DSSP and STR2 protein secondary structure prediction using a convolution-augmented Universal Transformer neural network.

System and Installation Requirements

The model was tested using the following configurations:

• Ubuntu 16.04.6

• CUDA 9.0

• CuDNN 7.6.5

• Python 3.6

• GPU: Nvidia K80

• Nvidia Driver: 440.64.00

To install the required Python packages, run:

$ pip install -r requirements.txt

SSP Universal Transformer

The ssp_universal_transformer/ directory contains the model code, trained models, examples of prediction results, and sample data. Please view the README in this directory for usage information and code details.

From the provided datasets, the model will take the following as input:

1. Amino acid sequence (max_len=700)
2. Position Specific Scoring Matrix (PSSM)
3. Target secondary structure labels

and output the predicted secondary structure sequences.

Example runs

Tensorflow output of trained model on test datasets. Includes raw counts for confusion matrix as well as accuracy scores.