A multimodal deep learning model to predict individual cancer patient survival probabilities

Dr Peter Hall, Dr Colin McLean, Prof Joanne Edwards

Can a deep learning framework that accommodates complex cancer ‘omics data, such as gene-expression data, with clinical phenotypic data enhance our ability to predict cancer patient survival?

Deep neural networks are versatile machine learning techniques, which make it possible to build frameworks which learn from multi-modal data that can outperform traditional modelling methods. Such methods require large amounts of data, which has limited their use to date. New data opportunities in Scotland will now enable us to enhance research patient cohorts containing detailed ‘omics data with rich NHS clinical data on patient characteristics, cancer characteristics, treatment information and a range of clinical outcomes. Combining real-world clinical data with existing research data will allow us to address this question, initially with the priority cancer types of colorectal cancer, mesothelioma and ovarian cancer.

The project will use software packages such as TensorFlow and Keras (R, Python). Methods will be explored for handling dimensional reduction of genomics, for example, by using unsupervised techniques, building clustered gene-expression networks, the use of gene and pathway layers in the neural network model.

Improving our ability to predict disease-specific survival and other relevant patient outcomes using the full breadth of data is important, not only for the discovery and development of novel biomarkers, but also as a tool to help guide clinicians and patients in their choice of treatments and to better understand their prognosis.

This cancer informatics project will lever clinical and biological informatics expertise in Edinburgh and Glasgow to develop and implement the use of deep learning applied to the full breadth of pertinent data.

Lab Websites