CHIMERA-DDR: A Machine Learning Framework for Classifying Heterogeneous Mismatch-Repair and Homologous-Recombination...

Area of Science:

• Oncology
• Genomics
• Computational Biology

Background:

• Current DNA damage repair (DDR) biomarkers use binary classifications, which are insufficient for tumors with complex, concurrent repair deficiencies.
• Understanding these complex deficiencies is crucial for accurate cancer subtyping and treatment selection.

Purpose of the Study:

• To develop a novel method for stratifying metastatic prostate cancer patients into more precise DDR subgroups.
• To identify molecular signatures associated with specific DDR phenotypes and clinical outcomes.
• To create a machine learning tool for accurate DDR subgroup classification.

Main Methods:

• Genomic analysis of 672 metastatic prostate cancer patients to identify DDR subgroups and molecular signatures.
• Development of CHIMERA DDR, a probabilistic machine learning tool using a nested Random Forest architecture.
• Clinical validation of the tool in 130 patients to assess immunotherapy response and classification performance.

Main Results:

• Stratified 672 patients into 11 DDR subgroups, identifying 51 key molecular signatures.
• Identified a distinct tumor-mutational-burden very-high subset with preserved genomic integrity and enhanced immunogenicity.
• CHIMERA DDR demonstrated exceptional classification performance (AUCs 0.919-0.999) in identifying MMRd and HRR mutant subtypes, including admixed phenotypes.

Conclusions:

• CHIMERA DDR enables precision therapeutic stratification by resolving complex DDR phenotypes beyond binary classifications.
• The tool accurately detects mismatch-repair-deficiency (MMRd) and homologous-recombination-repair (HRR) mutant molecular subtypes.
• This approach facilitates more accurate patient selection for targeted therapies, including immunotherapies.