RLMF-m6APred

This repository contains the released model weights for RLMF-m6APred, an RNA m6A site predictor that integrates a domain-adapted RNA language model with handcrafted sequence representations through residual language-guided modulation and fusion.

Web server: https://xiaof-xy--rlmf.modal.run

Submit a centered 41-nt RNA sequence to obtain a five-fold ensemble m6A prediction. RNA U is accepted and automatically normalized to T.

Contents

pretrained/rna_language_model/   Domain-adapted multi-species RNA backbone
weights/                         Five released folds for each dataset
  H_b/  H_k/  H_l/               Human brain, kidney, and liver
  M_b/  M_h/  M_k/  M_l/  M_t/  Mouse brain, heart, kidney, liver, and testis
  R_b/  R_k/  R_l/               Rat brain, kidney, and liver
manifests/weights_sha256.csv      File sizes and SHA-256 checksums

The release contains 55 downstream checkpoints (11 datasets x 5 folds). Each checkpoint is a tensor-only PyTorch state dictionary. Training logs, optimizer states, data splits, and experiment-search metadata are not included.

Download

Install the Hugging Face Hub client and download a fixed snapshot:

pip install -U huggingface_hub
hf download Yu-star/RLMF-m6APred --local-dir RLMF-m6APred-model

Inference and evaluation

The inference-only implementation and independent-test evaluation entry point are provided in the companion GitHub repository:

https://github.com/XiaoF-xy/RLMF-m6APred

Place or download this model repository so that the GitHub project can access:

pretrained/rna_language_model/
weights/<dataset>/fold_01.pt ... fold_05.pt

The five fold probabilities are averaged for independent-test evaluation.

Integrity

Use manifests/weights_sha256.csv to verify all 55 downstream checkpoints. The domain-adapted backbone is distributed as part of this model package.

Intended use

These weights are intended for research use, reproduction of the accompanying paper, and inference on 41-nt RNA sequences under the protocol described in the paper and companion code repository.

Limitations

Predictions may be affected by species, tissue, sequencing protocol, motif composition, and distribution shift. Outputs should not be interpreted as experimental validation or clinical evidence.

Citation

Citation information will be added after publication.

License

License information will be finalized before the public release. The licenses and terms of any upstream pretrained model components remain applicable.