Ab initio Protein Folding Using guided by Contact Prediction using Multi-Perspective Deep Convolutional Neural Networks

With the open-source tool DNCON22 as a reference, we developed four different deep convolutional neural networks (CNNs): 1) a basic deep CNN, 2) a dilated CNN, 3) a separable CNN, and 4) a basic deep CNN trained to predict contacts at the distance thresholds of 6, 8 and 10 Angstroms. These networks were trained and tested on the standard dataset of 1426 proteins discussed in the DNCON2 method. The architecture of the basic deep CNN method consists of 17 layers of 64 filters of size 3x3 and the last output layer with one 3x3 filter. After each layer, the activations are padded with zeros so that the input dimensions (300x300) are maintained through all the layers including the output of the last layer. The basic deep CNN model has a total of 625,921 trainable parameters: 64 3x3 filters on the 56 channels give along with 64 bias values result to 32,320 parameters, 16 layers of 64 3x3 filters on 64 channel activations with 64 bias values at each layer result in a total of 590,848 parameters, 128 parameters for batch normalization at each of the 17 layers result in 2,176 parameters, and one 3x3 filter in the last layer on 64 activation channels along with a bias results in 577 parameters. The architecture of the dilated CNN consists of 13 regular convolutional layers each with 64 3x3 filters followed by two dilated CNN layers. Both dilated CNN layers consist of 64 3x3 filters with a dilation rate of 2. The last layer is a single 3x3 filter. The dilated CNN model has 551,809 parameters total. Similarly, the separable CNN model has its first layer of 64 3x3 filters, followed by 15 depthwise separable CNN layers (SeparableConv2D in Keras) each with 64 3x3 filters, and the last layer with a depthwise separable CNN with 1 3x3 filter. This model has a total of 101,185 parameters. The fourth model is an extension of the basic deep CNN model trained to predict contacts at the thresholds of 6 and 10 Angstroms at the same time. To achieve this, we replace the last layer with three parallel CNN blocks each consisting of two convolutional layers - first layer with 32 3x3 filters and second with one 3x3 filter as the output. Each of these three outputs separately predict contacts at 6, 8, and 10 Angstroms. When calculating the binary cross entropy loss, we weight these outputs such that the weights are 0.25, 1.0, and 0.25 for 6 Angstroms predictor, 8 Angstroms predictor, 10 Angstroms predictor respectively. Predictions by the four methods are averaged to predict the final set of contacts. Finally, we used top 2L long-range and medium-range contacts (L is the length of a protein) to predict five models using the CONFOLD method.