Modelling cancer progression using Mutual Hazard Networks.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Cancer progression involves accumulating genomic events, but their chronological order is challenging to determine from cross-sectional data.
• Existing cancer progression models are limited to acyclic interactions and cannot capture cyclic or mutually exclusive relationships between events.

Purpose of the Study:

• To develop a novel machine learning algorithm, Mutual Hazard Networks (MHN), for inferring cyclic cancer progression models from cross-sectional data.
• To overcome the limitations of existing models in capturing complex event interactions.

Main Methods:

• MHN models genomic events based on their spontaneous fixation rates and their multiplicative effects on subsequent event rates.
• The algorithm infers cyclic progression models by analyzing co-occurrence patterns in cross-sectional genomic data.
• Model performance was evaluated using cross-validation on four independent datasets.

Main Results:

• MHN demonstrated superior performance compared to acyclic models in cross-validated model fit across all tested datasets.
• Application to The Cancer Genome Atlas glioblastoma dataset revealed a novel interaction: IDH1 mutations promote the fixation of TP53 mutations.
• This finding aligns with observations from consecutive biopsies, suggesting a specific temporal order of these key events.

Conclusions:

• MHN provides a powerful new tool for inferring cyclic cancer progression models, offering a more comprehensive understanding of cancer development.
• The algorithm can uncover complex, non-linear interactions between genomic events that are missed by traditional acyclic models.
• MHN has the potential to advance cancer research by revealing novel insights into disease progression and identifying new therapeutic targets.