Federated Learning for Mutational Signature Extraction in Healthcare

Sinding, Kevin Risgaard ; Rasmussen, Frederik

4. term

Software, Master

2024

2024-06-09

13

Cancer is a genetic disease caused by various factors, with each mutational process leaving a unique, identifiable signature within the genome. These mutational signatures provide valuable insights into the origins and development of cancer, aiding in the creation of targeted treatments. This study evaluates the use of federated learning (FL) for mutational signature extraction using Non-negative Matrix Factorization (NMF) and autoencoders (AE). The framework assesses performance on both synthetic and real-world genomic datasets, comparing FL methods to centralized approaches. The results show that FL achieves comparable accuracy in identifying mutational signatures but incurs increased computational time due to the distributed nature of the process. This suggests that FL is a viable alternative for privacy-preserving analysis, though it requires careful management of computational resources.

machine learning ; federated learning ; privacy preservation ; mutational signature extraction ; non-negative matrix factorization ; autoencoders

Download
View record in AAU Student Projects