PairCtrl

PairCtrl is a benchmark for genomic language models on gene pairs. On many gene-pair tasks, gene position alone (adjacency, strand, intergenic distance) or a close relative in the training data predicts the label. In PairCtrl the labels come from experiments rather than from genome layout. Every model is scored next to a position-only baseline called row zero, and the scores are also reported with gene position held fixed. The evaluation code is the pairctrl package (https://github.com/tejasvi/pairctrl).

Tasks

task	what is predicted	items	labels from	primary metric
ppi_exp	interaction between the proteins of two genes	1,783,791 pairs; 14 data sets of 4 genomes	two-hybrid, affinity-purification and cross-linking screens, curated complexes, small-scale studies	AUROC on the strict subset
operon_expr	two neighbouring genes on one transcript	12,636 adjacent pairs of 3 genomes, 6,268 of them labelled	full-length transcript reads (nanopore direct RNA, Cappable-seq)	gain over the position model
operon_regulondb_exp	the same, on the E. coli pairs of DGEB	4,314 pairs	RegulonDB transcription units with experimental evidence, rebuilt locally	gain over the position model
etgp_nodist	whether a regulatory element controls a gene	5,534 element-gene pairs in 3 data sets	CRISPR interference screens, as collected by DNALongBench	AUROC with loci held out

In ppi_exp every data set appears twice, once with degree-preserving negatives and once with negatives drawn from the screened space. The 41,070 positive pairs include 25,007 in the strict subset. A strict pair has genes more than 20 apart that do not share an operon, no ribosomal protein, and no neighbourhood, fusion or co-occurrence evidence in other genomes. In operon_expr a pair counts as a negative only when both genes are expressed. The inputs of etgp_nodist have a length that does not depend on the distance between element and gene, and whole loci are held out.

Files

• <task>/items.parquet: one row per item, with item_id, the label where it may be redistributed, position features, frozen folds or split columns, and the resampling units of the confidence intervals.
• <task>/scores.parquet: per-item scores of the baselines and of the leaderboard models. Zero-shot scores are cosine similarities of the two genes' embeddings. Supervised scores are out-of-fold predictions.
• scores_nc/<task>/scores_noncommercial.parquet: the per-item scores of six models whose weights have non-commercial licences, with their attribution in scores_nc/README.md.
• ppi_exp/relatedness.tsv and operon_expr/relatedness.tsv: MinHash similarity of each test genome to its nearest training genome, per split.
• genomes/genes.parquet and genomes/contigs.parquet: the 18,998 genes and the DNA of the 6 genome records that the tasks use. The columns accession, gene_a and gene_b of a task point to accession and gene_index here, so any model can be embedded.
• leaderboard/: the tables as CSV and JSON, and a static page, index.html.
• DATASHEET.md: the answers to the questions of "Datasheets for Datasets", with the maintenance plan.
• NOTICE.md, LICENSE, LICENSES/, SHA256SUMS.

Evaluation

Install the evaluation package with pip install pairctrl and download this dataset into a directory:

python -c "from huggingface_hub import snapshot_download; snapshot_download('tstomar/pairctrl', repo_type='dataset', local_dir='DATA_DIR')"

A submission is a parquet or CSV file with one score per item and the columns item_id and score. Then run

python -m pairctrl.evaluate --task ppi_exp --data DATA_DIR --submission scores.parquet --out metrics.json

Add --oof for out-of-fold predictions. The output lists the submission next to row zero and the distance rule, for each view and evaluation unit, with 95% bootstrap intervals. It reports AUROC, AUROC with positives compared only with negatives at the same gene position, AUPRC, the gain of the out-of-fold position model when the score is added, and lower and upper bounds on the share of the score's AUROC above 0.5 that is due to gene position. The intervals resample blocks of neighbouring genes within genomes for the bacterial tasks and loci for etgp_nodist.

Labels and inputs rebuilt locally

• operon_regulondb_exp: RegulonDB data may not be integrated into other databases without permission, so the labels are not included. Run python -m pairctrl.joins.regulondb TUSet.txt GeneProductSet.txt items.parquet labels_regulondb.parquet on your own RegulonDB v14.5 download and place the output next to items.parquet.
• etgp_nodist: python -m pairctrl.joins.etgp_inputs hg19.fa items.parquet inputs.parquet builds the model inputs from your copy of the hg19 reference genome. Only coordinates are included here.

Not included

• Three data sets of ppi_exp from a structure-prediction screen (CC BY-NC-ND 4.0).
• Two positive controls built on STRING's database channel, and DGEB's own operon labels.

Licence and attribution

The data are released under CC BY 4.0, except where NOTICE.md names a different upstream licence. NOTICE.md lists every source with its licence and citation and reproduces the BioGRID notice. The tables taken from BacBench keep its Apache 2.0 licence, whose text is in LICENSES/. The scores in scores_nc/ follow the non-commercial licences of the models that produced them and are released under CC BY-NC-SA 4.0.

Limitations

• The zero-shot tasks use reference genomes that most models probably saw during pretraining. They test what a representation encodes, not generalisation to new genomes.
• An unreported pair in an interaction screen was not necessarily tested. Degree-preserving negatives control for hub proteins but not for every detection bias.
• The operon labels depend on the growth conditions of the read data.
• The benchmark set of etgp_nodist has only 7 loci, so the Gasperini data set, with 184 loci, carries most of the information.

Citation

Cite the release with the CITATION.cff of the code repository or the archive at https://doi.org/10.5281/zenodo.23605245.