The Pulse

Spectral Feedback Raises Stable Protein Outputs by 32.3%

UC Berkeley researchers propose Spectral Feedback, which reopens and resamples selected positions in a generated protein sequence. It lifts predicted stable designs by 32.3% for a pretrained model, less on stronger ones, with a structural t

AI.info Team · Written with AI · fact-checked · 2026-09-24

Protein diffusion models usually generate a sequence by progressively filling masked amino-acid positions, and the methods that steer them toward a goal mostly adjust token choices or search among candidate outputs along the way. A paper posted to arXiv on Thursday by four researchers at the University of California, Berkeley proposes a different route: generate a sequence, then reopen selected positions and sample them again. The authors report a 32.3% increase in stable proteins for a pretrained model, with gains that shrink on stronger starting systems and a structural trade-off on the strongest.

Choosing which positions to reopen is the whole problem

The method, called Spectral Feedback, targets inverse folding: designing an amino-acid sequence that folds into a given protein backbone. Rather than change the model’s weights or redirect its denoising trajectory, it treats the diffusion model as a black box. After an initial sequence, it picks a set of positions to mask again and runs the model to resample them, in a loop.

Shai Dickman, Mert Cemri, Landon Butler and Kannan Ramchandran framed the shift in emphasis in the paper’s abstract. Where earlier alignment methods decide which amino acid to place at each position, they wrote, the question they pose is a different one:

“While prior alignment methods focus on what token labels to assign to maximize a target reward, we instead treat which tokens to revisit as the central alignment problem.”

Shai Dickman, Mert Cemri, Landon Butler and Kannan Ramchandran, University of California, Berkeley, in the paper

The difficulty is that edits interact: the value of changing one position depends on which others are changed at the same time, so choosing positions one by one can miss the best combinations.

Sparse Fourier models guide the edits

Searching every possible combination of positions quickly becomes impractical. The team instead represents the value of candidate edit sets with a Fourier expansion and uses sparse recovery to estimate which combinations are promising from a limited number of reward evaluations. The approach rests on an empirical finding that these edit-value functions are sparse, with most of their signal in a small number of mostly low-order coefficients, which the authors link to the sparse interactions between amino acids documented in biology.

The researchers tested the method on protein backbones from the Megascale dataset, using a stability oracle that predicts the change in Gibbs free energy between a designed sequence and its wild-type counterpart. They also tracked self-consistency RMSD against the target backbone, using ESMFold-predicted structures, and sequence naturalness measured with ProtGPT2, to catch reward gains that come at the expense of other qualities.

The 32.3% gain is not the same across models

In the abstract, Spectral Feedback raises the share of stable proteins by 32.3% for a pretrained model, 24.8% when paired with Best-of-10 sampling, and 5.8% for DRAKES, a diffusion model fine-tuned with reinforcement learning. Those figures come from the authors’ oracle-based evaluation, not from proteins made and measured in a laboratory.

The paper’s results table shows the trade-off on DRAKES. With Spectral Feedback, its average evaluation reward rises from 0.99 to 1.38, the highest in the table, but average self-consistency RMSD increases from 1.17 to 1.53, while the share of designs meeting the paper’s RMSD threshold falls from 93.4% to 85.8%. For the pretrained model and for Best-of-10, the same share rises, to 94.2% and 96.7%. The authors suggest the DRAKES drop may reflect that model’s own sample distribution rather than their algorithm.

More inference compute, not a free correction

Spectral Feedback avoids retraining, but it is not free: it makes extra calls to the diffusion model and the reward oracle to assess possible edit sets. In the authors’ experiments it matches Best-of-10 in two feedback iterations, where Best-of-10 needs ten full reverse processes and Spectral Feedback one full pass plus partial ones. They list as limitations the large number of edit-set samples that sparse recovery requires and value functions that are non-deterministic, which makes them harder to approximate and costly to evaluate.

The authors name DNA design and text generation as the next domains to try, and want to see how the results scale with larger vocabularies and longer sequences.