qepsilon

Repository for Protein quality assessment with graph convolution guided by a loss function designed for high quality decoys

Dependencies

Python 3.9.16

Numpy 1.23.5

Pandas 1.5.2

Matplotlib 3.6.2

Pytorch 1.13.1

Cuda 11.6

Pytorch Lightning 1.9.0

Pytorch Geometric 2.2.0

Biopython 1.81

Sklearn 1.2.0

torchsampler 0.14.1

Initial Setup

First install dependencies in a conda environment and clone the repository-

  git clone https://github.com/soumyadip1997/Q-epsilon.git
       
  cd Q-epsilon

Data Preprocessing and Train-Val-Test split

It can be done either from

1. Scratch

or

2. Use the already prepared Train,Val and Test set and Preprocessed data

From scratch

First we download the 3D structure decoy files.

CASP9 - https://zenodo.org/record/7697275/files/CASP9.zip?download=1

CASP10 - https://zenodo.org/record/7697275/files/CASP10.zip?download=1

CASP11 - https://zenodo.org/record/7697275/files/CASP11.zip?download=1

CASP12 - https://zenodo.org/record/7697275/files/CASP12.zip?download=1

CASP13 - https://zenodo.org/record/7697275/files/CASP13.zip?download=1

CASP14 - https://zenodo.org/record/7697275/files/CASP14.zip?download=1

Then we unzip the folders using

       unzip <CASP_FILE> -d Q-epsilon/

where <CASP_FILE> represents the downloaded CASP zip files

Then we make a new folder by the name "Features" inside Q-epsilon

       mkdir Features

Make sub directories under Features-

       sh create.sh

Then we extract the features of all the decoys by running the scripts in Feat_script directory-

For extracting one hot encodings of all the atoms of decoys

  python atom_feat.py --decoy-location Q-epsilon/ --output-location Q-epsilon/Features/

For extracting the gdtts of all decoys

  python gdtscores.py --decoy-location Q-epsilon/ --output-location Q-epsilon/Features/

For extracting the same and different residue atom neighbours and all the residue neighbours of all decoys

  python neigh.py --decoy-location Q-epsilon/ --output-location Q-epsilon/Features/

For extracting the number of atoms present inside each residue of all decoys

  python res_number.py  --decoy-location Q-epsilon/ --output-location Q-epsilon/Features/

For extracting transformer feature for each residue of all decoys

  python transformer_feat.py --decoy-location Q-epsilon/ --output-location Q-epsilon/Features/

For extracting the LDDT of all decoys run the following script in each of the CASP directory

  python calc_lddt.py --table-loc *.xz --output-location Q-epsilon/Features/ --group-info output.csv

Parameters

--decoy-location Location of the unzipped CASP folders

--output-location Location for storing the output features having the required sub-directories

--table-loc Location of the table.pkl.xz of each CASP

--group-info Location of group info present in output.csv

Then we need to create Train,Val and Test Set. Since we are training with both GDTTS and LDDT scores we have two sets of train, validation and test datasets.

GDTTS

Run the scripts inside Train_Val_Test directory.

To create the Train and Validation set first run

  python create_possible_list.py --same-res-atom-neigh Q-epsilon/Features/GRAPHNEIGH/Same_Res_Index_ --res-neigh Q-epsilon/Features/GRAPHNEIGH/Residue_Neigh_ --gdtts Q-epsilon/Features/GDT_TS/gdtts_ --atom-one-hot Q-epsilon/Features/ATOM/atom_one_hot_ --path-res-trans Q-epsilon/Features/TRANS/Trans_

Then run

  create_train_val.py --output-path Q-epsilon

To create the Test set first run

  python process_CASP13_14.py --decoy-location Q-epsilon/

Then run

  python create_possible_list_CASP13.py --same-res-atom-neigh Q-epsilon/Features/GRAPHNEIGH/Same_Res_Index_ --res-neigh Q-epsilon/Features/GRAPHNEIGH/Residue_Neigh_ --gdtts Q-epsilon/Features/GDT_TS/gdtts_ --atom-one-hot Q-epsilon/Features/ATOM/atom_one_hot_ --path-res-trans Q-epsilon/Features/TRANS/Trans_ --output-path Q-epsilon
  
  python create_possible_list_CASP14.py --same-res-atom-neigh Q-epsilon/Features/GRAPHNEIGH/Same_Res_Index_ --res-neigh Q-epsilon/Features/GRAPHNEIGH/Residue_Neigh_ --gdtts Q-epsilon/Features/GDT_TS/gdtts_ --atom-one-hot Q-epsilon/Features/ATOM/atom_one_hot_ --path-res-trans Q-epsilon/Features/TRANS/Trans_ --output-path Q-epsilon

This will create the test sets for CASP13 and 14 respectively.

Parameters

--same-res-atom-neigh - Location of same residue atom neighbours

-res-neigh - Location of residue neighbours

--gdtts - Location of gdtts

--atom-one-hot - Location of one hot atom embeddings

--path-res-trans - Location of residue features

--output-path - Location for saving the files generated by create_train_val.py

--decoy-location - Location of the CASP Folders

LDDT

Since for LDDT we are using the same target set as DeepUMQA(https://academic.oup.com/bioinformatics/article/38/7/1895/6520805) we are going to use Train_LDDT.npy, Val_LDDT.npy as the train and validation sets and Test_CASP13_LDDT.npy and Test_CASP14_LDDT.npy as the test sets.

Alternative: Downloading the preprocessed data and using the already prepared Train-Val-Test File

To save time all the above features can be downloaded from the links below-

transformer feature for residue - https://colostate-my.sharepoint.com/:u:/r/personal/soumya16_colostate_edu/Documents/TRANS.zip?csf=1&web=1&e=zYacPF

one hot encodings of atoms - https://zenodo.org/record/8271199/files/ATOM.zip?download=1

same and diferent residue atom neighbours - https://zenodo.org/record/8277479/files/GRAPHNEIGH.zip?download=1

residue neighbours - https://zenodo.org/record/7703112/files/NEIGH_RES.zip?download=1

number of atoms inside each residue - https://zenodo.org/record/8271199/files/Atomfreq.zip?download=1

gdtts - https://zenodo.org/record/8271199/files/GDT_TS.zip?download=1

LDDT- https://zenodo.org/record/8271212/files/LDDT.zip?download=1

Then unzip the files and store them inside Features Folder

  unzip <Filename.zip> -d Q-epsilon/Features/

where <Filename.zip> are the different files that we downloaded (ATOM.zip,ATOMfreq.zip,GDT_TS.zip,GRAPHNEIGH.zip,RES_NEIGH.zip)

Then use the Train_Val_Test.zip file which contains all the information for the train,validation and test sets.

Unzip the Train_Val_Test.zip.

  unzip Train_Val_Test.zip -d Q-epsilon/

It contains the following files

Train.npy - Train set

Val.npy - Validation Set

Test_CASP13_new.npy - Test Set with CASP13 decoys

Test_CASP14_new.npy - Test Set with CASP14 decoys

valid_targets_CASP13.npy - CASP13 targets

valid_targets_CASP14.npy - CASP14 targets

Training

GDTTS

Training is done in two steps:

First the GCN runs with L1-Loss for 50 epochs and the best model is saved as best_model_l1.ckpt

  python Train.py --batch-size 70 --epochs 50 --workers 12 --seed 42 --devices 1 --nodes 1 --loss-type 0 --train-set Train.npy --val-set Val.npy --test-set Test_CASP13_new.npy  --gdtts Features/GDT_TS/gdtts_ --atom-one-hot Features/ATOM/atom_one_hot_ --same-res-atom-neigh Features/GRAPHNEIGH/Same_Res_Index_  --diff-res-atom-neigh Features/GRAPHNEIGH/Diff_Res_Index_  --res-neigh Features/GRAPHNEIGH/Residue_Neigh_ --path-res-trans Features/TRANS/Trans_ --res-no Features/Atomfreq/atomfreq_ --save-model best_model_l1.ckpt

Then the same GCN runs with our $\epsilon$ modified L1-Loss for another 10 epochs using the best model obtained from the L1-Loss i.e. best_model_l1.ckpt. The final model is saved as best_model.ckpt

  python Train.py --batch-size 70 --epochs 10 --workers 12 --seed 42 --devices 1 --nodes 1 --loss-type 1 --train-set Train.npy --val-set Val.npy --test-set Test_CASP13_new.npy  --gdtts Features/GDT_TS/gdtts_ --atom-one-hot Features/ATOM/atom_one_hot_ --same-res-atom-neigh Features/GRAPHNEIGH/Same_Res_Index_  --diff-res-atom-neigh Features/GRAPHNEIGH/Diff_Res_Index_  --res-neigh Features/GRAPHNEIGH/Residue_Neigh_ --path-res-trans Features/TRANS/Trans_ --res-no Features/Atomfreq/atomfreq_ --load-model best_model_l1.ckpt --save-model best_model.ckpt

LDDT

For training with LDDT scores we use the pretrained GDTTS model.

  python Train_LDDT.py --workers 12 --batch-size 70 --save-model best_model_lddt.ckpt --load-model best_model.ckpt  --devices 1 --lr 0.001 --epochs 50 --loss-type=1 --train-set Train_LDDT.npy --val-set Val_LDDT.npy --gdtts Features/LDDT/LDDT_ --atom-one-hot Features/ATOM/atom_one_hot_ --same-res-atom-neigh Features/GRAPHNEIGH/Same_Res_Index_  --diff-res-atom-neigh Features/GRAPHNEIGH/Diff_Res_Index_  --res-neigh Features/GRAPHNEIGH/Residue_Neigh_ --path-res-trans Features/TRANS/Trans_ --res-no Features/Atomfreq/atomfreq_

Parameters

--batch-size - Batch size

--epochs - Number of epochs for training

--workers - Number of workers

--seed - Setting the seed value of pytorch

--devices - Number of gpus to be used

--nodes - Number of nodes to be used

--loss-type - Specify which loss function to be used i.e. 0 for L1-Loss and 1 for $\epsilon$ modified L1-Loss

--train-set - Train set location

--val set - Validation set location

--test-set - Test set location

--gdtts - Location with the common prefix of GDTTS/LDDT files

--atom-one-hot - Location with the common prefix for the one hot encodings of atom files

--same-res-atom-neigh - Location with the common prefix for the same residue atom neighbour files

--diff-res-atom-neigh - Location with the common prefix for the different residue atom neighbour files

--res-neigh - Location with the common prefix for residue neighbours

--path-res-trans - Location with the common prefix for the transfomer features of residues

--res-no - Location with the common prefix for the Atomfreq features

--load-model - Location to a model for loading

--save-model - Location for saving a model

Testing

We use the best model saved by the $\epsilon$ modified L1-Loss network for testing.

To save training time, the best GDTTS model can be downloaded from https://zenodo.org/record/7697220/files/best_model.ckpt?download=1 and the best LDDT model can be downloaded from https://zenodo.org/record/7760820/files/best_model_LDDT.ckpt?download=1

After training for a total of 60 epochs or downloading the model run the following for testing on CASP13/CASP14-

GDTTS

  python Test.py --workers 12 --model-path best_model.ckpt --devices 1 --nodes 1 --test-set Test_CASP13_new.npy/Test_CASP14_new.npy  --gdtts Features/GDT_TS/gdtts_ --atom-one-hot Features/ATOM/atom_one_hot_ --same-res-atom-neigh Features/GRAPHNEIGH/Same_Res_Index_  --diff-res-atom-neigh Features/GRAPHNEIGH/Diff_Res_Index_  --res-neigh Features/GRAPHNEIGH/Residue_Neigh_ --path-res-trans Features/TRANS/Trans_ --res-no Features/Atomfreq/atomfreq_ --result-file result_13_GDTTS.csv/result_14_GDTTS.csv

LDDT

  python Test.py --workers 12 --model-path best_model_LDDT.ckpt --devices 1 --nodes 1 --test-set Test_CASP13_LDDT.npy/Test_CASP14_LDDT.npy  --gdtts Features/LDDT/LDDT_ --atom-one-hot Features/ATOM/atom_one_hot_ --same-res-atom-neigh Features/GRAPHNEIGH/Same_Res_Index_  --diff-res-atom-neigh Features/GRAPHNEIGH/Diff_Res_Index_  --res-neigh Features/GRAPHNEIGH/Residue_Neigh_ --path-res-trans Features/TRANS/Trans_ --res-no Features/Atomfreq/atomfreq_ --result-file result_13_LDDT.csv/result_14_LDDT.csv

This saves all the results to the file specified by --result-file argument which in this case is either result_13.csv or result_14.csv.

Parameters

--workers - Number of workers

--model-path - Location to a model for loading

--devices - Number of gpus to be used

--nodes - Number of nodes to be used

--test-set - Test set location

--gdtts - Location with the common prefix of GDTTS/LDDT files

--atom-one-hot - Location with the common prefix for the one hot encodings of atom files

--same-res-atom-neigh - Location with the common prefix for the same residue atom neighbour files

--diff-res-atom-neigh - Location with the common prefix for the different residue atom neighbour files

--res-neigh - Location with the common prefix for residue neighbours

--path-res-trans - Location with the common prefix for the transfomer features of residues

--res-no - Location with the common prefix for the Atomfreq features

--result-file - Location of the file with filename for storing result

Plotting and calculating Result

For calculating the pearson and spearman correlation scores and plotting data run the following -

GDTTS

  python plot_score.py --result-file result_13_GDTTS.csv/result_14_GDTTS.csv --targets valid_targets_CASP14.csv/valid_targets_CASP13.csv --plot-name CASP13.pdf/CASP14.pdf

LDDT

  python calc_score.py --result-file Result_CASP14_LDDT.csv/Result_CASP13_LDDT.csv --decoys CASP14_LDDT.csv/CASP13_LDDT.csv

Parameters

--result-file - Location to the file produced by Test.py

--decoys - Location to the file that contains the decoy ids of CASP13 and CASP14 (CASP14_LDDT.csv/CASP13_LDDT.csv)

--targets - Location to the file that has target information

--plot-name - Location to save the plots

Other Files And Additional Information

prep_CASP15_decoy.py - Separates the multimeric complexes of CASP15 into their respective single sequence decoys wrt single sequence targets.

Test_CASP15_new.npy - Test set of CASP15 decoys

To get the alphafold generated decoys on a subset of CASP15 fasta files - https://zenodo.org/record/8270670/files/alphafold_CASP_output.zip?download=1

Test_alphafold_new.npy - Test set consisting of decoys produced by Alphafold2 when run on the Fasta Files of CASP15 targets.

Local lDDT scores have been calculated using the command line tools of SWISS_MODEL( https://www.openstructure.org/docs/2.5/actions/#comparing-two-structures ).

CASP15 - https://zenodo.org/record/8271136/files/CASP15.zip?download=1