Integrative modeling of read depth and B-allele frequency improves single-cell copy number calling from targeted DNA...

Copy number variations (CNVs) drive cancer initiation and progression, but resolving them at single-cell resolution from targeted DNA sequencing panels remains challenging. The Mission Bio Tapestri platform generates two complementary signals for CNV inference: sequencing depth and B-allele frequency (BAF) from heterozygous variants; however, existing methods such as karyotapR rely primarily on read depth, potentially missing allele-specific events invisible to depth-only approaches. Here we introduce scPloidyR, a hidden Markov model (HMM) that jointly models read depth and BAF at amplicon resolution for single-cell copy number calling from Tapestri data. scPloidyR fits independent per-chromosome Markov chains with copy number states as hidden variables, factorizes emission probabilities into depth and BAF likelihoods, and learns parameters via Baum-Welch expectation-maximization with Viterbi decoding. We compared scPloidyR with the established karyotapR Gaussian Mixture Model (GMM) through two simulation studies that evaluates BAF noise, variant density, amplicon density, sample size, and heterozygosity rate, and through application to a public Tapestri five-cell-line mixture dataset. In simulations, scPloidyR substantially outperformed karyotapR on class-balanced metrics (macro-F1: 0.472 vs. 0.264; alteration F1: 0.902 vs. 0.383 in simulation study 1) when allelic information was available. Adding just one heterozygous variant per amplicon increased scPloidyR accuracy from 0.548 to 0.899 for copy number gains. However, when BAF information was absent, karyotapR outperformed scPloidyR, and high BAF noise substantially degraded joint-model performance. On real data, scPloidyR produced more spatially coherent and biologically plausible copy number profiles. These results establish that joint depth-BAF modeling provides a clear advantage for single-cell CNV calling when allelic information is available, while depth-only methods remain preferable when such information is absent.