Addressing biases and limitations in feature attribution for circRNA modification profiling.

Li et al. (CircRM: Profiling circular RNA modifications from nanopore direct RNA sequencing. Brief Bioinform 2026;27:bbaf726.) introduced Circular RNA Modifications (CircRM), a computational framework employing eXtreme Gradient Boosting and SHapley Additive exPlanations (SHAP) to profile RNA modifications in circular RNAs, achieving high predictive accuracy. However, we argue that strong predictive performance does not validate the biological reliability of the resulting feature-importance rankings. In heterogeneous feature spaces, tree-based models exhibit inherent biases, favoring continuous, high-cardinality variables-such as genomic position-over sparse sequence patterns, potentially obscuring true biological determinants. Furthermore, reliance on SHAP introduces theoretical vulnerabilities; recent findings on attribution limitations indicate that baseline sensitivity can decouple explanations from local mechanistic behavior. To address these analytical pitfalls, we advocate for a robust framework incorporating Highly Variable Gene Selection and Feature Agglomeration to mitigate multicollinearity, complemented by model-agnostic non-parametric methods such as Spearman's rho and Kendall's tau. Adopting these strategies ensures that computational profiling yields biologically actionable insights rather than reflecting statistical artifacts.