PaperlayerView exact revision on arXiv

Source-linked AI summary

Spectral Feedback for Test-Time Alignment of Protein Diffusion Models

Shai Dickman, Mert Cemri, Landon Butler, Kannan Ramchandran

arXiv:2609.30456v1cs.AI

TL;DR

In protein inverse-folding alignment, the value of editing a position can depend on which other positions are edited simultaneously. Spectral Feedback uses sparse Fourier approximations to select edit positions and improves alignment rewards across several protein models.

  • Problem

    For test-time protein inverse-folding alignment, choosing edit positions is challenging because an edit’s value depends on which other positions are edited simultaneously.

    Source: §1 · p. 3

  • Method

    Spectral Feedback learns a sparse Fourier approximation of edit-set values and optimizes it to select positions for re-masking and re-sampling.

    Source: §4.2 · p. 7, §6 · p. 10

  • Results

    On Megascale protein inverse folding, Spectral Feedback improves alignment rewards across several protein models and scales best in compute and latency versus greedy or gradient-based edit-selection baselines.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Protein sequence edit-set value functions are empirically sparse in the Fourier basis, consistent with biological epistasis.

    Source: §1 · p. 2, §6 · p. 10

    Spectral Feedback requires many edit-set samples for sparse recovery to learn a sparse Fourier representation.

    Source: §6 · p. 10

Abstract

from arXiv · show

Reward maximization alignment methods for discrete diffusion models have primarily focused on steering the reverse process, either by influencing token logits or by selecting favorable sequences at intermediate steps. These approaches largely treat inference as a unidirectional process, lacking mechanisms for revisiting undesirable token selections. We introduce Spectral Feedback, an algorithm that selects edit-positions in a feedback loop, allowing the model to iteratively correct its own generations. This approach leverages the mask structure of discrete diffusion models by re-masking and re-sampling tokens, analogous to image editing methods that reintroduce noisy latents and re-run the reverse process. 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. Selecting edit-positions is challenging because edit effects are interdependent: the impact of modifying one token depends on which others are edited simultaneously. We define an edit-set as a set of token positions to re-mask and re-sample. Motivated by prior work on sparse interactions in biological systems, we find empirically that edit-set value functions for protein inverse folding admit sparse Fourier representations. This structure enables Spectral Feedback to efficiently learn and optimize the value functions for edit-position selection. Spectral Feedback is model-agnostic and can be applied to pretrained, test-time aligned, and fine-tuned diffusion models. For all of these models, the algorithm improves alignment performance without modifying the underlying generative process. Applied to inverse folding with a protein stability reward oracle, it achieves a 32.3% increase in stable proteins for a pretrained model, 24.8% for Best-of-10, and 5.8% for a state-of-the-art RL fine-tuned diffusion model.

1 Introduction

Spectral Feedback aligns discrete diffusion models by selecting edit positions for re-masking and re-sampling, using sparse Fourier approximations to handle interactions among edits. The introduction reports improved stability alignment across pretrained, Best-of-N sampling, and RL fine-tuned models.

Introduction

  • Re-noising feedback generates a sequence, re-masks a subset of positions, and resamples them, unlike token-logit steering or tree search over denoising trajectories.Earlier re-masking approaches commonly impose independence assumptions and rely on other alignment techniques, such as importance sampling or classifier-free guidance.Source: §1 · pp. 1–2

Formulation

  • The method maximizes an edit-set value function because an edit’s value can depend on which other positions are edited simultaneously, making exhaustive search over 2^L sets intractable.The introduction proposes removing independence assumptions so that re-masking can serve as an alignment method on its own.Source: §1 · pp. 2–3

Sparsity analysis

  • Protein edit-set value functions have empirically sparse Fourier expansions, with some targets dominated by first-order effects and others involving higher-order interactions.Most spectral energy is concentrated in lower-order terms, and the observed structure is consistent with biological epistasis.Source: §1 · pp. 2–3

Empirical results

  • Spectral Feedback reaches Best-of-10 reward in two feedback iterations and Best-of-50 in five, and improves alignment reward across pretrained, Best-of-N, and RL fine-tuned models.The evaluation tracks stability reward (∆∆G), self-consistency RMSD, and ProtGPT2 log-likelihoods to detect over-optimization.Source: Figure 2 · p. 2, §1 · p. 3

Method

  • Spectral Feedback queries a small number of edit sets, fits a sparse Fourier approximation, selects its highest-valued set, and re-invokes the reverse process with those positions re-masked.It treats the diffusion model as a black box and does not change model weights or the denoising trajectory, enabling use with pretrained, fine-tuned, or test-time aligned models.Source: §1 · p. 3

Empirical results

  • On Megascale protein-backbone inverse folding, Spectral Feedback improves reward across underlying models and scales best in compute and latency against greedy or gradient-based edit-set selection baselines.The stability oracle predicts ∆∆G for designed sequences relative to their wild-type variants.Source: §1 · p. 3

2 Preliminaries

The section defines discrete diffusion notation and the feedback procedure: choose positions to re-mask, then resample them with the reverse process. It frames edit-set selection as optimizing joint reward under a cardinality bound.

2 Preliminaries

  • Discrete diffusion progressively masks tokens through a CTMC and reverses from a fully masked sequence by progressively unmasking tokens.Transitions are constrained to single-token flips, while multiple tokens may be updated per step under an independence assumption.Source: §2 · p. 4
  • Feedback selects a subset of sequence positions to re-mask, then generates a new sequence by running the reverse process from the partially masked sequence.The procedure alternates edit-position selection and resampling to seek positions that maximize a value function.Source: §2 · p. 4
  • Resampling initializes the reverse process at its first time-step and runs it to completion; using a later step matched to the masking pattern and noise schedule is left for future work.Source: §2 · p. 4
  • The total step count N is the number of proteins generated, including the initial sequence, so the feedback loop performs N −1 iterations.Source: §2 · p. 4
  • The cardinality bound k balances reward improvement against reverse-process cost and the combinatorial complexity of searching candidate edit-sets.Prior methods select positions randomly or independently [Lee et al., 2025, Reid et al., 2022, Gruver et al., 2023], whereas this method searches for optimal sets using their joint reward structure.Source: §2 · p. 4

3 Related Work

Spectral Feedback aligns protein diffusion models at test time by selecting positions to reopen, rather than changing model weights or token logits. It treats edit-set selection as a combinatorial problem and uses sparse Fourier recovery to model interactions among positions.

Protein Sequence Design and Inverse Folding

  • Spectral Feedback complements ProteinMPNN and discrete diffusion models by operating at test time without changing weights, including on DRAKES-style fine-tuned models [Wang et al., 2025].DRAKES fine-tunes a discrete diffusion inverse-folding model with a differentiable reward signal, requiring model-specific training.Source: §3 · p. 4

Inference-Time Alignment for Discrete Generative Models

Sequence Editing

  • Prior editors use uniform noising (Lee et al. [2025]), reward-gradient position sampling (Gruver et al. [2023]), or random edits (DiffusER [Reid et al., 2022]); Spectral Feedback optimizes interacting edit sets.Self-Refine [Madaan et al., 2023] and multi-agent pipeline optimization use LLM-generated feedback rather than reward oracles.Source: §3 · p. 5

Sparse Boolean Function Recovery and Epistasis

  • Building on sparse higher-order epistasis [Poelwijk et al., 2019, Aghazadeh et al., 2021, 2020], Spectral Feedback recovers edit-set reward functions and studies their sparsity across protein targets.Higher-order Fourier recovery avoids LASSO's explicit enumeration of interaction features, which scales exponentially with order.Source: §3 · p. 5

4 Spectral Feedback

Spectral Feedback iteratively selects edit positions and re-samples them with a discrete diffusion model, using sparse Fourier structure in edit-set value functions to guide which tokens to revisit.

4.1 Spectral Function Approximation

  • Sparse Fourier structure makes value functions tractable to approximate, and Spectral Feedback uses ProxySPEX [Butler et al., 2025] to learn that representation efficiently.ProxySPEX uses Gradient Boosted Trees and samples binary masks with independent Bernoulli entries of parameter γ.Source: §4.1 · p. 6, §4.2 · p. 7
  • 0.84 average R2 for favg versus 0.73 for fmax under fewer than 1,000 coefficients; most coefficients are order at most 3, with interaction patterns differing by protein.Both example proteins reach R2 of at least 0.7; first-order coefficients dominate for 2KVV, while higher-order interactions are more significant for R6-560.Source: Figure 3 · p. 5, §4.1 · p. 6

4.2 Spectral Feedback Algorithm

  • For each feedback iteration, Spectral Feedback samples candidate edit sets, evaluates their rewards, fits a sparse Fourier approximation, and optimizes it to select at most k positions.The selected set maximizes the learned approximation, with candidate rewards assigned using the chosen value function.Source: Figure 4 · p. 6, §4.2 · p. 7
  • The algorithm re-masks selected amino acids and runs the partially masked sequence through the reverse process, requiring fewer iterations than full-mask initialization.Selection and model sampling alternate in a feedback loop using an off-the-shelf discrete diffusion model.Source: §4.2 · pp. 6–7
  • The value function is favg for direct pretrained or RL fine-tuned diffusion sampling, but fmax for greedy methods such as Best-of-N or Beam Search.fmax reflects that only the highest-valued sample matters under greedy test-time alignment.Source: §4.2 · p. 7

5 Experiments

Experiments test Spectral Feedback across protein inverse-folding models and edit-selection baselines. It improves oracle rewards, with efficiency and structural-quality tradeoffs varying across methods.

5.1 Experimental Setup

  • Experiments use Megascale protein backbones to assess stability, scRMSD, and naturalness; success requires evaluation ∆∆G > 0 and scRMSD < 2.Stability is measured by ∆∆G, scRMSD compares the target backbone with the predicted fold, and naturalness is sequence log-likelihood.Source: §5.1 · pp. 7–8
  • Tests cover pretrained and DRAKES ProteinMPNN models, plus pretrained Best-of-N; the method uses favg for pretrained and DRAKES models and fmax for Best-of-N.The authors suggest Spectral Feedback may extend to other alignment methods because their target distributions are similar.Source: §5.1 · p. 8

5.2 Baseline Edit Selection Methods

  • Baseline edit selectors use token gradients, isolated-mask expected reward, the best observed edit set, or randomized 1-flip hill climbing.Hill climbing accepts reward-improving additions or removals but is inherently serial, limiting parallelization to single-set reward calculations.Source: §5.2 · pp. 8–9

5.3 Results on Protein Inverse Folding

  • Spectral Feedback raises alignment and evaluation-oracle rewards for all three base methods; DRAKES + Spectral achieves the highest evaluation ∆∆G (1.38), with a scRMSD tradeoff.The increase in average scRMSD for DRAKES limits gains in protein success rate.Source: Table 1 · p. 9, §5.3 · p. 9
  • Spectral Feedback matches Best-of-10 performance in 2 iterations, compared with Best-of-10’s 10 full reverse processes.Feedback uses one full reverse process followed by partial processes that fill only a subset of tokens.Source: §5.3 · p. 9
  • At one feedback iteration, Spectral Feedback and hill climbing scale similarly, but Spectral achieves a better latency–reward tradeoff; exclusion and gradient methods are faster but weaker.Although Spectral adds ProxySPEX fitting and optimization overhead relative to argmax selection on the same samples, that overhead becomes insignificant in the latency tradeoff.Source: §5.3 · pp. 9–10, Figure 6 · p. 10

6 Discussion · A Hyperparameter Settings · A.1 Computational Resources

Spectral Feedback treats edit-position selection as the central test-time scaling problem, using sparse Fourier recovery to guide iterative re-sampling; the discussion notes its limitations and possible extensions.

6 Discussion

  • Spectral Feedback selects informative subsets of sequence positions to re-mask and re-sample through the diffusion reverse process.It frames edit-set selection as the central scaling target for protein sequence generation with stability reward oracles and protein language models.Source: §6 · p. 10
  • Because an amino acid’s edit value depends on which other positions are edited, the method models rewards over subsets and exploits sparse Fourier structure consistent with biological epistasis.Sparse recovery approximates the value function for optimization in a black-box meta-loop.Source: §6 · p. 10
  • On Megascale inverse folding, the method improves alignment reward across protein models while maintaining structural self-consistency that increases success rates.Source: §6 · p. 10
  • Non-deterministic value functions can hinder reaching larger R2 values and incur extra protein-model calls to simulate a reverse-process step.Source: §6 · p. 10
  • Sparse recovery requires many edit-set samples, but in this high-sample regime Spectral Feedback scales better than competing methods.Source: §6 · p. 10
  • Training a value function for partially masked sequences could improve runtime and enable approximations to reach greater R2 values.Source: §6 · p. 10
  • Future work could explore alternative sparse-recovery scaling and apply the algorithm to DNA design or text generation, including larger vocabularies and longer sequences.Source: §6 · p. 10

A.1 Computational Resources

  • Experiments used NVIDIA GH200, H100, and L40S GPUs; timing runs used GH200s with 32 GB RAM and 16 CPU cores.CPU parallelism sped up the Gurobi solver for sparse Fourier optimization.Source: §A.1 · p. 14

A.2 Protein Design Methods · A.3 Feedback Mechanisms · B Discrete Diffusion Models

The methods specify ProteinMPNN model choices and feedback edit-selection procedures, then formalize discrete diffusion through forward and reverse processes, including the common random-masking formulation.

A.2 Protein Design Methods

  • The experiments use pretrained ProteinMPNN and its RL fine-tuned version from DRAKES, with KL-penalty β = 0.001 for the fine-tuned model.Source: §A.2 · p. 14
  • Best-of-N improves monotonically with N but has diminishing returns, so Spectral Feedback is combined with it to improve alignment beyond those limits.Source: §A.2 · p. 14

A.3 Feedback Mechanisms

  • Feedback experiments select k = 20 edit-positions, use D = 8192 mask samples, and run three iterations, balancing improving reward against computational cost.Figure 7 presents validation curves for feedback-loop parameters and shows diminishing returns on a log scale.Source: §A.3 · p. 14, Figure 7 · p. 14
  • ProxySPEX fits a GBT ensemble, while first-order LASSO models edit-set value as a sparse linear function and uses cross-validation for edit-position selection.LASSO selects the top-k positive estimated coefficients; Exclusion and Gradient baselines use the same 20 edit-positions for fair comparison.Source: §A.3 · pp. 14–15

B Discrete Diffusion Models

  • Discrete diffusion evolves probability masses over a finite sequence space through a forward CTMC governed by rate matrices, then reverses from a known terminal distribution to recover the target distribution.The reverse process learns the score function, the ratio p_t(y)/p_t(x), and the forward rates can be chosen to define the terminal distribution.Source: B Discrete Diffusion Models · pp. 15–16
  • Random masking is a common forward process: tokens transition into masks and remain masked, while reverse sampling starts fully masked and progressively unmasks tokens that then stay unmasked.Source: B Discrete Diffusion Models · p. 16
  • The reverse process restricts transitions to single-token flips to avoid tracking exponentially many state transitions, while assuming token updates are independent enables multiple updates per step.Source: B Discrete Diffusion Models · p. 16

C Compute Complexity … D.2 Sparsity Experiments

Spectral Feedback’s compute costs scale with its diffusion, mask-value, and reward-oracle calls, while sparse Fourier structure makes edit-set optimization tractable and is observed across two alignment oracles.

C Compute Complexity

  • Spectral Feedback makes M × N + D × (N −1) diffusion-model calls and n × D × (N −1) alignment-reward-oracle calls.Best-of-N instead makes M × N diffusion-model calls and N reward-oracle calls.Source: C Compute Complexity · p. 16
  • Its modularity also lets Spectral Feedback improve other alignment methods, including Beam Search and RL.The paper points to Figure 12 for these applications.Source: C Compute Complexity · p. 16
  • Best-of-10 Spectral Feedback reaches better performance than scaling Best-of-N and can identify rare, more stable proteins after Best-of-N levels off.Figure 8 uses D = 8192, k = 20, and 5 feedback iterations.Source: Figure 8 · p. 16, C Compute Complexity · p. 16

D.1 Sparse Fourier Optimization

  • The optimization rewrites sparse Fourier parity terms in the Möbius basis, then expresses the objective as a polynomial over binary edit-set indicators.Source: §D.1 · p. 17
  • For a linear integer program, each monomial gets a binary variable, with subset-activation, deactivation, and cardinality constraints encoding feasible edit sets.Source: §D.1 · p. 17
  • At most s · 2^d + n decision variables and s · 4^d + n · 2^d + 1 constraints make the program tractable when sparsity s and degree d are small, even for large n.Source: §D.1 · p. 17

D.2 Sparsity Experiments

  • Using pretrained-model value functions with n = 64 empirical samples, SPEX found sparse Fourier structure for both ΔΔG and ProtGPT2 log-likelihood: few low-order coefficients achieved high R2 faithfulness.The analysis used a 100k compute budget, maximum interaction order 5, and both fmax and favg value functions.Source: §D.2 · p. 18

D.3 High-Sensitivity Value Function Ablation

This ablation explores exponentially tilted value functions that preserve the optimizer while emphasizing high-value subsets. Stronger tilting improves focus on the argmax but weakens Fourier sparsity, creating a tradeoff for spectral recovery.

D.3 High-Sensitivity Value Function Ablation

  • High global R2 alone does not guarantee argmax recovery, although the authors do not believe their value functions exhibit this pathological failure mode.They motivate follow-up objectives more directly aligned with maximization despite strong experimental results using the global-R2-driven approximation.Source: §D.3 · p. 20
  • An exponentially tilted objective preserves the optimizer while making L2 approximation more sensitive to high-value subsets.The maximizer is normalized to value 1, while lower-value subsets are downweighted toward exp(−β); larger β emphasizes accurate approximation near the top of the value function.Source: §D.3 · p. 20
  • For β from 0.25 to 8, spectral sparsity remains similar to the unnormalized function, but larger β shifts energy to higher-order terms and makes the spectrum increasingly dense.As β →∞, the tilted function approaches a Dirac delta at the maximizer, which is spectrally dense.Source: §D.3 · p. 20
  • A potential improvement is to learn tilted objectives for multiple β values from the same samples, then cross-validate the highest β meeting a prespecified approximation quality.This would target better argmax alignment while respecting the loss of Fourier sparsity at larger β.Source: §D.3 · p. 20

E Spectral Feedback scRMSD Discussion · F Additional Scaling and Alignment Studies · F.1 Spectral Feedback Applications

Spectral Feedback can improve alignment reward while worsening structural consistency when the underlying model and reward favor structurally inconsistent samples. Its broader applications include extending alignment methods, though feedback can require substantial compute.

E Spectral Feedback scRMSD Discussion

  • DRAKES shows increased scRMSD with Spectral Feedback, reducing inverse-folding accuracy and limiting success because successful proteins require scRMSD < 2.Figure 11 shows the worse scRMSD tradeoff for DRAKES when Spectral Feedback is applied.Source: E Spectral Feedback scRMSD Discussion · p. 21, Figure 11 · p. 21
  • DRAKES has higher average scRMSD than the pretrained model (1.17 vs 1.10), reflecting high-reward DRAKES samples with scRMSD between 8 and 10.The passage also reports that DRAKES has lower average log-likelihood, indicating overfitting to the alignment oracle and more out-of-distribution proteins.Source: E Spectral Feedback scRMSD Discussion · p. 21, Figure 11 · p. 21
  • Spectral Feedback may move samples toward higher stability reward but lower structural consistency; preserving auxiliary properties depends on reward quality and the underlying model distribution.The discussion proposes regularized or multi-objective rewards that explicitly include structural consistency or correlated statistics to prevent this trade-off.Source: E Spectral Feedback scRMSD Discussion · p. 22

F.1 Spectral Feedback Applications

  • Spectral Feedback improves a broad set of alignment techniques, including arbitrary methods such as Best-of-N or DRAKES when their reward gains diminish.Figure 12 reports results with k = 20, D = 8192, and 5 feedback iterations.Source: §F.1 · p. 22, Figure 12 · p. 22
  • Feedback can use substantial compute when many samples are used for edit selection, but is useful when a target alignment model reaches insufficient reward values.The passage emphasizes making each feedback step as efficient as possible.Source: §F.1 · p. 22
  • Compared with competing edit-position selection methods, Spectral Feedback with ProxySPEX scales well in alignment performance and latency.The passage points to Figure 6 for this comparison.Source: §F.1 · p. 22

F.2 Additional Alignment Oracles · G High-Order Interactions Case Study (∆∆G Alignment)

Spectral Feedback is evaluated with ProtGPT2 log-likelihood and a multi-objective reward alongside the ∆∆G oracle; the multi-objective examples show protein-dependent tradeoffs and suggest a way to balance objectives and reduce overfitting.

F.2 Additional Alignment Oracles

  • The paper examines Spectral Feedback beyond ∆∆G alignment using two additional oracles: ProtGPT2 log-likelihood and a multi-objective reward combining log-likelihood with ∆∆G.This tests flexibility across reward oracles as well as protein models.Source: §F.2 · p. 22
  • The ProtGPT2 log-likelihood oracle scores diffusion-generated sequences with ProtGPT2, a language model trained to capture protein amino-acid sequence distributions, using log-likelihood as a naturalness measure.Prior works using this oracle include Gruver et al. [2023] and Cemri et al. [2024]; ProtGPT2 was trained as described by Ferruz et al. [2022].Source: §F.2 · p. 22
  • Figure 13 presents a pretrained-model alignment trajectory using the ProtGPT2 oracle and first-order LASSO, with k = 10, D = 2048, and n = 16 oracle calls per value-function call.The supplied caption gives the setup but not trajectory values.Source: §F.2 · p. 22, Figure 13 · p. 23
  • R6-560 shows a direct tradeoff between ∆∆G and ProtGPT2 log-likelihood, while 2KRU attains higher log-likelihood when alignment primarily targets ∆∆G.These examples demonstrate that the tradeoff can differ by protein.Source: §F.2 · p. 23
  • The multi-objective reward is r(x) = α·∆∆G(x)+0.005·(1−α)·ProtGPT2(x), with α balancing the objectives and 0.005 rescaling ProtGPT2.The rescaling makes α more linear in weighting the importance of the two oracles.Source: §F.2 · p. 23
  • Balancing objectives with a multi-objective oracle may regularize Spectral Feedback against overfitting; the paper suggests balancing ∆∆G with scRMSD to avoid high-scRMSD proteins.The scRMSD example refers to proteins identified in Figure 11.Source: §F.2 · p. 23

G.1 Comparing ProxySPEX with LASSO … Iteration 5

ProxySPEX outperformed first-order LASSO on two protein backbones, with results suggesting that the advantage of high-order recovery depends on the backbone and its interaction structure. The appendices trace distinct ProxySPEX and LASSO sequence updates across feedback iterations for R6-560 and 2KVV.

G.1 Comparing ProxySPEX with LASSO

  • The authors attribute varying recovery performance to backbone-dependent interaction structure, while noting that first-order interactions dominate for most test-set proteins.They suggest α-helices in 2KVV may reflect more local interactions, whereas R6-560 has more significant high-order interactions; LASSO slightly outperforms ProxySPEX for DRAKES.Source: §G.1 · p. 24
  • ProxySPEX’s final reward exceeded LASSO’s by 8.3% for 2KVV and 24% for R6-560.Figure 15 compares the strategies for the pretrained model with edit selections limited to at most 20 positions.Source: §G.1 · p. 23, Figure 15 · p. 24

G.2 Target Protein: R6-560 (r6_560_TrROS_Hall) / G.3 Target Protein: 2KVV (v2K43S_2KVV)

  • The passages show two initial sequence strings and two result strings; one result is accompanied by seven listed substitutions at positions 4, 9, 15, 26, 41, 50, and 52.Source: §G.2 · p. 25, §G.3 · p. 26

Iteration 3

  • The passages report four distinct resulting sequences, without establishing their iteration or assignment to R6-560 or 2KVV.The sequences are APPPRVVTVLVAVTRADGRTELRRVTLTGLPRELRPGDVVTLPETGQRATIVEILP; APPPRVVTVEVAVTRADGKTELVTVTLRDTPRELRPGDVVTLPETGQKATIVRIIP; ERWIEERELMAATGLARSTITRLRRTRLEEGRHYRRVSKDGRDSPDATILYNLEAILELLA; and ERWIEERELMKETGLARSTITRLRRTKLKEGIHYRRVSADGKPSPDATILYNLEAIKELLA.Source: Iteration 3 · pp. 25–27

Iteration 4

  • The passages report four resulting sequences, with the first and second differing from each other and the third and fourth likewise differing; they do not establish iteration or backbone assignments.In passage order: APPPRVVEVLVAVTEADGRTVLRRVTLTGLPRELRPGDVVTLPETGEEATIVEILP; AAPPRVVTVEVAVTRADGKTELVTVTLTDLPRELKPGDVVTLPETGQKATIVRIIP; ERWIEERELMAATGLRRSTIARLRRTALEEGRHYRRVSADGRDSPDARILYNLEAILALLA; ERWIEERELMKETGLARSTITRLRRTRLKEGIHYRRVSADGKDSPDATILYNLEAIKALLA.Source: Iteration 4 · pp. 25–27

Iteration 5

  • The three cited passages each report a result sequence: APPPRVVEVLVAVTEADGRTVLRRVTLTGLPRELRPGDVVTLPETGERATIVEILP; AAPPRVVTVEVAVTRADGKTELVTVTLTDLPRELRPGDVVTLPETGQKATIVRIIP; and ERWIEERELMAATGLARSTLAALRRTALEEGRHYRRVSADGRDSPDARILYNLEAILALLA.Source: Iteration 5 · pp. 25–26

Related papers

Loading 2609.30456v1…