Benchmark competitiveness and generalization are not equivalent: a critical commentary on HybridSP.

Abstract

This Letter comments on the article by Wang et al. on HybridSP and argues that strong benchmark performance should not be interpreted as direct evidence of robust generalization. While the model is interpretable, computationally efficient, and methodologically valuable, its reported results may be inflated by benchmark-specific coefficient tuning and by comparisons across scoring pipelines with different underlying tasks. We further note that performance drops on bias-reduced and more challenging datasets already suggest limits to transferability. We therefore advocate for stricter validation frameworks to support broader claims regarding generalization in protein-ligand scoring.