MiRformer: a dual-transformer encoder framework for predicting miRNA-mRNA interactions

This repository contains the code for MiRformer, a deep learning model for predicting miRNA-mRNA interactions using dual-transformer encoder architecture.

Command-line Usage

The main entry point for training and evaluation is scripts/DTEA_model.py.

Evaluation

To test MiRformer model, use the following command. This example uses the DTEA base model with a custom configuration and a minimal test dataset.

python scripts/eval_DTEA_model.py \
    --mirna_max_len 24 \
    --mrna_max_len 520 \
    --device "cuda" \
    --embed_dim 1024 \
    --ff_dim 4096 \
    --num_heads 8 \
    --num_layers 4 \
    --predict_span \
    --predict_binding \
    --predict_cleavage \
    --use_longformer \
    --test_path "test_data/sample_data.csv"

Arguments:

• --mirna_max_len: max mirna length.
• --mrna_max_len: max mrna length.
• --device: Device to use (e.g., cpu, cuda).
• --embed_dim: embedding dimension, default to 1024.
• --ff_dim: feed-forward dimension, default to 4096.
• --num_heads: number of head to use in each transformer encoder.
• --num_layers: number of layers to use in each transformer encoder.
• --predict_span: flag to be set True for model to predict seed region spans in mRNA inputs.
• --predict_binding: flag to be set True for model to predict interaction between miRNA-mRNA pairs.
• --predict_cleavage: flag to be set True for model to predict cleavage sites in mRNA inputs.
• --use_longformer: flag to be set True when using longformers. When false, standard multi-head transformers are used.
• test_path: path to test dataset

Minimal Test Dataset

A minimal test dataset is provided in test_data/sample_data.csv to validate that the software runs correctly.

Location: test_data/sample_data.csv

Format: The CSV file requires the following columns:

• Transcript ID
• miRNA ID
• miRNA sequence
• mRNA sequence
• seed start
• seed end
• label (0 for negative, 1 for positive)