,000 human genes. The dataset is hosted on AWS Open Data and is licensed under MIT.

Use Cases

• Benchmarking protein fitness prediction models based on the collection of 200+ experimental assays.
• Training models for pathogenicity prediction based on clinical annotations for over 3,000 human genes.
• Evaluating sequence-function relationships based on the dataset's ~3 million mutated protein sequences.

Strengths

• Contains over 200 curated high-throughput experimental assays.
• Includes clinical annotations from experts for mutants in over 3,000 human genes.
• Covers approximately 3 million mutated protein sequences.

Limitations

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

Provenance

Source

aws_open_data

Collection Method

Curated collection of experimental assays and clinical annotations.

Time Range

null

Freshness

Last update date is unknown; freshness unverified.

Geography

null