We thank the reviewer for their thoughtful comments on our work. We are glad the reviewer appreciates the impact of this work in multi-conformation structure analysis, the strength of the multi-scale guidance strategy, and the clarity of the writing. As the reviewer highlights, atomic model building into cryo-EM maps is an active research area which stands to benefit from the recent advances in biomolecular structure prediction, especially to overcome the limitations of manual model building and low resolution maps. Through several additional experiments on real data, as described below, we hope to showcase the applicability of our approach for broadening the scope of cryo-EM map interpretation.

Additional Validation on Experimental Data. We thank the reviewer for the suggestion to further evaluate our approach on experimental data. We assess our method on 3 additional protein systems associated with real cryo-EM density maps, and whose deposited models are not part of the Boltz-1 training set. For each experimental map, we obtain 25 samples from each method, choose the best sample as assessed by map-model fit (real-space correlation coefficient), and report its all-atom RMSD and TM-score with respect to the deposited model. For ModelAngelo, we obtain its single prediction and report model completeness and TM-score. Metrics are presented in the table below, and we summarize the results here:

• Pma1: We use two experimental maps of this ATPase [EMD-64135, 64136], corresponding to its active and inhibited states [PDB: 9UGB, 9UGC]. Unguided Boltz only samples the active state, whereas guidance also samples the inhibited state with an RMSD of 2.01 Å. AlphaFold3 predictions are not accurate with respect to either state.
• CYP102A1: We use two experimental maps of this oxygenase [EMD-27534, 27536], of resolutions 4.4 Å and 6.5 Å, corresponding to its open and closed states [PDB: 8DME, 8DMG]. Guidance improves the RMSD of the closed state from 7.89 Å to 2.39 Å over unguided Boltz-1, whereas for the lower resolution open state map, a more modest improvement from 7.93 Å to 4.33 Å is observed. ModelAngelo only models 19% of the 4.4 Å map and none of the 6.5 Å map.
• YbbAP: We use two experimental maps of this transporter [EMD-51292, 51291], corresponding to states in which ATP is bound or unbound [PDB: 9GE7, 9GE6]. Unguided Boltz only samples the bound state, which guidance additionally improves to an RMSD of 1.31 Å. The unbound state is also obtained through guidance. Since the unbound state map is of medium resolution at 4.05 Å, ModelAngelo only models 55% of the structure.

Our results on these systems, along with the P-glycoprotein example in the manuscript (Figure 5), demonstrate our method’s ability to recover diverse conformations from real experimental maps. We will include these new results, including visualizations, in our revised manuscript.

ModelAngelo Performance. ModelAngelo is trained solely on cryo-EM maps of resolution 4 Å or better. This accounts for its poor performance on lower resolution maps, especially those of P-glycoprotein and CYP102A1. In contrast, our method’s use of a powerful generative prior aids in fitting atomic structure into lower resolution density.

Memorization vs Generalization. We thank the reviewer for suggesting checks of test set contamination. We compute the highest structure match score between the tested deposited models and PDB entries contained in the Boltz-1 training set. The structures of the CYP102A1 and YbbAP systems have no matching entries. One state of Pma1 has a match score of 20%, while the other has no matching entries. Finally, the match scores of 3 of the 4 P-Glycoprotein states (Figure 5) are between 40% and 48%, while the last state has no matching entries. Therefore, most structures have little to no overlap with the training dataset, validating that the generalization capabilities of structure prediction models extend to the case of inference-time guidance. We further emphasize that even in cases of structural overlap, as in P-glycoprotein, unguided Boltz-1 fails to sample multiple states, making guidance a useful intervention.

We thank the reviewer for their thoughtful comments on our work. We are glad the reviewer appreciates the use of AlphaFold3 as a strong biophysical prior for atomic modeling into cryo-EM maps, as well as the novelty of the proposed multiscale guidance strategy. While we show through our experiments that our approach broadens the scope of cryo-EM model building, we also hope that our flexible inference-time method will benefit from the parallel advances in biomolecular structure prediction. We elaborate on these aspects below.

Robustness of the Prior. The reviewer is correct that our method inherits both the strengths and limitations of the underlying structure prediction model. We aim to take advantage of the strong biophysical priors learned by AlphaFold3/Boltz-1, which is beneficial especially when fitting atomic models to lower resolution maps (Figure 5). Conversely, for complexes that are badly mispredicted by the base model, including full antibody complexes and DSL1 which we note in the Discussion section, guidance is unable to overcome this bias. However, a key strength of our inference-time approach is that it can be applied to any structure prediction model with a diffusion-based sampler. We thus envision that the continuous developments in structure prediction will further boost the ability of our guidance strategy to model highly dynamic complexes.

Validation Metrics. We thank the reviewer for the suggestion to expand the validation metrics. Map-model fit, as assessed by real-space correlation coefficient as in Supplementary Table B1, is one approach to model validation and selection. Additional measures of model quality such as MolProbity score or Q score would indeed serve as important validation, and we will include these in our revised manuscript. We will additionally conduct analyses of Boltz-1 confidence scores, which may serve as another mechanism for a user to detect failure cases.

Training with Cryo-EM Maps. While the flexibility of inference-time guidance is a strength of our approach, fine-tuning Boltz with map conditioning, potentially incorporating our guidance algorithm, represents an interesting avenue for future work. We thank the reviewer for this suggestion.

We thank the reviewer for their thoughtful comments on our work. We are glad the reviewer appreciates the clarity of the writing and the strength of the results. Our primary motivation is to fit atomic structures into cryo-EM maps of dynamic complexes, which often requires time-consuming manual effort especially on lower resolution maps where current automated methods tend to fail. By leveraging the powerful generative prior of AlphaFold3, our method provides a new framework for cryo-EM model building that aims to overcome these limitations. We present several additional experiments on real data to further showcase the general applicability of our approach.

Additional Validation on Experimental Data. We thank the reviewer for the suggestion to further evaluate our approach on experimental data. We assess our method on 3 additional protein systems associated with real cryo-EM density maps, and whose deposited models are not part of the Boltz-1 training set. For each experimental map, we obtain 25 samples from each method, choose the best sample as assessed by map-model fit (real-space correlation coefficient), and report its all-atom RMSD and TM-score with respect to the deposited model. For ModelAngelo, we obtain its single prediction and report model completeness and TM-score. Metrics are presented in the table below, and we summarize the results here:

• Pma1: We use two experimental maps of this ATPase [EMD-64135, 64136], corresponding to its active and inhibited states [PDB: 9UGB, 9UGC]. Unguided Boltz only samples the active state, whereas guidance also samples the inhibited state with an RMSD of 2.01 Å. AlphaFold3 predictions are not accurate with respect to either state.
• CYP102A1: We use two experimental maps of this oxygenase [EMD-27534, 27536], of resolutions 4.4 Å and 6.5 Å, corresponding to its open and closed states [PDB: 8DME, 8DMG]. Guidance improves the RMSD of the closed state from 7.89 Å to 2.39 Å over unguided Boltz-1, whereas for the lower resolution open state map, a more modest improvement from 7.93 Å to 4.33 Å is observed. ModelAngelo only models 19% of the 4.4 Å map and none of the 6.5 Å map.
• YbbAP: We use two experimental maps of this transporter [EMD-51292, 51291], corresponding to states in which ATP is bound or unbound [PDB: 9GE7, 9GE6]. Unguided Boltz only samples the bound state, which guidance additionally improves to an RMSD of 1.31 Å. The unbound state is also obtained through guidance. Since the unbound state map is of medium resolution at 4.05 Å, ModelAngelo only models 55% of the structure.

Our results on these systems, along with the P-glycoprotein example in the manuscript (Figure 5), demonstrate our method’s ability to recover diverse conformations from real experimental maps, especially at resolutions where other methods fail. We will include these results, including visualizations, in our revised manuscript.

Validation by RMSD. We agree with the reviewer that samples from the unguided model may also correspond to valid states of a protein. However, our goal is to specifically obtain atomic structures that are consistent with cryo-EM data. Our motivation in baseline comparisons against unguided Boltz-1 and AlphaFold3 is not to suggest that these structures are necessarily incorrect, just that they do not correspond to the target cryo-EM state in the absence of guidance. Guidance provides a mechanism through which the powerful predictive capabilities of these pretrained models can be directed towards the states a cryo-EM practitioner desires from their data.

Validation on Synthetic Data. While a noiseless synthetic map does correspond perfectly to its target structure, it is not certain a priori that Boltz-1 will be able to sample this structure. Indeed, Boltz-1 never samples the inward conformation of the STP10 system (Section 4.2). As the reviewer suggests, our synthetic experiments can thus be seen as validation that our proposed guidance strategy is effective at driving the sampling process toward a structure consistent with a map.

Multiscale Guidance. We apologize for the lack of clarity surrounding nomenclature of the multiscale guidance strategy. The terms “global” and “local” are meant to describe the scale of the conformational changes driven by these two guidance phases. The global phase is most effective at guiding coarser domains of the structure, since the Wasserstein distance metric permits a strong gradient even between a sample atom and map point that are far apart. However, since the point cloud is only a rough approximation of the map, the density errors applied during the local phase are more adept at guiding loops and secondary structure elements into the right locations. While the global guidance phase is indeed responsible for the majority of reduction in RMSD, we find that the local guidance phase remains important for accurate model fitting. Conducting an ablation on the inhibited state of Pma1, global guidance alone reduces RMSD from 7.6 Å to 3.0 Å, while the addition of local guidance further improves accuracy to 2.0 Å. We will include ablations on real data in our revised manuscript.

Novelty. We agree with the reviewer that the novelty of the method primarily lies within this application domain. That said, we believe our multiscale guidance strategy may have broader relevance and hope that it will inspire other diffusion guidance approaches in 3D small molecule generation or computer vision.

We thank the reviewer for their thoughtful comments on our work. We are glad the reviewer appreciates the clarity of the writing and the value of the synthetic experiments. As the first method to leverage the state-of-the-art AlphaFold3 as a prior for cryo-EM model building, we believe our method will broaden the structural interpretation of experimental data, and hope to showcase its applicability through the additional experiments we detail in our response below.

Additional Validation on Experimental Data. We thank the reviewer for the suggestion to further evaluate our approach on experimental data. We assessed our method on 3 additional protein systems associated with real cryo-EM density maps, and whose deposited models are not part of the Boltz-1 training set. For each experimental map, we obtain 25 samples from each method, choose the best sample as assessed by map-model fit (real-space correlation coefficient), and report its all-atom RMSD and TM-score with respect to the deposited model. For ModelAngelo, we obtain its single prediction and report model completeness and TM-score. We additionally expand our set of baselines to include ADP-3D, and describe these evaluations further in the ADP-3D Baseline section below. Metrics are presented in the table below, and we summarize the results here:

• Pma1: We use two experimental maps of this ATPase [EMD-64135, 64136], corresponding to its active and inhibited states [PDB: 9UGB, 9UGC]. Unguided Boltz only samples the active state, whereas guidance also samples the inhibited state with an RMSD of 2.01 Å. AlphaFold3 predictions are not accurate with respect to either state.
• CYP102A1: We use two experimental maps of this oxygenase [EMD-27534, 27536], of resolutions 4.4 Å and 6.5 Å, corresponding to its open and closed states [PDB: 8DME, 8DMG]. Guidance improves the RMSD of the closed state from 7.89 Å to 2.39 Å over unguided Boltz-1, whereas for the lower resolution open state map, a more modest improvement from 7.93 Å to 4.33 Å is observed. ModelAngelo only models 19% of the 4.4 Å map and none of the 6.5 Å map.
• YbbAP: We use two experimental maps of this transporter [EMD-51292, 51291], corresponding to states in which ATP is bound or unbound [PDB: 9GE7, 9GE6]. Unguided Boltz only samples the bound state, which guidance additionally improves to an RMSD of 1.31 Å. The unbound state is also obtained through guidance. Since the unbound state map is of medium resolution at 4.05 Å, ModelAngelo only models 55% of the structure.

Our results on these systems, along with the P-glycoprotein example in the manuscript (Figure 5), demonstrate our method’s ability to recover diverse conformations from real experimental maps, especially at resolutions where other methods fail. We will include these results, including visualizations, in our revised manuscript.

ADP-3D Baseline. We thank the reviewer for the suggestion to compare ADP-3D, which we evaluate on P-glycoprotein (Figure 5), Pma1, and YbbAP. CYP102A1 produced errors due to its large size. For each cryo-EM map, we provide ADP-3D with the ModelAngelo prediction as the required initial structure, and choose the best of 5 samples by map-model fit. For Pma1 and YbbAP metrics, please see the table in the Additional Validation on Experimental Data section above. For P-glycoprotein, we reproduce Table 2 of the manuscript below for the reviewer’s convenience. We find that ADP-3D improves TM-score over ModelAngelo for Pma1 and P-glycoprotein maps, due to some additional modeling of the incomplete initial structure. However, we also observe very poor RMSD scores, especially for YbbAP and P-glycoprotein, where completely incorrect predictions are made for large portions of the structure.

Accuracy vs Resolution. We appreciate the reviewer’s suggestion to analyze all-atom RMSD as a function of map resolution. In the table below, we report the mean and standard deviation of 10 samples guided to the inward state of the STP10 system (Figure 3), across varying resolutions of the synthetic map. We find that performance is best at resolutions better than 6 Å. We additionally observe that RMSD slightly increases going from a 2 Å to a 5 Å map, likely due to smoother gradients in the local guidance phase, suggesting that our method may be improved by adopting a frequency marching schedule.

Extending to Boltz-2. We agree with the reviewer that extending our approach to Boltz-2 represents an important area of future work. Since Boltz-2 introduces significant code changes and was released after the NeurIPS deadline, we have not yet integrated our guidance method. The improved predictive capabilities of Boltz-2 may indeed broaden the applicability of our guidance approach.