This repository contains the official implementation of our paper titled SAINT-Angle: self-attention augmented inception-inside-inception network and transfer learning improve protein backbone torsion angle prediction.

Protein structure provides valuable insights into how proteins function within living organisms as well as how proteins interact with one another. Prediction of protein backbone torsion angles (φ and ψ), therefore, is a key subproblem in protein structure prediction. However, existing experimental methods for determining backbone torsion angles are costly and time-consuming. Hence, the concerned community is focusing on developing computational methods for predicting torsion angles.

In this paper, we present SAINT-Angle, a highly accurate method for protein backbone torsion angles prediction. SAINT-Angle uses a self-attention based deep learning architecture called SAINT which was previously developed in our lab for the protein secondary structure prediction. We extended and improved the existing SAINT architecture as well as used transfer learning for predicting the backbone torsion angles. We conducted a thorough analysis and compared the performance of SAINT-Angle with contemporary state-of-the-art prediction methods on a collection of publicly available benchmark datasets, namely, TEST2016, TEST2018, CAMEO, and CASP. The experimental results suggest the notable improvements that our proposed method has achieved over the best alternate methods.

In order to run SAINT-Angle, create a workspace directory. Download the inference.py script and place it inside the workspace directory.

You may create a separate Conda environment for running SAINT-Angle and install TensorFlow (2.6.0 version) and Keras (2.6.0 version). We have tested our code with the aforementioned versions of TensorFlow and Keras. For faster inference, we recommend install the GPU version of TensorFlow.

You will find here all the pretrained model weights essential for running SAINT-Angle. Download these model weights and place them in a folder named models inside the workspace directory. There are, in total, 12 model weights files (with .h5 extension). The details on each model are given as follows.

Here, Base means the feature set consisting of PSSM, HMM, and PCP features from SPOT-1D. Win10, Win20, and Win50 mean the window features from SPOT-Contact. ProtTrans means the extracted features from ProtTrans [Github]. ESIDEN means the feature set consisting of PSSM, AA, PCP, DC, RE, PSSP, and RBP features from ESIDEN.

In order to predict backbone torsion angles using SAINT-Angle, you may use any of the given models. Also, there are 2 ensembles of different combinations of given models available in SAINT-Angle for prediction purpose. The details on each ensemble are given as follows.

By providing the appropriate command line argument (discussed in the subsequent section), you may use the aforementioned (ensemble of) models for torsion angles prediction.

1. Open up a terminal window and navigate to your workspace directory using cd command (if necessary).

2. To run SAINT-Angle for predicting protein backbone torsion angles, type in the terminal the following command.

   python inference.py

   This command runs the inference.py with default arguments using models/model_1.h5 as model and datasets/TEST2016_A as dataset.

3. You may change the model that you will use for the inference by providing the model parameter.

   -m MODEL_NAME or --model MODEL_NAME sets the inference script to use either models/MODEL_NAME.h5 model (in case of inference with single model) or MODEL_NAME ensemble model (in case of inference with ensemble of models) for the prediction. You must use one of the following MODEL_NAME to specify the model that you want use for the inference.

   model_1	model_2	model_3	model_4	model_5	model_6	model_7
   model_8	model_9	model_10	model_11	model_12	ensemble_3	ensemble_8

4. You may change the dataset that you will use for the inference by providing the dataset parameter.

   -d DATASET_NAME or --dataset DATASET_NAME sets the inference script to use features from the datasets/DATASET_NAME dataset to predict the backbone torsion angles of protein(s) belonging to DATASET_NAME dataset.

5. --verbose sets the inference script to print out detailed messages and --output sets the inference script to output predicted angles in predictions folder inside workspace directory (discussed in the subsequent section).

6. -h or --help sets the inference script to display help message.