lociPARSE: a locality-aware invariant point attention model for scoring RNA 3D structures

by Sumit Tarafder and Debswapna Bhattacharya

published in Journal of Chemical Information and Modeling

Codebase for our locality-aware invariant Point Attention-based RNA ScorEr (lociPARSE).

Installation

pip install lociPARSE

Or

git clone https://github.com/Bhattacharya-Lab/lociPARSE.git
cd lociPARSE
pip install .

Typical installation time should take less than a minute in a 64-bit Linux system.

Usage

Instructions for running lociPARSE:

from lociPARSE import lociparse
lp = lociparse()
score = lp.score("R1108.pdb")

Additional functionality

score.pMoL.show() #Returns pMoL value
score.pNuL.show() #Returns a list of pNuL values
score.pNuL.show(1) #Returns the pNuL value of 1st nucleotide
score.save("score.txt") #Saves the scores

1. Given an RNA pdb "R1108.pdb" as input, lociPARSE predicts both molecular-level lDDT (pMoL) and nucleotide-wise lDDT (pNuL) score.

2. Use show() function to print the pMoL or pNuL values.

3. Save the output in a provided filename of your choice("score.txt"). First line shows pMoL score. Each of the subsequent lines sepcify 2 columns: column-1: nucleotide index in PDB and column-2: predicted nucleotide-wise lDDT (pNuL) score.

Inference time for a typical RNA structure (~70 nucleotides) should take a few seconds.

Datasets

• The lists of IDs used in our training set, test sets and validation set used in ablation study are available here.
• Training set and test set of 30 independent RNAs were taken from trRosettaRNA.
• CASP15 experimental strctures and all submiited predictions were downloaded from CASP15.
• The set of 60 non-redundant RNA targets TS60 for hyperparameter optimization was in-house curated. See (https://doi.org/10.1093/biomethods/bpae047) for more details.

Training and evaluation materials

If you want to train or evaluate lociPARSE, please follow these initial steps:

• Download the necessary materials from here and place it in the root directory(/lociPARSE)

  wget https://zenodo.org/records/12729167/files/Materials.tar.gz
  

• Extract the Material.tar.gz folder

  tar -xvzf Materials.tar.gz --strip-components=1
  

• Make sure if you have installed appropriate torch version compatible with the CUDA version installed in your machine for GPU training. See here for more (https://pytorch.org/get-started/locally/).

Training lociPARSE

If you wish to train lociPARSE from scratch on our training set, please follow these steps:

• Download our training dataset Train.tar.gz from here and place it inside Input/Dataset folder.

• Extract the training dataset

  tar -xzvf Train.tar.gz
  

• Run the following command to train our architecture

  chmod a+x lociPARSE_train.sh && ./lociPARSE_train.sh > log.txt
  

  It will take approximately 16 hours to finish feature generation and 50 epochs of training on a single A100 gpu.

• The best model saved on validation loss will be placed inside the Model folder as "QAmodel_retrained.pt".

Evaluation of lociPARSE

If you want to generate our reported results in the paper from the provided predictions, follow these steps:

• To generate Tables 1-5, please run the following commands one by one.

  cd Evaluate
  python3 QA_eval.py Test30_CASP15 0
  

• You will find the corresponding results inside Evaluate/Results folder.

• To generate Supplementary Figures S1-S2, please run the following commands.

   cd Evaluate
   python3 draw.py
  

• Generated figures will be inside Evaluate/Figures folder.

If you want to predict the scores by lociPARSE from scratch and re-evaluate, follow these steps:

• Download our test datasets Test.tar.gz from here and place it inside Input/Dataset folder.

• Extract the folders

  tar -xzvf Test.tar.gz
  

• To predict and evaluate results on our two test sets Test30 and CASP15 (Tables 1-5), please run the following command.

  chmod a+x evaluate.sh && ./evaluate.sh Test30_CASP15 Model/QAmodel_lociPARSE.pt
  

• You will find the corresponding results inside Evaluate/Results folder.