Context-aware simulation enables systematic optimization of long-read mapping parameters

Long-read mapping performance is critical for downstream genomic analyses but remains sensitive to parameter selection. We present CycSim, a context-aware long-read simulator that learns sequence-context-dependent error profiles from empirical data and generates realistic simulated reads. CycSim more faithfully recapitulated real long-read characteristics than existing simulators, providing a high-fidelity simulation framework with known ground truth. Using this framework, we identified a Cyclone-specific parameter set that achieved 2.78-fold faster mapping than an ONT-oriented baseline while maintaining comparable variant-calling performance. For SV-oriented optimization, CycSim-guided refinement improved mapping efficiency by 8.14–34.16% across ONT, HiFi, and Cyclone HG002 datasets, increased SV F1 scores by 0.57–1.75 percentage points, and showed consistent improvements across independent benchmark datasets and different SV callers. Together, these results demonstrate the utility of CycSim for platform- and analysis-goal-specific algorithm development, benchmarking, and parameter optimization.