Genomics England uses Amazon SageMaker to predict cancer subtypes and patient survival from multi-modal data

This article discusses a collaboration between Genomics England and AWS to develop multi-modal machine learning models for predicting cancer subtypes and patient survival using genomic and imaging data.

Specifically, the article covers:

• The datasets used, including genomic and whole slide imaging data from The Cancer Genome Atlas (TCGA) for breast cancer and gastrointestinal cancers
• Three multi-modal machine learning frameworks implemented:
  1. PORPOISE - an attention-based network combining pathology images and genomic features
  2. Hierarchical Extremum Encoding (HEEC) - a novel architecture from AWS using tree ensembles and hierarchical representations
  3. Hierarchical Image Pyramid Transformer (HIPT) - a self-supervised approach for the imaging modality
• The architecture deployed on AWS using SageMaker for data processing, model training, and model deployment
• Best practices such as separating development and production environments, and using CI/CD pipelines for automation

The collaboration aimed to enhance Genomics England's understanding of cancer biomarkers and biology using multi-modal data.