Retention Is Not Preservation: Budget-Matched Evaluation of Isofunctional Protein Rewriting

Generative protein models are increasingly capable of controllable editing—such as shortening enzymes for viral delivery or diversifying sequences to evade antibodies—while preserving biological function. We introduce ISOFORM, a conditional generative model for constrained protein editing that jointly optimizes target length, sequence identity, and explicit functional constraints. ISOFORM achieves state-of-the-art performance, successfully retaining nearly all annotated functional motifs during severe length compression and identity-controlled divergence. However, through extensive experiments across our model and existing state-of-the-art systems, we uncover a fundamental dilemma hidden by aggregate evaluation metrics: a severe decoupling between functional motif retention and 3D structural preservation. Even when achieving SOTA retention of functional residues, current models yield significantly lower global structural similarity and severely distort the 3D spatial geometry of the retained active sites. Because aggregate scores mask this ubiquitous structure-function decoupling, we advocate for decoupled, structure-aware evaluation protocols. Our findings highlight the necessity of evaluating 3D motif geometry alongside sequence retention, establishing a rigorous foundation for controllable protein design.