Machine learning-predicted chromatin organization landscape across pediatric tumors

Area of Science:

• Genomics
• Cancer Biology
• Computational Biology

Background:

• Structural variants (SVs) significantly impact oncogenesis by altering 3D genome folding.
• Whole-genome sequencing enables large-scale SV profiling in tumors, but experimental validation of their functional impact is challenging.

Purpose of the Study:

• To predict the impact of somatic SVs on genome folding using a machine learning model.
• To identify tumor types and genomic regions most affected by SV-driven folding disruptions.
• To functionally interpret SVs by integrating epigenetic data and prioritizing those affecting active enhancers.

Main Methods:

• Utilized a convolutional neural network (Akita) to predict genome folding disruptions from SVs in 61 pediatric tumor types.
• Applied dimensionality reduction to SV disruption scores to identify recurrently disrupted regions.
• Integrated epigenetic data with a modified Activity-by-Contact scoring method to prioritize SVs near active enhancers.

Main Results:

• Demonstrated significant variability in SV-induced genome folding disruptions across tumor types, with lymphomas, sarcomas, metastatic, and germline cell tumors showing the most disruption.
• Identified five recurrently disrupted genomic regions enriched for high-impact SVs, some harboring tumor-associated genes and regulators despite low mutation rates.
• Prioritized SVs with disrupted contacts at active enhancers, highlighting known oncogenes and novel candidate loci involved in tumorigenesis.

Conclusions:

• Machine learning effectively predicts SV-driven genome folding alterations in pediatric cancers.
• Identified novel SVs, disrupted loci, and potential genetic mechanisms contributing to cancer development.
• This framework facilitates future research linking SV-mediated regulatory changes to cancer pathogenesis.