Embedding name: RNA-FM-full-transcript

Modality: RNA/mRNA sequence / full mature transcript language-model embedding
Algorithm: Transformer RNA language model
Embedding source URL: Local generation from Ensembl release 114 GRCh38 mature RNA transcript sequences and master_gene_table.csv
Original model/code URL: https://github.com/ml4bio/RNA-FM

Public files

embeddings.npz: compressed NumPy archive containing the embedding matrix under key embeddings.
genes.tsv: row metadata table. Row order matches the embedding matrix.
metadata.json: machine-readable provenance and processing metadata.

Matrix

Shape: 19976 genes × 640 dimensions
Dtype: float32

Detailed processing notes

Source and interpretation

This embedding is a gene-level mature RNA transcript representation derived from RNA-FM. It was generated locally from the mature transcript sequence universe prepared from Ensembl release 114 GRCh38 cDNA sequences and the canonical transcript IDs in master_gene_table.csv.

The model used was RNA-FM / rna_fm_t12 with the official pretrained checkpoint RNA-FM_pretrained.pth. For each matched gene, the mature transcript sequence was represented as raw RNA sequence. No Helix-specific E codon-start markers were inserted. RNA-FM final layer 12 token representations were extracted, BOS/EOS tokens were excluded, token embeddings were mean-pooled within each chunk, and long transcripts were combined by sequence-length weighted mean pooling across chunks.

This embedding should therefore be interpreted as a full mature transcript RNA language-model representation. It differs from Helix-mRNA because it uses raw RNA sequence without CDS-aware E markers, and it differs from CDS-only DNA embeddings because it includes UTR sequence from the mature transcript.

Provenance fields

source_name: RNA-FM full mature transcript
original_model_or_method: RNA-FM / rna_fm_t12 RNA language model
input_sequence_type: mature RNA full transcript
input_sequence_source: Ensembl release 114 GRCh38 cDNA-derived mature RNA sequences and master_gene_table.csv canonical transcript IDs
genome_build: GRCh38
sequence_reference: Ensembl release 114
model_checkpoint: RNA-FM_pretrained.pth
model_checkpoint_sha256: 5b5d7d87b37c291ef42c140ef9edf7aea29f255fa2a4fd435f776c52e93d5e99
model_code_url: https://github.com/ml4bio/RNA-FM
model_interface: fm.pretrained.rna_fm_t12()
model_output_dimension: 640
final_layer_used: 12
chunk_len: 900
min_chunk_len: 0
batch_size: 1
pooling_strategy: mean pooling over final layer 12 token embeddings, excluding BOS/EOS tokens, with sequence-length weighted mean across chunks
training_or_finetuning: none

Sequence and transcript decisions

The source gene set came from master_gene_table.csv. Canonical transcript IDs from the master table were used to select one mature transcript sequence per gene from the Ensembl release 114 GRCh38 cDNA FASTA.

RNA-FM reused the raw mature RNA sequence universe prepared during the Helix-mRNA workflow. However, it used the raw RNA FASTA without Helix E codon-start markers. This decision prevents Helix-specific input formatting from contaminating the RNA-FM representation.

Genes without a matched mature transcript sequence, without a matched CDS sequence from the shared preparation workflow, or where sequence preparation failed were excluded from this embedding.

Processing workflow

For each gene in the mature RNA transcript table:

1. Read the canonical transcript ID from master_gene_table.csv.
2. Reuse the mature transcript sequence table prepared from Ensembl release 114 GRCh38 cDNA.
3. Convert mature cDNA-derived sequence to raw RNA sequence.
4. Use the raw RNA sequence without Helix E markers.
5. Load RNA-FM / rna_fm_t12 from the local checkpoint cache.
6. Split long transcript sequences into 900 nt chunks.
7. Run RNA-FM with batch_size=1.
8. Extract final layer 12 token representations.
9. Exclude BOS/EOS tokens from pooling.
10. Mean-pool token embeddings within each chunk.
11. Combine chunk vectors with sequence-length weighted mean pooling.
12. Save the final matrix as float32.

Important implementation decisions

The standard RNA-FM loader attempted to use an external CUHK checkpoint URL that could return 403. The final workflow copied the HuggingFace snapshot checkpoint into the torch hub checkpoint cache so fm.pretrained.rna_fm_t12() could load the official checkpoint locally and reproducibly.

No model training or fine-tuning was performed. The final matrix was generated only by local inference from the official pretrained checkpoint.

Summary of final mapping:

Source identifiers / mature transcript sequences: 19,976
Dimensions: 640
Rows generated before duplicate collapse: 19,976
Missing/failed sequence genes: 121
Unmapped source rows: 0
Ambiguous source rows: 0
Final mapped genes: 19,976
Master table rows: 20,097
Master coverage: 99.398%
Duplicate final mappings averaged: 0

Processing and provenance metadata

This section preserves information moved out of metadata.json during catalog normalization. The loading-facing metadata is intentionally compact; detailed source, mapping, processing, quality-control, and reproducibility records remain here.

Source catalog commit: ab37742040d20b523aebd188d3079871a0c2b534.

Canonicalized field names

Previous field	Compact metadata field
`embedding_name`	`display_name`
`num_genes`	`n_genes`
`embedding_dimension`	`n_dimensions`
`matrix_shape[0]`	`n_genes`
`matrix_shape[1]`	`n_dimensions`

Derived standardized fields

species was set to Homo sapiens from the validated human Ensembl ENSG identifier namespace.
master_coverage_fraction was computed as n_genes / 20097.

Legacy processing and provenance record

{
  "all_zero_rows": {
    "count": 0,
    "documented": true,
    "expected": false,
    "reason": "No exact all-zero rows were detected in embeddings.npz.",
    "report_file": null
  },
  "batch_size": 1,
  "chunk_len": 900,
  "created_at": "2026-07-08T10:59:52",
  "former_processing_embedding_id": "rnafm_full_transcript_v1_1",
  "input_sequence_source": "Derived from Helix mature mRNA preparation based on Ensembl release 114 GRCh38 cDNA/canonical transcript mapping",
  "input_sequence_type": "mature mRNA full transcript",
  "min_chunk_len": 0,
  "model": "RNA-FM / rna_fm_t12",
  "model_checkpoint_sha256": "5b5d7d87b37c291ef42c140ef9edf7aea29f255fa2a4fd435f776c52e93d5e99",
  "public_file_structure": {
    "embeddings": "embeddings.npz",
    "genes": "genes.tsv",
    "metadata": "metadata.json",
    "readme": "README.md"
  },
  "runtime_seconds": 3351.46,
  "sequence_processing": "Helix E codon-start markers removed; T converted to U; sequences restricted to A/C/G/U/N",
  "software": {
    "python": "3.10",
    "torch": "2.2.2+cu118"
  },
  "status": "processed_qc_passed"
}