acceptodds
Under review as a conference paper at ICLR 2027

Proteina-Pronto: Fast De Novo Protein Design and Structure Prediction via Distillation

Abstract

Accurate flow-matching protein design models and diffusion-based structure prediction models are bottlenecked by the hundreds of network evaluations required at inference time. This restricts large-scale discovery campaigns that require generating and scoring between and candidates per target. We introduce Proteina-Pronto, a framework for accelerating atomistic generation and folding-based evaluation. For generation, we replace low-temperature stochastic sampling with a deterministic ODE using Temporal Score Rescaling, then couple the modality-specific denoising schedules to a shared global time for diffusion distillation. This enables a 16x reduction in sampling steps for both unconditional sequence-structure co-design and target-conditioned binder design, with design performance comparable to the original teacher models. For folding-based evaluation, we distill Boltz-2 diffusion module and accelerate its computationally expensive Pairformer trunk through representation distillation, achieving a 3.9x end-to-end wall-clock speedup while retaining approximately 99% of its original lDDT on the PDB test set. Together, these accelerations enable more efficient inference-time scaling with best-of-N sampling and guided search, yielding up to 6x as many unique successful designs on hard targets within the same total GPU budget as the original pipeline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
distillation, folding model, protein design

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.20 over 5 reviews, above 96% of the others.
Reviewer_gtzLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the high computational cost of protein generation and structure prediction, where iterative sampling and expensive model architectures limit the throughput of large-scale protein design campaigns. The authors propose Proteina-Pronto, a framework that accelerates both protein generation and folding-based evaluation through distillation, aiming to preserve model quality while substantially reducing inference costs.

For protein generation, the authors first replace low-temperature stochastic sampling with a deterministic ODE using Temporal Score Rescaling (TSR). They then introduce a schedule-coupling formulation that maps the separate denoising schedules of backbone coordinates and residue-level latent variables onto a shared global time, enabling the application of diffusion distillation to multimodal product-space models. Several distillation approaches are investigated, including Terminal Velocity Matching (TVM) and π-Flow. For structure prediction, the authors accelerate Boltz-2 by distilling its diffusion module into a four-step sampler and replacing its computationally expensive Pairformer trunk with a smaller architecture trained through representation distillation.

The framework is evaluated across unconditional atomistic protein generation, target-conditioned binder design, and biomolecular structure prediction. The distilled protein generators achieve performance comparable to their original 256-step samplers using only 16 steps. For structure prediction, the accelerated Boltz-2 retains approximately 99% of the original model's lDDT on the PDB test set while achieving a 3.9 times faster end-to-end inference. By combining accelerated generation with faster structure-based evaluation, Proteina-Pronto enables more efficient inference-time scaling, producing up to approximately six times as many unique successful binder designs per GPU-hour on challenging targets. Overall, the study demonstrates that distillation can substantially improve the computational efficiency of protein design pipelines while largely preserving the predictive and generative performance of the original models.

Critical strengths
  1. Substantial Computational Efficiency Gains with Strong Experimental Validation

The paper demonstrates substantial inference acceleration across three protein modeling systems: La-Proteina, Proteina-Complexa, and Boltz-2. The distilled generative models reduce sampling from 256 to 16 steps while maintaining competitive protein generation and binder-design performance. For structure prediction, Boltz-2distill achieves a 3.9× end-to-end speedup while retaining approximately 99% of the original model's lDDT. Importantly, the evaluation extends beyond individual model efficiency to complete protein design workflows, demonstrating up to approximately 6× more unique successful binder designs per GPU-hour. The extensive experiments, including ablation studies, structural quality assessments, and downstream evaluations, provide convincing evidence of the framework's practical value.

  1. Effective Adaptation of Distillation Techniques to Protein Design and Structure Prediction

The paper addresses several practical challenges in accelerating protein generative models. For La-Proteina and Proteina-Complexa, the authors introduce a schedule-coupling formulation that enables distillation of multimodal generation processes with asymmetric denoising schedules. They also propose self-forcing self-conditioning to reduce the training–inference mismatch in few-step generation, supported by clear improvements in binder-design performance. For Boltz-2, they combine diffusion distillation with representation distillation to accelerate both structure generation and the computationally expensive Pairformer trunk. Although the individual techniques build on established methods, their effective adaptation and integration across different protein modeling tasks constitute a meaningful technical contribution.

Critical weaknesses
  1. Potential Limitations in Structural Diversity Due to Fixed Modality-Coupling Schedules

The proposed multimodal distillation framework restricts the independently conditioned denoising processes of backbone coordinates and residue-level latents to a predefined one-dimensional temporal trajectory. Although the original teacher models also use asymmetric schedules during inference, their two-dimensional time-conditioning formulation allows alternative denoising trajectories. By distilling only a fixed trajectory, Proteina-Pronto may limit the range of protein conformations accessible through different backbone–side-chain generation sequences, particularly for structurally complex proteins. While the authors demonstrate that structural diversity is largely preserved relative to the original sampling baseline, they do not investigate whether alternative coupling schedules could expand the accessible conformational space. A systematic comparison of generation quality and structural diversity across different coupling schedules would help clarify this potential limitation.

  1. Insufficient Validation of Binder Ranking Consistency in the Distilled Folding Model

The paper demonstrates that Boltz-2distill retains approximately 99% of the original Boltz-2's lDDT while substantially reducing inference time. However, preserving average structure prediction accuracy does not necessarily guarantee that the distilled model maintains the same ranking of candidate binders as the original model. This distinction is particularly important because Boltz-2distill is used as a reward model for inference-time search, where the relative ordering of candidates directly influences which designs are selected. Although the authors provide downstream evaluations using independent folding models, they do not directly assess ranking consistency between Boltz-2 and Boltz-2distill. An analysis of rank correlations, top-k candidate overlap, and agreement in selecting high-quality binders would strengthen the evidence that inference acceleration preserves the reliability of candidate prioritization.

Questions for authors
  1. Impact of Fixed Modality-Coupling Schedules on Structural Diversity

The proposed distillation framework restricts backbone coordinates and residue-level latents to a predefined one-dimensional denoising trajectory. Although the authors demonstrate that structural diversity is largely preserved relative to the original teacher model, could this fixed coupling schedule exclude protein conformations that might be accessible through alternative backbone–side-chain denoising trajectories? Have the authors evaluated different coupling schedules, such as uniform or alternative asymmetric schedules, to determine their effects on structural diversity, co-designability, and binder-design success rates?

  1. Binder Ranking Consistency Between Boltz-2 and Boltz-2distill

The paper reports that Boltz-2distill retains approximately 99% of the original Boltz-2's lDDT. However, maintaining average structure prediction accuracy does not necessarily imply preserving the relative ranking of candidate binders. Given that Boltz-2distill is used as a reward model during inference-time search, have the authors evaluated whether it produces rankings consistent with the original Boltz-2? In particular, could the authors report rank correlations (e.g., Spearman's ρ), top-k candidate overlap, and agreement in identifying high-quality binders under the same candidate pool?

I would be happy to reconsider and potentially increase my overall rating if the authors can provide convincing evidence addressing these two concerns, particularly through additional analyses demonstrating that the fixed coupling schedule does not substantially constrain structural diversity and that Boltz-2distill preserves reliable binder ranking and candidate selection.

2 comments
AuthorsOpenReview ↗

Author response I

We thank the reviewer for taking the time to review our work.

While the authors demonstrate that structural diversity is largely preserved relative to the original sampling baseline, they do not investigate whether alternative coupling schedules could expand the accessible conformational space. A systematic comparison of generation quality and structural diversity across different coupling schedules would help clarify this potential limitation.

We follow the exponential-backbone/quadratic-latent coupling established by La-Proteina. Its extensive schedule ablation (Appendix G.2, Table 8) shows that only tested configurations generating the backbone faster than the latents reach 50% all-atom co-designability, with the selected pairing achieving the highest co-designability while maintaining high structural diversity. This scheduling choice was subsequently adopted by Proteina-Complexa and RFD4-Proteina. These results lead us to expect modest gains, if any, from further tuning within the studied schedule families, although broader gains in conformational coverage remain untested. We acknowledge that our current students are specialized to a fixed coupling path and do not retain the teacher’s flexibility to change paths at inference. A possible extension would condition the student on schedule parameters and distill trajectories across a family of monotone paths. We leave this extension to future work and focus here on accelerating the empirically validated teacher configuration.

Insufficient Validation of Binder Ranking Consistency in the Distilled Folding Model

We agree that average lDDT alone does not establish ranking consistency. We therefore evaluated the same 1,900 binder candidates, 100 for each of 19 targets, with original Boltz-2 and our four-step distilled model. The candidates were generated using Proteina-Pronto_TSR with 256 sampling steps. We computed Spearman’s rank correlation separately within each target and averaged these correlations across targets. To contextualize the results, we also repeated the original Boltz-2 evaluation with different random seeds, providing a reference for its stochastic repeatability.

MetricMean ρ: teacher–studentMean ρ: teacher–teacher, different seeds
Complex pLDDT0.8260.910
Mean interface PAE0.7360.794
Binder Cα scRMSD0.7620.797

These results provide direct evidence of substantial within-target ranking agreement beyond preservation of average structure-prediction accuracy. The teacher–student correlations approach the teacher-repeat reference particularly closely for interface PAE and binder scRMSD, while pLDDT shows a larger remaining gap. We added these results in section H2 “Rank-agreement between Boltz-2 and Boltz-2-distill”.

I appreciate the authors for the earnest rebuttal. When I consider the scope of this research, my concerns are resolved clearly via the rebuttal. So, I'd like to raise my rate.

Reviewer_XqXROpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Proteina-Pronto, which uses TSR and few-step diffusion distillation to reduce the computational costs of both protein generation and evaluation while preserving practically effective sampling behavior. The saved computational budget is then used to generate more candidates, addressing an important bottleneck in the protein design pipeline.

The experiments explicitly demonstrate a 16-fold reduction in sampling steps at a generation quality comparable to that of the 256-step teacher, along with a 3.9-fold end-to-end speedup in structure prediction while maintaining a relatively high lDDT, supporting the effectiveness of the method.

Critical strengths
  1. The paper provides evidence of accelerated generation as well as extensive structural analyses, such as Table 5 and Figure 6. These results support the claim that the model maintains both success rate and diversity, rather than collapsing to a small set of structures that are easy to generate.
  2. The paper also implements a complete protein design pipeline, demonstrating benefits in a practical workflow and providing a clear breakdown of where these gains come from.
Critical weaknesses
  1. Ablation studies on the TSR hyperparameters appear to be missing, or their results are not clearly highlighted in the paper.
  2. Proteina-Pronto both removes triangle attention and applies distillation, but the experiments do not clearly show how much performance distillation recovers after triangle attention is removed. Such experiments would help readers understand the specific contribution of distillation.
Questions for authors
  1. Could the authors provide ablation studies on the TSR hyperparameters, showing their effects on folding and end-to-end metrics and assessing sensitivity to these hyperparameters? Could the authors also clarify the specific contribution of distillation after triangle attention is removed?
2 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for taking the time to review our paper.

Ablation studies on the TSR hyperparameters appear to be missing, or their results are not clearly highlighted in the paper.

For protein generation, Section G.2 and Table 6 already report TSR hyperparameter ablations for unconditional monomer generation, varying the backbone and latent scaling parameters at 32, 64, and 128 sampling steps. These results characterize the effects on co-designability, structural diversity, and secondary-structure composition. For folding, we have now evaluated additional TSR settings for the Boltz-2 teacher, using the 200-step ODE sampler with full MSAs and varying only k and σ_TSR:

kσ_TSRPDB-test lDDTCASP15 lDDT
1.2 (default)25600.8210.444
1.525600.8220.445
1.22560.8220.442
1.2160.8150.438

Teacher lDDT changes little when increasing k from 1.2 to 1.5 or reducing σ_TSR from 2560 to 256, while reducing σ_TSR to 16 produces a modest decrease on both datasets. These teacher-sampling ablations complement Section F.2.1 and Figure 11, which explain the use of noise-dependent TSR scaling to stabilize distillation.

We added results for all metrics, both test sets, and the different TSR parameters in section H.5 “Sensitivity to TSR Hyperparameters in Folding”.

How much does distillation recover after triangle attention is removed

We agree that this ablation is important. We evaluated the pretrained Boltz-2 model after simply removing triangle attention from all 64 Pairformer blocks, retaining the remaining teacher weights without further training. Keeping the original diffusion module and its 200-step sampler fixed across configurations, this removal reduces PDB-test lDDT from 0.82 to 0.665; our representation distillation restores it to 0.818. Similarly, TM-score drops from 0.84 to 0.7 and recovers to 0.842 after distillation. On CASP15, lDDT follows the same pattern: 0.443 for the original trunk, 0.333 after removal without training, and 0.44 after representation distillation. These results demonstrate that trunk distillation recovers nearly all the accuracy lost through triangle-attention removal, independently of diffusion-module distillation. We added this ablation in Section H.6 “Contribution of Trunk Distillation after Removing Triangle Attention”, to make the contribution of representation distillation explicit.

Thank you for the additional experiments. My concerns about TSR hyperparameter ablations and the contribution of trunk distillation are largely resolved. In particular, the triangle-attention removal ablation clearly demonstrates that representation distillation recovers most of the lost accuracy. The folding ablations also suggest limited sensitivity to the tested TSR settings. A remaining limitation is that these sensitivity results concern teacher sampling rather than the distilled model and full downstream pipeline, so robustness of end-to-end design performance is not yet directly established. This does not materially change my positive assessment of the practical acceleration and extensive evaluation. I therefore maintain my rating of 4.

Reviewer_AqpMOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In the field of protein generation, this paper achieves acceleration through techniques such as SDE to ODE conversion, time-scheduled coupling, and diffusion distillation. In the field of folding, the authors achieve acceleration through diffusion distillation and representation distillation. Simultaneously, the authors also adjusted subsequent search strategies to further accelerate the process.

Critical strengths

This paper presents very thorough experiments, including SDE vs. TSR-ODE, different distillation algorithms (TVM, MeanFlow, π-Flow, DMD), TSR hyperparameters, and self-conditioning switches. I have great respect for this work.

Critical weaknesses

1.The training steps for DMD are contradictory. On page 18, Table 2 and the main text state that DMD trained 20k steps. However, on page 24, F.1.1, the same method is recorded as 40k steps. Please confirm if this is a typo.

  1. Page 18, A.2 compares the structural diversity of the 8-step distillation method, but this is inconsistent with Table 5 on page 32: MeanFlow 176 vs 213, TVM 202 vs 229, π-Flow 207 vs 207. I need a reasonable explanation.

  2. Missing reference a. SeedFold: Scaling Biomolecular Structure Prediction b. SimpleFold-Turbo: Adaptive Inference Caching Yields 14-fold Acceleration of Flow-matching Protein Structure Prediction c. Physically Valid Biomolecular Interaction Modeling with Gauss-Seidel Projection d. Reshaping biomolecular structure prediction through strategic conformational exploration with HelixFold-S1 e. Towards Sub-Second Molecular Docking as a Structural Primitive: A Quantized Consistency Diffusion Framework

Questions for authors
  1. Please explain the mismatch between the two experimental data sets mentioned above and supplement the relevant citations.

  2. On page 38 (PDF page 38), Figure 21, I was surprised to find that the lddt of the 4step ODE boltz-2 is only around 0.518. Could you try using a Protenix-Mini? I don't think it should be that low.

If you can resolve my concerns regarding a few experimental data points, I'm willing to raise the score to 4.

1 comment
AuthorsOpenReview ↗

Author response

We thank the reviewer for taking the time to review our paper.

A few inconsistencies (training steps for DMD are contradictory, A.2 compares the structural diversity of the 8-step distillation method, but this is inconsistent with Table 5)

Thank you for catching these inconsistencies. The evaluated DMD checkpoint was trained for 20k steps, as correctly stated in Table 2. Likewise, the eight-step diversity values in Table 5 agree with Figure 3 and are correct: 229 structural clusters for TVM, 213 for MeanFlow, and 207 for π-Flow. Both discrepancies arose because the appendix retained descriptions and results from earlier runs; we overlooked updating these passages when incorporating the final results. We corrected these inconsistencies in the paper.

Missing references

Thank you for pointing out these relevant works. We added these references, with brief descriptions, in the related work section (added text shown in red).

I was surprised to find that the lddt of the 4step ODE boltz-2 is only around 0.518. Could you try using a Protenix-Mini? I don't think it should be that low.

The low lDDT reflects the use of OpenStructure’s stereochemistry checks, which are enabled throughout our submission. Following the reviewer’s suggestion, we evaluated Protenix-Mini at four sampling steps using the same MSAs and compared it with the full-MSA four-step Boltz-2 ODE baseline (step-scale=1.0) and our distilled model. For each model, we rescored the same predictions with stereochemistry checks enabled and disabled:

Model, all 4 stepsPDB-test lDDT (checks enabled)PDB-test lDDT (checks disabled)
Boltz-2 ODE, full MSA0.5190.825
Protenix-Mini0.5170.790
Our distilled model0.8110.824

Both undistilled ODE baselines achieve substantially higher lDDT when stereochemistry checks are disabled, whereas our distilled model’s score changes only slightly. These results show that high unchecked lDDT can coexist with local stereochemical defects, explaining the unexpectedly low checked lDDT despite largely preserved global structure.

We include stereochemical checks because all-atom structure prediction should preserve physically plausible local geometry alongside structural similarity. Unchecked lDDT can remain high despite severe bond-length or bond-angle distortions and steric clashes; the original lDDT formulation introduced these checks specifically to account for such violations (Mariani et al., 2013). Reporting both settings makes this distinction explicit: our distilled model retains high lDDT when these geometric constraints are taken into account.

Consistent with our findings, concurrent work also reports that accelerated folding predictions can retain high unchecked lDDT despite substantial stereochemical violations, with markedly lower scores once stereochemistry checks are enabled (paper “Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2”).

We have added the Protenix-Mini comparison, the results with stereochemistry checks enabled and disabled, and a discussion of the scoring protocol in Section H.4, “Four-step Folding: Protenix-Mini and Stereochemistry Checks”.

Reviewer_cDgEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Proteina-Pronto accelerates La-Proteina and Proteina-Complexa through deterministic low-temperature TSR sampling, coupled modality schedules, and distillation. For Boltz-2, it combines a four-step diffusion student with a distilled trunk without triangle attention, achieving a 3.9×3.9\times end-to-end speedup while retaining approximately 99% of the teacher's PDB lDDT.

Critical strengths
  1. Accelerating both generation and evaluation addresses a practical bottleneck. Measuring unique successful designs against total compute demonstrates benefits beyond fewer network calls.
  2. Using TSR to moderate velocity scaling at high noise levels is a useful adaptation for stable low-temperature distillation. Schedule-aware parameterization and self-forcing self-conditioning address further task-specific difficulties.
  3. The appendix provides meaningful ablations of self-conditioning, TSR parameters, loss weighting, and trunk versus diffusion distillation. Self-forcing raises 16-step binder-design performance from 60 to 180 unique successes and from 8 to 14 solved targets (Section C.2). Structural-quality metrics and DMD failures expose problems that backbone designability misses.
Critical weaknesses
  1. Ligand generalization. Section G.7 notes that CASP15 interface and ligand metrics cover only 12--14 targets. The size of the PDB test set and of its ligand-bearing subset is never stated, so the remaining ligand evidence cannot be weighed. Report both, and consider a cutoff-filtered Runs N' Poses evaluation.

  2. Inference flexibility (folding). The default π\pi-Flow student fixes segmentation during training. The 1- and 2-step students are separately trained, as are the PDD students, one per maximum block length. No single folding checkpoint is shown across budgets, and training cost scales with the number of budgets supported. TVM already provides this on the generation side; the folding section should match it or explain why it cannot.

  3. Framework clarity. Generation requires coupled schedules and self-conditioning; folding uses EDM-specific parameterization, loss weighting, and trunk distillation. Presenting these as two related methods sharing deterministic teacher construction and distillation would clarify their scope.

  4. Trunk novelty. https://arxiv.org/html/2510.18870v2 established triangle attention's cost and the effectiveness of removing it. Section 2 cites it and Section F.3.1 attributes it explicitly, but Section 3.2 introduces the student without reference; add the citation at the point of use. The contribution is distilling an existing Boltz-2 trunk rather than training one from scratch. The 59% figure is useful Boltz-2 profiling. The student retains the sequence track, which PairMixer drops entirely, so the two architectures are not the same.

  5. Quality--throughput tradeoffs. Sixteen-step TVM achieves 77.4% all-atom co-designability versus 82.2% for the 256-step SDE baseline, while preserving structural diversity (Table 5). BoltzDesign strict protein-target successes fall from 99 to 73 at equal sample count, despite higher successes per GPU-hour (Table 8); Table 1 reports only the ratios. Foreground the absolute counts. Quality is also non-monotonic in step count, with TVM scoring higher at 32 steps than at 256, which deserves comment. Downstream success remains an in-silico proxy.

  6. Baseline fairness. Figure 21 uses a smaller MSA for the undistilled four-step ODE baseline. The 0.811 versus 0.518 lDDT gap is too large for MSA depth alone to explain, but this is the comparison that isolates distillation and should be rerun matched.

Questions for authors
  1. Can you evaluate Runs N' Poses systems released after Boltz-2's June 1, 2023 training cutoff, excluding student-training overlap and preferably stratifying by training-set similarity?

  2. Can any single folding checkpoint run at other budgets, such as 1--10 network calls? If not, report one PDD checkpoint's supported range and the training cost of covering several budgets.

  3. Does the undistilled four-step ODE baseline's lDDT deficit persist with matched MSAs? This would isolate distillation's benefit.

4 comments
AuthorsOpenReview ↗

Author response I

We thank the reviewer for the thoughtful review.

The size of the PDB test set and of its ligand ... is never stated, ... ligand evidence cannot be weighed. Report both, consider a cutoff-filtered RunsNPoses eval.

For our 4-step distilled model, global structure metrics are evaluated on 540 PDB targets and ligand metrics on 250 targets. The CASP15 evaluation has 66 targets, including 14 contributing interface metrics and 12 contributing ligand metrics. We added these dataset sizes explicitly in Sec. G.7 (new text shown in red) and the caption of Figure 21.

We thank the reviewer for suggesting the Runs N’ Poses benchmark. We curated a subset with systems released on or after 1 June 2023, Boltz-2’s training cutoff, excluding systems that lack an MSA, contain metal complexes, or exceed H100 GPU memory. The resulting subset contains 968 systems and 1,069 ligands. None of these systems was used as a training target or template for student training. Ions and crystallization artifacts are modeled but not scored.

We evaluated every system with both Boltz-2 at 200 steps (step scale 1.5) and Boltz-Pronto at 4 steps. Both used three recycling iterations, five samples, no templates, and the benchmark’s own MSAs. For each system, we evaluated the prediction ranked first by the model’s confidence head. A ligand is successful when lDDT-PLI > 0.8 and BiSyRMSD<2A.

The table reports ligand success rates, stratified by the benchmark-provided similarity score to reference structures published before the training cutoff:

SimilarityLigandsBoltz-2 (200 steps), success (%)Boltz-Pronto (4 steps), success (%)
0–20352.92.9
20–30345.98.8
30–404814.618.8
40–507428.420.3
50–6010340.842.7
60–7017665.360.2
70–8019178.075.9
80–10037689.483.8
Not available3237.540.6
All ligands1,06964.160.9

Boltz-Pronto tracks the success rate of the Boltz-2 teacher closely, even on ligands with low similarity to the training data.

Inference flexibility (number of steps) for distilled folding model...

We agree that the PiFlow students reported in the submission use fixed segmentations. PDD supports inference flexibility by training with variable block lengths, allowing the same checkpoint to operate at multiple step budgets. To demonstrate this, we evaluated our two existing PDD checkpoints, trained with maximum block lengths of 256 and 128 intervals on a 256-interval grid, corresponding to minimum budgets of one and two steps, respectively. The table reports PDB-test lDDT for the confidence-ranked top prediction among five samples. Each row uses the same checkpoint and distilled trunk at every budget, without additional training.

PDD checkpoint1 step2 steps4 steps8 steps16 steps
Minimum 1 step0.7740.7900.8000.8010.804
Minimum 2 steps—0.7990.8090.8100.811

Both checkpoints improve in lDDT as the inference budget increases, with diminishing gains beyond four steps. Backbone lDDT and TM-score remain broadly stable, while interface and ligand metrics show smaller, non-monotonic changes. Thus, folding distillation does not require training a separate model for every supported budget: each PDD checkpoint supports multiple budgets at a fixed training cost, and the two checkpoints offer different minimum-step and accuracy tradeoffs.

We retain four-step π-Flow as the default model in our main results because it achieves the strongest overall performance at this budget in our evaluations. The additional PDD results demonstrate that a single folding checkpoint can also trade inference compute for accuracy across multiple step budgets, without retraining. At four steps, the PDD checkpoint supporting a minimum of two steps nearly matches π-Flow in PDB-test lDDT (0.809 versus 0.811), while π-Flow retains an advantage in interface and ligand metrics. These results provide users with a choice between our strongest four-step configuration and a checkpoint supporting flexible inference budgets.

We added results for all metrics and number of steps in {1, 2, 4, 8, 16} in Section H.3 “Inference flexibility with PDD”.

The hardware and training configuration are already specified in §F.2.5: the four-step π-Flow student uses 32 NVIDIA H100 GPUs, an effective batch size of 32, and 60k training steps. PDD follows the same hardware and training setup, with the PDD-specific settings described in §F.2.6. The missing details are the total training steps for the two PDD checkpoints: 110k and 320k for the checkpoints supporting a minimum of one and two sampling steps, respectively. Each checkpoint supports multiple inference budgets without further training, using the same shared distilled trunk. We added the missing training details (number of training steps for each PDD model) in section F.2.6 (text shown in red).

AuthorsOpenReview ↗

Author response II

Presenting core methodology for diffusion distillation for folding and design would clarify scope

We presented generation and folding separately because they require substantially different adaptations for distillation. Atomistic generation models such as La-Proteina and Proteina-Complexa rely on stochastic low-temperature sampling, recurrent self-conditioning, and distinct denoising schedules for backbone coordinates and residue latents. Boltz-2 supports deterministic ODE sampling, but its low-temperature scaling proved unstable for distillation in our experiments. Its EDM formulation also requires different parameterization and training choices, including clean-coordinate prediction and noise-dependent loss weighting. Accelerating the folding trunk additionally requires a separate representation-distillation stage. Separate subsections allow us to explain these task-specific choices clearly. That said, we agree that both share the same core approach to diffusion acceleration: construct a stable deterministic low-temperature teacher and distill it. We revised the opening of section 3, adding a comment on the common core approach for diffusion distillation (shown in red in the main text).

trunk novelty (removing triangular attention)

We agree that removing triangle attention is an established architectural simplification. Our choice follows both prior work in PairMixer and our profiling of Boltz-2. Our contribution here is adapting a pretrained Boltz-2 trunk through representation distillation, independently of diffusion-module distillation, and combining the two distilled components at inference without joint training. We added the PairMixer citation at the point of use in Section 3.2.

Quality-throughput tradeoffs (co-designability and diversity)... Foreground the absolute counts for Boltz2...

Co-designability versus diversity. Thank you for highlighting this distinction. Although all-atom co-designability decreases from 82.2% to 77.4%, our primary metric is what we call “diversity”: the number of unique co-designable structures produced within a fixed number of generated samples. We compute it by clustering all co-designable samples at a TM-score threshold of 0.5 and reporting the number of resulting clusters (rather than a diversity ratio, adopted by other works). This distinction matters because a model that always produces the same co-designable protein would achieve 100% co-designability but only one successful cluster, making it of little practical use for generating diverse candidates. Our diversity metric captures both successful generation and structural uniqueness: the 16-step TVM student produces 246 co-designable clusters versus 247 for the 256-step SDE baseline, despite its lower per-sample co-designability. Thus, “comparable” refers primarily to the number of distinct successful samples obtained with 16x fewer sampling steps. To make this clear, we modified the text in multiple places: Abstract and intro mentions performance comparable to the teacher when comparing unique successes, and the results Section 4.1 discusses diversity / unique successful samples as well as raw co-designability.

BoltzDesign success rates versus throughput. For design campaigns operating under a fixed compute budget, the relevant practical outcome is the number of successful designs obtained within that budget. Our acceleration therefore targets successful designs per GPU-hour, accounting for both design and evaluation costs. Table 8 reports the accompanying tradeoff: across the protein targets, the distilled model produces fewer successes at equal sample count-73 versus 99 under the strict criterion and 100 versus 111 under the relaxed criterion-but achieves approximately 2.2× and 2.6× more successful designs per GPU-hour, respectively. These results demonstrate improved computational efficiency rather than uniformly preserved per-design success rates, and success remains an in-silico proxy. We now mention these absolute numbers in the main text in section 4.3, alongside the results normalized by compute.

Non-monotonic performance. Thank you for highlighting this behavior. Non-monotonic unique-success counts are also observed without distillation. For the TSR teacher, Figure 3 and Table 5 show that structural diversity decreases from 271 co-designable clusters at 32 steps to 240 at 256 steps. The original SDE baseline improves through 256 steps, the largest budget reported in the submission. Additional SDE evaluations at 400 and 600 steps yield 233 and 229 co-designable clusters, respectively, compared with 247 at 256 steps. Thus, increasing the sampling budget under fixed sampler settings does not consistently increase the measured number of unique successful structures. This behavior is not specific to distillation or deterministic samplers (TSR), as it is also observed for the original SDE sampler.

AuthorsOpenReview ↗

Author response III

Baseline fairness. Figure 21 uses a smaller MSA for the undistilled four-step ODE baseline. The 0.811 versus 0.518 lDDT gap is too large for MSA depth alone to explain, but this is the comparison that isolates distillation and should be rerun matched.

Figure 21 reported an earlier four-step ODE evaluation using an MSA subsampled to 1,024 sequences. We agree that this baseline should be updated to use full MSAs, matching the other configurations. The full-MSA evaluation gives essentially unchanged lDDT: 0.519 versus 0.518 on the PDB test set, and 0.236 versus 0.238 on CASP15. Thus, matching the MSA setting preserves the observed lDDT advantage of the distilled model under the submission’s evaluation protocol, which includes stereochemistry checks. We replaced the four-step ODE results in Figure 21 with the full-MSA evaluation (and its caption), and updated the corresponding numbers in Section G.7 (changes in red).

Thank you for the new experiments and answers to my various questions! I still support acceptance and will maintain my score.

Reviewer_cgDSOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses inference cost in protein design campaigns, where generating and scoring 10^4–10^6 candidates per target is bottlenecked by two models that each require hundreds of network evaluations at inference: a flow-matching generator (La-Proteina / Proteina-Complexa) and a diffusion-based folding model (Boltz-2). The authors propose Proteina-Pronto, a distillation framework with three parts. First, they replace low-temperature stochastic sampling with a deterministic temperature-scaled-ODE (TSR) sampler that reproduces the quality benefit of low-temperature sampling without retraining. Second, they show that the asymmetric noise schedules of a multimodal product space (Cα coordinates and per-residue latents) can be coupled through a single global time, under which the teacher velocity reduces to a per-modality rescaling; this makes existing single-time distillation methods (TVM, π-Flow, MeanFlow, DMD) directly applicable to the product-space generator. Third, for folding they distill the Boltz-2 diffusion module to four steps with π-Flow and separately distill the Pairformer trunk into a triangle-attention-free student via representation distillation, yielding Boltz-2_distill. Reported results: TVM at 16 steps is described as comparable to the 256-step SDE on unconditional generation; π-Flow reaches 77.4% backbone designability in a single step; Boltz-2_distill retains lDDT 0.811 on the PDB test set versus 0.820 for the low-temperature teacher, at 3.9× end-to-end and 6.0× GPU-time speedup; and combining accelerated generation with accelerated scoring yields up to 6.1× more unique successful designs per GPU-hour on hard binder targets.

Critical strengths
  • The folding-acceleration half is independently top-tier work. The authors profile Boltz-2 inference before optimizing it (Fig. 20: diffusion 41%, Pairformer 32%, CPU featurization 16%, with the trunk share rising 12%→44% with complex size), then distill the two identified components separately and report a clean 2×2 ablation of their individual and joint contributions (Fig. 21, Tab. 7: trunk-only lDDT 0.818 at 1.1×, π-Flow-only 0.814 at 2.8×, combined 0.811 at 3.9×). Evaluation spans two sets (PDB test and CASP15) with bootstrap 95% CIs and stereochemistry checks enabled in the lDDT computation.
  • Speed claims are reported in a way that resists inflation. Tab. 7 separates GPU-only (6.0×) from end-to-end wall-clock (3.9×) using CUDA-event timing with featurization/IO broken out, and the paper explicitly discloses that after acceleration the dominant bottleneck shifts to CPU featurization (~42%). This is the right way to report acceleration and is rarer than it should be.
  • The schedule-coupling derivation (Eq. 7–9) is correct, compact, and load-bearing. Reducing the teacher velocity under asymmetric modality schedules to an α′_y(t) rescaling is a small observation, but it is what allows off-the-shelf single-time distillation to be transplanted onto a product-space generator without retraining the teacher — so every student here sits on top of public teacher weights. The student parameterization (Eq. 11) and the (k,σ) annealing that initializes the student at the teacher's first-order approximation are likewise well motivated.
Critical weaknesses

The headline "6×" is not currently assessable: the judge of a "unique successful design" is never defined, and may differ between the two compared arms. Location: abstract; §4.4 and Fig. 8 (right) — "replacing the 200-step Boltz-2 scorer with our 4-step Boltz-2_distill, Proteina-Pronto_TSR generation yields 6.1× as many unique successful designs per GPU-hour as the SDE sampler scored with regular Boltz-2"; cf. §E.1, where binder success is defined by AF2-Multimer re-folding (ipAE < 7, complex pLDDT > 90, scRMSD < 1.5), and §D ("scored using the folding model selected for each experiment"). Problem: §4.4 justifies Fig. 8 (left) on the grounds that "evaluation costs are identical across methods"; the fact that costs differ on the right implies the scorer itself was swapped. No section states who adjudicates success in Fig. 8 (right) (or Fig. 23 right, Fig. 25). If the two arms are counted by different judges, 6.1× is a ratio between "successes as counted by Boltz-2" and "successes as counted by Boltz-2_distill", which are not the same quantity: any miscalibration of the 4-step student's confidence head in the permissive direction — plausible, and not tested here — inflates the numerator independently of generator quality. Why it threatens a claim: this number appears in the abstract, introduction and conclusion. It is not a demonstrated flaw — the authors may have adjudicated everything with AF2-Multimer — but as written the reader cannot distinguish a real throughput gain from a shift in the decision rule. Minimal fix: state the success criterion and adjudicator for Fig. 8 (right); if the adjudicator was swapped, re-score all already-generated designs with one common independent judge (AF2-Multimer) and redraw the curve — no new generation needed — and report the Boltz-2 vs Boltz-2_distill success-call agreement (confusion matrix) on a common design set. Failing that, drop "6×" and decompose into "2× from generation (Fig. 8 left, fixed judge) and 3× from evaluation cost (Tab. 7, independently established)".

The 16-step "comparable" claim hides a measurable −4.8pp co-designability regression. Location: abstract and §4.1, against Tab. 5. All-atom co-designability is 77.4% for TVM at 16 steps versus 82.2% for the 256-step SDE. At N = 500 the binomial SE is ≈1.7pp, so the gap is ≈2.8 SE — not sampling noise. Diversity (246 vs 247) and designability (99.2 vs 98.8) genuinely are equivalent, so the precise statement is "equal diversity and designability, ~5pp loss in all-atom co-designability at 16× fewer steps". Minimal fix: replace "comparable" with the two numbers; no new experiments.

The 1-step headline reports the metric the model was not built for. Location: §1 and §4.1 — "Single-step generation achieves 77.4% backbone designability when sequences are redesigned with ProteinMPNN" — against the same row of Tab. 5, which gives all-atom co-designability 6.8% and Cα co-designability 14.4%. The La-Proteina line exists to do atomistic co-design; a designability measured after discarding the model's own sequence and side chains and re-designing with ProteinMPNN does not represent the 1-step student's intended capability, where it fails ~93% of the time. The numbers are honestly tabulated, but the introduction quotes only the favourable one, and introduction numbers are the ones readers retain. Minimal fix: report 6.8% alongside, or scope the sentence ("the 1-step student is useful as a backbone generator; atomistic co-design requires ≥4 steps"). No new experiments.

Questions for authors

Who adjudicates "unique successful design" in Fig. 8 (right), Fig. 23 (right) and Fig. 25, and under what criterion? Specifically: are the successes attributed to the Proteina-Pronto_TSR + Boltz-2_distill arm and to the SDE + Boltz-2 arm counted by the same judge? If they are not, can you re-score all already-generated designs with AF2-Multimer alone and redraw the GPU-hour curves? This requires no new generation and is, for me, the single blocking issue. What is the agreement between Boltz-2 and Boltz-2_distill as success callers on a common set of designs (confusion matrix, or at least per-target agreement rate)? If the student's confidence head is unbiased relative to the teacher, the concern in Q1 largely evaporates and the 6.1× stands as reported.

Were the PDB test structures — and their sequence/structure homologues — excluded from the distillation training mixture and from template fine-tuning? What exclusion criterion was used (date cutoff, cluster-level split)? Given the λ_gt = 0.1 ground-truth distogram term in Eq. (30), a one-sentence confirmation would settle this.

2 comments
AuthorsOpenReview ↗

Author response I

We thank the reviewer for thoroughly reviewing our work.

On the use of different folding model for evaluation in Fig. 8 (right), use of a common evaluator, ...

Thank you for highlighting this ambiguity. Figure 8 (right) indeed uses Boltz-2 and Boltz-2_distill, respectively, to evaluate candidates. Our goal is to make the distilled folding model a practical replacement for the original model, including direct candidate filtering using its own predictions and confidence scores. The original experiment therefore evaluated the throughput of pipelines using each folding model directly.

Nevertheless, we agree that a common final judge provides an important additional test of whether the gains persist beyond each model’s own assessment. We therefore re-evaluated all candidates passing their pipeline’s initial Boltz-2 or Boltz-2_distill filter with AlphaFold2, applying identical AF2 success criteria across pipelines and counting only designs passing both stages. Because AF2 is run only on initially retained candidates, its additional computational cost is limited. Including generation, initial screening, and AF2 verification, Proteina-Pronto_TSR with Boltz-2_distill still produces 6.1x as many unique AF2-validated successful designs per target on hard targets at a matched compute budget, relative to SDE generation with original Boltz-2; the corresponding gain on easy targets is 3.1x.

Although AF2 validation reduces the absolute success counts, these counts depend on the structure predictor and filtering criteria: identical numerical confidence thresholds need not have equivalent selectivity across models, as illustrated by the model-specific filter calibration in Latent-X technical report (Latent-X: An Atom-level Frontier Model for De Novo Protein Binder Design; Appendix B, Table S1). Differences in confidence calibration and predicted structures may therefore contribute to the reduction. Crucially, the throughput advantage persists under identical AF2 validation criteria, demonstrating the benefit of the accelerated model also when used in a two-stage pipeline with independent validation.

We added these results to the paper, in Section H1 “Re-evaluating Successes under Boltz-2 and Boltz-Pronto under AF2”. We also added a sentence clarifying this in Section 4.4 (added text shown in red).

The 16-step "comparable" claim hides a measurable −4.8pp co-designability regression.

Thank you for highlighting this distinction. Although all-atom co-designability decreases from 82.2% to 77.4%, our primary metric is what we call “diversity”: the number of unique co-designable structures produced within a fixed number of generated samples. We compute it by clustering all co-designable samples at a TM-score threshold of 0.5 and reporting the number of resulting clusters (rather than a diversity ratio, adopted by other works). This distinction matters because a model that always produces the same co-designable protein would achieve 100% co-designability but only one successful cluster, making it of little practical use for generating diverse candidates. Our diversity metric captures both successful generation and structural uniqueness: the 16-step TVM student produces 246 co-designable clusters versus 247 for the 256-step SDE baseline, despite its lower per-sample co-designability. Thus, “comparable” refers primarily to the number of distinct successful samples obtained with 16x fewer sampling steps. To make this clear, we modified the text in multiple places: Abstract and intro mentions performance comparable to the teacher when comparing unique successes, and the results Section 4.1 discusses diversity / unique successful samples as well as the decrease in raw co-designability.

The 1-step headline reports the metric the model was not built for... Designability vs. co-designability

Both the introduction and Section 4.1 explicitly describe the one-step result as backbone designability after ProteinMPNN sequence redesign; we do not claim strong one-step atomistic co-design. Nevertheless, we appreciate that reporting the corresponding all-atom co-designability alongside this result would make the distinction more prominent. We clarified this explicitly in the intro (line shown in red in the updated manuscript).

AuthorsOpenReview ↗

Author response II

Who adjudicates "unique successful design" in Fig. 8 (right) ... What is the agreement between Boltz-2 and Boltz-2_distill ... ?

Please see our response above regarding “who adjudicates successes” in Fig. 8 (right), and how results change once we re-evaluate initially deemed successes (with Boltz2 or Boltz2-distill as initial filter) using AF2. On the agreement between Boltz2 and our distilled variant. We agree that average lDDT alone does not establish ranking consistency. We therefore evaluated the same 1,900 binder candidates, 100 for each of 19 targets, with the original Boltz-2 and our four-step distilled model. The candidates were generated using Proteina-Pronto_TSR with 256 sampling steps. We computed Spearman’s rank correlation separately within each target and averaged these correlations across targets. To contextualize the results, we also repeated the original Boltz-2 evaluation with different random seeds, providing a reference for its stochastic repeatability.

MetricMean ρ: teacher–studentMean ρ: teacher–teacher, different seeds
Complex pLDDT0.8260.910
Mean interface PAE0.7360.794
Binder Cα scRMSD0.7620.797

These results provide direct evidence of substantial within-target ranking agreement beyond preservation of average structure-prediction accuracy. The teacher–student correlations approach the teacher-repeat reference particularly closely for interface PAE and binder scRMSD, while pLDDT shows a larger remaining gap. We added these results in section H2 “Rank-agreement between Boltz-2 and Boltz-2-distill”.

Test set used to evaluate folding models, excluded from training data, ...

We thank the reviewer for raising this point. Our student training uses the publicly available portion of Boltz-2’s training data, following the same public-data preparation and sampling pipeline. We do not have access to the additional non-public training datasets used by the teacher. Thus, our student is trained on a subset of the teacher’s training data and receives no additional experimental structures outside that training pool.

Our original PDB evaluation used the publicly released Boltz-1 evaluation set, evaluating 540 targets. Upon auditing this set against Boltz-2’s training data, we found that 177 targets predate Boltz-2’s training cutoff, of which 170 are exact PDB-ID matches to the training data. Consequently, this benchmark is not fully held out for Boltz-2 or our distilled model, although both were evaluated on the same targets and our student uses only the public portion of the teacher’s training data.

To assess whether our conclusions persist on post-cutoff structures, we excluded all 177 pre-cutoff targets and evaluated both models on the 363 remaining targets, released after Boltz-2’s Jun 1, 2023 training cutoff. This is a date-based evaluation split; we do not claim that it excludes all sequence or structure homologues of training examples. The results are reported below. Entries show means, with 95% confidence intervals in brackets.

MetricBoltz-2, 200 steps, step scale 1.5Boltz-Pronto, 4 steps
lDDT0.789 [0.773, 0.805]0.783 [0.765, 0.800]
Backbone lDDT0.857 [0.843, 0.872]0.860 [0.845, 0.875]
TM-score0.804 [0.780, 0.827]0.810 [0.786, 0.834]
DockQ > 0.230.592 [0.533, 0.650]0.607 [0.549, 0.664]
DockQ > 0.490.475 [0.417, 0.534]0.477 [0.419, 0.533]
lDDT-PLI0.512 [0.456, 0.566]0.510 [0.456, 0.563]
L-RMSD < 2 Å0.483 [0.419, 0.546]0.496 [0.434, 0.558]
L-RMSD < 5 Å0.554 [0.489, 0.618]0.599 [0.535, 0.661]

On this post-cutoff subset, Boltz-Pronto retains approximately 99.2% of the teacher’s lDDT, with comparable mean performance across the remaining metrics. Our central conclusion therefore remains unchanged: the four-step distilled model closely preserves the original teacher’s prediction accuracy, including when evaluation is restricted to post-cutoff structures.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: Additional results for folding model distillation

We added a comparison with DeCAF, a concurrent diffusion-distillation method based on Boltz-1, in Section 4.3 and Figure 7 of the main paper, with extended evaluations on the PDB test set, CASP15, and PoseBusters in Section H.7 (Figures 29 and 30). On the PDB test set, our four-step Boltz-Pronto achieves 0.811 all-atom lDDT, compared with 0.429 and 0.582 for five- and ten-step DeCAF, respectively. Even our one-step model achieves 0.774 lDDT. We also include DeCAF’s original Boltz-1 teacher as a reference.

We additionally benchmark runtime and improve the DeCAF implementation. Our four-step model takes 8.0 seconds per target, compared with 55.3 seconds for released DeCAF at five and ten steps, a 6.9× speedup. By integrating optimized kernels and mixed precision, we accelerate DeCAF by approximately 3x without observed accuracy degradation. Even against this improved baseline, our model remains 2.1-2.3x faster while achieving substantially higher folding accuracy. These results are discussed in Section 4.3, with implementation details and a timing breakdown breakdown per component in Section H.7, Table 10.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.