RECURRRENT NEURAL NETWORK BASED MODEL FOR AUTISM SPECTRUM DISORDER PREDICTION USING CODON ENCODING

Abstract

Deep learning methods are noteworthy tools that go together with traditional machine learning techniques to enable computers learn from data and create smarter applications. Deleterious gene classification is an important task in a standard computational framework for biomedical data analysis. As gene sequences are high dimensional and do not represent explicit attributes for computational modelling, extracting features from them becomes a complex task. Recently neural deep learning architectures automatically extract valuable features from input patterns. The principal idea of this work is to exploit the power of Recurrent Neural Networks (RNN) to learn sequential patterns through high-level information associated with observed signals which in turn can be used for classification. Classification of affected genes that cause disease like Autism-spectrum disorder (ASD) is a noteworthy challenge in biomedical research. Long Short Term Memory (LSTM) units go well with sequence-based tasks with long-term dependencies and hence this work examines a stacked LSTM architecture for classifying genes causing ASD. The model is trained and tested with two hand crafted datasets and a codon encoded dataset. Experiments revealed the superiority of these advanced recurrent units compared to the traditional Deep Neural Networks and Bi-directional RNNs distinctively with codon encoded dataset