RNARepLearn

This package is used to train RNA sequence + structure graph embeddings. Aggregating the sequence and structural features of RNAs, we create embeddings for each nucleotide in a RNA sequence. These embeddings serve as alternative to traditional 1-hot encoding and can be used as input to RNA-related prediction tasks.

Setup

Setup conda environment

A conda environment file is provided conda/env.yml. This environment should contain all necessary package dependencies.

RNARepLearn installation:

After activating the conda environment run:

git clone https://github.com/nicolasgoedert97/RNARepLearn.git
cd RNARepLearn
pip install .

RFAM preprocessing

The rfam submodule can be used to download and process data from the RFAM database

cd rfam
snakemake --config-file config.yml

Efficient loading/saving of training data

This package uses the bioio package (https://github.com/mhorlacher/bioio) to save the training data to tfrecords.

git clone https://github.com/mhorlacher/bioio
cd bioio
pip install .

Data preprocessing

Alternatively to the rfam module, to process training data for immediate use, the following file structure is desireable:

<Data-directory>
  |_____<Dataset1>
        |_________<Dataset1>.fa / <Dataset1>.csv
  |_____<Dataset2>
  .
  .

Ideally sequece information is provided in fasta or tsv format. Two snakemake workflows are provided to process the data: preprocess/5utr and preprocess/fold. The fold workflow takes up fasta files and the 5utr workflow takes .csv files as input, containing columns: seq for the sequence information and 5utr for the mean ribosome load. These workflows both compute the secondary structure, based on the RNA sequences. For the latter, additionally the 5'UTR mean ribosome load information of each sequence is labeled for the given RNA sequence. This creates a subfolder called ´tfrecord´ for each dataset.

<Data-directory>
  |_____<Dataset1>
        |_________<Dataset1>.fa / <Dataset1>.csv
        |_________tfrecord
                  |____<Dataset1>.tfrecord
                  |____<Dataset1>.tfrecord.features.json
                  |____<Dataset1>.index
  |_____<Dataset2>
  .
  .

The data can then be loaded as follows:

from RNARepLearn.datasets import GFileDataset, GFileDatasetUTR
datasets = ["<Dataset1>","<Dataset2>"]
data = GFileDataset("<Data-directory>",datasets)

Example training setup

import torch
import torch_geometric

# Path to rfam base folder
rfam_dir = "../rfam/data/raw/processed/release-14.8"
rfams = ["RF00001","RF00005"]

from RNARepLearn.datasets import GFileDataset

dataset = GFileDataset(rfam_dir, rfams)

from RNARepLearn.utils import train_val_test_loaders

train_loader, val_loader, test_loader = train_val_test_loaders(dataset, 0.8, 0.1, 0.1)

from RNARepLearn.models import TestModel
model = TestModel()

from RNARepLearn.train import MaskedTraining
training = MaskedTraining(model, 10, 15, writer)

training.run(train_loader, val_loader)

Command line excecution

Define datasets and other parameters inside config.gin

rnareplearn --gin config.gin

Example data

In the example_data directory some explanatory notebooks are contained. Additionally, the related example data and pretrained models are provided here: https://www.dropbox.com/scl/fo/hen0xms1f8tux9yzef56r/h?rlkey=piy0yzbp5oeykiqfoe4dk3u7m&dl=0 .