RibonanzaNet

Training code for RibonanzaNet, preprint: https://www.biorxiv.org/content/10.1101/2024.02.24.581671v1.

Example notebooks

You may not want to retrain RibonanzaNet from scratch and rather just use pretrained checkpoints, so we have created example notebooks:
finetune: https://www.kaggle.com/code/shujun717/ribonanzanet-2d-structure-finetune
secondary structure inference: https://www.kaggle.com/code/shujun717/ribonanzanet-2d-structure-inference
chemical mapping inference: https://www.kaggle.com/code/shujun717/ribonanzanet-inference

Data Download

You just need train_data.csv, test_sequences.csv, and sample_submission.csv from https://www.kaggle.com/competitions/stanford-ribonanza-rna-folding/data

Environment

Create the environment from the environment file env.yml

conda env create -f env.yml

Install ranger optimizer

conda activate torch

git clone https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
cd Ranger-Deep-Learning-Optimizer
pip install -e .

How to run

First activate environment conda activate torch

Set up accelerate with accelerate config in the terminal or with --config_path option

For an example of a accelerate config file, see accelerate_config.yaml

Training

accelerate launch run.py --config_path configs/pairwise.yaml

Inference

accelerate launch inference.py --config_path configs/pairwise.yaml

Process raw prediction into submission file for Ribonanza

python make_submission.py --config_path configs/pairwise.yaml

Configuration File

This section explains the various parameters and settings in the configuration file for RibonanzaNet

Model Hyperparameters

• learning_rate: 0.001
  The learning rate for the optimizer. Determines the step size at each iteration while moving toward a minimum of the loss function.

• batch_size: 2
  Number of samples processed per GPU per batch.

• test_batch_size: 8
  Batch size used for testing the model per GPU per batch.

• epochs: 40
  Total number of training epochs the model goes through.

• dropout: 0.05
  The dropout rate for regularization to prevent overfitting. It represents the proportion of neurons that are randomly dropped out of the neural network during training.

• weight_decay: 0.0001
  Regularization technique to prevent overfitting by penalizing large weights.

• k: 5 1D Convolution kernel size

• ninp: 256
  The size of the input dimension.

• nlayers: 9
  Number of RibonanzaNet blocks.

• nclass: 2
  Number of classes for classification tasks.

• ntoken: 5
  Number of tokens (AUGC + padding/N token) used in the model.

• nhead: 8
  The number of heads in multi-head attention models.

• use_flip_aug: true
  Indicates whether flip augmentation is used during training/inference.

• gradient_accumulation_steps: 2
  Number of steps to accumulate gradients before performing a backward/update pass.

• use_triangular_attention: false
  Specifies whether to use triangular attention mechanisms in the model.

• pairwise_dimension: 64
  Dimension of pairwise interactions in the model.

Data Scaling

• use_data_percentage: 1
  The fraction of data used from the dataset (1= full data training).

• use_dirty_data: true
  Indicates whether to include training data that has only one of 2A3/DMS profiles with SN>1.

Other Configurations

• fold: 0
  The current fold in use if the data is split into folds for cross-validation.

• nfolds: 6
  Total number of folds for cross-validation.

• input_dir: "../../input/"
  Directory for input data. Put train_data.csv, test_sequences.csv, and sample_submission.csv here.

• gpu_id: "0"
  Identifier for the GPU used for training. Useful in single-GPU setup.

---

File structure

logs has the csv log file with train/val oss, models has model weights and optimizer states, oofs has the val predictions