Leveraging Auxiliary Potentials in RFDiffusion for the Design of NA-Binding Proteins

The de novo design of nucleic acid-binding proteins presents unique challenges, distinct from small-molecule or protein-protein interface design, primarily due to the large size and complex electrostatics of the NA target. While platforms like RFDiffusion offer a powerful starting point, effective strategies for applying them to this problem remain underexplored. Here, we systematically investigate and compare two distinct design paradigms─de novo generation and motif-scaffolding─across an expanded benchmark of 11 nucleic acid targets spanning diverse binding modes. We evaluate several enhancements to the standard framework, including a rigid-body-averaged potential and a px0-guided potential. Our results reveal a nuanced trade-off landscape. The px0-guided potential tends to produce low-clash interfaces with favorable binding metrics, often enriched in β-sheets. However, this comes at the cost of foldability, yielding designs with compromised single-chain stability as assessed by Boltz2. Rigid-body averaging alone confers no advantage in direct interface metrics but selectively improves NA interface coherence upon refolding, suggesting it may act as a coordination regularizer. Critically, combining both strategies (px0 + averaging) offers a promising balance, retaining favorable interface metrics while improving computational foldability scores. In motif-scaffolding across 11 diverse complexes, auxiliary potentials consistently improve outcomes, yet absolute performance varies unpredictably with the target context; motif preservation ranges from near-native to severely disrupted, underscoring the stochastic nature of the process. Comparison with RFdiffusion3 reveals that newer models can produce superficially appealing metrics but often generate uniformly wrapped complexes that lack target specificity and structural diversity. Our results, while based on computational metrics, suggest that modular, tunable potentials, particularly in combination, offer advantages over monolithic end-to-end approaches. This study provides a practical comparative assessment of design strategies for NA-binding proteins that may inform future efforts in this area.