mRNA-FM

A 12-layer BERT-style transformer pre-trained on 45 million mRNA coding sequences (CDS) using codon (3-mer) tokenisation.

Architecture

Parameter	Value
Layers	12
Attention heads	20
Embedding dimension	1280
FFN dimension	5120
Vocabulary size	73
Positional encoding	Learned
Architecture	ESM-1b-style pre-LN Transformer
Max sequence length	1024 codon tokens (1022 codons = 3066 nucleotides)

Vocabulary: <cls> (0), <pad> (1), <eos> (2), <unk> (3), 64 standard RNA codons (indices 4-67), 4 null-padding tokens (68-71), <mask> (72).

Pretraining

• Objective: Masked language modelling (codon-level, 15% masking rate)
• Data: RefSeq -- 45 million mRNA coding sequences
• Source checkpoint: mRNA-FM_pretrained.pth from cuhkaih/rnafm

Tokenisation

mRNA-FM uses codon (3-mer) tokenisation: the input sequence is split into consecutive non-overlapping codons (triplets) and each codon is mapped to a single token. The model therefore only accepts sequences whose length is a multiple of 3.

Input sequences must use RNA notation (U, not T). Convert before tokenising:

seq = seq.replace("T", "U")

Parity Verification

Hidden-state representations verified identical (max abs diff = 0.00) to the original implementation at all 13 representation levels (embedding + 12 transformer layers). Verified on GPU (CUDA) with PyTorch 2.7 / transformers 4.57.6. SDPA numerical differences are expected (~3e-4 max diff over 12 layers) and are not a correctness issue.

Related Models

See the full RNA-FM collection.

Model	Training data	Embedding dim	Notes
RNA-FM	23.7 M ncRNA	640	Character tokenisation
mRNA-FM	45 M CDS	1280	This model

Usage

Embedding generation

import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model = AutoModel.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model.eval()

# Sequences must be RNA (U not T) and length divisible by 3 (codons)
sequences = [
    "AUGGGGUGCGAUCAUACCAGCACUAAUGCCCUCCUGGGAAGUCCUCGUGUUGCA",
    "AUGCUAGCUAGCUAGCUAUG",
]
enc = tokenizer(sequences, return_tensors="pt", padding=True)

with torch.no_grad():
    out = model(**enc)

cls_emb   = out.last_hidden_state[:, 0, :]   # (batch, 1280) -- CLS token
token_emb = out.last_hidden_state             # (batch, n_codons+2, 1280) -- per-codon

# Intermediate layers
out_all = model(**enc, output_hidden_states=True)
layer6_emb = out_all.hidden_states[6]

CDS-aware embedding (mRNA sequences)

For mRNA sequences with a CDS track, use batch_encode_with_cds to apply T→U conversion, extract only the coding region, chunk to codon boundaries, and encode — all in one call.

import numpy as np
import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model = AutoModel.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model.eval()

# Binary CDS track: 1 at the first nucleotide of each codon in the CDS, 0 elsewhere
sequences = ["ATGCTAGCTAGCTAGCTATGCTAGCTAGCTAGCT"]
cds = [np.array([0]*5 + [1, 0, 0]*9 + [0]*2)]  # example

enc, chunk_counts = tokenizer.batch_encode_with_cds(
    sequences, cds, return_tensors="pt", padding=True, add_special_tokens=True
)
with torch.no_grad():
    out = model(**enc)

# chunk_counts[i] = number of chunks produced for sequences[i]
# mean-pool non-special tokens for each sequence:
hidden = out.last_hidden_state  # (total_chunks, seq_len, 1280)

MLM logits

import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained("Taykhoom/mRNA-FM", trust_remote_code=True)
model.eval()

enc = tokenizer(["AUG<mask>GCUAUG"], return_tensors="pt")
with torch.no_grad():
    logits = model(**enc).logits   # (1, n_codons+2, 73)

Fine-tuning

Standard HF conventions. Use the CLS token embedding (out.last_hidden_state[:, 0, :]) as input to a classification or regression head for sequence-level tasks. Mean-pool over codon positions (excluding CLS and EOS) for codon-level aggregation.

Implementation Notes

The original implementation uses F.multi_head_attention_forward (eager). This HF port adds attn_implementation="sdpa" and attn_implementation="flash_attention_2" support, which were not part of the original codebase.

Each codon token represents exactly one triplet of nucleotides. The tokeniser splits the raw sequence into non-overlapping codons; any trailing nucleotides that do not form a complete codon are silently discarded.

Citation

@article{chen2022_rnafm,
  title   = {Interpretable {RNA} Foundation Model from Unannotated Data for Highly Accurate {RNA} Structure and Function Predictions},
  author  = {Chen, Jiayang and Hu, Zhihang and Sun, Siqi and Tan, Qingxiong and Wang, Yixuan and Yu, Qinze and Zong, Licheng and Hong, Liang and Xiao, Jin and Shen, Tao and King, Irwin and Li, Yu},
  journal = {arXiv preprint arXiv:2204.00300},
  year    = {2022},
  doi     = {10.48550/arXiv.2204.00300}
}

Credits

Original model and code by Chen et al. Source: GitHub. The HF conversion code was authored primarily by Claude Code and reviewed manually by Taykhoom Dalal.

License

MIT, following the original repository.