acceptodds
Under review as a conference paper at ICLR 2027

FARFlow: Learning Finite-Span Flow Maps for Few-Step Protein Backbone Generation

Abstract

Diffusion- and flow-based models generate high-quality protein backbones, but reducing sampling cost requires reliable predictions over finite time intervals. We introduce Finite-Span Anchoring and Replay Flow (FARFlow), a solution-function-based method for few-step backbone generation. FARFlow separates the requested prediction span from the input noise level through an unscaled, zero-anchored additive condition. This representation leaves the local-velocity condition unchanged at zero span while providing a smoothly varying correction for finite-span requests. Direct endpoint anchoring supervises predictions using paired clean backbones, and fresh-noise replay extends this supervision to inputs reconstructed from predicted endpoints. Under constrained budgets, FARFlow achieves higher designability than compared baselines while retaining reasonable structural diversity. On the standard multi-length benchmark, it maintains high generation quality across all evaluated lengths, achieving aggregate designability comparable to higher-budget multi-step models with substantially fewer network evaluations.

Primary area:
generative models
Keywords:
protein backbone generation, few-step generation, flow matching, solution-function learning, de novo protein design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iQuROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FARFlow for few-step protein backbone generation, building on Proteína's Cα-coordinate Transformer and the SoFlow solution-function framework. The method combines three components: (1) zero-anchored span conditioning, which ensures that the added span condition vanishes at zero span; (2) direct endpoint anchoring, which supervises clean-endpoint predictions using paired ground-truth backbones; and (3) fresh-noise replay, which re-noises detached endpoint predictions and supervises the resulting predictions against the same clean backbones to mitigate the training–inference mismatch. Experiments demonstrate high computational designability at 20–100 NFE. However, the generated backbones show relatively high similarity to reference structures in AFDB, indicating limited structural novelty by this metric.

Critical strengths
  • The paper introduces an unscaled, zero-anchored span representation to support both local velocity learning and finite-span backbone prediction. It also employs Fresh-noise Replay to construct training inputs from the model's own predictions, thereby mitigating the distribution mismatch between training and inference.
  • On the 60-128 residue benchmark, FARFlow-100 achieves 95.8% designability, which is higher than QFlow (88.5%) and RMF (83.2%). Ablation experiments show that retraining with the model's predicted skeleton has better performance than training with the real skeleton with noise (79.2% vs. 76.4%), indicating that Replay can bring additional performance improvement.
Critical weaknesses
  • Table 1 only compares QFlow and RMF, where QFlow is the unrectified base version in ReQFlow and not the final accelerated model.As the direct basis FARFlow Proteina, also lack of quantitative comparison under the same NFE.In Table 2, RFdiffusion achieves 94.4% designability using only 50 NFE, while FARFlow-50 achieves only 84.4% and requires 100 NFE to achieve 94.8%. Although FIG. 5 demonstrates the sampling speed advantage of FARFlow, the time-consuming comparison of QFlow and RMF is not included.
  • The paper is missing key training and computation configurations, including the source, Size, and protein length distribution of the training dataset, whether the model is trained from scratch, hyperparameters such as learning rate, Batch Size, number of training steps, and loss weight, as well as GPU configuration and training time. The training overhead of adding Anchor and Replay is also not reported. This lack of information affects the reproducibility of the experiment and also makes it difficult to evaluate the actual computational cost of the method.
  • The paper mainly evaluates FARFlow at 20–100 NFE, without reporting generation results at 1, 2, 5, or 10 NFE.The experiments only cover backbones up to 250 residues. Figure 3a shows that FARFlow-50's designability decreases from 92% at 150 residues to 80% at 250 residues, motivating further evaluation on longer proteins.The AFDB maxTM of FARFlow-100 in Table 2 is 0.735, which is higher than 0.664 of RFdiffusion and 0.600 of Genie2, indicating that its generated structure is more similar to the known structure. The relationship between high designability and low structural novelty has not been fully analyzed in papers.
Questions for authors
  • Why does Table 1 select the unrectified QFlow and not compare the final ReQFlow? Could the authors provide quantitative results for ReQFlow and Proteína at 20, 50, and 100 NFE, and compare the generation quality and the actual running time under the same evaluation conditions?
  • Could the authors provide details on the training dataset, model initialization, key hyperparameters, GPU configuration, and training time? How much do Anchor and Replay add to the training calculation cost respectively?
  • Could the authors provide representative generation results at 1, 2, 5, and 10 NFE, as well as for proteins with 300–500 residues, to evaluate the method's performance under extremely low NFE budgets and on longer proteins? Moreover, could the authors further analyze whether there is a trade-off between FARFlow's high designability and its relatively lower structural novelty?
1 comment

Dear Reviewer iQuR,

Thank you again for your constructive comments. As the author–reviewer discussion period is drawing to a close, we would like to kindly ask whether our responses and revised manuscript have addressed your concerns.

Following your suggestions, we have added matched-NFE comparisons with ReQFlow and Proteína, additional sampling-time measurements, detailed training configurations, and measurements of the computational overhead of Anchor and Replay. We have also evaluated FARFlow at 1, 2, 5, and 10 NFE and on longer backbones of 300–500 residues. Furthermore, we have expanded our analysis of the trade-off between designability, structural diversity, and database novelty, explicitly acknowledging the limitations of our method.

These additions have been incorporated into the revised manuscript, with changes highlighted in blue. We would greatly appreciate any further feedback and would be happy to clarify any remaining questions before the discussion closes.

Best regards,

The Authors

Reviewer_4jo9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FARFlow, a solution-function-based framework for efficient protein backbone generation. FARFlow introduces zero-anchored span conditioning to improve finite-span prediction and endpoint anchoring with fresh-noise replay to improve clean-backbone reconstruction from analytic and model-induced noisy inputs.

Critical strengths
  1. The zero-anchored conditioning mechanism can be integrated into Proteina's existing workflow.
  2. The distribution mismatch between training and inference is indeed the key challenge in zero/few-shot protein generation.
  3. The empirical result is promising.
Critical weaknesses
  1. Limited evaluation of long-chain and very-low-step generation. The evaluation is restricted to proteins with length of at most 250 residues. One of the important advantages of Proteina is its ability to generate high-quality, long-chain protein backbones, including structures with up to approximately 800 residues.
  2. The experiments are restricted to 20, 50, and 100 NFE. Results at lower budgets, e.g. 1, 2, 4, 8, and 16, would help determine whether the model learns accurate large-span transformations
  3. The obstacle of sampling speed is a bit exaggerated. Faster models are surely better but Proteina removes the expensive tri layers is already order of magnitude faster. The real bottleneck is the structure prediction method, which is required to obtain the full protein after inverse folding.
  4. The paper evaluates only unconditional monomeric backbone generation. Conditional generation tasks, particularly motif scaffolding, are important applications in practical protein design and have been evaluated by related backbone generation methods.
Questions for authors
  1. How much of the improvement attributed to zero-anchored span conditioning comes specifically from subtracting the zero-span embedding?
  2. How does FARFlow perform with fewer than 20 network evaluations, particularly at 1, 4, 8, and 16 NFE?
  3. Can the proposed training strategy be extended to motif scaffolding while preserving motif geometry and generation efficiency?
  4. What is the end-to-end speedup when including ProteinMPNN-based inverse folding and ESMFold-based structure validation?
  5. Are the improvements from fresh-noise replay consistent across independent training runs?
1 comment

Dear Reviewer 4jo9,

Thank you again for your constructive comments. As the author–reviewer discussion period is drawing to a close, we would like to kindly ask whether our responses and revised manuscript have addressed your concerns.

Following your suggestions, we have isolated the contribution of zero-span embedding subtraction, verified the benefits of fresh-noise replay across five independent training runs, and extended our evaluation to lower NFE budgets. With tuned sampling, FARFlow achieves 89.0% and 99.8% designability at 8 and 16 NFE, respectively, using the same non-distilled checkpoint. We have also measured end-to-end pipeline costs and clarified the current limitations regarding long-chain generation and motif scaffolding.

These additions have been incorporated into the revised manuscript, with changes highlighted in blue. We would greatly appreciate any further feedback and would be happy to clarify any remaining questions before the discussion closes.

Best regards,

The Authors

Reviewer_8oA5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds on Proteina's backbone generative model that operates on CA coordinates and applies the SoFlow few step generation objective to it. In addition they add additional things, amoung them zero-anchored additive span conditioning, direct endpoint anchoring and fresh-noise replay. They test designability across a range of NFEs and show good performance.

Critical strengths

The authors tackle a very relevant problem, making protein design models cheaper to run. Their designability gain versus the budget is interesting, for example at 100 NFEs. The throughput meaured is a useful metric, and the limitations are clearly mentioned.

Critical weaknesses

[W1] it seems like the headline claims are not fully supported: eg the replay vs control differs only by 3 percentage points without any replicas, and the largest contributor claim for the zero anchoring is a bit convolved with the span scale change, frequence base change etc, so it is not really clear if zero anchoring is actually the contributing factor. More isolation and reporting with replicas would make this stronger.

[W2] Few step baselines seem badly tuned since they only produce noise, and at actual interesting few steps (1-20) the authors either do not evaluate or the difference is actually super small, so the advance is not really clear.

Questions for authors

What are FARFlow’s training data, model size, initialization and training compute, and what are those of the baselines?

Please provide seeds, CIs and larger samples for Table 3

How do tuned few-step samplers for Proteína, SPID and FrameFlow/Genie2 compare at 20/50 NFE?

Can you give cluster-based diversity, PDB novelty and secondary-structure statistics against natural references?

What are the one-step and 2–8-step results?

How were the ρ, λ̃ and time-grid settings chosen, and how sensitive are results?

1 comment

Dear Reviewer 8oA5,

Thank you again for your constructive comments. As the author–reviewer discussion period is drawing to a close, we would like to kindly ask whether our responses and revised manuscript have addressed your concerns.

Following your suggestions, we have expanded the ablation studies to five independent training runs with confidence intervals, isolated the effects of individual span-encoding components, and added evaluations at 1–8 NFE and comparisons with additional baselines at 20/50 NFE. We have also provided further training details and analyses of structural diversity, database novelty, and secondary-structure composition.

These additions have been incorporated into the revised manuscript, with changes highlighted in blue. We would greatly appreciate any further feedback and would be happy to clarify any remaining questions before the discussion closes.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.