MOSAIC enables selective transfer and fairness-aware generalization in cross-species DNA methylation prediction

DNA methylation prediction increasingly relies on data collected across multiple species and modification types, yet these datasets are often highly heterogeneous, severely imbalanced, and strongly long-tailed. Separate models preserve task-specific patterns but miss transferable biological information, whereas indiscriminate data pooling can improve average performance while degrading underrepresented tasks. MOSAIC is a prompt-aware sparse mixture-of-experts framework for cross-species and cross-modification DNA methylation prediction. It combines a shared DNABERT-2 backbone with species and methylation-type prompts that guide sparse expert routing across related biological contexts, preserving task-specific specialization while enabling selective information sharing. On a 20-task benchmark spanning 4mC, 5hmC, and 6mA, MOSAIC showed the strongest overall performance-fairness profile among controlled baselines and representative predictors, achieving a higher worst-task Matthews correlation coefficient, a smaller head-tail performance gap, and a higher Jain's fairness index. Ablation, routing, and perturbation analyses further suggested that these gains were consistent with selective sharing aligned with biologically plausible sequence and routing patterns rather than indiscriminate parameter sharing. External zero-shot evaluation further showed measurable transfer within the evaluated OOD settings. A publicly accessible web server is provided to support practical methylation prediction and routing inspection.