pysster

pysster implements convolutional neural networks using TensorFlow to classify and interpret biological sequence and annotated secondary-structure data (DNA, RNA, and artificial epistatic sequences), enabling motif discovery and prediction of functional regions such as RNA A-to-I editing sites.

Key Features:

• Convolutional Neural Networks (CNNs): Implements CNN architectures tailored to capture local dependencies and patterns in sequential biological data.
• TensorFlow backend: Uses TensorFlow for model implementation and training.
• User-defined deep network construction: Supports construction of custom deep neural network architectures for sequence analysis.
• Supported sequence types: Handles DNA, RNA, annotated secondary-structure sequences, and artificial epistatic sequence datasets.
• Training, evaluation and interpretation workflows: Provides routines for model training, performance evaluation, and interpretation of learned representations.
• Visualization of predictions: Produces visualizations of network predictions to aid interpretation of model outputs.
• Motif learning: Learns sequence motifs and optionally structure motifs for classification tasks.
• Hyper-parameter optimization and defaults: Includes sensible default parameters and a hyper-parameter optimization procedure.

Scientific Applications:

• RNA A-to-I editing classification: Applied to classify RNA regions subject to adenosine-to-inosine (A-to-I) editing.
• Generalizability on artificial epistatic datasets: Used to build and evaluate neural network models on artificial epistatic sequence datasets.

Methodology:

Implements convolutional neural networks using TensorFlow, supports user-defined architectures, motif learning, model visualization, and hyper-parameter optimization.