Multi-channel PolyConvNet

1D convolutional neural net for predicting the lamellar period of copolymers based on sequence of beads.

Team Borides

Members: Ruijie Zhu, Kastan Day, Aria Coraor, Seonghwan Kim, Jiahui Yang

Directories

.
├── data                                             # input data file for feature generation
├── features                                         # folder containing scripts used to generate features and all features used to train the neural net
├── Multi-channel PolyConvNet.ipynb                  # code used to train / test the Multi-channel PolyConvNet
├── Multi-channel PolyConvNet VAE.ipynb                  # code used to train / test the Multi-channel PolyConvNet with VAE features
├── models                                           # folder containing all trained models
├── LICENSE
└── README.md

ML Features

1. Sliding window features

29-dimensional feature used to capture the activation of polymer sequence

2. Kernels

• Exponential kernel: 30-dimensional feature used to capture the interaction at two ends
• Cosine kernel: 15-dimensional feature used to capture the periodicity of sequence

3. VAE features

4-dimensional feature generated using the Variational Autoencoder model

4. Interaction parameter

Multi-channel PolyConvNet

The model consists of a series of convolution layer and fully connected layers that extract patterns from the polymer sequence.

Model Performance

Computational Efficiency

Feature Generation	Time (min)
Sliding Window Features (4 channels)	0.5
Kernel Features	0.08
VAE Features	30

Model Training/Validation	Time (min)
Training	1
Validation	0.02

• All runtimes reported using Theta GPU