Imputation-Aware Optimization of Ultra-Low-Density SNP Panels: Synthetic Benchmarking and Evaluation in a Camellia oleifera Family

Ultra-low-density single-nucleotide polymorphism panels require marker selection that accounts for subsequent imputation. We formulate imputation-aware panel optimization (IAPO) by combining recoverability, training-only utility, and chromosome balance in a monotone submodular objective. We clarify the conditions linking coverage to imputation error and verify a ridge-prediction perturbation bound. Ten independent synthetic replicates, each with 720 individuals and 3000 dosage-coded loci, give a 192-marker prediction correlation of 0.347±0.092 (mean ± standard deviation). The paired advantage over coverage-only selection is 0.058 (95% confidence interval 0.010–0.106). Alternative architectures, reference-based imputation, population shifts, and candidate-ranking metrics expose important dependencies; 192 markers are not a reproducible optimum. A separate reanalysis of published C. oleifera data evaluates 164 offspring with normalized segregation-state genotypes and observed oil content and fruit yield. Across two repetitions of five-fold within-family validation, the 192-marker IAPO phenotype correlations are 0.067±0.112 for oil content and 0.153±0.145 for fruit yield. Prediction is modest and method rankings depend on trait and imputer. This empirical workflow test does not establish transfer across families or validate polyploid dosage calling.