GeoTarget-RNA-6D-Vital

GeoTarget-RNA-6D-Vital Models

This repository contains the trained model checkpoints for the GeoTarget-RNA thesis project. GeoTarget-RNA is a hybrid Graph Neural Network designed to predict miRNA-mRNA interactions by traversing the 2D/3D structural topology of the interactions.

Available Models

1. student_6d_lupi.pt (Recommended for Inference)

   • Description: A lightning-fast, 6D-Vital LUPI (Learning Using Privileged Information) student model. It only requires 6 explicit biophysical features at inference time.
   • Performance: 0.8458 APS (miRBench test), 0.9985 APS (miRTarBase external benchmark).
   • Features Required: Duplex Energy, Flanking AU, Pum2 Motif, ARE Motif, FOX Motif, Seed Wobbles.

2. hybrid_gnn_v28.pt

   • Description: The original V28 Hybrid GNN model. It fuses RNA-FM 64D per-token embeddings with 24 explicit thermodynamic rules.

3. teacher_22d_biophysics.pt

   • Description: The heavy biophysics teacher model used to train the 6D student during the LUPI distillation phase. It achieved 0.8607 APS but is computationally heavier for inference.

Intended Use & Limitations

• Intended Use: Predicting binary interaction probabilities (binding vs non-binding) for human miRNA and mRNA 3'UTR sequences. The model excels at identifying non-canonical targets containing bulges or seed mismatches that traditional sequence-based models miss.
• Limitations: The model requires the secondary structure (MFE dot-bracket string) to be pre-computed via ViennaRNA to form the graph topology.

How to use

import torch

# Load the 6D student model
model = torch.load("student_6d_lupi.pt", map_location="cpu")
model.eval()

print("GeoTarget-RNA model successfully loaded!")

For the complete inference pipeline and feature extraction scripts, visit our GitHub repository: https://github.com/reogine/GeoTarget-RNA