CatIF-RL: Activity-Oriented Enzyme Sequence Design by Steered Inverse Protein Folding

Abstract

Protein inverse folding models are designed to generate amino acid sequences compatible with a given backbone structure, but they are not explicitly optimized for specific biological functions. Here, we present CatIF-RL, a framework that steers a graph-based denoising diffusion inverse folding model toward designing enzyme variants with improved predicted catalytic activity. CatIF-RL first adapts the inverse folding model to enzyme structural data, then introduces activity-oriented preference signals using predicted catalytic constant (kcat) as the optimization objective, enabling specialization through generative data set curation and group-relative policy optimization (GRPO). This process iteratively shifts the sequence distribution toward higher predicted kcat while constraining sequence divergence to sequences that remain compatible with the input structure. On the held-out benchmark, CatIF-RL achieves an approximately 4-fold increase in predicted kcat relative to native enzymes as scored by the surrogate predictor, outperforming representative inverse-folding methods, while keeping sequence recovery (0.55) and predicted structural quality within the range of existing inverse-folding methods, and supporting motif-preserving partial sequence design. CatIF-RL establishes a practical framework for activity-oriented enzyme design and provides a generalizable strategy for steering structure-conditioned protein generation toward functional optimization.