acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Consistent Flow Distillation for Uncertainty-aware Protein Structure Generation

Abstract

Diffusion and flow-matching models have achieved remarkable performance in protein structure generation and ensemble modeling, yet their practical use is often computationally constrained at scale by expensive multi-step sampling and costly iterative refinement. Recent work has explored applying distillation to protein generative models. However, directly transplanting the methods in image generation to 3D proteins overlooks fundamental geometric and physical constraints, most notably SE(3) gauge freedom and frame-dependent supervision, leading to single-step collapse with poor accuracy and structural quality. To address this challenge, we develop a geometry-consistent distillation(GCD) framework that makes one-step protein generation stable and well-posed. Our formulation captures local backbone correlations by explicitly coupling neighboring residues and further integrates an SE(3)-invariant corruption interface and a residue-local, frame-aligned distillation objective. For flow-matching teachers, we further derive the corresponding harmonic-prior score–velocity identity, connecting the teacher’s vector-field parameterization to score-based distillation. Empirically, naive Euclidean distillation collapses in the one-step setting, whereas GCD yields stable and competitive one-step students and transfers across multiple distillation paradigms. These results show that geometry-consistent supervision enables aggressive acceleration of protein generative models while retaining strong structural fidelity and substantial ensemble diversity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein structure generation, de novo protein design, Flow Matching distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_d4pfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Geometry-Consistent Distillation (GCD), which adapts SiD to accelerate sequence-conditioned protein conformational ensemble generation for AlphaFlow and ESMFlow. The authors motivate the paper by the fact that typical Euclidean SiD as used in images, for example, results in structural collapse. Therefore, the paper proposes a method similar to SiD in which harmonic prior and Kabsch alignment is used for noising. They also derive the corresponding harmonic score-velocity identity and compute both the auxiliary projector loss and the generator score-distillation loss in residue-local coordinate frames. The paper evaluates 1 and few-step distilled students on the 100-target PDB ensemble benchmark and the 82-target ATLAS MD benchmark.

Critical strengths

The general issue identified with a simple application of SiD to a model like AlphaFlow is correct and the general problem of accelerating these models is definitely interesting and relevant.

Critical weaknesses

1.Paper's Motivation:

  • One of the main issues of the paper in my opinion is that it seems to be solving an AlphaFlow-specific issue. AlphaFlow is especially slow at inference as it requires injecting the noisy structure xtx_t into AlphaFold2's template stack before the Evoformer trunk, which results in many Evoformer passes to generate a small ensemble. However, more recent architectures such as AlphaFold 3 [1], Boltz-1 [2], which run the expensive trunk once and roll out only a lightweight diffusion/structure head. Even the more recent AlphaFlow alternative, AlphaFlow-Lit [3] achieves a speedup over AlphaFlow without distillation by simply caching the Evoformer trunk.

  • Additionally, method loses relevance when considering all-atom and sequence–structure co-generation models. While the introduction section motivates the method via de novo protein design and inverse folding, all experiments are restricted to sequence-conditioned generation. The specific distillation approach in GCD also does not transfer to modern all-atom diffusion models (such as AlphaFold 3), or joint sequence–structure co-generation models (such as DISCO [4]). Have the authors considered how a 3-network bilevel distillation framework remain relevant for more recent approaches that rely on cached-trunk architectures or all-atom diffusion?
  1. Issues with the methodology and ablations:
  • Training GCD-SiD requires maintaining 3 full AlphaFold/ESMFold models simultaneously in GPU memory and backpropagating through the teacher and projector at every generator update. This cost is significant and prohibitive in many cases. More importantly, the paper fails to include any training compute metrics in the results.
  • Additionally, I found the ablations to be confounding multiple factors. Moving from SiD (0.49% Precision reported in Table 1a) to GCD-SiD (81.10% Precision) is essentially changing three things at one: (1) using the harmonic prior (2) applying Kabsch RMSDAlign and (3) computing losses in residue-local frames. The AlphaFlow teacher was originally trained with the harmonic prior and RMSDAlign so feeding unaligned isotropic Gaussian noise as is done in the SiD baselines is simply OOD for the teacher. There is no ablation isolating the local-frame score loss from simply matching the teacher's native corruption process. It would be helpful to include the performance of Euclidean SiD when using the teacher's native harmonic prior and RMSDAlign ) without the local-frame objective.
  1. Baselines are incomplete and experimental results are overstated:
  • AlphaFlow's native 1-step distilled baseline (AlphaFlow-Distilled) and few-step teacher truncation are not included in the Tables 1, 2, and 3.
  • There also seems to be a mode collapse in GCD-DMD and GCD-SC which contradicts transferability claims made in the paper. In the abstract and in Section 4, the paper claim that GCD transfers effectively across distillation paradigms (GCD-DMD and GCD-SC): “The table also includes GCD-DMD and GCD-SC as additional controls, demonstrating that the geometry-consistent formulation can be incorporated into multiple distillation paradigms.” However, Table 1a shows that both GCD-DMD and GCD-SC seem to experience significant mode collapse as the diversity drops to 2.35% for GCD-DMD and 2.59% for GCD-SC from the 18.47% Alphaflow teacher model.
  • Finally, the parametrization, training and sampling of the few step GCD models are not explained. Tables 1, and 2 show results for 2-step and 4-step distillation, however, the paper never specifies how these few-step models are parameterized or trained. In Section 3 and Algorithm 1 the student generator is written as a 1-step map Sg=Gθ(A,M)S_g = G_\theta(A, M) without noisy structure xtx_t or timestep tt arguments. Therefore, it isn’t clear to me how this model can be used for few-step distillation as well.

[1] Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w

[2] Boltz-1 Democratizing Biomolecular Interaction Modeling Jeremy Wohlwend, Gabriele Corso, Saro Passaro, Mateo Reveiz, Ken Leidal, Wojtek Swiderski, Tally Portnoi, Itamar Chinn, Jacob Silterra, Tommi Jaakkola, Regina Barzilay bioRxiv 2024.11.19.624167; doi: https://doi.org/10.1101/2024.11.19.624167

[3] Li, S., Li, M., Wang, Y., He, X., Zheng, N., Zhang, J., & Heng, P.-A. (2024). Improving AlphaFlow for Efficient Protein Ensembles Generation. https://arxiv.org/abs/2407.12053

[4] Rector-Brooks, J., Lambert, T., Skreta, M., Roth, D., Long, Y., Li, Z.-Q., Zhang, X., Cretu, M., Li, F.-Z., Ganapathy, T., Jin, E., Bose, A. J., Yang, J., Neklyudov, K., Bengio, Y., Tong, A., Arnold, F. H., & Liu, C.-H. (2026). General Multimodal Protein Design Enables DNA-Encoding of Chemistry. https://arxiv.org/abs/2604.05181

Questions for authors

See questions above.

3 comments

We thank the reviewer for emphasizing the computational challenges of iterative protein generation and for asking how the proposed formulation relates to more recent architectures. We clarify the ensemble-generation objective, the actual training costs, and the controlled baseline comparisons.

1. Motivation and applicability beyond AlphaFlow

Our work addresses a different question from accelerating single-structure prediction: preserving a conditional protein conformational ensemble distribution during one-step distillation. The novelty of our work is one-step ensemble-preserving distillation, rather than single-step generation alone.

The computational bottleneck addressed by GCD is not limited to repeated Evoformer computation. Diffusion and flow-matching models typically generate samples by progressively denoising an initial noise sample toward a plausible structure, a process that often requires tens to hundreds of sampling steps. AAlphaFlow-Lit accelerates this process by caching conditioning features and avoiding redundant computation across sampling steps. In contrast, GCD reduces the number of sampling steps required for generation through distillation. These approaches target complementary aspects of inference efficiency: AlphaFlow-Lit reduces the computational cost per step, while GCD reduces the number of steps. Consequently, the two approaches are compatible in principle and could be combined to achieve further acceleration.

Beyond inference efficiency, the reviewer also raises the question of whether GCD can be extended to newer all-atom and joint sequence–structure generative models. Importantly, AlphaFlow/ESMFlow and our distilled students already generate all-heavy-atom protein structures, including both backbone and side-chain atoms, through their AlphaFold/ESMFold structure modules. The pseudo-β\beta coordinates in our formulation define the stochastic flow state rather than limit the resolution of the generated structures. For AF3-style models, which operate on atom-coordinate diffusion states, extending GCD would require adapting the distillation objective to the corresponding denoising/score parameterization and atom-level geometric representation. Joint sequence–structure generators additionally require handling discrete sequence variables. Thus, while the specific formulation depends on the teacher's generative representation, the underlying principle of geometry-consistent distributional distillation is not inherently restricted to AlphaFlow-like architectures.

2. Training cost and the projector–generator scheme

Compared with standard teacher–student distillation, our bilevel setup introduces an additional projector network, which increases training memory requirements. However, AlphaFlow's native supervised distillation requires complete teacher rollouts to construct training targets, whereas GCD-SiD queries the teacher at student-induced corrupted states without requiring full teacher trajectories.

We additionally measured peak GPU memory and training throughput under the same single-H200, batch-size-1 configuration:

MethodPeak allocated memory (GiB)Estimated time (24k samples)Estimated time (160k samples)
AlphaFlow supervised distillation (online 10-step teacher)9.6152.3 h348.7 h
GCD-SiD10.77 (projector) / 32.52 (generator)34.2 h228.1 h

GCD-SiD requires higher peak memory during generator updates because gradients propagate through the frozen teacher and projector. Nevertheless, its measured training throughput corresponds to approximately 35% less training time than online supervised distillation under the matched configuration. The 24k-sample setting corresponds to our actual GCD-SiD training budget, while 160k samples is the training budget used by the officially released AlphaFlow-Distilled model.

3. Teacher-matched corruption and the requested Cartesian control

The collapsed Euclidean SiD baseline and GCD use the same harmonic prior, RMSD-aligned interpolation, frozen AlphaFlow teacher, and corrupted teacher inputs xtx_t. Here, 'Euclidean' refers to the distillation discrepancy, not the noising process. In particular, the Cartesian baseline does not use isotropic Gaussian noise or omit RMSD alignment. The two formulations differ in how their teacher–projector–generator residuals are constructed: the baseline compares predictions directly in unframed global 3D coordinates, whereas GCD uses residue-local frame alignment together with its geometric loss construction.

Thus, the teacher-native harmonic-prior and RMSD-aligned Cartesian control requested in the review is already the reported Euclidean SiD baseline, which gives 0.49/0.44/0.00 Precision/Recall/Diversity. These comparisons support the importance of geometry-consistent distillation discrepancies rather than an out-of-distribution corruption mismatch.

4. Native distillation and reduced-step teacher baselines

We additionally evaluated the original AlphaFlow teacher at reduced sampling steps and the publicly released native supervised AlphaFlow-Distilled model. All numbers below are median PDB-ensemble Precision/Recall/Diversity (%) under the same evaluation protocol.

MethodStepsPrecisionRecallDiversity
AlphaFlow teacher184.0279.642.74
AlphaFlow teacher283.3281.279.11
AlphaFlow teacher482.2480.6314.50
AlphaFlow teacher1080.1979.7818.47
AlphaFlow-Distilled182.7580.119.60
GCD-SiD181.1080.4718.31
GCD-SiD282.1681.6617.39
GCD-SiD479.5380.0021.92

These results distinguish one-step ensemble preservation from one-step structure prediction alone. Direct one-step teacher sampling gives high precision (84.02) but only 2.74 diversity. AlphaFlow-Distilled, which uses supervised targets obtained from complete teacher rollouts, achieves 82.75/80.11/9.60. In comparison, one-step GCD-SiD increases diversity by 8.71 percentage points and recall by 0.36 points, with a 1.65-point precision trade-off. Its diversity remains close to the ten-step teacher. At two steps, GCD-SiD also produces a substantially broader ensemble than the truncated two-step teacher (17.39 versus 9.11 diversity), while retaining comparable precision and recall.

5. Generality across distillation objectives

The GCD-DMD and GCD-SC experiments are not ablations of SiD; they use different training objectives and update rules. SiD uses an auxiliary projector to approximate the student-side score and constructs a score-identity-based distribution-matching objective; DMD adopts KL-based distribution matching; and consistency distillation enforces consistency across a teacher trajectory. The resulting fidelity–diversity trade-offs can differ. Indeed, GCD-DMD and GCD-SC retain high Precision/Recall (83.21/79.59 and 83.49/79.57) but produce narrower ensembles (2.35 and 2.59 diversity). This differs from the Cartesian SiD baseline, whose Precision/Recall fall to 0.49/0.44, indicating severe structural degeneration. Thus, the geometric construction supports accurate structural generation across these objectives, with GCD-SiD providing the strongest one-step ensemble preservation among the evaluated variants.

6. Parameterization and few-step training

Our few-step sampling follows the original AlphaFlow sampling procedure, iteratively applying the same time-conditioned student network GθG_\theta. Starting from harmonic noise x(0)∼q0x^{(0)}\sim q_0, each step predicts a clean structure and updates the intermediate state:

\qquad \hat x_k=[\widehat S_k]_\beta,$$ $$x^{(k)}= \frac{\tau_k}{\tau_{k-1}} \operatorname{RMSDAlign}(x^{(k-1)},\hat x_k) +\left(1-\frac{\tau_k}{\tau_{k-1}}\right)\hat x_k,$$ where $d(\cdot)$ denotes pairwise structural distances and ${\tau_k}_{k=0}^{K}$ is the sampling schedule. For training, we uniformly sample $J\in{1,\ldots,K}$ and apply the GCD objective to the prediction at step $J$, with the preceding $J-1$ steps executed without gradient tracking. The student prediction is detached during projector training, while generator training backpropagates through the selected step with fixed teacher and projector parameters. At inference, all $K$ steps are performed to obtain the final structure $\widehat S_K$.
Reviewer_3dMSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper applied Score identity Distillation to AlphaFlow and ESMFlow for few-step protein conformational ensemble generation, and showed enhanced performance compared with the original multi-step base model.

Critical strengths

The paper showed interesting and strong empirical results. Comparing with the multi-step AlphaFlow, the one-step GCD-SiD achieves a 5x speedup with marginally increased precision, recall, and diversity.

Critical weaknesses
  1. There is limited meaningful methodological contribution from this paper. This paper applies SiD to protein conformational ensemble generation and uses standard techniques in the field such as harmonic priors, RMSD alignment, and local frame losses. The paper does provide several theoretical results to support this application, but there are inconsistencies between the derivations and the actual implementation. Specifically, the score velocity identity is derived assuming a Gaussian corruption kernel, while the implementation applies RMSD/optimal alignment to the sampled noise, which could make the resulting distribution no longer Gaussian. In addition, the score is derived in Cartesian coordinates, while the generator is trained using a local frame loss.

  2. Insufficient comparison against baselines. The paper does not compare against direct one-step AlphaFlow sampling or AlphaFlow's existing supervised single pass distillation provided in the original paper. Currently, it is unclear whether the proposed SiD based method offers any improvement over using other distillation approaches.

Questions for authors

In Algorithm 1, should there also be an input for stochasticity for the generator G(A, M)?

2 comments

We appreciate the reviewer's careful assessment and recognition of the strong empirical results. We clarify why the contribution goes beyond a direct combination of AlphaFlow and SiD, how the reference-path derivation relates to the implemented loss, and what the native one-step baselines show.

1. Why GCD is more than an implementation of SiD for AlphaFlow

The contribution of GCD is not the individual use of harmonic priors, RMSD alignment, or residue-local frames, but the incorporation of geometric consistency into distribution-matching distillation for protein structures. When Cartesian SiD is directly applied to protein structure generation, the teacher–projector–generator residuals depend on global coordinate frames that are not determined by AlphaFlow's rigid-motion-invariant structural input. This can lead to inconsistent optimization signals even when the teacher-native corruption process is retained. GCD addresses this ambiguity by constructing the distillation residuals in residue-local frames, rather than simply adding FAPE-based supervision to an existing distillation objective.

Gaussian-reference derivation and local-frame adaptation. We acknowledge that the Gaussian-reference score identity does not directly extend to the endpoint-dependent RMSD alignment used in the implemented corruption. The reference identity motivates the SiD residual construction, but does not establish an exact score identity for the aligned distribution.

The local-frame formulation addresses a separate issue: the dependence of Cartesian distillation residuals on arbitrary global rigid transformations. We show that the implemented generator loss eliminates this dependence.

For a predicted structure SS, let Qi(S)Q_i(S) and ci(S)c_i(S) denote the rotation and origin of residue frame ii, and let xj(S)x_j(S) be the pseudo-β\beta coordinate of residue jj. Its representation in frame ii is

Φij(S)=Qi(S)⊤[xj(S)−ci(S)].\Phi_{ij}(S)=Q_i(S)^\top[x_j(S)-c_i(S)].

For any rigid transformation T=(Q,t)∈SE(3)T=(Q,t)\in SE(3), the residue coordinates and local frames transform as xj(TS)=Qxj(S)+tx_j(TS)=Qx_j(S)+t, ci(TS)=Qci(S)+tc_i(TS)=Qc_i(S)+t, and Qi(TS)=QQi(S)Q_i(TS)=QQ_i(S). Therefore,

\Phi_{ij}(TS) &=(QQ_i(S))^\top \left[Qx_j(S)+t-Qc_i(S)-t\right]\\ &=\Phi_{ij}(S). \end{aligned}$$ For the reported $\alpha=1$ setting, let $a=\operatorname{vec}\Phi(S_\phi)$, $b=\operatorname{vec}\Phi(S_\psi)$, and $U=\operatorname{vec}\Phi(S_g)$. The implemented generator loss can be written as $$\ell_\theta=(a-b)^\top W(U,a)(b-U), \qquad W(U,a)=A_{\mathrm{loc}}^\top D(U,a)A_{\mathrm{loc}},$$ where $A_{\mathrm{loc}}$ applies the local structural operator and $D(U,a)$ contains the positive normalization weights computed from local-coordinate differences. Because $a$, $b$, and $U$ are individually invariant to global rigid transformations, both residuals and their weighting matrix remain unchanged. Consequently, for arbitrary $T_\phi,T_\psi,T_g\in SE(3)$, $$\ell(T_\phi S_\phi,T_\psi S_\psi,T_gS_g)=\ell(S_\phi,S_\psi,S_g).$$ Thus, the complete generator loss is independent of the global rigid pose of each predicted structure. This removes the coordinate-frame dependence of the Cartesian distillation residuals while retaining the teacher–projector–generator objective used in GCD.

2. Native one-step baselines and ensemble preservation

We additionally evaluated the original AlphaFlow teacher at reduced sampling steps and the publicly released native supervised AlphaFlow-Distilled model. All numbers below are median PDB-ensemble Precision/Recall/Diversity (%) under the same evaluation protocol.

MethodStepsPrecisionRecallDiversity
AlphaFlow teacher184.0279.642.74
AlphaFlow teacher283.3281.279.11
AlphaFlow teacher482.2480.6314.50
AlphaFlow teacher1080.1979.7818.47
AlphaFlow-Distilled182.7580.119.60
GCD-SiD181.1080.4718.31
GCD-SiD282.1681.6617.39
GCD-SiD479.5380.0021.92

Direct one-step teacher sampling gives high precision but very limited diversity. Native supervised distillation achieves 82.75/80.11/9.60, whereas GCD-SiD obtains 81.10/80.47/18.31. The latter therefore improves one-step diversity by 8.71 percentage points, with 0.36 points higher recall and a 1.65-point precision trade-off. Thus, GCD not only prevents collapse but produces a one-step student that preserves target-specific conformational variability.

3. Stochastic input and few-step student notation

The notation Gθ(A,M)G_\theta(A,M) in Algorithm 1 omits the stochastic input for simplicity. In implementation, the student is initialized with harmonic noise z∼q0z\sim q_0, and the corresponding pairwise distance matrix d(z)d(z) is provided as structural input. Thus, one-step generation is more explicitly written as

Sg=Gθ(A,M;d(z),tinit),z∼q0.S_g=G_\theta(A,M;d(z),t_{\mathrm{init}}),\qquad z\sim q_0.

Different noise samples produce different conformations for the same protein sequence. We will clarify this notation in Algorithm 1.

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

the paper proposed GCD, a method that distill AlphaFlow and ESMFlow from their 10-tep into one or few step students. It adopts SiD with a few adjustments such as frame aligned loss, a harmonic chain prior corruption path etc.

The authors claim that this approach does not collapse wheras Euclidean SiD does and shows 5-10x speedups.

Critical strengths

The application targeted here (emsemble generation speedups) is useful and the approach seems sound. The authors evaluate two teachers in two data regimes with several step counts, and they are also honest about the limitations compared to MD in general.

Critical weaknesses

[W1] `There are some critical gaps in the claims, such as the collapse of the baseline: the authors only compare their full final method to simple naive SiD. The authors should do careful component ablation (alignment only, harmonic prior only, frame loss only, preconditioner only, combinations of these) to see where the gain actually comes from. For example, the teacher is trained with a harmonic prior and aligned noise from my understanding, so it makes sense that naive isotropic/unaligned SiD would not work.

[W2] The authrs claim "competitive one step", "substantial ensemble diversity" and "generalisation across objectives". However these things do not seem to be supported by the data: there is substantial degradation compared to teacher in W2/%PC-sim and other metrics. In addtiion while the models reach similar precision/recall to the teacher, the diversity is only 2.35 and 2.59 vs the teacher’s 18.47. This does not seem like ensemble diversity/generalisation across objectives. The authors also claim "uncertainty-aware", but show no evidence of that.

[W3] There are quite a few baselines and metrics that seem critical to evaluate, including the original AlphaFlow distilled models, SPID and other methods, and then local geometry metrics such as Molprobity scores, rotamer/ramachandran outliers etc.

Questions for authors

What exactly is S̃ in Eq. (9) and (10), and where is the frame-aligned loss derived?

Why do GCD-DMD and GCD-SC lose diversity, and does this tell us something about SiD specifically?

How does the AlphaFlow-distilled baseline, and a 2–3 step teacher at matched wall-clock, compare?

What are the training data, α, steps and compute?

3 comments

We thank the reviewer for their constructive comments on the corruption process, the geometric origin of the observed training instability, and the evaluation of ensemble diversity. We address these concerns and provide additional comparisons with native AlphaFlow distillation and reduced-step teacher sampling.

1. Teacher-matched corruption and the cause of collapse

The collapsed Euclidean SiD baseline and GCD use the same harmonic prior, RMSD-aligned interpolation, frozen AlphaFlow teacher, and corrupted teacher inputs xtx_t. Here, 'Euclidean' refers to the distillation discrepancy, not the noising process. In particular, the Cartesian baseline does not use isotropic Gaussian noise or omit RMSD alignment. The two formulations differ in how their teacher–projector–generator residuals are constructed: the baseline compares predictions directly in unframed global 3D coordinates, whereas GCD uses residue-local frame alignment together with its geometric loss construction.

To distinguish geometric supervision from a generic regression failure or score mismatch, we additionally tested a Laplacian-preconditioned harmonic-reference score without residue-local frame alignment. It does not rescue training: the loss remains highly oscillatory and fails to converge reliably. In contrast, the residue-local formulation trains stably. Since AlphaFlow's structural input is invariant to global rigid transformations, Cartesian residuals can depend on arbitrary coordinate frames, leading to inconsistent optimization signals. GCD eliminates this dependence by constructing the distillation residuals in residue-local frames. These results support the importance of geometric consistency in stabilizing distillation, rather than attributing the observed collapse to a mismatch in the corruption process.

2. One-step ensemble preservation and native baselines

We first clarify that 2.35 and 2.59 are the diversity values of GCD-DMD and GCD-SC, rather than our primary GCD-SiD result. The latter achieves 81.10/80.47/18.31 Precision/Recall/Diversity in one step, compared with 80.19/79.78/18.47 for the ten-step AlphaFlow teacher.

We additionally evaluated the original AlphaFlow teacher at reduced sampling steps and the publicly released native supervised AlphaFlow-Distilled model. All numbers below are median PDB-ensemble Precision/Recall/Diversity (%) under the same evaluation protocol.

MethodStepsPrecisionRecallDiversity
AlphaFlow teacher184.0279.642.74
AlphaFlow teacher283.3281.279.11
AlphaFlow teacher482.2480.6314.50
AlphaFlow teacher1080.1979.7818.47
AlphaFlow-Distilled182.7580.119.60
GCD-SiD181.1080.4718.31
GCD-SiD282.1681.6617.39
GCD-SiD479.5380.0021.92

These results distinguish one-step ensemble preservation from one-step structure prediction alone. Direct one-step teacher sampling gives high precision 84.02 but only 2.74 diversity. AlphaFlow-Distilled, which uses supervised targets obtained from complete teacher rollouts, achieves 82.75/80.11/9.60. In comparison, one-step GCD-SiD increases diversity by 8.71 percentage points and recall by 0.36 points, with a 1.65-point precision trade-off. Its diversity remains close to the ten-step teacher. At two steps, GCD-SiD also produces a substantially broader ensemble than the truncated two-step teacher, while retaining comparable precision and recall.

Local backbone geometry. We also report quantitative Ramachandran statistics to complement the qualitative plots in the manuscript.

MethodStepsFavored (%) ↑Outliers (%) ↓Evaluable coverage (%)
AlphaFlow teacher187.637.5596.16
AlphaFlow teacher292.173.7199.82
AlphaFlow teacher493.472.2799.88
AlphaFlow teacher1090.353.3999.75
AlphaFlow-Distilled188.446.6097.73
GCD-SiD184.769.5797.16
GCD-SiD291.633.8299.80
GCD-SiD491.483.2899.82

One-step GCD-SiD has a lower favored fraction and more Ramachandran outliers than native AlphaFlow-Distilled, identifying a local-geometry trade-off alongside its broader conformational ensemble. The few-step students substantially improve these statistics: the two- and four-step models reach 91.63% and 91.48% favored fractions, with 3.82% and 3.28% outliers, respectively, compared with 90.35% favored and 3.39% outliers for the ten-step teacher. Together with the PDB ensemble results, these measurements show that the few-step students retain broad conformational ensembles while recovering local backbone geometry close to the teacher.

Both works build on SiD, but their successful operating regimes differ. SPID studies designable protein backbone generation, including unconditional, fold-class-conditioned, and motif-scaffolding settings. Its Proteína experiments report very low one-step designability, with strong results obtained through few-step sampling and inference-time noise modulation. Our work instead studies sequence-conditioned conformational ensemble generation, in which the goal is to capture multiple plausible structures of the same protein sequence. In this setting, GCD-SiD preserves substantial target-specific ensemble diversity with a single network call.

We reported the measured end-to-end inference time in the paper. Over the complete tested generation pipeline, AlphaFlow median per-target time decreases from 19:29 to 4:10 on one A100 (4.67×4.67\times), while ESMFlow decreases from 8:34 to 0:54 on one H200 (9.52×9.52\times).

3. Generality across distillation objectives and the meaning of diversity

Our methodological contribution is a distillation-specific geometric formulation that inserts geometry inside the teacher–projector–generator distribution-matching update on samples from the current student distribution. The GCD-DMD and GCD-SC experiments are not ablations of SiD; they use different training objectives and update rules. SiD is derived from Fisher-divergence minimization, using an auxiliary projector to estimate the student-side score. During generator optimization, gradients propagate through the inputs of both the frozen teacher and projector. In contrast, DMD minimizes a reverse-KL-based distribution-matching objective without backpropagating through the teacher networks. Consistency distillation instead enforces consistency between predictions along teacher-guided sampling trajectories. These differences in training objectives and gradient computation can lead to distinct distillation performance. Indeed, GCD-DMD and GCD-SC retain high Precision/Recall (83.21/79.59 and 83.49/79.57) but produce narrower ensembles. This differs from the Cartesian SiD baseline, whose Precision/Recall fall to 0.49/0.44, indicating severe structural degeneration. Thus, the geometric construction supports accurate structural generation across these objectives, with GCD-SiD providing the strongest one-step ensemble preservation among the evaluated variants.

Here, 'uncertainty-aware' means using a generated ensemble to represent the range of plausible conformations for the same protein sequence. The uncertainty is reflected in conformational heterogeneity—whether the samples cover multiple experimentally observed or dynamically accessible states—rather than parameter uncertainty or a single confidence score. We quantify it through PDB conformer coverage and diversity, together with MD flexibility, distributional, contact, and exposure observables.

Importantly, our one-step GCD-SiD achieves Precision/Recall/Diversity of 81.10/80.47/18.31, compared with 80.19/79.78/18.47 for the ten-step teacher. This demonstrates that one-step distillation can retain substantial target-specific conformational variability rather than acting only as a high-precision mode predictor.

4. Meaning of S~\widetilde S and frame-aligned distillation losses

In Eqs. (9)-(10), S~\widetilde S denotes the residue-local representation of a predicted protein structure, rather than a separate structural prediction. Following the FAPE formulation, each residue ii defines a local frame with rotation Ri(S)R_i(S) and origin ci(S)c_i(S). A structural point pj(S)p_j(S) is expressed in this frame as

Φij(S)=Ri(S)⊤(pj(S)−ci(S)).\Phi_{ij}(S)=R_i(S)^\top\bigl(p_j(S)-c_i(S)\bigr).

Frame-aligned discrepancies are constructed by comparing corresponding (i,j)(i,j) residue from two predicted structures, with each structure expressed in its own residue-local frames. This removes dependence on their global rotations and translations.

In Eq. (9), S~ψ\widetilde S_\psi and S~g\widetilde S_g denote the local structural representations of the projector and generator predictions. The projector is trained to match the generator's detached structural prediction using FAPE-style supervision. In Eq. (10), x~ϕ\widetilde x_\phi, x~ψ\widetilde x_\psi, and x~g\widetilde x_g denote the corresponding residue-local representations of the predicted pseudo-β\beta coordinates. Their differences form the teacher-projector and projector–generator residuals used in the SiD distillation objective.

The reported one-step AlphaFlow GCD-SiD uses α=1.0\alpha=1.0, 1,500 generator and 1,500 projector optimizer updates, an effective global batch of 8, and 24,000 example presentations across both phases on 2 H200 GPUs, for 34.88 GPU-hours. The student and projector are initialized from the pretrained teacher, while the teacher remains frozen.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.