ml-project-2-chopin-deux

This README describes the main scripts and data-files present in this project folder.

The relevance of each directory and file will be apparent when the main goal of this project are kept in mind: Comparing the classification accuracy of our protein surface-based machine learning algorithm to commonly used methods, mainly Rosetta.

Compare the model with Rosetta

Install virtual enviroment using requirements.txt

Step 1: Clone the git repository to your local device.

Step 2: Open /scripts/Test_set_and_comparison.ipynb and run the notebook.

Step 3: Find the comparison at the end of the notebook.

*note: if for some reason, the test data does not work, it is also provided in the google drive below

Used scripts for main goal Comparison

To perform a fair comparison, identical datasets are required. To do this, one needs to know which residues were used for the single amino acid retrieval in the RosettaSurf publication.

• In data/rosetta_slurm there are slurm files for every amino acid. scripts/data_comparison_prep/get_rsurf_resi.py returns a csv of all residues used by Andreas in RosettaSurf. It reads his slurm commands to get these. data/rosetta_res.csv is complete and contains 200 of each of the 20 AA's. Also columns for pdb_id, aa type and position.

• In scripts/data_comparison_prep/get_specific_aa.py, the test and training set is created to compare the performance of Chopin software to the RosettaSurf. The test set is what we will perform predictions on, the ~200 * 20 residues from 1495 pdb's. All other residues from those pdb's will serve as a training set. The pdbs are loaded in ascending order by pdb, chain, and then residue.

• The /scripts/data_comparison_prep/zander.py adds new features to the data set, including geometric angles between the residues.

• /scripts/RosettaSurf_benchmark.ipynb is the script from the original RosettaSurf github which contains the data for the Rosetta comparison.

Generating and retrieving surfaces using dMaSIF

• /scripts/data_comparison_prep/find_pdbs.py allows you to retrieve the relevant pdbs from the cluster after maSIF has generated the relevant surface data files.

---

Additional scripts for checks

• /scripts/data_comparison_prep/get_pdb_set.py A simple script to compare the pdb's used in the RosettaSurf comparison to the training and test set of dMaSIF.

• After generating surfaces with dMaSIF, we checked how many were succesfully retrieved, see /scripts/data_comparison_prep/pdb_id_check.py. It saves a list of pdb_ids for which no surfaces were generated.

• The /scripts/masif_plugin/ folder contains scripts to visualise protein surfaces generated using the MaSIF framework in PyMol. Most of these scripts are from the original MaSIF publication (Gainza et al., 2020). The /scripts/masif_plugin/dmasif_pymol.py script is written by Arne Schneuing and allows for visualization of our surfaces.

Helpers and other auxiliary files

• The scripts/data_comparison_prep/get_aa_helper.py contains helper functions that are then used in scripts/data_comparison_prep/get_specific_aa.py to obtain training and testing dataset for comparison with RosettaSurf.
• The scripts/data_comparison_prep/get_aa_loader.py is used to load, sort, and pre-clean the dataframe containing Pdbs used in Rosetta and the data points containing surfaces' features by removing the unnecessary columns.

---

Procedure to train the model

Step 1: Get the training data from the google drive (https://drive.google.com/drive/folders/16iuOe2GBY9-flGjP7FfbjYjIabSq2Uy1?usp=sharing) Place all the directories in the "data" directory.

Step 2: Open /scripts/train.ipynb on a cluster and use the tensorboard to monitor the progress. For instructions on how to open a notebook and the cluster and a tensorboard on a cluster see the instructions below:

Tensorboard

module load gcc
module load python
source amino_env/python_env/bin/activate

Change directory to the scripts directory.

Get hostname (it is a number) with:

hostname -I | awk '{print $1}'

Fill it in here:

tensorboard --logdir=runs --host hostname --port 8888

On you local device: ssh -L 5555:hostname:8888 [email protected]

Jupyter notebook

jupyter notebook --no-browser --port=1234 --ip=$(hostname -i)
ssh -NL 1234:hostname:1234 [email protected]

Copy the url that will appear in remote terminal and paste it in your browser.