syjiang/ncRNA-pretrain-47M

Non-coding RNA Pretraining Corpus (47.2M sequences)

A large, deduplicated, cluster-indexed corpus of non-coding RNA (ncRNA) for pretraining RNA foundation models. Sequences are aggregated from five complementary repositories, unified to the RNA alphabet, exactly deduplicated, and clustered to provide a balanced per-epoch sampling index.

• 47,154,121 unique ncRNA sequences (32.48 billion nucleotides)
• 21,118,814 MMseqs2 linclust clusters (sampling index, not a reduction)
• 30,145,573 sequences of length ≤ 512 nt (structure-annotation subset)
• Length range 16–8,192 nt, alphabet ACGU (+ IUPAC ambiguity <0.1%)

Sources

Source	Sequences (kept)	Nucleotides (G)
RNAcentral	40,119,088	24.46
NCBI nt (ncRNA)	6,507,049	7.80
Rfam	428,720	0.05
Ensembl 109	98,523	0.17
Ensembl Genomes Bacteria 56	741	0.00

Sequences were merged in the order RNAcentral → nt → Rfam → Ensembl → Ensembl Bacteria, and exact duplicates were removed keeping the first occurrence. The source field therefore records the first repository (in this order) that contained each unique sequence, i.e. its novel contribution.

Composition (top RNA types)

RNA type	Sequences
rRNA	22,950,711
tRNA	5,976,688
lncRNA	3,609,604
misc_RNA	3,310,876
pre_miRNA	633,146
snoRNA	579,986
sRNA	539,291
ncRNA	504,481
snRNA	479,658
SRP_RNA	448,941
hammerhead_ribozyme	233,359

Schema

Each row is one sequence:

field	type	description
id	string	source accession (RNAcentral URS / GenBank / Rfam / Ensembl)
source	string	rnacentral / nt / rfam / ensembl / ensembl_bacteria
rna_type	string	biotype where available (RNAcentral / Ensembl), else unknown
length	int32	sequence length (nt)
cluster_id	int32	MMseqs2 cluster index for balanced sampling
is_representative	bool	whether this sequence is its cluster representative
sequence	string	RNA sequence over {A,C,G,U} (+ rare IUPAC codes)
description	string	original header text (truncated to 200 chars)

Balanced sampling (recommended)

The corpus is dominated by ribosomal RNA. To counter this skew, sample one sequence per cluster_id per epoch rather than drawing uniformly. The full corpus is still seen over training, while each epoch stays balanced across the sequence-similarity landscape.

from datasets import load_dataset
ds = load_dataset("syjiang/ncRNA-pretrain-47M", split="train", streaming=True)
for ex in ds:
    seq = ex["sequence"]          # e.g. "AUACGUAGGUGACAAGCGUU..."
    ...

Construction pipeline

1. Aggregate ncRNA from the five sources above.
2. For NCBI nt, recover ncRNA by description-level annotation (ncRNA keyword whitelist, mRNA/protein/CDS and XM_/NM_ exclusion, high-confidence rescue).
3. Uppercase and unify the nucleotide alphabet (collapse T/U duplicates).
4. Keep sequences of length 16–8,192 nt; remove exact duplicates (seqkit rmdup).
5. Cluster with MMseqs2 linclust at --min-seq-id 0.7 -c 0.8 to build the per-epoch sampling index.
6. Emit sequences in the RNA alphabet.

Source releases

• RNAcentral: current release
• NCBI nt (ncRNA): NCBI BLAST nt FASTA distribution
• Rfam: current release
• Ensembl 109: release-109
• Ensembl Genomes Bacteria 56: Ensembl Genomes Bacteria release-56

License

Released under CC-BY-4.0. Underlying records remain subject to the licenses of RNAcentral, NCBI, Rfam and Ensembl; please cite the original databases.

[CITE: RNAcentral, NCBI nt, Rfam, Ensembl, MMseqs2, seqkit]