Diffusion guidance

Diffusion guidance refers to inference-time methods that steer a diffusion process toward desired properties, constraints, or observations. It is a general concept that applies to both protein design, structure prediction, and sequence-based protein design, of which all-atom diffusion is just one application.

Details

Xie et al (1) outline three broad types of guidance used by diffusion models:

Score guidance, in which gradients from classifiers, constraints, or rewards are used to nudge the diffusion path.
Path-integral reweighting, such as Feynman-Kac potentials, which use importance weights to update trajectories.
Invariant correctors, such as Metropolis-adjusted Langevin methods, which mix within a biased marginal without changing the trajectory weights.

However, other search algorithms such as Beam search and Monte Carlo Tree Search have been used in conjunction with diffusion models (2).

Xie Y, Winkler L, Sun L, Lewis S, Foster AE, Luna JJ, et al. Enhanced Diffusion Sampling: Efficient Rare Event Sampling and Free Energy Calculation with Diffusion Models. 2026; Available from: https://arxiv.org/abs/2602.16634

Didi K, Zhang Z, Zhou G, Reidenbach D, Cao Z, Cha S, et al. Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute. 2026; Available from: https://arxiv.org/abs/2603.27950

Quartz 4

Explorer

Diffusion guidance

Details

Protein design

Inference-time scaling of de novo designed proteins is more effective for harder targets

Guided sequence-only diffusion outperforms autoregressive fine-tuned PLMs on designing high-fitness sequences

Multimodal sequence-structure guidance outperforms unimodal guidance in cases where both need to be designed

Stronger diffusion guidance reduces diversity of generated outputs

Structure prediction

Diffusion-based structure prediction can be steered by modifying the conditioning embeddings rather than the latent space, and such embeddings can be used for subsequent iterations

DL structure tools introduce chiral errors such as swapped chiral centers, D-amino acids, and cis-amide bonds

Enhanced diffusion with metadynamics-like potentials can sometimes convergence slower than unbiased diffusion

Ensembles can be modeled by structure prediction NNs using experimental data via guidance sampling

Multistate Bennett acceptance ratio can be used to reweigh one or more samples from a guided diffusion trajectory

Protein backbone design diffusion models can be repurposed for fitting structures into electron density

Confidence metrics for diffusion-based structure prediction methods can be improved with minimal changes to conditioning representations

Guidance potentials can be added to diffusion-based structure prediction for enhanced sampling of protein conformations