Dual-Constraint Protein Design: Validating Repulsion Guidance and Hotspot Preservation in RFdiffusion

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: tiny / short paper (2-4 pages excluding references; extended abstract format)

Keywords: Generative Protein Design, Geometric Deep Learning, RFdiffusion

TL;DR: We extend the RFdiffusion framework to validate Negative Design capabilities, demonstrating that repulsion guidance can effectively steer protein generation away from forbidden "coldspots" without compromising structural validity.

Abstract: Generative Artificial Intelligence, particularly geometric deep learning, has revolutionized structural biology by enabling the \textit{de novo} design of proteins with specific functions. SE(3)-equivariant diffusion models, such as RFdiffusion, have emerged as state-of-the-art tools for generating protein backbones that satisfy complex geometric constraints. However, current research predominantly focuses on "positive design'' tasks, scaffolding functional motifs (hotspots) to ensure binding affinity. Less attention has been paid to "negative design,'' or the ability to explicitly inhibit off-target interactions through geometric repulsion (coldspots). Achieving high specificity in therapeutic design requires a generative process that can simultaneously optimize for desired interactions while strictly avoiding forbidden spatial regions. In this work, we present a systematic evaluation of RFdiffusion's steerability under dual geometric constraints. We first validate the model's efficacy in standard hotspot scaffolding, confirming high-fidelity recovery of functional sites. We then extend this analysis to "coldspot'' generation by implementing repulsion guidance mechanisms to enforce negative geometric constraints. Our experiments demonstrate that repulsion guidance effectively modulates the diffusion trajectory, producing valid backbones that respect excluded volumes. We further provide a comparative analysis of the structural distributions between hotspot-constrained and repulsion-guided samples. These results quantify the distributional shift induced by negative constraints, validating the model's capacity for precise, multi-objective structure generation in complex biological environments.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Shentong_Mo1

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 62