RFdiffusion revolutionized de novo protein design by bringing denoising diffusion probabilistic models to the world of structural biology. The algorithm supports a seldom-explored “guiding potentials” feature, which allows users to define additional criteria to guide the generative protein design process. Presently, potentials must be specified with a technical, PyTorch implementation that is arithmetically fully differentiable with respect to the three-dimensional coordinates of the backbone carbons. Owing to this complexity, we have observed few publicly accessible advances in the use of guiding potentials, despite potential applications for research groups with challenging design goals and limited computing power. By default, RFdiffusion comes with guiding potentials that promote contact formation and mini-protein compactness. These methods, though simple, significantly improve the quality of de novo designs. We hypothesized that building more advanced potentials could add further diversity and control to the generative diffusion process, offering a computationally efficient alternative to fine-tuning and/or retraining diffusion models. Here, we present four new guiding potentials to enhance RFdiffusion’s generative toolset for designing mini-proteins and small molecule binders. First, the fixed contact potential allows for fine-tuned control over mini-protein binding sites throughout the entire denoising trajectory. Second, the boundary potential supports implicit membrane modeling for design projects near cellular membranes. Third, the post-translational modification (PTM) potential enables pseudo-all-atom understanding by approximating steric clashes with PTMs. Fourth, the small molecule orientation potential helps control how a protein backbone folds around small molecule ligands. For each potential, we demonstrate increased control and prevalence of favorable design features in RFdiffusion-generated models. Furthermore, we demonstrate that these structural characteristics persist even after sequence design with LigandMPNN and validation with AlphaFold 3, indicating strong designability and foldability. Lastly, we share tips to make custom potentials more accessible to all protein designers.
Narang et al. (Sun,) studied this question.