acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Hybrid-Coordinate Flow Matching for Accurate and Physically Plausible Apo-to-Holo Protein Generation

Abstract

Predicting holo-like protein conformations from apo structures is essential in structure-based drug design. Most existing apo-to-holo models operate in Cartesian space and achieve strong global accuracy. However, high global accuracy does not ensure physically valid local backbone geometry. Internal-coordinate models offer local control and validity, but regressive generation can amplify angular errors—the lever-arm effect. We introduce CrypticFlow, a hierarchical flow-matching model that combines these two complementary representations. An ESM-3-conditioned E(3)-equivariant graph network first generates sparse Cα anchors in Cartesian space. A conditional midpoint Transformer then recursively fills the gaps in local internal coordinates. During recursive generation, the decoder constrains the distance between each midpoint and any sequence-adjacent parent anchor to the corresponding apo Cα–Cα distance. On the D3PM test split, CrypticFlow achieves state-of-the-art median and mean Cα RMSD. Across five independent training models, the median and mean direct RMSD are 1.333 ± 0.056 Å and 2.766 ± 0.041 Å, respectively. The quality of local geometry also improves substantially. Our model achieves 99.58 ± 0.02% adjacent-distance fidelity and 95.87 ± 0.25% pseudo-angle fidelity, measured within 0.2 Å and 20° of the holo reference. Then, we reconstruct the predicted Cα-only traces into all-atom structures using the cg2all coarse-to-all-atom backmapping method and evaluate their physical and geometric quality with MolProbity. Our model achieves a MolProbity score of 3.034 ± 0.054, outperforming the evaluated baselines. These results show that a hierarchical flow-matching model with hybrid coordinates can improve global conformational accuracy and structural fidelity together.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow-matching, AI, generative model, protein, hierachial, recursive

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_LaNzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CrypticFlow, which is a hierarchical flow matching model for apo-to-holo protein generation. This method first generates coarse Ca anchors in Cartesian space using EGNN, then iteratively fills in residues in local internal coordinates. Finally, it also applies some hard rules to ensure geometry quality. In the D3PM test set benchmark, CrypticFlow achieves both strong holo state prediction and great local geometry preservation compared with other baseline methods.

Critical strengths
  1. Modeling the apo-to-holo states using a hierarchical approach, first generating coarse coordinates in Cartesian space, then middle residues with local internal coordinates, is a good way to address poor local geometric fidelity in purely Cartesian models and global error accumulation in chain-wide internal-coordinate generation.
  2. Well-designed experiments and ablations to show the effectiveness of some components, including but not limited to the ESM-3 conditioning, local internal coordinates, and different anchor strides, which makes the results more convincing.
Critical weaknesses
  1. The method is developed in the absence of ligands, which means it only works for the protein monomer apo-to-holo conformation change. In real SBDD, people may care more about the conformational change when proteins bind to different ligands. This is not an issue for the method proposed, but it can make the problem statement and application weaker.
  2. The model is trained and evaluated on the D3PM dataset. It would be interesting to show how it performs on an external validation set and on classical apo-to-holo targets to demonstrate its generalization ability.
Questions for authors
  1. Have you thought about doing an exposure-bias training to solve the potential train-inference mismatch, where during training it always sees the ground-truth coarse Ca structures, but in inference it would rely on its own generated structures? An example work is https://arxiv.org/pdf/2602.04883, which also does hierarchical protein generation, and they show that exposure-bias training is helpful for solving this problem.
  2. The midpoint flow is stochastic, so the model could in principle generate multiple holo-like conformations for the same apo input. Have the authors evaluated whether repeated samples form a meaningful conformational ensemble e.g., in terms of structural diversity or coverage of alternative holo states?
3 comments
AuthorsOpenReview ↗

Response to Reviewer LaNz — Q1. Exposure-bias training

Yes, we tried this. As you point out, in standard training the midpoint flow sees only ground-truth parent contexts. We therefore applied the two methods of PAR that you suggested to CrypticFlow and retrained it with seed 1, keeping every other setting unchanged.

  • SS (scheduled sampling): During training, the parent context of the midpoint flow (the top-level anchors and the coarser-level midpoints) is replaced with the model's own predictions with a given probability. The probability increases from 0 to 0.5 over the first 25k steps and stays at 0.5 afterwards (162k steps in total).
  • NCL (noisy context learning): With probability 0.5, isotropic Gaussian noise is added to the ground-truth parent Cα coordinates. We matched the noise scale to how far the generated anchors actually deviate at inference: σ = 1.05 Å per coordinate, so that the RMSD of the noise equals the median error of the generated anchors on the validation set (1.81 Å).
  • SS+NCL: Both are applied; the noise is added only to contexts that remain ground truth.

Test results (150 pairs, seed 1):

ModelMedian RMSD (Å) ↓Mean RMSD (Å) ↓< 2 Å (%) ↑Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓
CrypticFlow1.3092.73758.099.5795.933.038
+ SS1.2572.75059.399.5495.183.121
+ NCL1.4842.81059.399.4094.483.131
+ SS+NCL1.3272.79158.799.3993.733.216
  • Accuracy did not improve. Median RMSD changed by −0.05 Å with SS and by +0.02 Å with SS+NCL, both within the seed-to-seed s.d. of CrypticFlow over five seeds (0.056 Å), and worsened by 0.17 Å with NCL. The fraction of pairs below 2 Å changed by +0.7 to +1.3 pp (1–2 pairs), comparable to the seed-to-seed s.d. (0.73 pp).
  • Local geometry and structure quality worsened with all three methods. Cα angle validity dropped by 0.75–2.20 pp (seed-to-seed s.d. 0.25 pp), and distance validity by 0.03–0.18 pp. MolProbity also rose from 3.038 for CrypticFlow to 3.121 (SS), 3.131 (NCL) and 3.216 (SS+NCL); lower is better.

This appears to be because, in CrypticFlow, the train–inference mismatch does not amplify errors. At inference, the midpoint flow receives as parents the top-level anchors generated by the model, which differ from the ground truth. If the mismatch were a problem, the error of the final structure, after the midpoints are filled in, would exceed the error of the anchors. However, on the 150 test pairs (five seeds), the RMSD of the final structure was almost identical to the RMSD measured on the anchor residues alone (median difference +0.013 ± 0.004 Å). The midpoint stage thus adds no error even when given incorrect anchors, and the final error is set by the error of the top-level anchors. Because SS and NCL change only the training of the midpoint stage and leave the training of the EGNN that produces the anchors unchanged, they cannot reduce this error. We will add these results to the appendix of the revision.

AuthorsOpenReview ↗

Response to Reviewer LaNz — Q2. Do repeated samples form an ensemble?

We evaluated this. Using each of the five trained models reported in the paper (seeds 1, 3, 42, 88, 99), we generated 20 samples for each of the 150 test pairs. Values are mean ± s.d. over the five seeds, and RMSD is the direct Cα RMSD without superposition, as in the paper.

  • Structural diversity. We define the diversity of a target as the mean pairwise Cα RMSD among its 20 samples. Over the 150 targets, the median is 0.020 ± 0.003 Å and the 90th percentile is 0.22 ± 0.05 Å. Compared with the median apo–holo RMSD of 4.46 Å on the test set, the samples are effectively the same structure.
  • Cause. The top-level EGNN anchor flow is a deterministic ODE that starts from the apo structure, so all 20 samples share identical top-level anchors (0.0 Å difference). Stochasticity enters only through the midpoint flow, which generates the midpoints with these fixed anchors as parents. That is, once the EGNN has determined the top-level anchors, the overall shape of the structure is constrained, and the stochasticity of the midpoint stage alone does not appear to produce different conformations.
  • Coverage. Here the oracle is the single sample, out of the 20, with the lowest Cα RMSD to the true holo structure (direct RMSD as defined in the paper: no superposition, over the commonly observed residues). Even so, it gives almost the same accuracy as a single sample.
  • Local geometry and structural quality are also unchanged. Adjacent Cα–Cα distance validity has the same five-seed mean for both choices (within a seed they differ by at most one bond out of 54,509), and Cα angle validity is 0.050 pp higher for the oracle. MolProbity is 0.007 lower for the oracle. Scoring the same structures again shifts it by as much, so this is not a difference we would read as an improvement in quality.
CrypticFlow Single sampleCrypticFlow Best of 20 (oracle)
Median RMSD (Å) ↓1.333 ± 0.0561.330 ± 0.060
Mean RMSD (Å) ↓2.766 ± 0.0412.756 ± 0.042
< 2 Å (%) ↑58.53 ± 0.7358.80 ± 0.56
Cα–Cα distance valid (%) ↑99.58 ± 0.0299.58 ± 0.02
Cα angle valid (%) ↑95.87 ± 0.2595.92 ± 0.25
MolProbity ↓3.031 ± 0.0523.024 ± 0.054

Therefore, CrypticFlow in its current form does not appear to generate structural diversity or alternative holo states. It is not an ensemble generator but a nearly deterministic predictor that maps one apo structure to one holo-like structure. We will state this explicitly in the limitations section of the revision.

I appreciate the authors' efforts to address these concerns. I will maintain the score of 3.

Reviewer_c5tUOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CrypticFlow, a hierarchical hybrid-coordinate framework for apo-to-holo protein conformation generation. The method first employs an E(3)-equivariant network to generate sparse Cα anchors in Cartesian space and then recursively completes the remaining residues using a Transformer operating on local internal coordinates. By combining Cartesian modeling of global conformational changes with internal-coordinate modeling of local geometry, the framework aims to improve structural fidelity while mitigating the lever-arm effect associated with chain-wide internal-coordinate generation. On the fixed 150-pair D3PM test set, CrypticFlow achieves a median/mean Cα RMSD of 1.333/2.766 Å, compared with 2.12/3.34 Å for the strongest reported FDBM setting. The method also improves local distance and angle fidelity and achieves a MolProbity score of 3.034 after cg2all reconstruction, indicating improved local structural quality in addition to global conformational accuracy.

Critical strengths
  1. The hybrid-coordinate formulation provides a well-motivated way to balance global and local structural modeling. Separating coarse global motion in Cartesian space from local refinement in internal coordinates is intuitive and technically meaningful. The hierarchical construction is also designed to mitigate error accumulation and the lever-arm effect associated with chain-wide internal-coordinate generation.

  2. The evaluation considers both global conformational accuracy and local structural plausibility. This provides a more comprehensive assessment than relying on RMSD alone for apo-to-holo prediction. The improvements in adjacent-distance fidelity, pseudo-angle fidelity, and MolProbity provide complementary evidence that the gains in global RMSD are achieved while maintaining local backbone geometry.

Critical weaknesses
  1. The strongest concern is the limited external validation. All central claims are based on the same fixed 150-pair D3PM test split. The paper does not evaluate actual cryptic-pocket recovery, ligand docking utility, or an external apo–holo dataset. The authors explicitly acknowledge that the model generates a ligand-free Cα trace rather than a ligand-specific ensemble and that D3PM does not evaluate pocket recovery or docking utility. This weakens the connection between the impressive structural metrics and the motivating SBDD application.

  2. Baseline comparisons are not capacity matched. The authors state that CrypticFlow contains 34.08M parameters and is substantially larger than the published bridge baselines. Moreover, some baseline numbers are taken from previous reports, while local/all-atom analyses rely on available stored predictions. This makes it difficult to determine how much of the global RMSD gain arises from the hierarchical representation versus model capacity, ESM-3 conditioning, and other architectural differences.

Questions for authors
  1. Can the authors evaluate whether improved holo RMSD translates into improved pocket recovery or downstream docking on an external cryptic-pocket benchmark? This would materially affect my assessment because SBDD is the primary motivation.

  2. Can the authors provide a parameter/capacity-matched Cartesian baseline using the same ESM-3 conditioning and a comparable training budget? This would help isolate the contribution of hierarchical hybrid coordinates.

  3. Why you don't use ligand information ?

5 comments
AuthorsOpenReview ↗

Response to Reviewer c5tU — Q1. External cryptic-pocket benchmark: pocket recovery and downstream docking

We agree that SBDD is the main motivation, and we measured pocket-level metrics directly on an external cryptic-pocket benchmark.

External benchmark. We used CryptoBench as the external evaluation set. From its 5,493 original records we removed every entry overlapping D3PM-Large by PDB, UniProt, sequence or domain, and verified the common residues and the ligand in the actual structures, which leaves 2,517 records over 650 apo PDBs.

Evaluation regime. We applied the same criterion used to build D3PM-Large — an apo–holo global Cα RMSD of at least 3 Å — to this benchmark as well, giving 212 records over 59 apo PDBs as the evaluation set (Sesame; Miñán et al., 2025, Appendix A.1.1). This selection looks only at apo and holo and uses no prediction.

Pocket recovery.

MethodGlobal Cα RMSD (Å)Pocket Cα RMSD (Å)MolProbity
Apo (input) (baseline)4.2274.1242.256
GT holo (reference)——2.209
CrypticFlow4.059 ± 0.0213.816 ± 0.0233.461 ± 0.035
FDBM5.286 ± 0.1414.928 ± 0.1924.341 ± 0.060
ABM5.318 ± 0.1085.027 ± 0.2304.367 ± 0.021
SBAlign5.065 ± 0.0434.860 ± 0.0474.653 ± 0.007
EGNN15.638 ± 0.18914.226 ± 0.2574.943 ± 0.008

The RMSD is a direct RMSD without refitting, reported as the mean ± seed-to-seed standard deviation over five seeds (1, 3, 42, 88, 99). The pocket is the set of Cα atoms within 8 Å of a holo ligand heavy atom, and values are averaged within an apo PDB and then over apo PDBs. MolProbity was measured after the same fixed-Cα cg2all reconstruction, on the 211 records over 58 apo PDBs on which apo, GT holo and all five models succeeded.

CrypticFlow is the only method better than the apo input in both global and pocket RMSD, and the improvement is larger at the pocket (−0.307 Å) than globally (−0.168 Å) — meaning it did not move the whole protein uniformly but moved the ligand-binding site more. The other four methods move further away from holo on both.

Downstream docking. Within the same ≥ 3 Å regime (212 records over 59 apo PDBs), the docking targets are the 154 records and 37 apo PDBs whose ligand has at least 12 heavy atoms, is fully resolved, and has at most 45 heavy atoms.

The docking conditions are identical for every receptor. We run known-site docking defined by the crystal structure with gnina v1.3.3, with one box fixed per record and shared unchanged by every receptor (the bounding box of the holo ligand heavy atoms plus 4 Å on each face). Every model receptor is a predicted Cα trace converted to all atoms by the same cg2all reconstruction, and the ligand is a conformer freshly embedded with RDKit ETKDGv3 (the crystal pose is never used as input). Pose RMSD is symmetry-corrected and measured without refitting, and the aggregation weighting is the same as in the pocket table.

ReceptorMedian top-1 pose RMSD (Å)Median best-of-20 RMSD (Å)2 Å success rate
Holo crystal structure (positive control)2.1471.17550.2%
GT holo backbone → reconstructed (ceiling)4.1702.90811.1%
Apo (input)6.4734.6422.7%
CrypticFlow5.7814.3605.4%
ABM6.4934.7300.0%
SBAlign6.8624.8440.0%
FDBM7.0264.8370.0%
EGNN8.3445.9790.0%

With the holo structure predicted by CrypticFlow as the receptor, both top-1 and best-of-20 are closer to the correct pose than with the apo input and than with the other four generative models. top-1 is the pose ranked 1st by CNN rescoring; best-of-20 is the closest of the 20 generated poses, i.e. the ceiling under perfect ranking. CrypticFlow is the only generative model that produced any pose succeeding at the 2 Å criterion.

Two controls set the scope of this comparison. Even the ceiling, which passed through the same cg2all reconstruction, reaches only an 11.1% success rate at 2 Å. Getting the backbone exactly right is therefore not enough for Cα-level prediction to recover most of the docking success; side-chain reconstruction is the remaining bottleneck. The gap to the real crystal structure (50.2%) is the size of that shortfall. This result therefore shows that improvements in backbone accuracy propagate into smaller ligand-placement errors; it does not claim to solve docking success itself.

AuthorsOpenReview ↗

Response to Reviewer c5tU — Q2 (part 1). Capacity- and pretraining-matched baselines, inside the training distribution

We agree with the criticism and scaled all four baselines to the same parameter count as CrypticFlow, fed them the same ESM-3 features, and retrained them from scratch. The data (D3PM-Large training split), the number of epochs (1,000), the effective batch size (8) and the checkpoint-selection criterion (validation direct Cα RMSD) were all matched.

With the baselines scaled to CrypticFlow's size (D3PM-Large test, 150 pairs)

MethodParametersGlobal Cα RMSD (Å) ↓Median (Å) ↓< 2 Å (%) ↑Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓
holo (upper bound)—001001001002.291
apo input (unchanged)—6.7434.4610.0099.7095.592.392
CrypticFlow34.08M2.766 ± 0.0411.333 ± 0.05658.53 ± 0.7399.5895.873.034 ± 0.054
ABM33.69M2.5471.06968.0038.4690.213.802
FDBM33.69M4.1903.3287.3337.7384.743.962
SBAlign33.75M3.6752.46722.0011.1847.104.612
EGNN35.07M4.1892.61438.0016.8259.884.388

Bold marks the 1st place among the models that generate structures (apo and holo are reference lines, not models).

Global Cα RMSD uses the same definition as the main text — a common frame obtained by aligning holo to apo once, with no per-prediction refitting. The median and the sub-2 Å fraction are computed from the same direct RMSD. Distance and angle validity also follow the main text (an adjacent Cα distance within 0.2 Å of holo; an angle over three consecutive Cα atoms within 20°). MolProbity was measured after converting the Cα structures of all five models to all atoms with the same cg2all reconstruction. CrypticFlow is the mean ± standard deviation over 5 seeds; the baselines are a single seed (seed 1).

We state the result that goes against us first — with capacity matched, ABM beats us on global RMSD. Mean direct RMSD 2.547 Å against our 2.766 Å, and it also leads on the median (1.069 vs 1.333 Å) and on the sub-2 Å success rate (68.00 % vs 58.53 %). Our seed-to-seed standard deviation is 0.041 Å, so this difference is not explained by seed noise. Once the parameter count is matched, the claim that we are the best in global accuracy does not hold, and we will add this to the appendix.

The same ABM, however, falls far behind us on the quality of the structure. The fraction of adjacent Cα distances it gets right is only 38.46 %, less than half of our 99.58 %, and Cα angle validity is 90.21 % against 95.87 %. MolProbity scored after all-atom reconstruction is 3.802 against our 3.034 ± 0.054. It places atoms at more accurate positions than we do, yet does not generate structures that satisfy accurate local geometry.

Within this test set all five models are better than the apo input. The apo structure, left untouched, has a global RMSD of 6.743 Å, while the five models are at 2.547–4.190 Å. We close 59 % of that gap. This fact is used again in part 2, where we compare against the external benchmark.

This distinction is what we claim. Global RMSD says only how close a predicted structure is to the answer on average; it says nothing about whether that structure is usable in the downstream steps. Scaling the baselines by 60× lets them catch up and surpass us on global RMSD, but it does not close the gap in local geometry.

The results of moving the same setting to an external benchmark not used in training are given in part 2, and the compute required for training and inference in part 3.

AuthorsOpenReview ↗

Response to Reviewer c5tU — Q2 (part 2). The same comparison on the external benchmark (CryptoBench)

Carrying the same four checkpoints over to an external benchmark not used in training changes the picture. On CryptoBench we re-evaluated all five models on the same cohort as our common response (the 212 pairs with apo–holo global Cα RMSD ≥ 3 Å, over 59 apo PDBs) and with the same weighting (averaged within an apo PDB, then over apo PDBs).

After matching capacity and the ESM-3 condition, on the external benchmark (CryptoBench)

MethodParametersGlobal Cα RMSD (Å) ↓Median (Å) ↓< 2 Å (%) ↑Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓
holo (upper bound)—001001001002.209
apo input (unchanged)—4.2273.9740.0099.8897.292.256
CrypticFlow34.08M4.059 ± 0.0214.006 ± 0.0531.51 ± 0.6299.67 ± 0.0492.28 ± 0.143.461 ± 0.035
ABM33.69M4.3274.7421.8936.5388.994.037
FDBM33.69M4.5154.5150.4740.5689.933.890
SBAlign33.75M4.7644.6440.0011.0646.814.644
EGNN35.07M8.74411.3380.0014.6154.094.668

Bold marks the 1st place among the models that generate structures (apo and holo are reference lines, not models).

Among the generative models we are the best on both accuracy and local geometry. Every column but one is ahead of all four baselines — including the global accuracy we had conceded on the internal test set. The one column that does not separate is the sub-2 Å success rate (ABM 1.89 % against our 1.51 ± 0.62 %), which lies within our seed-to-seed standard deviation, so no ordering can be claimed.

ABM, which was ahead of us inside, is reversed outside. With the same checkpoint, a global RMSD of 2.547 Å on the D3PM-Large test becomes 4.327 Å on CryptoBench, while ours goes from 2.766 to 4.059 Å. The global-accuracy advantage ABM gains once capacity is matched holds only inside the training distribution.

More tellingly, none of the four baselines beats the apo input. The apo structure, left untouched, has a global RMSD of 4.227 Å, while the four capacity-matched models are at 4.327–8.744 Å, all worse — moving the structure took them further from the answer. The only generative model below that line is CrypticFlow (4.059 Å). On the test set of part 1 all five models were better than the apo input (6.743 Å); once the training distribution is left, that property remains only for us.

The gap in local geometry is not filled by capacity. The best adjacent-distance validity among the baselines is FDBM's 40.56 %, less than half our 99.67 %. Scaling the parameters by 60× and attaching ESM-3 leaves the gap of part 1 intact. SBAlign shows most clearly that capacity is not the bottleneck — going from 0.54M to 34M moves distance validity from 11.3 to 11.1 % and MolProbity from 4.65 to 4.64, that is, essentially not at all.

The conditions needed to read the table: only CrypticFlow uses 5 seeds, and the baselines are a single seed (seed 1). MolProbity was measured after reconstructing the Cα structures of all five models to all atoms with the same cg2all CalphaBasedModel-FIX, and aggregated over the 211 commonly successful pairs, excluding the one pair that fails identically for every model (CB002742, duplicate atom labels). The FIX model fills in the remaining atoms while holding Cα fixed, so global RMSD, distance validity and angle validity are unaffected by the reconstruction.

Reviewer_DusiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CrypticFlow, a hierarchical flow-matching framework for Apo-to-Holo protein conformation prediction. The method combines an E(3)-equivariant graph neural network for coarse Cartesian anchor generation with a Transformer that recursively generates intermediate residues using local internal coordinates. Additional geometric constraints are introduced to preserve local backbone geometry.

Experiments on D3PM demonstrate improvements in Cα RMSD and local structural fidelity over existing baselines. However, I have substantial concerns regarding the attribution of these improvements, the methodological novelty, and the lack of experimental evidence supporting the intended biological applications.

Critical strengths

S1. Meaningful geometric motivation. The paper addresses an important limitation of Cartesian protein generation methods, where accurate global coordinates do not necessarily guarantee valid local backbone geometry. The proposed hierarchical combination of coordinate representations is conceptually reasonable.

S2. Strong benchmark performance. CrypticFlow achieves competitive global RMSD and substantially improves local geometric fidelity on D3PM. The inclusion of all-atom reconstruction and MolProbity evaluation provides additional evidence beyond conventional backbone RMSD.

S3. Extensive technical analysis. The paper includes detailed geometric derivations, architectural descriptions, and several ablation studies. The investigations of hierarchical error propagation and local-coordinate representations are particularly useful.

Critical weaknesses

W1. The substantial differences in model capacity and pretrained conditioning undermine the fairness of the baseline comparisons.

CrypticFlow contains 34.08M trainable parameters, approximately 60 times more than the major bridge-based baselines, and additionally incorporates pretrained ESM-3 representations. The ESM-3 ablation in Table 8 shows that removing these features increases validation mean RMSD by 0.687 Å, indicating that pretrained conditioning contributes substantially to the reported performance. Without capacity-matched and conditioning-matched baselines, it remains unclear how much of the improvement can be attributed to the proposed hierarchical hybrid-coordinate architecture.

W2. The internal ablations do not establish a consistent advantage of the proposed internal-coordinate representation.

The local Cartesian control in Table 10 achieves better mean RMSD (3.037 vs. 3.064 Å), pseudo-angle fidelity (96.36% vs. 95.21%), and MolProbity scores (2.931 vs. 3.025) than Minimal Local ICS. Although Local ICS improves adjacent-distance fidelity and reconstructed bond/angle geometry, these results suggest a trade-off rather than a general advantage in structural accuracy or physical plausibility. Furthermore, Table 9 shows that fixing Apo-derived radii improves distance fidelity by only 0.15 percentage points while slightly worsening mean RMSD and MolProbity, leaving the practical contribution of this additional constraint insufficiently justified.

W3. The methodological novelty and contribution of hierarchical generation are insufficiently isolated.

The framework combines several established components, including equivariant graph networks, flow matching, conditional Transformers, and multiscale generation. In particular, the coarse-to-fine generation strategy is conceptually related to PAR (Qu et al., 2026), which also employs hierarchical protein generation with flow-based decoding. While the proposed anchor-relative internal-coordinate representation is a meaningful adaptation, the existing experiments do not convincingly establish which architectural components are responsible for the improvements. The stride-1 control also differs in normalization settings, while the full-chain internal-coordinate comparisons are not parameter-matched, limiting their ability to isolate the benefit of hierarchy.

W4. There is a substantial disconnect between the cryptic-pocket motivation and the experimental validation.

The Introduction emphasizes cryptic-pocket opening and structure-based drug discovery, including the KRAS(G12C) switch-II pocket, yet the experiments are restricted to the general D3PM Apo-to-Holo benchmark. No cryptic-pocket-specific evaluation, pocket recovery metric, or downstream docking experiment is provided. Moreover, the model predicts a single ligand-independent Cα trace, without directly generating the side-chain conformations essential for ligand recognition. Although the authors acknowledge these limitations, the current results do not establish whether the predicted conformations are useful for identifying or exploiting cryptic binding sites.

W5. The evidence for physical plausibility and robustness remains incomplete.

Despite improvements over the baselines, the reconstructed structures remain substantially worse than the ground-truth Holo controls in all-atom validation, including clashscore (95.35 vs. 38.3) and Ramachandran outlier rate (3.12% vs. 0.35%). The geometry decoder also requires sphere-feasibility corrections and numerical fallbacks, but their impact on the resulting structures is not sufficiently characterized. In addition, the midpoint Transformer is trained with ground-truth parent anchors but uses generated parents during inference; although the paper recognizes this mismatch and investigates full-chain error propagation, it lacks a direct robustness analysis of perturbations to the generated anchors. These issues limit the strength of the physical-plausibility claims.

Questions for authors

See weakness

9 comments
AuthorsOpenReview ↗

Response to Reviewer Dusi — W1. Capacity and pretraining condition

The criticism is correct. We measured the two factors separately.

Capacity. We scaled all four baselines to the same parameter count as CrypticFlow, fed them the same ESM-3 features, and retrained them from scratch under identical conditions — the same data (D3PM-Large training split), the same number of epochs (1,000), the same effective batch size (8), and the same checkpoint criterion (validation direct Cα RMSD).

MethodParametersIntegration stepsGlobal Cα RMSD (Å) ↓Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓Cα clashscore ↓
CrypticFlow34.08M52.766 ± 0.04199.5895.873.034 ± 0.05495.35 ± 6.64
ABM33.69M1002.54738.4690.213.802193.6
SBAlign33.75M1003.67511.1847.104.612420.6
FDBM33.69M1004.19037.7384.743.962237.4
EGNN35.07M14.18916.8259.884.388358.2

CrypticFlow is the mean ± standard deviation over 5 seeds; the baselines are a single seed (seed 1). The integration steps are the same setting as Table 5 of the paper.

We state the result that goes against us first — with capacity matched, ABM beats us on global RMSD. 2.547 Å against our 2.766 Å, and it also leads on the median (1.069 vs 1.333 Å) and on the sub-2 Å success rate (68.00 % vs 58.53 %). Our seed-to-seed standard deviation is 0.041 Å, so this is not seed noise. Once the parameter count is matched, the claim that we are the best in global accuracy does not hold, and we will add this to the appendix.

The same ABM, however, falls far behind on the quality of the structure. The fraction of adjacent Cα distances it gets right is 38.46 %, 2.6× lower than our 99.58 %, and Cα angle validity is 90.21 % against 95.87 %. Scored after all-atom reconstruction, MolProbity is 3.802 against our 3.034 ± 0.054. It places atoms at more accurate positions than we do, yet its adjacent Cα–Cα distances are not valid. Such a Cα trace breaks down in bond lengths and clashes once it is reconstructed to all atoms, and that is what surfaces in MolProbity and in docking (W4).

In addition, at the same capacity we use far less compute. CrypticFlow finishes 1,000 epochs in 13.3 GPU-hours on one RTX 6000 Ada. At this size ABM and FDBM do not fit on a single 48 GB card; they required four RTX 6000 Blackwell cards and 255 GPU-hours (19×), and SBAlign took 208 GPU-hours on eight H100s. Inference on the same card (one a5000, 150 pairs run sequentially) is 63× faster per pair — 0.249 s against ABM's 15.765 s.

In summary, what a capacity-matched comparison lets us claim is not global accuracy but that we produce usable structures far more cheaply.

Pretraining condition. ESM-3 is also pretrained on structural data, so we shared the concern that structural information about the evaluation proteins might leak in through the embeddings. We therefore trained, with the same five seeds, a control in which the ESM-3 embedding is replaced by ESM-2, which is pretrained on sequence alone (esm2_t33_650M_UR50D), and a control with the language-model features removed entirely.

SettingMean RMSD (Å) ↓Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓
CrypticFlow (ESM-3)2.766 ± 0.04199.58 ± 0.0295.87 ± 0.253.034 ± 0.054
ESM-3 → ESM-22.765 ± 0.01999.59 ± 0.0196.66 ± 0.102.966 ± 0.022
No language-model features3.555 ± 0.04899.51 ± 0.0391.67 ± 0.253.376 ± 0.041

We do depend on the language-model features themselves — removing them degrades mean RMSD from 2.766 to 3.555 Å. That gain, however, is not due to ESM-3. Swapping in ESM-2, which has no structural pretraining, gives 2.765 Å, essentially the same. Structural leakage does not explain it; the gain is explained by information learned from sequence.

Local geometry does not depend on the language model either. With the language-model features removed entirely, Cα–Cα distance validity is 99.51 %, almost the same as our 99.58 %. It is the one quantity that moves with neither capacity nor pretraining condition, and that is the basis for our claim that it comes from the representation.

AuthorsOpenReview ↗

Response to Reviewer Dusi — W2 (part 1). Consistency of the internal-coordinate advantage

The criticism is correct — inside the training distribution it is a trade-off. We took the very pair you cite from Table 10 (Minimal Local ICS vs Local Cartesian, which differ only in the midpoint representation), extended it from three seeds to five (1, 3, 42, 88, 99), and re-measured it on the same 150-pair test split.

First, what the three names mean. All three use the same hierarchical generation and differ only in which coordinates the midpoint — the residue placed between two parent Cα atoms — is written in.

  • Minimal Local ICS — the midpoint is written in local internal coordinates (distance and direction from the parents). This is the minimal setting: our representation only, with the fixed apo-derived radii, the auxiliary losses and the sphere-feasibility correction switched off.
  • Local Cartesian — the same minimal setting with the midpoint instead written as Cartesian (x, y, z) in the same parent frame; it differs from the above in the representation alone.
  • CrypticFlow — our full model: Minimal Local ICS with all three switched on.
MetricCrypticFlow (reference)Minimal Local ICSLocal Cartesian
Mean direct RMSD (Å) ↓2.766 ± 0.0412.734 ± 0.0212.703 ± 0.032
Cα angle valid (%) ↑95.87 ± 0.2596.01 ± 0.0796.99 ± 0.03
MolProbity ↓3.034 ± 0.0543.017 ± 0.0722.924 ± 0.049
Cα–Cα distance valid (%) ↑99.58 ± 0.0299.46 ± 0.0293.17 ± 0.51

The Cartesian side leads on three of the four metrics. Adding seeds leaves the direction you saw in Table 10 unchanged, and we have no basis for claiming our representation is generally better in global accuracy or all-atom quality.

The same holds for the fairer control matched to CrypticFlow in losses and training settings (CrypticFlow_local_cartesian) — median RMSD 1.247 vs 1.333 Å, pseudo-angle 97.01 vs 95.87 %, MolProbity 2.928 vs 3.034, all in favour of Cartesian, with only Cα–Cα distance validity reversed at 93.41 vs 99.58 %. The covalent geometry of the reconstructed structure is nevertheless better on our side (bond-length RMSD 0.0242 vs 0.0317 Å, bond-angle RMSD 2.82 vs 3.32°, the same direction in all five seeds).

One may still ask whether this trade-off also holds outside the training distribution. The comparison above was made entirely within D3PM-Large. We therefore applied the same two models at the same five seeds to the external dataset CryptoBench with no further training. The regime, metric definitions and weighting are as in our common response (apo–holo global Cα RMSD ≥ 3 Å, 212 records over 59 apo PDBs; averaged within an apo PDB, then over apo PDBs).

MethodGlobal Cα RMSD (Å) ↓Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓
Apo (input, unchanged)4.2399.8897.292.26
CrypticFlow4.06 ± 0.0299.67 ± 0.0492.28 ± 0.143.46 ± 0.04
Minimal Local ICS4.05 ± 0.0299.13 ± 0.0592.54 ± 0.173.44 ± 0.08
Local Cartesian4.30 ± 0.1320.15 ± 1.4457.64 ± 11.514.49 ± 0.10

On external data the trade-off does not appear. The three metrics on which Local Cartesian led inside D3PM — global RMSD, angle validity and MolProbity — all reverse outside it. Its global RMSD is 4.30 Å, worse even than the apo input that was not moved at all (4.23 Å); angle validity splits 92.54 vs 57.64 % and MolProbity 3.44 vs 4.49. The distance-validity gap of 6.29 pp inside becomes 79 pp outside (99.13 vs 20.15 %). This is not a bargain in which one quantity is given up for another: outside the training distribution it is a loss on all four metrics.

So the trade-off you saw in Table 10 is not a property of the representation but a property visible only inside the training distribution. On the same D3PM pairs the Cartesian side fits global accuracy and all-atom quality slightly better, but that advantage does not transfer to proteins unseen in training. The Cα–Cα distance validity that the internal-coordinate side preserves, by contrast, is kept almost intact outside it (99.58 → 99.13 %).

We also state the facts that go against us. On external data the best local-geometry values still belong to the apo input (distance 99.88 %, angle 97.29 %, MolProbity 2.26). Our claim is that among models that move the structure we damage it least, not that we beat leaving the input untouched. Minimal Local ICS — which has the fixed radii, the auxiliary losses and the sphere correction all switched off — performs essentially the same as the full CrypticFlow here. What generalizes outside the training distribution is therefore the midpoint representation, not the rest of our design.

AuthorsOpenReview ↗

Response to Reviewer Dusi — W2 (part 2). Contribution of the fixed radii, and the scope of our claim

The second point, the contribution of the fixed radii, is also essentially nil. You are right. To verify this we trained a control (Predicted radii, the same five seeds) in which every path by which the apo distance enters as a radius is removed — both the copying and the upper bound. The result is a distance validity of 99.56 %, a 0.02 pp difference from CrypticFlow (99.58 %), with MolProbity moving by +0.042. The figure of 99.58 % is therefore not produced by copying the radii. We will revise the wording that reads as if it credited the copying constraint.

Where that figure actually comes from is shown by a ladder of generation schemes. We reproduce here the table we prepared in answering Reviewer GXZT's Q3 — all four settings use the same five seeds and the same 150-pair test set.

ModelAdjacent radiusMidpoint generationCα–Cα distance valid (%) ↑MolProbity ↓
Stride 1 (pure EGNN flow matching)n/aNo hierarchy; all residues in global Cartesian67.34 ± 2.353.117 ± 0.087
CrypticFlow_local_cartesiann/aHierarchical, local Cartesian93.41 ± 0.512.928 ± 0.040
Predicted radiiPredicted by the modelHierarchical, local ICS99.56 ± 0.013.077 ± 0.041
CrypticFlowCopied from apoHierarchical, local ICS99.58 ± 0.023.034 ± 0.054

What we claim as our contribution is the combination itself — hierarchical local ICS. Without the hierarchy, when a single EGNN generates all residues directly in global Cartesian coordinates by flow matching (Stride 1), distance validity reaches only 67.34 %. With the hierarchy and the midpoint written in local Cartesian coordinates it is 93.41 %, and with the same hierarchy and only the midpoint written in local ICS it is 99.58 %. Since 99.56 % is reached without copying the radii at all, this figure arises when the hierarchy and local ICS are present together. Neither one alone produces it.

And the global accuracy paid for it is small. Along the same ladder the mean direct RMSD is 2.450 ± 0.014 Å for Stride 1, 2.703 ± 0.033 Å for local Cartesian and 2.766 ± 0.041 Å for CrypticFlow — all within 0.32 Å of one another. While distance validity splits by 32 pp, global accuracy moves only that much. Inside the training distribution we give up a little accuracy and buy a great deal of local geometry; outside it, as the CryptoBench table in part 1 shows, even that bargain disappears.

This is the scope of what we can claim. Inside the training distribution, what the hybrid coordinate system buys is Cα–Cα distance validity and the covalent geometry of that Cα trace once it is reconstructed to all atoms — bond-length RMSD 0.0242 vs 0.0317 Å and bond-angle RMSD 2.82 vs 3.32°, ahead in all five seeds. It does not lead in global accuracy (RMSD) or MolProbity. Outside the training distribution the scope widens — in the CryptoBench table of part 1 we lead Local Cartesian on all four metrics, and only the internal-coordinate family beats the apo input in global RMSD. We will describe this distinction between the two regimes in the appendix.

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

The paper proposed crypticFlow predicts a holo Cα trace from an apo trace. An ESM-3-conditioned EGNN flows a stride-4 set of Cα anchors in Cartesian space. A Transformer then fills midpoints in a local internal-coordinate frame tied to the two parent anchors and an apo reference, copying the apo Cα–Cα distance when the parent is sequence-adjacent. On the fixed 150-pair D3PM test split, five seeds give median direct RMSD 1.333 ± 0.056 Å and mean 2.766 ± 0.041 Å, with high adjacent-geometry fidelity and a cg2all MolProbity score of 3.034 ± 0.054.

Critical strengths
  1. The protocol is careful: a fixed D3PM split, direct RMSD in the pre-aligned frame, five seeds, checkpoint selection by validation RMSD, and means with sample standard deviations. The median gap versus the best reported FDBM (2.12 Å) and the sub-2 Å success gap (58.5% versus 48%) are large enough to matter.

  2. The hierarchy is a direct response to the lever-arm problem, and stride 4 caps each residue at two midpoint refinements. Full-chain internal-coordinate controls are the right comparison, and the paper reports them.

  3. Local geometry and unrelaxed all-atom quality move together with global RMSD. Adjacent-distance fidelity is 99.58 ± 0.02% within 0.2 Å and pseudo-angle fidelity is 95.87 ± 0.25% within 20°, against roughly 32% and 82% for FDBM. MolProbity 3.03 beats the stored bridge outputs (about 4.08–4.62).

  4. Inference is one five-step Euler trajectory per level and is reported as faster than the bridge baselines despite 34.08M parameters. The limitations section states the capacity mismatch, the ligand-free setting, and the absence of a pocket or docking metric.

Critical weaknesses
  1. Baselines in Table 1 are published aggregates. Sesame has no stored per-target outputs, so local metrics omit it. Capacity is not matched. One sample per target is not a conformational ensemble, which is what cryptic-pocket use actually needs.

  2. Adjacent-distance fidelity is partly built in. Sequence-adjacent radii are copied from the apo trace, and apo and holo Cα–Cα distances are both near 3.8 Å, so agreement within 0.2 Å of the holo reference is a weak test. The authors say the aggregate gain from fixed radii is modest and that these scores are fidelity to the reference, not physical validity. The 99.58% versus 32% comparison still reads as a much larger geometric win than the mechanism supports. Pseudo-angles and MolProbity are the fairer local metrics.

  3. A local-Cartesian midpoint control is slightly better on RMSD. Global accuracy and the internal-coordinate state are therefore separable: the Cartesian anchor stage, ESM-3 features, and model size can explain the RMSD, while internal coordinates explain local fidelity. That control belongs in the main table, with and without ESM-3, at matched parameter count.

Questions for authors
  1. What are median RMSD, sub-2 Å success, angle fidelity, and MolProbity for the local-Cartesian control and for the no-ESM-3 control, on the same five seeds?

  2. On pairs with large apo–holo pocket RMSD, what is the pocket-residue RMSD and the change in pocket volume, separate from global RMSD?

  3. If adjacent radii are predicted instead of copied, by how many points do the 99.58% distance number and the MolProbity score move?

3 comments
AuthorsOpenReview ↗

Response to Reviewer GXZT — Q1. Controls on the same five seeds

We trained both controls with the same five seeds as CrypticFlow (1, 3, 42, 88, 99) and evaluated them on the test set (150 pairs). For the language-model features, we also report a control that replaces ESM-3 with ESM-2, which is pretrained on sequences only.

  • Values are mean ± s.d. over the five seeds.
  • RMSD is direct Cα RMSD in the shared pre-aligned frame.
  • MolProbity and the covalent-geometry RMSDs are computed on all-atom models reconstructed with cg2all, without relaxation.
ModelMedian RMSD (Å) ↓Mean RMSD (Å) ↓< 2 Å (%) ↑Cα–Cα distance valid (%) ↑Cα angle valid (%) ↑MolProbity ↓Bond-length RMSD (Å) ↓Bond-angle RMSD (°) ↓
CrypticFlow1.333 ± 0.0562.766 ± 0.04158.53 ± 0.7399.58 ± 0.0295.87 ± 0.253.034 ± 0.0540.0242 ± 0.00062.82 ± 0.07
CrypticFlow_local_cartesian1.247 ± 0.0582.703 ± 0.03358.93 ± 1.4693.41 ± 0.5197.01 ± 0.042.928 ± 0.0400.0317 ± 0.00123.32 ± 0.09
No ESM-32.507 ± 0.0923.555 ± 0.04843.47 ± 1.6699.51 ± 0.0391.67 ± 0.253.376 ± 0.0410.0297 ± 0.00073.41 ± 0.05
ESM-3 → ESM-21.321 ± 0.0422.765 ± 0.01958.00 ± 0.8299.59 ± 0.0196.66 ± 0.102.966 ± 0.0220.0234 ± 0.00042.70 ± 0.04

Distance valid: fraction of adjacent Cα–Cα distances within 0.2 Å of the holo value. Angle valid: fraction of Cα triplet angles within 20° of the holo value, at the same residues. MolProbity: the geometric quality score of the reconstructed all-atom model itself (clashes, Ramachandran and rotamer statistics), not a comparison against holo. Bond-length/bond-angle RMSD: deviation from standard covalent geometry, as reported by MolProbity.

CrypticFlow_local_cartesian. Keeping the losses and training settings identical to CrypticFlow, we replaced local ICS with the midpoint's Cartesian coordinates (x, y, z) in the same parent frame, whose origin is the midpoint of the two parent Cα atoms. Accuracy is nearly unchanged (seed-paired differences: median RMSD −0.09 ± 0.09 Å, sub-2 Å +0.4 ± 1.1 pp), and Cα angle validity (+1.1 pp) and MolProbity (−0.11) are slightly better. What it loses is bond-level geometry. In all five seeds, adjacent Cα–Cα distance validity drops by 6.2 pp, and the reconstructed all-atom models have larger covalent bond-length RMSD (+0.0074 ± 0.0010 Å) and bond-angle RMSD (+0.50 ± 0.09°). The MolProbity score is computed from clashes, Ramachandran and rotamer statistics and does not include bond lengths or angles, so this loss does not appear in the score.

Language-model features (ESM-2, No ESM-3). Because ESM-3 was also pretrained on structural data, structural information about the evaluated proteins could have entered through its embeddings. To test this, we replaced the ESM-3 embeddings with those of ESM-2 (esm2_t33_650M_UR50D), which is trained on sequences only; the 1280-dimensional embeddings were zero-padded to the 1536-dimensional input, and every other setting was kept. Switching to ESM-2 did not change performance (seed-paired differences: median RMSD −0.01 ± 0.08 Å, sub-2 Å −0.5 ± 1.1 pp, MolProbity −0.07 ± 0.06, angle validity +0.8 ± 0.2 pp; covalent bond-length and bond-angle RMSDs slightly lower in all five seeds). We will add this result to the appendix. In contrast, removing the language-model features worsens every accuracy and quality metric in all five seeds (median RMSD +1.17 Å, sub-2 Å −15.1 pp, angle validity −4.2 pp, MolProbity +0.34, bond-length RMSD +0.0055 Å, bond-angle RMSD +0.59°).

CrypticFlow therefore relies on pretrained language-model features, but the benefit is explained by information learned from sequences rather than by the structural pretraining of ESM-3. And what the local-ICS representation buys is not accuracy but bond-level geometry: predicting the midpoint in local Cartesian coordinates instead leaves median RMSD and sub-2 Å essentially unchanged while adjacent Cα–Cα distance validity alone drops by 6.2 pp.

AuthorsOpenReview ↗

Response to Reviewer GXZT — Q2. Pairs with a large pocket change

We have added a pocket-level analysis on the D3PM-Large test split.

Pocket definition and evaluation set. D3PM pairs carry no binding-site label, so we defined each pocket from the ligand actually present in the holo structure.

  • For each test pair, we listed all non-polymer components of the holo mmCIF entry, excluding water, ions, glycans and common crystallization additives (e.g., SO4, GOL, EDO, PEG, DMS, NAG).
  • A ligand counted as a candidate if it met all three conditions:
    • ≥6 heavy atoms;
    • contact within 4.5 Å with the evaluated holo chain;
    • ≥5 residues of that chain within 8 Å.
  • Pocket residues are the residues with any heavy atom within 8 Å of a ligand heavy atom in the holo structure.
  • To keep the pocket unambiguous, we used only the 85 pairs with exactly one candidate ligand; 55 pairs had several candidates (multiple candidate binding sites) and 10 had none.
  • The selection uses only the holo structure and no prediction.
  • Pockets contain 5–94 residues (median 42).
  • Among these 85 pairs, we define the 22 pairs with apo–holo pocket-aligned RMSD ≥ 4 Å as "pairs with large pocket change" and report them.
  • No model receives ligand information, and apo, holo and all predictions are evaluated on the same residues.

Metrics.

  • Pocket-aligned Cα RMSD
    • We superimpose the pocket Cα atoms of a structure onto the holo pocket Cα atoms (Kabsch; rotation + translation) and compute the RMSD over those atoms.
    • Because the fit uses only pocket residues, any rigid displacement of the whole pocket (e.g., from a domain motion) is removed, leaving only the change in pocket shape.
  • Pocket volume error
    • We compute the convex-hull volume of the pocket Cα atoms (Qhull via SciPy) for apo, holo and each prediction, using the same residues.
    • The error is |V_pred − V_holo|.
    • The apo reference, |V_apo − V_holo|, equals the magnitude of the apo→holo volume change.
    • Because all compared methods output only Cα atoms, this is not a true cavity volume but a proxy for how much the pocket backbone opens or closes.
  • Global Cα RMSD: same definition as in the main results (common frame of the preprocessed pair, no per-prediction refit).

How per-method values were obtained.

  • CrypticFlow: per-protein mean over five training seeds (one prediction each). The seed-to-seed standard deviation of the 22-pair mean pocket-aligned RMSD is 0.06 Å.
  • FDBM, ABM, SBAlign: the saved single predictions used in the paper's Cα table.
  • EGNN: the coordinates of the submitted EGNN control were not saved, so we retrained five EGNNs with the same architecture (1,000 epochs) and averaged them in the same way as CrypticFlow.

Results: 22 pairs with large pocket change (mean (median))

MethodPocket-aligned RMSD (Å) ↓Volume error (ų) ↓Global RMSD (Å) ↓Pairs with lower pocket-aligned RMSD than apo
Apo (input)6.43 (5.40)1647 (1245)7.04 (6.16)—
CrypticFlow2.31 (1.17)332 (157)2.73 (1.74)22/22
FDBM3.02 (2.19)425 (267)3.38 (2.62)20/22
ABM3.10 (2.44)375 (377)3.40 (2.87)20/22
SBAlign4.64 (4.01)908 (672)4.87 (3.69)17/22
EGNN6.80 (6.76)1480 (1325)10.37 (9.76)8/22

Findings.

  1. Pocket shape. CrypticFlow lowers the pocket-aligned RMSD from 6.43 Å (apo) to 2.31 Å (median 1.17 Å), the lowest among the compared methods, and is below apo in all 22 pairs.
  2. Pocket volume. CrypticFlow reduces the volume error from 1647 ų (apo) to 332 ų (median 157 ų), the lowest in both mean and median.

We will add this analysis, the selection rules and the list of 85 pairs to the appendix. As limitations, the evaluation set is restricted to holo structures with a single candidate ligand, and the Cα hull is a backbone-level proxy rather than an all-atom cavity measure.

AuthorsOpenReview ↗

Response to Reviewer GXZT — Q3. Predicting the adjacent radii instead of copying them

We built a control, Predicted radii, in which the model predicts the adjacent radii instead of copying them from the apo structure. We trained it with the same five seeds as CrypticFlow (1, 3, 42, 88, 99) and evaluated it on the test set (150 pairs).

Predicted radii setting. All other settings are identical to CrypticFlow; we removed every path through which apo-measured Cα–Cα distances enter the radii.

  • Adjacent radii are not fixed to the apo distances. The model predicts all radii, including those at the final level of the hierarchy.
  • The radius ceiling is also removed. CrypticFlow bounds each radius by 1.1 times the apo contour length of the corresponding span, whereas Predicted radii predicts the distance itself with no upper bound.
  • Apo distances therefore enter the radii neither as values nor as a ceiling. The apo structure is used only as an input condition, as in every other model.

Results.

  • Values are mean ± s.d. over the five seeds.
  • Distance valid: fraction of adjacent Cα–Cα distances within 0.2 Å of the holo value.
  • MolProbity is computed on all-atom models reconstructed with cg2all, without relaxation.
  • Rows are ordered by increasing distance validity.
ModelAdjacent radiiMidpoint generationCα–Cα distance valid (%) ↑MolProbity ↓
Stride 1 (pure EGNN flow matching)n/aNo hierarchy; all residues in global Cartesian67.34 ± 2.353.117 ± 0.087
CrypticFlow_local_cartesiann/aHierarchical, local Cartesian coordinates93.41 ± 0.512.928 ± 0.040
Predicted radiiPredictedHierarchical, local ICS99.56 ± 0.013.077 ± 0.041
CrypticFlowCopied from apoHierarchical, local ICS99.58 ± 0.023.034 ± 0.054

Predicting the radii instead of copying them lowers distance validity by 0.02 ± 0.02 pp (99.58 → 99.56%) and changes the MolProbity score by +0.042 ± 0.051 (seed-paired differences). Copying the radii therefore has a small effect on both numbers.

Why distance validity stays high. Comparing the same metric across generation schemes shows that the high value comes from hierarchical generation with local ICS, not from copying the radii.

  • When a single EGNN generates all residues in global Cartesian coordinates by flow matching, without hierarchy (Stride 1, pure EGNN flow matching), distance validity is 67.34%.
  • With the hierarchy but with midpoints predicted as local Cartesian coordinates, in a frame centred on the two parent Cα atoms (CrypticFlow_local_cartesian — the same losses and training settings as CrypticFlow, differing only in the midpoint representation), it is 93.41%.
  • With midpoints predicted in local ICS, it is 99.56% even without copying the radii (Predicted radii).

In the revision, we will add the Predicted radii control and report this result next to the sentence comparing 99.58% with the baselines.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Common response — external dataset (CryptoBench)

To test whether the model generalizes beyond D3PM, we applied the models trained on D3PM-Large, without any retraining, to CryptoBench (Škrhák et al., 2024).

Evaluation set

  • To remove proteins that overlap with the training data, we compared both the apo and the holo structure of each CryptoBench record against the apo and holo structures of all 1,591 D3PM-Large pairs (train/val/test). A record was excluded if any of the following held:
    • the same PDB ID and chain, or the same UniProt accession;
    • full-length sequence identity ≥30% with ≥80% coverage of both sequences (MMseqs2);
    • domain-level similarity: identity ≥30%, alignment length ≥80 residues, ≥80% coverage of the shorter sequence, and E-value ≤10⁻³.
  • 2,517 records (650 distinct apo PDB entries) pass this filter and the structural checks. The checks require ≥50 matched Cα atoms between apo and holo, ≥5 pocket Cα atoms, and a holo ligand.
  • D3PM-Large was curated to contain only pairs whose apo–holo Cα RMSD exceeds 3 Å (Miñán et al., 2025). To evaluate under the same condition, we applied this curation criterion to CryptoBench. This leaves 212 records (59 distinct apo PDB entries) with an apo–holo Cα RMSD of at least 3 Å. The selection uses only the apo and holo structures; no prediction was consulted.

Metrics

  • Global Cα RMSD: as in the main text, measured in the shared pre-aligned frame (holo aligned once to apo), without a post-hoc prediction-to-target Kabsch fit.
  • Cα–Cα distance validity and Cα angle validity: as in the main text.
    • Distance: the fraction of adjacent Cα pairs within an unbroken segment whose distance is within 0.2 Å of holo.
    • Angle: the fraction of consecutive Cα triplets within the same segments whose angle is within 20° of holo.
  • MolProbity score: the Cα structures of all methods are reconstructed to all-atom models with the same procedure (cg2all) and scored with MolProbity. It is computed on the 211 records (58 apo PDB entries) for which reconstruction and scoring succeeded for every structure.
  • Because one apo structure can carry several ligand records, every metric is first averaged within each apo PDB entry and then averaged with equal weight across apo PDB entries.

Compared methods CrypticFlow, FDBM, ABM, SBAlign, and EGNN were all trained on the D3PM-Large training split with five seeds (1, 3, 42, 88, 99). Checkpoints were selected by validation direct Cα RMSD, and values are mean ± standard deviation over the five seeds.

Results: CryptoBench, apo–holo Cα RMSD ≥ 3 Å (212 records, 59 apo PDB entries)

MethodGlobal Cα RMSD (Å) ↓Cα–Cα distance validity (%) ↑Cα angle validity (%) ↑MolProbity ↓
Apo (input, unmoved)4.2399.8897.292.26
CrypticFlow4.06 ± 0.0299.67 ± 0.0492.28 ± 0.143.46 ± 0.04
FDBM5.29 ± 0.1425.64 ± 1.4277.33 ± 1.834.34 ± 0.06
ABM5.32 ± 0.1125.75 ± 0.3977.06 ± 0.614.37 ± 0.02
SBAlign5.06 ± 0.0411.28 ± 0.3445.71 ± 0.154.65 ± 0.01
EGNN15.64 ± 0.196.25 ± 0.1731.12 ± 0.734.94 ± 0.01

On this external dataset, which was not used for training, CrypticFlow is the only method that improves on the apo input in global Cα RMSD; all four compared generative models move further from holo than the apo input does. Among the generative models, CrypticFlow also has the highest Cα distance and angle validity and the best MolProbity score.

The apo input, however, leads us on local geometry and on MolProbity: a structure that was not moved cannot break its own geometry. What we claim is a relative advantage among models that do move the structure, together with an improvement over the apo baseline in global accuracy.

We will add this external evaluation, the leakage-filter criteria, and the record list to the appendix.

AuthorsOpenReview ↗

Common response — summary of the new experiments

Thank you, and an overview of what we added during the discussion period

We thank all four reviewers. Most of your comments pointed to places we had not measured. During the discussion period we turned each of them into an experiment. We have revised the wording of the paper to match what those experiments showed.

The new experiments, and where each result can be found:

New experimentWhat we measuredWhere the result is
External generalizationOn CryptoBench, after removing every record that overlaps the training data and applying the same criterion as D3PM-Large (apo–holo Cα RMSD ≥ 3 Å), 212 records / 59 apo PDB, with no retraining: global and local metrics and MolProbitycommon response - external dataset
Pocket recovery and docking on the external benchmarkOn the same cohort: pocket Cα RMSD, pocket volume error, and downstream dockingc5tU Q1, Dusi W4 (part 2)
Pocket-level evaluation on D3PMPocket Cα RMSD and pocket volume error on the 85 pairs with a single unambiguous binding siteDusi W4 (part 1), GXZT Q2
Capacity- and pretraining-matched baselinesThe four baselines scaled to 34M, given the same ESM-3 features, and retrained from scratch; evaluated inside and outside the training distributionDusi W1, c5tU Q2 (part 1) and (part 2)
Midpoint-representation controlA model that differs only in writing the midpoint in local Cartesian coordinates, trained on the same five seedsGXZT Q1, Dusi W2 (part 1)
Radius controlA model with every path by which the apo distances enter the radii removed, trained on five seedsGXZT Q3, Dusi W2 (part 2)
Language-model-feature controlsESM-3 replaced by ESM-2, and the language-model features removed altogether, each trained on five seedsGXZT Q1
Capacity-matched hierarchy-free controlsThe two full-chain ICS controls widened to our parameter count and retrainedDusi W3 (part 2)
Physical plausibility and anchor robustnessClashscore and Ramachandran outliers reported separately, and the final error measured with isotropic noise of σ = 0.25–2.00 Å injected into the top-level anchorsDusi W5
Exposure-bias trainingScheduled sampling, noisy context learning, and both together, retrainedLaNz Q1
Sample-ensemble evaluation20 samples per target on the 150 test pairs, measuring sample diversity and coverage of alternative holo statesLaNz Q2
ComputeFor the same five capacity- and ESM-3-matched models: training GPU-hours normalized to 1,000 epochs, and inference time on the 150 pairsc5tU Q2 (part 3)

All of the added results will go into the appendix of the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.