GNN Model Usage

Purpose

• This model was developed to practice development of ncRNA sequence generation
• It is highly experimental, and should not be used in any real-world environments where accuracy is paramount

Methodology

• We utilize a GATv2Conv (Graph Attention Network v2) architecture to predict RNA sequences that would fold towards a given structure
• The model takes a secondary structure in dot-bracket notation as input and predicts a nucleotide (A, U, G, C) at each position in a single forward pass (non-autoregressive)
• The idea is that a researcher might want to have a ncRNA with a specific structure that will fold into a desired shape
• They might know what structure would make this shape, but they do not know the sequence that makes that structure

Training Data

• Trained on synthetic RNA subsequences (25–50 nt) from three RFam families: RF04135 (medium conservation), RF04185 (low conservation), and RF04331 (high conservation)
• Sequences generated by cmemit from Infernal covariance models; secondary structures predicted by RhoFold and stored as .ct files
• 1,000 sequences per family, 10,000 subsequences total

Training Details

• Four GATv2Conv layers with 4-head attention, hidden dimension 128, LayerNorm, ELU activation, residual connections, and dropout (0.3)
• Adam optimizer with learning rate 1e-3, weight decay 1e-3, gradient clipping at norm 1.0
• ReduceLROnPlateau scheduler (factor 0.5, patience 5 epochs)
• Early stopping with patience of 15 epochs prevents overfitting
• Train/validation split is a seeded random 80/20 split
• Maximum 200 epochs, batch size 32

Usage

• For proper usage, view the GNN_Training_Example.ipynb file

AI Disclosure

• The code in this repo was written with assistance from Claude Code CLI tool.

RhoDesign Model Usage

The RhoDesign_Model/ folder contains the scripts used to process RNA data, prepare the fine-tuning dataset, fine-tune the RhoDesign model, and test the fine-tuned model on a single PDB file.

Workflow

The full RhoDesign workflow is:

1. Use the RhoFold pipeline to create or convert the generated subsequence dataset into the required structure files, including SSCT, PDB, and related outputs.
2. Prepare the RhoDesign training dataset from the RhoFold output folder.
3. Fine-tune the RhoDesign model using the prepared dataset.
4. Test the fine-tuned model on a single PDB file.

Files in RhoDesign_Model/

Rhofold Prediction+Folder Seperation.py

This script is used first. It runs or organizes the RhoFold prediction pipeline and creates the needed output files, including SSCT, PDB, and other prediction-related files.

These outputs are later used for preparing the RhoDesign fine-tuning dataset.

python "Rhofold Prediction+Folder Seperation.py"

prepare_rhodesign_daset for training.py

This script prepares the dataset for RhoDesign fine-tuning.

It extracts and organizes data from the output folder generated by the RhoFold predictions. This preparation step was needed because the fine-tuning dataset used approximately 1000 samples from each RNA family.

python "prepare_rhodesign_daset for training.py"

finetune_rhodesign_model.py

This is the main fine-tuning script.

It uses the prepared dataset to fine-tune the RhoDesign model on the processed RNA structure data.

python finetune_rhodesign_model.py

design_single_pdb_spyder.py

This script is used after fine-tuning.

It tests the fine-tuned RhoDesign model on a single PDB file and generates an RNA sequence design for that structure.

python design_single_pdb_spyder.py

Suggested Run Order

From inside the repository:

cd RhoDesign_Model

python "Rhofold Prediction+Folder Seperation.py"
python "prepare_rhodesign_daset for training.py"
python finetune_rhodesign_model.py
python design_single_pdb_spyder.py

Before running the scripts, update any dataset paths, output folder paths, model checkpoint paths, and configuration paths so they match your local environment.