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

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Circular RNA Modifications (CircRM) is a computational framework for profiling RNA modifications in circular RNAs.
• The framework uses eXtreme Gradient Boosting and SHapley Additive exPlanations (SHAP) for high predictive accuracy.

Purpose of the Study:

• To evaluate the biological reliability of feature-importance rankings from computational RNA modification profiling.
• To address inherent biases in tree-based models and theoretical vulnerabilities in SHAP explanations.
• To propose a more robust analytical framework for biologically actionable insights.

Main Methods:

• Critique of CircRM's reliance on tree-based models and SHAP.
• Advocacy for Highly Variable Gene Selection and Feature Agglomeration to mitigate multicollinearity.
• Integration of model-agnostic non-parametric methods (Spearman's rho, Kendall's tau).

Main Results:

• High predictive performance does not equate to biological reliability of feature importance.
• Tree-based models can favor certain variable types, potentially masking true biological drivers.
• SHAP explanations may be sensitive to baseline choices, decoupling from mechanistic behavior.

Conclusions:

• A robust computational framework is needed to ensure biological relevance in RNA modification profiling.
• Proposed methods enhance the reliability of identifying true biological determinants in circular RNAs.
• The revised approach yields actionable insights rather than statistical artifacts.