CREMA: recovery of high clonal fraction mosaic chromosomal alterations missed by phase-based detection

Background Mosaic chromosomal alterations (mCAs), a type of clonal hematopoiesis, are large somatic gains, losses, and copy-neutral events associated with age-related genomic instability and cancer susceptibility. Phase-based mCA detection methods such as MoChA are highly sensitive for low clonal fraction events, including events present in approximately 1% of blood cells, because they leverage long-range haplotype imbalance. However, when clonal fraction is very high, heterozygous genotypes can be misclassified as homozygous because allelic imbalance becomes extreme. This misclassification of phase-informative heterozygous sites can reduce sensitivity, particularly for high clonal fraction copy-neutral loss of heterozygosity (CN-LOH), which retains near-normal log R ratio (LRR) and can resemble constitutional autozygosity. Sensitive detection approaches for high clonal fraction mCAs are needed as these events are more likely to manifest in phenotypic consequences and confer elevated risk of hematologic malignancies. Results We developed CREMA (Clonal Recovery using Expectation-Maximization Algorithms), a standalone post-processing framework that identifies and classifies high clonal fraction mCA candidates within runs of homozygosity (RoH) or other segmentation outputs. CREMA recenters LRR, models central B-allele frequency (BAF) distributions using expectation-maximization, selects one- or two-band models using Akaike information criterion, and applies a geometric classifier in BAF-derived clonal fraction and LRR space. Across test analyses, RoH nomination followed by CREMA post-processing identified high clonal fraction mCA candidates, predominately CN-LOH events, that were not present in the MoChA call set. The largest contribution was observed in a myelofibrosis cohort enriched for high clonal fraction 9p CN-LOH involving the JAK2 region, where RoH plus CREMA recovered 232 additional CN-LOH events, corresponding to 11.1% of autosomal mCAs of at least 2 Mb. In two population-based PLCO datasets, CREMA recovered smaller targeted sets of additional high clonal fraction CN-LOH events, adding 0.5% and 0.4% to the autosomal mCA call sets. Conclusions CREMA complements phase-based mCA detection by using RoH or other segmentation methods for candidate generation and BAF/LRR modeling for candidate-level classification. This workflow improves detection of high clonal fraction events, particularly CN-LOH, while remaining applicable to high clonal fraction gains and losses.