Frozen biomolecular benchmark data. Start with the eight main categories below; supporting materials and experimental results have their own browsing entrances.

Main data · Structures, MSA, features and labels · Predictions and experiment results · Find the right table

Access: Public dataset page; data access requires manual approval. Existing approved users have been granted matching access to the two companion datasets. New requests are reviewed separately for each dataset. Sign in with an approved account to browse or download.

Single-chain / single-entity tasks

Category	Source	Direct data	Count	What is counted
Protein	ProteinGym	Browse	53,776 query rows; 217 assays	Frozen single-chain query variants across 217 protein assays
RNA	RNAGym	Browse	7,912 query rows; 32 assays	Frozen single-chain query variants across 32 RNA assays
Ligand-Regression	TDC ADMET / AnewDDE	Browse	5,881 test; 23,498 train_val	9 ADMET regression endpoints; test and train_val are separate; Spearman
Ligand-Classification	TDC ADMET / AnewDDE	Browse	10,498 test; 41,932 train_val	13 ADMET classification endpoints; test and train_val are separate; AUROC

Complex tasks

Category	Source	Direct data	Count	What is counted
Protein-Protein	AbAgym / BindingGym	Browse	700,083 measurements	All curated AbAgym / BindingGym measurements before paper-cohort sampling. Current paper cohort: 27 assays
Protein-Ligand	Platinum	Browse	857 measurements	All curated source measurements before evaluation sampling
Protein-RNA	NaBE	Browse	495 measurements	All curated source measurements before evaluation sampling
Protein-DNA	NaBE	Browse	1,458 measurements	All curated source measurements before evaluation sampling

Single-chain describes the prediction unit; a sample may include a physical complex. DNA appears in Protein–DNA. ADMET test and train_val remain separate in the split selector. The data split in other browsers groups rows for storage; scientific train and test splits are specified separately.

These two groups follow the NM Brief summary-table categories. Source measurements, frozen queries and final paper cohorts have separate counts. A source pool is the curated set of available measurements before assay eligibility checks and evaluation sampling. Cohort and snapshot differences.

Paper cohort overview

The figure counts assays in the paper cohort. The table above counts all curated source measurements before sampling. Composition and counts.

Supporting materials and experiments

What you need	Where to go
Structure files and observed chains	Open
MSA alignment sequences	Open
Per-assay structure/MSA/feature links	Open
Feature arrays and values	Open
Ground-truth labels for evaluation	Open
Original files and provenance	Open
Few-shot scores and predictions	Open
Matched zero-shot results	Open
Higher-order / multiround results	Open
Exact paper assay membership	Open

The current PPI collection and historical complex browser contain AbAgym and BindingGym measurements. SKEMPI2 was removed from the current publication after the few-shot usage audit. Usage audit.

No data files were deleted. Companion data tables are copied from a frozen main-repository revision; raw-asset links continue to point to the manually gated main repository. Technical table names are explained individually in the navigation guide. Table row counts cannot be summed into a unique biological sample count.

Documentation

The eight primary configurations are available through the Dataset Viewer or the Hugging Face datasets client. For example, load_dataset("yanjiehuang/MutaBench", "Protein", split="data", token=True). Use an approved account and keep credentials in your local Hugging Face authentication store.

Downloads last month: 684

Total file size:

94 GB