Deep learning analysis of large-scale sequence data in health and disease

As is the case in many other information processing domains, deep learning frameworks are emerging as powerful and highly effective tool in Bioinformatics and health analytics, including in the analysis of genetics, proteomics and clinical text data. Key to the success of deep learning solutions is their ability to extract complex problem-relevant features from raw input representations. Often, these input representations are sequential in nature (e.g. the representation of a DNA sequence, a string representation of a protein’s amino acid sequence, and unstructured textual data relating to clinical health records or biomedical literature).

Recent research in the domain of Natural Language Processing (NLP) has demonstrated the power of large scale artificial neural networks to represent, process and extract information from text data. In particular, pre-training has emerged as an effective strategy for training such networks – conceptually, pretraining involves training a model using a very large amount of unlabelled data and a very general task, before fine-tuning that model on a more specific task.

In this project, we will adapt NLP deep learning models to the task of extracting useful information from heterogeneous sources of sequence data relating to health and disease. Although highly successful in NLP, there has been little work to date applying the large-scale sequence processing and pretraining framework to other kinds of sequence structure, including the various kinds of sequence structure found in pharmacology, biochemistry, genetics and proteomics. Our goal is to develop deep learning solutions that allow us to better understand and represent these kinds of data, and the clinically relevant relationships that can exist between them.