EnQA

A 3D-equivariant graph neural network for protein structure accuracy estimation. It can predict the quality of both tertiary and quaternary structures.

Requirements:

biopandas==0.3.0dev0
biopython==1.79
numpy==1.21.3
pandas==1.3.4
scipy==1.7.1
torch==1.10.0

Install Transformer protein language models by the following command:

pip install git+https://github.com/facebookresearch/esm.git

equivariant_attention (Optional, used by models based on SE(3)-Transformer only)

pdb-tools (Optional, used by models with multiple chains only)

You may also need to set execution permission for utils/lddt and files under utils/SGCN/bin.

Note: Currently, the dependencies support AMD/Intel based system with Ubuntu 21.10 (Impish Indri). Other Linux-based system may be also supported but not guaranteed.

EnQA-MSA (recommended for estimating the quality of AlphaFold predicted tertiary structures)

usage: EnQA-MSA.py [-h] --input INPUT --output OUTPUT

Predict model quality and output numpy array format.

optional arguments:
  -h, --help       show this help message and exit
  --input INPUT    Path to input pdb file.
  --output OUTPUT  Path to output folder.

The assumption is that the B-factor colulmn of the input PDB file stores the per-residue plddt scores predicted by AlphaFold. Below is a demo example of running EnQA-MSA:

python EnQA-MSA.py --input example/enqa-msa/1A09A.pdb --output example/output/

Model training for EnQA-MSA

First, generate feature files and embeddings from MSA-Transformer. Example for data_list_file which contains list for models.

# PDBs in <input_pdb_folder> and <reference_pdb_folder> should have same file name for the same target, for example: path/to/input/test.pdb and path/to/ref/test.pdb

python3 generate_data.py <input_pdb_folder> <reference_pdb_folder> <feature_save_folder> <data_list_file> 
python3 generate_embedding.py <reference_pdb_folder> <embedding_save_folder>

After all feature files are generated, here is how to train the model:

python3 train_enqa_msa.py --core_data <feature_save_folder> --attn <embedding_save_folder> --train <data_list_file for training> --validation <data_list_file for validation> --output <model_save_folder> --epochs 60

EnQA assisted with AlphaFold2 predictions

usage: python3 EnQA.py [-h] --input INPUT --output OUTPUT --method METHOD [--cpu] [--alphafold_prediction ALPHAFOLD_PREDICTION] [--alphafold_feature_cache ALPHAFOLD_FEATURE_CACHE] [--af2_pdb AF2_PDB]

Predict model quality and output NumPy array format.

optional arguments:
  -h, --help                  Show this help message and exit
  --input INPUT               Path to input pdb file.
  --output OUTPUT             Path to output folder.
  --method METHOD             Prediction method, can be "ensemble", "EGNN_Full", "se3_Full", "EGNN_esto9" or "EGNN_covariance". Ensemble can be done listing multiple models separated by comma.
  --alphafold_prediction      Path to alphafold prediction results.               
  --alphafold_feature_cache   Optional. Can cache AlphaFold features for models of the same sequence.
  --af2_pdb AF2_PDB           Optional. PDBs from AlphaFold predcition for index correction with input pdb when input PDB only contains partial sequence of the AlphaFold results.
  --cpu                       Optional. Force to use CPU.

Example usages

python3 EnQA.py --input example/model/6KYTP/test_model.pdb --output outputs/ --method EGNN_Full --alphafold_prediction example/alphafold_prediction/6KYTP/

If you want to run models based on the SE(3)-Transformer, then the Python package equivariant_attention is required and should be installed following Fabian's implementation.

Example:

python3 EnQA.py --input example/model/6KYTP/test_model.pdb --output outputs/ --method se3_Full --alphafold_prediction example/alphafold_prediction/6KYTP/  

Generating AlphaFold2 models for assisted quality assessment

For generating models using AlphaFold2, an installation of AlphaFold2 following its Official Repo is required. For our experiments, we use its original model used at CASP14 with no ensembling (--model_preset=monomer), with all genetic databases used at CASP14 (--db_preset=full_dbs), and restricts templates only to structures that were available at the start of CASP14 (--max_template_date=2020-05-14).

Model training for EnQA assisted with AlphaFold2 predictions

First generate 5 AlphaFold reference models per Generating AlphaFold2 models for assisted quality assessment. Then generate the labels and features after you have the predicted results from AlphaFold and the corresponding native PDBs:

python3 process.py --input example/model/6KYTP/test_model.pdb --label_pdb example/label/6KYTP.pdb --output outputs/processed --alphafold_prediction example/alphafold_prediction/6KYTP/

Code in train.py provides a basic framework to train the EGNN_full model with Pytorch, After all feature files for training and validation are generated, suppose the processed features files(in .pt format) are saved in path/to/train/ and path/to/validation/, here is an example to train the model:

python3 train.py --train path/to/train/ --validation path/to/validation --output outputs/ --epochs 60

Geometric feature generation

The featurizers from Spherical graph convolutional networks (S-GCN) are used to process 3D models of proteins represented as molecular graphs. Here we provide the voronota and spherical harmonics featurizer for Linux.

If you need to rebuild the voronota for a different system, please check out the S-GCN Repo.

Also, there are binaries built for featurizer under a different system. (Currently, only MacOS and Linux are supported)

Applying EnQA to protein complex (quaternary) structures with multiple chains

EnQA was trained on tertiary structures of singel-chain proteins to predict their quality. It was not trained on protein complex strucures. But interestingly it can be applied to evaluate the quality of protein complex strucutres by treating them as single-chain protein structures.

For EnQA-MSA, you can preprocess the input PDB with the mergePDB function we provided to convert it into a "merged single chain PDB" and make that as the input PDB.

For EnQA assisted with AlphaFold2, you can provide protein complexes as input, and no additional work is required.

Reference

Chen C, Chen X, Morehead A, Wu T, Cheng J. 3D-equivariant graph neural networks for protein model quality assessment. Bioinformatics. 2023 Jan 13:btad030. doi: 10.1093/bioinformatics/btad030.