DeepMiRT: miRNA Target Prediction with RNA Foundation Models

Model Description

DeepMiRT predicts miRNA-target interactions using RNA-FM embeddings and cross-attention. It ranks #1 on eCLIP benchmarks among 12 methods (AUROC 0.75) and achieves AUROC 0.96 on a comprehensive 813K-sample test set.

Architecture

Encoder: RNA-FM T12 (12-layer Transformer, pre-trained on 23M ncRNAs) — shared for both miRNA and target
Interaction: Cross-attention (2 layers, 8 heads) — target queries miRNA
Classifier: MLP (640 → 256 → 64 → 1)

Usage

pip install deepmirt

from deepmirt import predict

probs = predict(
    mirna_seqs=["UGAGGUAGUAGGUUGUAUAGUU"],
    target_seqs=["ACUGCAGCAUAUCUACUAUUUGCUACUGUAACCAUUGAUCU"],
)
print(f"Interaction probability: {probs[0]:.4f}")

Training

Data: miRNA-target interactions from multiple databases and literature mining
Two-phase: Phase 1 (frozen backbone, 27 epochs) → Phase 2 (unfreeze top 3 RNA-FM layers)
Hardware: 2x NVIDIA L20 GPUs, mixed precision (fp16)

Performance

Benchmark	AUROC	Rank
miRBench eCLIP (Klimentova 2022)	0.7511	#1/12
miRBench eCLIP (Manakov 2022)	0.7543	#1/12
Our test set (813K samples)	0.9606	#1/16

Files

epoch=27-val_auroc=0.9612.ckpt — Best model checkpoint (495 MB)
config.yaml — Training configuration

License

MIT

Downloads last month: -

liuliu2333
/

deepmirt