DeepMiRT: miRNA Target Prediction with RNA Foundation Models

Upload a CSV file with columns containing mirna and target in the column names.

Example format:

Model Architecture

DeepMiRT uses a shared RNA-FM encoder (12-layer Transformer, pre-trained on 23M non-coding RNAs) to embed both miRNA and target sequences into the same representation space. A cross-attention module (2 layers, 8 heads) allows the target to attend to the miRNA, capturing interaction patterns. The attended representations are pooled and classified by an MLP head (640 → 256 → 64 → 1).

miRNA  → [RNA-FM Encoder] → miRNA embedding  ─────────┐
                                                       ↓
Target → [RNA-FM Encoder] → target embedding → [Cross-Attention] → Pool → [MLP] → probability

Training

• Data: miRNA-target interactions from multiple databases and literature mining
• Two-phase training: Phase 1 (frozen backbone) → Phase 2 (unfreeze top 3 RNA-FM layers)
• Hardware: 2× NVIDIA L20 GPUs, mixed-precision (fp16)
• Best checkpoint: epoch 27, val AUROC = 0.9612

Performance

Citation

If you use DeepMiRT in your research, please cite:

@software{liu2026deepmirt,
  title={DeepMiRT: miRNA Target Prediction with RNA Foundation Models},
  author={Liu, Zicheng},
  year={2026},
  url={https://github.com/zichengll/DeepMiRT}
}

License

MIT License. See LICENSE.