VibeProteinBench

VibeProteinBench is a benchmark for evaluating language-interfaced protein design models across three tasks: protein recognition, rationale-guided engineering, and functional protein generation.

Configurations

• recognition — protein recognition tasks.
• engineering — human-eval-verified, rationale-guided protein engineering tasks.
• generation — functional protein generation tasks.

Each configuration is provided as a single test split in Parquet format.

Responsible AI (RAI) Metadata

Data limitations / known limitations

VibeProteinBench relies primarily on in silico validation metrics such as predicted structure confidence, energy changes, docking scores, GO prediction, and interface-quality measures, rather than experimental wet-lab validation. Therefore, benchmark success may not fully guarantee real-world biological function or experimental viability.

Bias and underrepresented populations

The benchmark is biased toward proteins, structures, annotations, and molecular targets available in public biological databases such as UniProt/Swiss-Prot, RCSB PDB, Gene Ontology, InterPro, ECOD, BRENDA, OGTFinder, PubChem, and related resources. Under-characterized protein families, organisms, or structures, may be underrepresented.

Personal or sensitive information

The dataset does not contain personal, human-identifiable, demographic, or private sensitive information. It consists of protein sequences, molecular targets, functional annotations, structural annotations, and benchmark queries derived from public biological databases.

Validated use cases and construct validity

The dataset is intended for evaluating language-interfaced protein design models across protein recognition, rationale-guided engineering, and functional protein generation. Construct validity is supported through expert-curated task design, mechanistic rationales, public biological annotations, and multi-faceted in silico validation, but not through experimental validation.

Potential societal impacts

This benchmark may support progress in computational biology, protein engineering, biotechnology, and therapeutic design by enabling more systematic evaluation of language-interfaced protein design models. However, improved protein design capability may also lower barriers to misuse, including harmful or toxic biological design, so responsible deployment and safeguards are important.

Synthetic data inclusion and source

The benchmark does not primarily consist of synthetic biological measurements or experimentally generated synthetic proteins. It includes constructed natural-language benchmark queries and evaluation tasks derived from natural protein sequences and experimentally-determined structures, expert-defined rationales, and computational preprocessing.

Source dataset provenance

The benchmark is derived from public biological and chemical resources including UniProt/Swiss-Prot, RCSB PDB, Gene Ontology, InterPro, ECOD, BRENDA, OGTFinder, PubChem, PubMed, and Semantic Scholar. Additional validation and annotation signals are obtained using computational tools such as Biopython, DSSP, Protenix-v1, PyRosetta, AutoDock Vina, pyKVFinder, GO-GPT, RDKit, and related pipelines.

Preprocessing / collection / annotation procedures

The dataset was curated by selecting proteins, GO terms, protein targets, and small-molecule targets from public databases with contamination controls such as temporal cutoffs and literature-based filtering. Queries were constructed for recognition, engineering, and generation stages using computed biochemical properties, structural annotations, expert-curated mechanistic rationales, and task-specific in silico evaluation criteria; human experts further reviewed tasks for realism, clarity, and verifiability.

License

Released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).