GRIMM-EC

GRIMM is a benchmark for predicting enzyme function (EC number) from biological sequence. Sequences are reviewed (SwissProt) prokaryotic proteins with EC annotations; partitions are stratified per label by UniRef50 cluster so that homologous sequences do not leak between train and evaluation splits.

Two modalities are provided: amino acids (per-protein SwissProt sequences) and nucleotides (per-CDS sequences from ENA).

⭐ Which version to use

EC_v2/ should be used for all new work, as it corrects important bugs in v1 data generation pipeline. EC_v1/ is maintained for reproducibility of existing citing work and should only be used to reproduce already-published v1 results

Structure

EC_v2/   amino_acids/   split_1 … split_5 / {train, validation, test1, test2}.tsv
         nucleotides/   split_1 … split_5 / {train, validation, test1, test2}.tsv
EC_v1/   amino_acids/   split_1 … split_5 / {train, validation, test1, test2}.csv
         nucleotides/   split_1 … split_5 / {train, validation, test1, test2}.csv

All files are tab-separated. v2 uses the .tsv extension; v1 retains the original .csv extension (tab-separated despite the name) for release stability — load v1 with pd.read_csv(path, sep="\t").

Columns — amino acids: Entry, EC number, EMBL, RefSeq, UniRef50, UniRef90, UniRef100, Sequence; nucleotides: Entry (EMBL CDS id), EC number, Sequence.

The 5 folds are not a standard k-fold: each is an independent train/valid/test partition that preserves UniRef50 clusters. Train and evaluate the 5 folds as independent models (individually or as an ensemble), not as rotating CV folds.

Splits

• train / validation / test1 — closed-set: evaluation sequences whose labels also appear in training, but held out by UniRef50 cluster.
• test2 — open-set: sequences from labels absent from training (out-of-distribution).

Labels

EC numbers (4th level). Proteins with multiple EC annotations are kept as a single compound label string (e.g. 1.1.99.1;1.2.1.8), distinct from its component labels — they are not expanded into separate rows.

Sizes (GRIMM-EC v2, average per fold)

Modality	train	validation	test1	test2
amino acids	~178,053	~28,719	~29,689	~959
nucleotides	~251,745	~42,557	~45,185	~1,755

237,421 proteins · 6,393 EC labels (1,321 compound) · 65,996 UniRef50 clusters. Sequences from UniProt release 2025_02.

Verified for v2 (all 5 folds, both modalities): 0 (sequence, EC) overlap between train and any evaluation split; 0 accession overlap between splits; test2 labels absent from train. Identical sequences carrying different EC labels may appear in different splits — this is intended cross-label difficulty under per-label stratification, not leakage.

Known limitations of GRIMM-EC v1 (fixed in v2)

EC_v1/ is retained for reproducibility of already-published results. Its test2 has been corrected to be fully disjoint from train/validation/test1; train/validation/test1 are unchanged from the original release. Remaining v1 limitations (all fixed in v2):

1. AA sequences are UniRef50 representatives, not per-protein SwissProt sequences (the nucleotide modality used real per-CDS sequences and is unaffected).
2. ~5.8% of AA test1 rows (and ~4.1% of validation) share an exact sequence with train, because 1–2 cluster labels were split by sequence rather than by cluster (nucleotides: ~0.4%).

See the GRIMM repo git history for full details.

Citation

Preprint: https://arxiv.org/abs/2602.16504

Code: https://github.com/Hoarfrost-Lab/grimm