Xu-AI4Science/marri-rri-network-data

MARRI genome-wide RRI candidate screen

A genome-wide table of candidate RNA--RNA interaction (RRI) pairs across 171,703 representative human RNAs (GENCODE v49 protein-coding and lncRNA transcripts, GENCODE tRNAs, and mature miRBase miRNAs, CD-HIT-EST-deduplicated at 80% identity -- the same representative-RNA atlas used in the MARRI paper's candidate interaction atlas). This is a sequence-complementarity-based candidate screen, generated as a faster drop-in replacement for the RIblast all-vs-all search described in the MARRI paper's Methods. It identifies plausible pairs for downstream scoring (e.g. with the MARRI model itself), not experimentally validated or MARRI-scored interactions.

Code: https://github.com/GainGod-Xu/MARRI/tree/main/genome_wide_rri

What this data is -- and isn't

• Is: a reverse-complement-k-mer-seeded, GPU/vectorized heuristic complementarity ranking of candidate partners for every one of the 171,703 representative RNAs, with up to 50 ranked partners per query and up to 5 non-overlapping candidate binding sites per partner, each centred in a 256-nt window pair.
• Isn't: a set of MARRI model predictions. gpu_score is a cheap local-complementarity heuristic (GC/CG +3, AU/UA +2, GU/UG +1, mismatch -2, summed over a 64-nt flank around the seed anchor) used only to rank and shortlist candidates during screening -- it is not the MARRI model's interaction probability, not a thermodynamic free energy, and not validated binding evidence. To get an actual interaction probability for any pair here, run it through the MARRI model (see the main repo).
• Isn't an RNAduplex-reranked result. This table used screening mode gpu50 (heuristic-only ranking); rnaduplex_energy is empty for every row. The code also supports a gpu100-duplex50 mode that reranks the top-100 heuristic hits with ViennaRNA's RNAduplex MFE, which was not run for this table.

Known limitation: metadata placeholders

gene_type1/2 and genomic_context1/2 are not populated in this run -- every row has the default values "others" and "Exon". The GENCODE/miRBase/CD-HIT metadata TSV (rna_entity_id -> gene_type, genomic_context) that would fill these in correctly was not available when this table was built. The rna_entity_id1/2 columns (e.g. GENCODE:ENST00000001008.6, MIRBASE:..., GENCODE_TRNA_SEQ:...) are correct and can be used to look up the real gene type / genomic context yourself from GENCODE v49 / miRBase if needed.

Columns

Column	Meaning
seq_id	Stable hash-derived ID for this (query, target, site) row
label	Always 0 in this table (placeholder; not a positive/negative training label)
seq1, seq2	256-nt windows (RNA alphabet, 5'->3') centred on the candidate site on the query and target RNA respectively
rel_start1/end1, rel_start2/end2	Half-open site coordinates relative to seq1/seq2 (i.e. position within the 256-nt window)
gene_type1/2, genomic_context1/2	Placeholder metadata -- see limitation above
gpu_score	Heuristic local-complementarity score used for ranking (see above); higher is more complementary. Observed range in this table: min -158, median 55, p99 353, max 406
seed_hits	Number of seed (12-mer) hits supporting this query-target pair before site enumeration
rnaduplex_energy	Empty for every row (screening mode gpu50 does not compute this; see above)
selection_mode	Always gpu50 in this table
rna_entity_id1, rna_entity_id2	Stable IDs into the representative-RNA atlas (GENCODE:<transcript>, MIRBASE:<id>, or GENCODE_TRNA_SEQ:<id>)
partner_rank	Rank (1-50) of the target among this query's ranked partners
site_rank	Rank (1-5) of this site among the partner's non-overlapping candidate sites
rna1_window_start/end, rna2_window_start/end	Absolute (0-based, half-open) coordinates of the 256-nt window within the full query/target RNA sequence
rna1_site_start/end, rna2_site_start/end	Absolute (0-based, half-open) coordinates of the candidate site itself within the full RNA sequence

Scale and provenance

• 171,703 queries x up to 50 partners x up to 5 sites = 13,377,052 retained rows (most query/partner/site slots are empty because a site either wasn't found or was pruned by the strict non-overlap rule).
• Generated 2026-10-16 via a 5-way parallel Slurm job array on the FASRC Kempner test partition (CPU-only -- this pipeline is bottlenecked by Python-side seed lookups, not GPU compute), after fixing a np.searchsorted dtype-promotion bug in the screening code that had been causing a ~15,000x per-seed-lookup slowdown. The fix only changes speed, not results: verified byte-identical output on a 100-query slice before and after the fix.

License and citation

Released under CC-BY-4.0. If you use this data, please cite the MARRI paper (citation to be added once published) and link back to https://github.com/GainGod-Xu/MARRI.