rna-llm-comparison

Code to build a secondary structure RNA predictor model based on LLM representations.

Prerequisites

git clone https://github.com/lucianozablocki/rna-llm-comparison
cd rna-llm-comparison/

With an Anaconda working installation, run:

conda env create -f environment.yml

This should install all required dependencies.

conda activate rna-llm

You will also need to download representations, check the links in Embeddings section

Generate train and test partitions for ARchiveII

We will use a family fold approach: for each one of the 9 families under consideration, one will be left out for test. This way, we have 9 train and test partitions. To generate them, run:

python gen_splits.py

This will create the partitions under data/archiveII_famfold.

Train/test scripts

Save the embeddings in the data directory (for example data/rnafm_ArchiveII.h5 format). Then edit the scripts/run_famfold.py to specify the embedding name (for example emb_name = "rnafm").

ArchiveII fam-fold train and test:

python scripts/run_famfold.py

bpRNA TR0/TS0 train and test:

python scripts/run_bprna.py

Only train the model

The train_model.py script contains a few command line parameters, an example to run the train for 15 epochs using RiNALMo-generated representations, with tmRNA family as held out test partition set is:

python train_model.py --train_partition_path data/archiveII_famfold/tmRNA/train.csv --test_partition_path data/archiveII_famfold/tmRNA/test.csv --out_path <directory_where_to_write_results>

Embeddings