GleghornLab/protein_concept_probing

Protein concept probing

Twenty-two protein annotation tasks for probing what a protein language model's internal states encode, from amino-acid composition to Gene Ontology terms. Thirteen tasks label a whole protein and nine label each residue. The human tasks are reviewed Homo sapiens UniProtKB entries shorter than 1,024 residues; four further tasks are xTrimoPGLM benchmark tables.

Config	Level	Task	Proteins	Labels	Train / validation / test
functional_sites	residue	multiclass	6,893	6	4,778 / 1,106 / 1,009
go_bp	protein	multilabel	12,255	304	8,591 / 1,888 / 1,776
go_cc	protein	multilabel	16,443	164	11,409 / 2,457 / 2,577
go_mf	protein	multilabel	11,504	123	8,010 / 1,786 / 1,708
interpro_active_site	protein	multilabel	718	58	482 / 116 / 120
interpro_binding_site	protein	multilabel	640	39	446 / 87 / 107
interpro_conserved_site	protein	multilabel	2,948	334	2,062 / 450 / 436
interpro_domain	protein	multilabel	8,003	364	5,307 / 1,371 / 1,325
interpro_family	protein	multilabel	3,601	239	2,097 / 678 / 826
interpro_homologous_superfamily	protein	multilabel	8,198	100	5,499 / 1,368 / 1,331
interpro_repeat	protein	multilabel	1,189	65	819 / 189 / 181
lipidation	residue	multiclass	978	25	654 / 129 / 195
localization	protein	multiclass	7,530	10	5,321 / 590 / 1,619
metal_ion_binding	protein	binary	7,242	1	5,346 / 599 / 1,297
peptide	residue	multiclass	4,977	4	3,485 / 700 / 792
phosphorylation	residue	multiclass	8,648	6	6,057 / 1,304 / 1,287
post_translational_modification	residue	multiclass	15,349	12	10,722 / 2,400 / 2,227
prot_param	protein	regression	18,126	31	12,747 / 2,777 / 2,602
secondary_structure	residue	multiclass	8,986	4	6,310 / 1,351 / 1,325
ssp_q3	residue	multiclass	11,479	3	9,731 / 1,083 / 665
ssp_q8	residue	multiclass	11,479	8	9,731 / 1,083 / 665
topology	residue	multiclass	4,777	17	3,378 / 754 / 645

Homology leakage

The human tasks keep their published cluster splits: MMseqs2 clusters at 20% identity (30% for InterPro active site, binding site, conserved site and repeat) are assigned whole to train plus validation or to test, and validation clusters are drawn at 50% identity from the rest, with 80% coverage of the target sequence. The four benchmark tasks keep their published test split, and validation is a random 10% of their train split, so they hold no homology guarantee between train and validation.

How it was built

1. Read each curated human table and its recorded train, validation and test assignment.
2. Turn its targets into one label column: a multi-hot vector over the InterPro entries or Gene Ontology terms, 31 ProtParam values, or one class index per residue in the contiguous order of the classes the table holds.
3. For Gene Ontology, keep the terms with more than six positive proteins in both validation and test.
4. For the four benchmark tables, split 10% of train off as validation with seed 184502, then drop sequences of 1,024 residues or more.

Columns

Column	Meaning
id	UniProt accession for the human tasks; a stable sequence-derived id for the benchmark tasks
sequence	amino-acid sequence
label	multi-hot list (multilabel), list of 31 floats (regression), integer (binary or multiclass), or one integer per residue (residue tasks)

Each config folder also holds metadata.json with the task level, task type and label names. The benchmark tables publish their classes but not the order of their integers, so their label names are class_<n>.

What a model trained on it may say

These are probing benchmarks. InterPro entries and many UniProt features are curated with computational evidence, so a probe here measures agreement with the annotations, not experimentally measured function. Residue tasks count unannotated residues as their own class (-), which dominates most of them.

Sources and licence

The human tasks are the curated tables of Figshare record 10.6084/m9.figshare.31169506, released under CC BY 4.0, built from UniProtKB (accessed 2025-09-09) and the InterPro entry list (accessed 2025-09-16). The four benchmark tasks (metal_ion_binding, localization, ssp_q3, ssp_q8) are from biomap-research on Hugging Face under Apache 2.0.

Files

<config>/train.parquet, <config>/validation.parquet, <config>/test.parquet and <config>/metadata.json for each of the 22 configs.