acceptodds
Under review as a conference paper at ICLR 2027

A2Holo: Dual-Stream Flow Matching of 3D Structure and a Protein Language Model for Conformational Transition Prediction

Abstract

Modelling a ligand-bound holo conformation from a ligand-free apo conformation is essential in structure-based drug discovery. We present A2Holo, which casts apo-to-holo prediction as flow matching between two experimental structures rather than generation from noise. It transplants the dual-stream joint attention of in-context image generation to residue-aligned protein language model embeddings and structure. Without ligand conditioning or stochastic sampling, A2Holo's single deterministic trajectory reaches a median Cα RMSD of 0.809 Å on the D3PM benchmark, the lowest among existing methods. Its TM-score exceeds even the best of 10 samples from every stochastic baseline on at least 144 of 150 pairs. Beyond performance, we directly measure the transplanted multimodal mechanism itself. Because sequence and structure are residue-aligned, its joint attention can be measured at residue resolution and tested for reproducibility across independently trained model variants. Within the structure stream, the attention a residue receives recovers moving residues better than the attention it emits, and this asymmetry is learned. A2Holo demonstrates a successful multimodal transplant from in-context image generation to protein conformational change, with strong deterministic apo-to-holo prediction. It also provides a quantitative account of how that mechanism organises residue-aligned sequence and structure around the residues that move.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein conformational change, multimodal attention, flow matching, structural biology, protein language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_T8CDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a way to predict holo structures from apo inputs. The work does this by using a protein language model conditioner and an MMDiT style flow trunk, and learns a vector field between apoo and holo structures. On the standard D3PM benchmark, the method significantly outperforms prior work.

Critical strengths

The paper's strongest claim is that the results are exceptionally strong, and a significant improvement over prior work.

Critical weaknesses

I have several methodological issues that I think make it difficult to immediately support acceptance:

  1. There's a significant body of work on this topic, and it's quite hard to understand what's driving the strong results. In paticular, works like [1] and [2] tackle similar (though non idetntical) problems. Historically, it's been challenging to directly push apo to holo structures, so I wonder what this method is doing differently and think that a rigorous ablation is important. To be clear, I don't mean listing the differences -- I think, to the paper's credit, this is quite clear -- I mean actually ablating each component to see what's enabling this.
  2. [1] found that there's significant overlap between the test set and Uniprot. Does this work use a correctly deduplicated split?
  3. My rough calculation estimates 365M parameters in this model (12 dual stream + 12 single stream layers). There are only 1200 apo-holo pairs. I find it very hard to imagine training a significant model with this mismatch, so I think this needs to be properly justified and explained.

If these questions were properly explored and answered, I would shift to acceptance.

[1] Zhang, Xuejin, et al. "Bending and binding: Predicting protein flexibility upon ligand interaction using diffusion models." NeurIPS 2023 Generative AI and Biology (GenBio) Workshop. 2023. [2] Corso, Gabriele, et al. "Composing unbalanced flows for flexible docking and relaxation." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  1. What is driving the strong results?
  2. Why can a very small dataset train a relatively large transformer?
7 comments
AuthorsOpenReview ↗

Overall response and revision map

We thank Reviewer T8CD for a careful and specific review. Figure, table, and appendix numbers below refer to the revised manuscript.

“This paper presents a way to predict holo structures from apo inputs... On the standard D3PM benchmark, the method significantly outperforms prior work.”

We thank you for the summary.

“The paper's strongest claim is that the results are exceptionally strong, and a significant improvement over prior work.”

“I have several methodological issues that I think make it difficult to immediately support acceptance...”

Because the exceptionally strong headline performance is also the point that most requires explanation, the revision ablates its components and repeats the comparison on a protein-disjoint split.

We investigated your concerns along four axes:

  • What drives the performance: component-level ablations of the training objective and conditioning.
  • Architecture and model size: controlled comparisons separating sequence information from raw parameter count.
  • Dataset overlap: explicit analysis of the public split followed by a protein-disjoint re-split with all compared methods retrained.
  • Attention measurements: residue-level joint-attention measurements across 9 independently trained models, plus one association on the protein-disjoint split.
AuthorsOpenReview ↗

What is driving the strong results?

“There's a significant body of work on this topic, and it's quite hard to understand what's driving the strong results.”

“Historically, it's been challenging to directly push apo to holo structures, so I wonder what this method is doing differently...”

“To be clear, I don't mean listing the differences ... I mean actually ablating each component to see what's enabling this.”

In short, the data point to the training recipe, above all the relative-displacement loss, to the sequence stream, and to added capacity only together with the language-model embedding. We agree that the different sources of the performance should be separated. We share your caution, since directly pushing apo structures to holo has long been difficult. Listing architectural differences is therefore not enough; we ablated the components directly and separately examined benchmark overlap and model capacity.

On the public split, the released configuration gives a median Cα RMSD of 0.809 Å; without the relative-displacement loss it gives 1.455 Å.

By contrast, changing the particular sequence representation does not worsen the median beyond the variation across retrainings. A fixed amino-acid-type vector gives 0.784 Å, ESM-2 gives 0.780 Å, ESM-C gives 0.757 Å, and removing the global token gives 0.772 Å. Five seeds of the released configuration span 0.761–0.836 Å.

Removing the sequence stream entirely gives 1.004 Å (Appendix A.14). Amino-acid identity remains available to that model through the structure stream’s one-hot node features, so this comparison removes the pretrained sequence representation and dual-stream interaction rather than all residue identity.

Section 4 then measures what the joint attention learns, beyond what accuracy alone shows. The revision moves the four reductions of the joint map, previously Figure A6 in the appendix, into the main text as Figure 5. Within the structure stream, received attention identifies moving residues more strongly than emitted attention, and this asymmetry is learned. After controlling for distance from the apo centroid, the partial rank correlation between received attention and total distance change is 0.046–0.145, exceeding matched untrained networks in all 9 trainings (p = 0.0039).

The Conclusion now adds one association on the protein-disjoint split: “pairs whose structure-stream attention marks the moving residues more closely show less off-target motion (Spearman −0.42 over 58 pairs, Holm-adjusted p = 0.035, Appendix A.9).” Appendix A.9 defines both quantities.

The public split also contains substantial overlap with training. Among 109 chain-sharing test pairs with a training partner, 105 have the same conformational change represented in training. Copying the displacement of the nearest training partner gives 0.438 Å median RMSD on those pairs, compared with 0.478 Å for A2Holo. Across the 123 test pairs that have a training partner (not the same set as the 123 UniProt-sharing pairs of Appendix A.9), A2Holo’s error correlates with the corresponding copying error at Spearman +0.868.

A same-size model that replaces joint attention with cross-attention, in which each stream attends only to the other, is one of the seven models in Appendix A.15, and A2Holo has the lowest median of the seven (Table A6). The sequence-stream ablation and the seven models are reported in Appendices A.14–A.15; the public-split training-component runs above are additional to the paper. The updated anonymous release links the checkpoints of the seven models (CHECKPOINTS.md), with their training options.

AuthorsOpenReview ↗

Model size, sequence information, and the small training set

“My rough calculation estimates 365M parameters in this model ... There are only 1200 apo-holo pairs.”

“Why can a very small dataset train a relatively large transformer?”

In short, the model does fit its training pairs closely, and in the 2×2 comparison below on the protein-disjoint split its extra capacity pays off only when the second stream carries the language-model embedding. Your estimate is close: A2Holo has 368.8M parameters (368,807,959), and the public split contains 1,291 training pairs.

The model fits that training set closely: across five seeds, the training median Cα RMSD is 0.157–0.212 Å with one Euler step. We therefore agree that training fit itself is not the informative question. The more relevant question is whether the additional capacity improves held-out prediction.

Appendix A.15 addresses this directly using a 2×2 comparison on the protein-disjoint split (one training per model; blocks are double-stream + single-stream). The structure-copy model keeps the 12+12 dual-stream architecture, but its second stream receives a copy of the structure input instead of the embedding, with no global token.

Second-stream contentSmaller model12+12 model
Ankh3 embedding4+4, 127.9M: 5.362 ÅA2Holo, 368.8M: 4.998 Å
No LM information0+12, 113.7M: 5.235 ÅStructure-copy, 367.8M: 5.283 Å

With language-model information, increasing from 4+4 to 12+12 lowers Cα RMSD in 37/58 protein-disjoint test pairs (p = 0.0037). Without language-model information, increasing from 0+12 to the size-matched 12+12 structure-copy model instead raises RMSD in 36/58 pairs (p = 8.4×10⁻⁴). The two effects differ in 38/58 pairs (p = 3.1×10⁻⁴).

Thus, in this 2×2 comparison, the larger dual-stream model pays off only when its second stream carries the language-model embedding.

The public-split depth experiments give the same caution against attributing performance simply to parameter count. The 4+4 model (127.9M parameters) does not differ significantly from 12+12: 12+12 has lower RMSD in 84/150 pairs and 4+4 in 66/150 (p = 0.087). Models with 8+8, 10+10, 14+14, and 16+16 blocks, spanning 248–489M parameters, give median RMSDs of 0.747–0.835 Å, comparable to the 0.761–0.836 Å range across five seeds of the released 12+12 model.

The released 12+12 configuration was fixed before these depth comparisons rather than selected after observing them.

We therefore do not argue that the dataset size is large relative to the parameter count.

Reviewer_dE1hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a dual-stream flow matching method that integrates 3D structure information and protein language model representations to predict holo structures from apo structures.

Critical strengths

A2Holo achieves the lowest RMSD among the compared methods on the D3PM benchmark.

Critical weaknesses
  1. Novelty: Joint modeling of protein sequences and structures for downstream tasks is already a common approach in the community, as demonstrated by HTP, DPLM-2, and PT-DiT. This paper largely builds upon this well-established concept and offers limited methodological novelty. Moreover, in previous works, apo-to-holo prediction has primarily been used as an example of paired-data translation to demonstrate the capabilities of proposed diffusion bridge frameworks. The biological significance of this task, particularly in the absence of ligand information, is not sufficiently justified.

  2. Ablation: The RMSD results in Table 1 are impressive, but the authors do not provide sufficient ablation studies to explain why their method outperforms previous baselines. Notably, even without the sequence stream, the model achieves a median RMSD of 1.004 Å, substantially outperforming previous methods. This raises the question of whether the considerably larger Transformer backbone, rather than the proposed dual-stream design, is the primary reason for the performance improvement.

Questions for authors

Data Leakage: The benchmark split may contain substantial protein-level overlap between the training and test sets. As reported in Appendix C of ApolloDiff, 93% of the test samples in the original SBAlign benchmark have an apo–holo pair with an identical UniProt accession in the training or validation set.

5 comments
AuthorsOpenReview ↗

Overall response and methodological positioning

We thank Reviewer dE1h for a careful and specific review. Figure, table, and appendix numbers below refer to the revised manuscript.

“This paper proposes a dual-stream flow matching method that integrates 3D structure information and protein language model representations...”

We thank you for the summary.

“Joint modeling of protein sequences and structures for downstream tasks is already a common approach in the community...”

We agree that multimodal sequence–structure modeling itself is not a new idea. Since models such as AlphaFold2 and ESMFold, simply combining sequence and structural information would not constitute the contribution we intend to claim. What remains less clear is how much the sequence modality actually contributes and how that contribution arises during structural prediction. This is the question we study here.

We revisited the related works you named. In DPLM-2, the Table 7 ablation removes sequence pre-training warm-up and synthetic structures, but the sequence modality remains present in every setting; it therefore does not directly compare a multimodal model with a structure-only counterpart. PT-DiT likewise does not remove or replace its sequence channel based on PCA-reduced ESM-2/AF2 embeddings; its closest comparison treats ESM2-15B as a separate baseline. We are also not aware of a multiple-training statistical analysis in either work that directly isolates the contribution of the multimodal interaction.

Our revision therefore emphasizes the measurement of sequence–structure interaction, rather than the mere presence of two modalities. Figure 4, new in the main text, shows the joint attention map for one test pair. This analysis is part of the contribution rather than a supplement to the model: the paper states “one modelling contribution and two measured ones,” and the revision moves the four reductions of the joint map, previously Figure A6 in the appendix, into the main text as Figure 5. Because the sequence and structure tokens are aligned by residue, the four quadrants of joint attention can be interpreted at residue resolution and compared across independently trained models.

Within the structure stream, the attention a residue receives identifies moving residues more strongly than the attention it emits, and this asymmetry is learned rather than imposed by the block layout. After controlling for distance from the apo centroid, the partial rank correlation between received attention and total distance change is 0.046–0.145, exceeding the mean of untrained initializations at the same depth in all 9 independently trained models (p = 0.0039). Across streams, “the signal sits at the structure-residue end of the cross quadrants, resolved in the sequence-to-structure one” (Section 4), again in all 9 trainings (p = 0.0039).

Appendix A.15 provides a complementary architectural control. At essentially fixed model size, replacing the language-model embedding with a copy of the structure input and removing the global token changes the model’s behavior substantially: A2Holo moves the apo structure farther than this structure-copy model in 54/58 protein-disjoint test pairs (p = 2.7×10⁻¹⁰).

“...in previous works, apo-to-holo prediction has primarily been used as an example of paired-data translation...”

We agree that this is a fair characterization of prior work. In our case, apo-to-holo is not simply used as an example task for a broader framework; it is the object of study itself. We examine how the two streams interact residue by residue during the conformational transition and evaluate not only prediction accuracy but also backbone geometry and the cost of producing a single pose.

On the protein-disjoint split in Appendix A.9, where every compared method is retrained, A2Holo remains more accurate per pair than each baseline, while Table A2 evaluates backbone geometry.

The revision now cites DPLM-2 and PT-DiT explicitly in the introduction.

AuthorsOpenReview ↗

Biological meaning of ligand-free apo-to-holo prediction

“The biological significance of this task, particularly in the absence of ligand information, is not sufficiently justified.”

We understand this concern, since without the ligand the model cannot know what the pocket will hold. The ligand-free setting is relevant when the binding partner is not known in advance, as in the search for cryptic pockets. Downstream structure-based workflows such as docking must also ultimately choose a receptor structure to begin with.

A single predicted bound form is a meaningful target if bound structures of the same apo protein tend to agree despite differences in ligand identity. We therefore measured this directly in D3PM.

Across the full 1,591-pair dataset, we identified 1,080 pairs of holo structures whose apo structures coincide (apo Cα RMSD ≤ 0.1 Å). Using displacement cosine ≥ 0.9 and displacement-magnitude ratio ≥ 0.75 to define the same conformational change, 92.8% agree.

Among the 711 comparisons with different ligand codes, agreement increases to 94.4%. For comparison, the 369 pairs with identical ligand codes agree in 89.7% of cases. Within this benchmark, the bound conformations associated with the same apo structure are therefore usually consistent even when ligand identity differs.

The exceptions are concentrated rather than broadly distributed. Only 31 of the 347 apo chains with at least two holo structures contain more than one distinct conformational change by this criterion, and these cases are predominantly associated with nucleotide-binding proteins.

We do not interpret this analysis as showing that ligand identity is generally irrelevant. A2Holo predicts one deterministic bound structure and cannot represent an ensemble or choose between ligand-specific states when distinct states exist. The manuscript states this scope (Appendix A.19), and the Conclusion identifies multi-state prediction and ligand conditioning as future directions.

AuthorsOpenReview ↗

What drives the performance?

“The RMSD results in Table 1 are impressive, but the authors do not provide sufficient ablation studies to explain why their method outperforms previous baselines.”

“...even without the sequence stream, the model achieves a median RMSD of 1.004 Å, substantially outperforming previous methods. This raises the question of whether the considerably larger Transformer backbone... is the primary reason...”

Raw model size does not explain the gain: on the protein-disjoint split, in this 2×2 comparison, the larger dual-stream model pays off only when its second stream carries the language-model embedding (Appendix A.15). We expanded the ablations at three levels: training components, sequence information, and model capacity; blocks are written as double-stream + single-stream. The public-split training-component, embedding-swap and 4+4 runs below are additional to the paper.

Training components. On the public split, removing the relative-displacement loss increases median Cα RMSD from 0.809 to 1.455 Å, the clearest degradation among the tested training components.

Replacing the sequence embedding with a fixed vector per amino-acid type, ESM-2, or ESM-C, or removing the global token gives 0.757–0.784 Å. These differences are comparable to the 0.761–0.836 Å range across five retrainings of the released configuration, so we do not claim that the specific Ankh3 embedding is uniquely responsible for the accuracy.

Removing the sequence stream. Retraining the structure-only model from scratch with the same recipe and stopping rule gives 1.004 Å versus 0.809 Å for A2Holo. Removing the sequence stream therefore costs 0.195 Å, larger than the 0.075 Å range across five retrainings of the released configuration. However, this ablation also removes the double-stream blocks and changes the model size, so it cannot by itself isolate the sequence modality.

Separating sequence information from capacity. Appendix A.15 therefore adds a 2×2 comparison on the protein-disjoint split (one training per model):

Second-stream contentSmaller model12+12 model
Ankh3 embedding4+4, 127.9M: 5.362 ÅA2Holo, 368.8M: 4.998 Å
No LM information0+12, 113.7M: 5.235 ÅStructure-copy, 367.8M: 5.283 Å

With language-model information, increasing from 4+4 to 12+12 lowers Cα RMSD in 37/58 test pairs (p = 0.0037). Without language-model information, increasing from 0+12 to the 12+12 structure-copy model instead raises RMSD in 36/58 (p = 8.4×10⁻⁴). The two effects differ in 38/58 pairs (p = 3.1×10⁻⁴).

Five of the seven models in Appendix A.15 are closely size-matched at 367.8–368.8M parameters, including the structure-copy and cross-attention controls, and A2Holo has the lowest median Cα and Cβ RMSD among the seven.

We also varied depth on the public split. Models with 8+8 to 16+16 blocks (248–489M parameters, not listed in the paper) give 0.747–0.835 Å (Table A4). The smaller 4+4 model does not differ significantly from 12+12: 12+12 is better on 84/150 pairs and 4+4 on 66/150 (p = 0.087).

Reviewer_WeW1OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

A2Holo predicts protein apo-to-holo conformational changes using a dual-stream flow model that jointly processes pretrained sequence embeddings and structural coordinates. Its main contributions are explicit sequence–structure interaction, geometry-aware training, and prediction with ten deterministic integration steps. Validation includes comparisons with existing methods on the SBAlign splits and architectural ablations. Code is available, trained checkpoints are not provided.

Critical strengths
  • Explicit interaction between pretrained sequence representations and structural coordinates provides a plausible mechanism for incorporating sequence-derived information into conformational prediction. This is the paper’s most compelling methodological contribution.

  • The reported prediction accuracy with only ten deterministic integration steps suggests practical value for applications requiring a single candidate structure. This is a useful computational property, although step count alone does not establish a runtime advantage.

Critical weaknesses
  • The experimental design does not substantiate the claimed generalization or isolate the architectural contribution. Among the 150 test pairs, 112 share at least one exact PDB-chain endpoint with training, including 33 whose holo target already appears in the training set. Moreover, ApolloDiff Appendix C reports substantial UniProt overlap across the SBAlign splits reused here. These results therefore do not establish generalization to unseen proteins; they need to be distinguished from prediction of additional conformations of known proteins. The probe evaluation raises a related concern: assigning folds through fold = used % N_FOLDS does not enforce protein-level separation unless protein grouping is implemented elsewhere. The ablation also confounds information, architecture, and capacity. Removing PLM conditioning simultaneously removes all 12 dual-stream blocks and reduces the parameter count from 367.6M to 112.5M. Consequently, the RMSD increase from 0.809 to 1.004 Å cannot isolate the contribution of PLM information or dual-stream fusion. Together, these issues weaken the evidence for both the scope of generalization and the proposed explanation for the performance gain.

  • The meaning of a deterministic holo prediction without ligand conditioning remains insufficiently defined. Different ligands can stabilize distinct bound conformations of the same protein, yet A2Holo produces a single structure for a given apo input. Such a predictor could reasonably target a representative or most probable bound conformation, but the paper needs to specify and validate that interpretation. Lower RMSD to one reference holo structure establishes reference-reconstruction accuracy; it does not demonstrate coverage of ligand-dependent states or establish that sampling is unnecessary. This distinction is central when comparing A2Holo with stochastic methods such as SBAlign, ABM, FDBM, and ApolloDiff. The methodological presentation, particularly Figure 2, should also make the conditioning information and prediction target clearer.

Questions for authors
  • What are the exact PDB-chain and UniProt overlaps between training, validation, and test sets? Is protein-level grouping enforced anywhere before the probe fold assignment? Please reconcile the implementation with the claimed protein-level separation.

  • What does the single predicted holo structure represent when ligand identity is unavailable and several distinct bound conformations exist? How are samples from stochastic baselines generated, selected, and scored, and what sampling or computational budgets are used?

  • Does the reported advantage persist on a protein-disjoint split and on proteins with multiple distinct bound conformations? The latter evaluation should compare the single prediction with sampled ensembles and, where feasible, ligand-conditioned models.

  • Can the authors separate the effects of PLM information, dual-stream fusion, and model capacity—for example, by retaining PLM conditioning while replacing dual-stream interaction with a simpler fusion mechanism at comparable parameter count?

6 comments
AuthorsOpenReview ↗

Overall response and revision map

We thank Reviewer WeW1 for a careful and specific review. The questions and comments were precise and technically well grounded, and they directly shaped the analyses added in the revision. Figure, table, and appendix numbers below refer to the revised manuscript.

“A2Holo predicts protein apo-to-holo conformational changes using a dual-stream flow model...”

We thank you for the accurate summary.

“Explicit interaction between pretrained sequence representations and structural coordinates provides a plausible mechanism...”

We appreciate this reading, as it closely matches the scope of the paper. The work is centered on making the interaction between sequence representations and structural coordinates explicit and on measuring how that interaction behaves during the apo-to-holo transition. Figure 4, new in the main text, shows this interaction for one test pair, and Section 4 evaluates this behavior across 9 independently trained models and 133 test pairs rather than relying on a single checkpoint. As the Conclusion states, A2Holo “provides not only a deterministic route from an apo structure to a holo one, but a measured account of how the two modalities are combined along that route,” and the revision moves the four reductions of the joint map, previously Figure A6 in the appendix, into the main text as Figure 5.

“The reported prediction accuracy with only ten deterministic integration steps suggests practical value...”

We agree that step count alone is not a runtime claim. Appendix A.10 (Table A3) reports measured wall-clock cost; we summarize the sampling and compute budgets below.

“The experimental design does not substantiate the claimed generalization or isolate the architectural contribution.”

“Together, these issues weaken the evidence for both the scope of generalization and the proposed explanation for the performance gain.”

We agree with both points and added analyses addressing each one. We do not use the public D3PM split to make an unseen-protein generalization claim, and the original sequence-removal ablation simultaneously changed sequence information, architecture, and parameter count.

We therefore provide four targeted analyses:

  • Protein-disjoint evaluation: we reconstructed the protein-disjoint split procedure of ApolloDiff Appendix C and retrained SBAlign, ABM, FDBM, and A2Holo on the resulting 1,425/45/58 train/validation/test split (Appendix A.9).
  • Architecture: we trained seven models on that split, A2Holo and six controls, to separate sequence information, fusion mechanism, and model capacity (Appendix A.15, Table A6).
  • Deterministic target: we clarified what the single ligand-free output represents and, in this response, quantified how often multiple holo structures associated with the same apo structure correspond to the same conformational change.
  • Probe separation: we stated the probe fold assignment exactly and, in this response, repeated the probe analysis with folds grouped by 30% sequence-identity clusters.

We also revised the introduction sentence “for a new target it is missing” to “while the structure at hand is often the unbound one.”

The detailed evidence for each point is given below.

AuthorsOpenReview ↗

Dataset overlap and protein-disjoint evaluation

“Among the 150 test pairs, 112 share at least one exact PDB-chain endpoint with training, including 33 whose holo target already appears in the training set.”

“What are the exact PDB-chain and UniProt overlaps between training, validation, and test sets?”

We reproduced these counts from the released split and confirm that they are correct. The public D3PM split is not protein-disjoint.

Overlap with trainingPublic test (150)Public validation (150)Rebuilt test (58)
Exact PDB chain1121070
Apo only / both / holo only79 / 18 / 1577 / 16 / 14–
Holo chain present33300
Apo chain present97930
Identical training pair000
UniProt accession119112 (74.7%)0
UniProt with train + validation123–0

We give the corresponding accuracy breakdown here: median Cα RMSD is 0.476 Å on the 112 chain-overlapping pairs versus 4.22 Å on the remaining 38. Of those 112, 109 have a training partner, defined as a training pair that shares an exact chain or a 30% sequence-identity cluster and aligns to at least half of the residues at ≥30% identity. For 105/109, the same conformational change is also present in training.

Copying the displacement of the nearest-apo training partner gives 0.438 Å median RMSD on those 109 pairs versus 0.478 Å for A2Holo. Across the 123 test pairs that have a training partner (not the same set as the 123 UniProt-sharing pairs), A2Holo’s RMSD correlates with the best-partner copying error at Spearman +0.868.

“Moreover, ApolloDiff Appendix C reports substantial UniProt overlap ... These results therefore do not establish generalization to unseen proteins...”

We agree that this distinction matters, and we do not claim generalization to unseen proteins. To measure the protein-disjoint setting directly, we reconstructed the split procedure described in ApolloDiff (Zhang et al., 2023, Appendix C).

Pairs are joined by single linkage if they share a UniProt accession or reach 0.7 sequence similarity or TM-score; test and validation each receive 20 whole clusters drawn at random with a fixed seed, plus the pairs of 10 single accessions taken from clusters whose other accessions stay in training. Test and validation pairs are restricted to 3–10 Å apo–holo Cα RMSD, and the 63 held-out pairs outside that range are dropped. This gives 1,425/45/58 train/validation/test pairs. No test pair shares a UniProt accession or PDB chain with training or validation; one test pair shares a PDB entry with such a pair through a different chain.

“Does the reported advantage persist on a protein-disjoint split...”

On the reconstructed split, A2Holo has the lowest median Cα RMSD, 4.998 Å against 5.115–5.656 Å for the retrained baselines. It has lower Cα RMSD in 38–43/58 pairs (p ≤ 0.0045), higher TM-score in 38–49/58 (p ≤ 8.0×10⁻⁵), and higher lDDT in 48–57/58 (p ≤ 1.8×10⁻⁷).

ApolloDiff releases neither code nor checkpoints, so we retrained SBAlign, ABM, and FDBM on the reconstructed split and evaluated all available methods with the same code. Each baseline draws 10 samples, scored under the three selection rules listed in the sampling comment below; A2Holo produces one deterministic trajectory.

Backbone geometry shows a larger separation. With geometry-based sample selection, the median consecutive Cα–Cα violation rate is 3.4% for A2Holo, against 67.1% for SBAlign, 35.2% for ABM and 38.5% for FDBM (Table A2).

“The probe evaluation raises a related concern ... does not enforce protein-level separation...”

“Is protein-level grouping enforced anywhere before the probe fold assignment?”

You are right that the original fold assignment does not separate pairs that share a protein, and there is no protein grouping elsewhere. Folds are assigned by apo–holo pair: every residue of a pair stays in one fold, but pairs sharing a protein can fall into different folds. The original wording “at the protein level” was therefore ambiguous. Appendix A.17 now states the assignment explicitly, and Figure A1 no longer says that folds are split by protein.

We also refit every probe using folds defined by 30% sequence-identity clusters. The Section 4.5 conclusion remains unchanged: the depth-averaged relative decodability ratio stays below one in all 9 trainings, at 0.342–0.552 for displacement magnitude and 0.623–0.854 for total distance change, versus 0.320–0.569 and 0.549–0.845 with pair-level folds. The paper reports the pair-level numbers.

Appendix A.9 reports the overlap counts (123 UniProt-sharing and 112 chain-sharing test pairs) and the protein-disjoint benchmark (Tables A1–A2); the other overlap figures above are additional to the paper.

AuthorsOpenReview ↗

What explains the architectural gain?

“The ablation also confounds information, architecture, and capacity...”

We agree with your diagnosis. Removing the PLM stream also removes all 12 double-stream blocks and reduces model size, so the public-split increase from 0.809 to 1.004 Å cannot by itself isolate sequence information, fusion mechanism, or capacity.

“Can the authors separate the effects of PLM information, dual-stream fusion, and model capacity—for example, by retaining PLM conditioning while replacing dual-stream interaction with a simpler fusion mechanism at comparable parameter count?”

We separated these three factors experimentally on the protein-disjoint split with seven models (58 test pairs; one seed per model; blocks are double-stream + single-stream; Appendix A.15, Table A6), including a same-size fusion control (cross-attention). A simpler fusion, token concatenation, is an additional run described below.

The central comparison is a 2×2 test of model size and second-stream content. With language-model information present, increasing the model from 4+4 to 12+12 blocks lowers Cα RMSD in 37/58 pairs (p = 0.0037). Without language-model information, increasing from the 0+12 model to the 12+12 structure-copy model instead raises Cα RMSD in 36/58 pairs (p = 8.4×10⁻⁴). The two effects differ in 38/58 pairs (p = 3.1×10⁻⁴).

The structure-copy control retains the 12+12 dual-stream architecture, but its second stream receives a copy of the structure input in place of the language-model embedding, and it has no global token. Five of the seven models in Table A6, including the structure-copy and cross-attention variants, are closely matched at 367.8–368.8M parameters. A2Holo has the lowest median Cα and Cβ RMSD among the seven.

The 367.6M and 112.5M counts printed by the released code exclude a 1.18M-parameter auxiliary head that is not used for structure prediction; the counts printed in Appendix A.15 (368.8M and 113.7M) include it.

At fixed language model and model size, one of the seven models replaces joint attention in every double-stream block with cross-attention, in which each stream attends only to the other (Table A6).

We also tested a simpler fusion that is neither size-matched nor in the manuscript: concatenating Ankh3 tokens into a shared-weight 0+12 single stream (113.7M parameters). Its median of 5.365 Å, which happens to equal the ESM-C row of Table A6, compares with 5.235 Å without the tokens, improving 22/58 pairs and worsening 36/58 (p = 0.0059).

At fixed 12+12 depth, A2Holo moves the apo structure farther than the structure-copy model in 54/58 pairs (p = 2.7×10⁻¹⁰). The two models without language-model embeddings have median displacements of 0.71–0.73 Å, compared with 1.12–2.02 Å for the five models that use them.

We do not interpret this as evidence that Ankh3 itself is uniquely responsible for the accuracy. On the public split, replacing Ankh3 with ESM-2, ESM-C, or a fixed vector for each amino-acid type gives 0.757–0.784 Å (additional runs, not in the paper), compared with 0.809 Å for the released model and 0.761–0.836 Å across its five seeds.

The architectural result supported by these controls is that the larger dual-stream model pays off only when its second stream carries the language-model embedding, while raw parameter count alone does not explain the improvement. Section 4 complements these controlled accuracy experiments by measuring the learned sequence–structure interaction directly across 9 independent trainings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.