acceptodds
Under review as a conference paper at ICLR 2027

Fold'EM: Direct atomic structure inference from raw CryoEM particles

Abstract

Single-particle cryo-electron microscopy (cryo-EM) has become a widely adopted technique for biomolecular structure determination. The conventional cryo-EM computational pipeline first combines many particle images to reconstruct an electrostatic potential (ESP) map and then fits an atomic model to the recovered map. Density reconstruction has high sample complexity, requiring large numbers of particle images and making structure determination high-cost and low-throughput, particularly for structurally heterogeneous samples. Downstream atomic model building, in turn, becomes increasingly difficult as the resolution of the reconstructed map deteriorates. Protein structure prediction models provide strong sequence-derived priors on atomic structure, and experiment-guided approaches can use these priors to recover structures consistent with experimental measurements. Yet, in cryo-EM, such priors are typically integrated only after density reconstruction during atomic model fitting. We introduce Fold'EM, an inference-time framework that combines priors from protein structure generative models directly with raw cryo-EM particles to determine atomic models from a small number of particle images, bypassing both intermediate density reconstruction and downstream model building against the reconstructed map. Across synthetic benchmarks and experimental cryo-EM data, Fold'EM recovers accurate atomic structures both with known particle orientations and in an ab initio setting where orientations are inferred jointly with structure. In heterogeneous datasets, Fold'EM further resolves distinct conformational states from mixed particle populations without separately reconstructing a density map and building an atomic model for each state. We believe these results open new avenues for structure determination in the low-sample regime and for characterizing low-population conformational states directly from cryo-EM particles.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo EM, experiment-guided AlphaFold, AlphaFold

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

The paper proposes a novel cryo-EM reconstruction method leveraging AlphaFold's hidden-state to provide informative protein-structure priors during reconstruction, rather than relying on volumetric priors. A key novel component of the proposed method is using this prior not only for the final density reconstruction, but as a prior during the EM-like iterations which alternate the nuisance variable estimation (poses and class assignments) with the density reconstruction.

Critical strengths

The paper tackles an important problem with well motivated and quite ideas. The proposed method is thoroughly evaluated over synthetic and real data, with and without available poses (ab-initio) and in both homogeneous and heterogeneous settings. The experimental section shows promising results, in particular the high data-efficiency with respect to previous classical and deep learning methods.

Critical weaknesses

Unfortunately, I found different parts of the manuscript difficult to parse. While the manuscript is fairly well written, some parts of the presentations could be significantly improved.

First, the authors often introduce variables and equations in the prose without previously defining all terms. While their meaning is often possible to infer with more context, after reading the rest of the section, this makes the manuscript much harder to parse. I think this offers an opportunity to significantly improve the presentation and the accessibility of the paper.

On a similar note, I struggled to get the full picture of the proposed method until the end of the manuscript, as the main idea is spread through multiple pages. Algorithm 1 summarizes the method but is pushed at the end of the appendix, could be moved to the main paper? Also Fig. 1 is supposed to give a schematic overview of the method at the beginning, but I didn't find it very useful. For example, I struggled to connect it with the (already unclear) equations in the text. Defining clearly each variable at the beginning and placing them in the figure to clarify each component's input/output would help clarifying the whole picture. Where do step 1 and step 2 from Sec. 3 fit exactly in this diagram? How are the subfigures (A) and (B) connected? Also, subfigures (C) and (D) currently look like the outputs of (A) and (B), which is confusing: maybe this figure could be split in two separate ones?

Questions for authors

Sec 2: what is a\textbf{a}? what is ∣a∣|\textbf{a}|? s\textbf{s} and z\textbf{z}? these variables are used without being previously defined. I see some get defined only in Sec 3.1.

What is the "experimental data term" L(Z)L(\textbf{Z}) in Eq. 2? the variable Z\textbf{Z} doesn't appear in Eq. 2. Did you mean g(H)g(\mathcal{H}) instead of g(H)g(\textbf{H}) maybe? it seems you often make use of some overloaded notation with no explanation.

End of page 5: what is ρ\rho exactly? this variable was never defined before. Is it the ground truth density for a single class? what is ρ(H)\rho(\textbf{H})? is it the same as just ρ\rho in the previous text? Also, maybe the unnumbered equation at the bottom is obvious, but a short derivation (in the appendix) would make the manuscript much easier to understand.

I am not sure I fully get the point of the conclusion of Sec 2: is this argument suggesting tracking also the covariance (rather than just a single consensus map) of the reconstructed density? In the conclusions, it is state that "under pose or class uncertainty, a consensus density need not be a sufficient statistic"; maybe this could be framed as a stand-alone theorem with a rigorous proof referenced in the main manuscript? I wonder if presenting this argument as part of the prose, as currently done, is the clearest choice. Also, I might be missing a step: why is this necessary an argument in favor of particle-domain inference rather than density maps (plus the missing sufficient statistics)?

Sec 3.1: why not using methods like RELION for ab-initio pose estimation?

Can you number of equations? even if not referenced directly int he paper, it makes it easier to refer to them when discussing the paper.

4 comments

presentation and notation

We thank the reviewer for the careful reading. The points below are addressed in the revised manuscript, with all changes highlighted in red. The question on the conclusion of Sec. 2 needs a longer answer, which we give in a separate comment.

1. Overview of the method, Algorithm 1 and Fig. 1. We have added a compact summary of the procedure to Sec. 3.2 (Algorithm 1), showing the initialization, the alternation between structure inference (Step 1) and nuisance refinement (Step 2), and the final decoding, with references to the corresponding equations. The full pseudocode, including the guided reverse-diffusion loop, remains in Appendix D (Algorithm 2). Regarding Fig. 1: panel (A) shows the whole of Algorithm 1, namely the initialization of poses and classes followed by the alternation of Step 1 and Step 2; panel (B) zooms into Step 1, showing how the particle likelihood is back-propagated into the intermediate Pairformer state H\mathbf H; panels ( C ) and (D) are an example result (PDB 9B9J). We have revised the caption to state this explicitly and to explain the labels used in the figure and updated panel (B) to include the notation

2. Undefined symbols in Sec. 2 (a\mathbf a, ∣a∣|\mathbf a|, s\mathbf s, z\mathbf z). We have revised the Notation paragraph of Sec. 2 so that every symbol is defined at its first use. a\mathbf a is the AlphaFold3 input specification (polymer sequences and any ligands), and ∣a∣|\mathbf a| is its number of tokens: one per standard residue or nucleotide, and one per atom for ligands and modified residues. s∈R∣a∣×cs\mathbf s\in\mathbb R^{|\mathbf a|\times c_s} and z∈R∣a∣×∣a∣×cz\mathbf z\in\mathbb R^{|\mathbf a|\times|\mathbf a|\times c_z} are the single (per-token) and pair (per-token-pair) representations produced by the Pairformer trunk.

3. The data term L(Z)L(\mathbf Z) and overloaded notation. We apologize for the ambiguity. Calligraphic symbols denote ensembles and bold symbols their members: X={Xj}j=1n\mathcal X=\lbrace\mathbf X^j\rbrace_{j=1}^n, H={Hj}j=1n\mathcal H=\lbrace\mathbf H^j\rbrace_{j=1}^n, and Z=g(H)={g(Hj)}j=1n\mathcal Z=g(\mathcal H)=\lbrace g(\mathbf H^j)\rbrace_{j=1}^n, with gg applied member-wise. The learned-geometry analysis concerns a single ensemble member jj, with Z=g(H)\mathbf Z=g(\mathbf H) (superscript jj dropped). L(Z)L(\mathbf Z) is that member's contribution to the experimental data term of Eq. 2 (Eq. 3 in the revised manuscript), i.e. the expected log-likelihood of the particles assigned to it: L(Z)=∑k: ck=jEX∼p(X∣Z)log⁡p(yk∣X,ϕk)L(\mathbf Z)=\sum_{k:\ c_k=j}\mathbb E_{\mathbf X\sim p(\mathbf X\mid\mathbf Z)}\log p(\mathbf y_k\mid\mathbf X,\boldsymbol\phi_k)

Particles assigned to other members do not depend on Z\mathbf Z, so LL differs from the full data term only by a constant, and the two have the same gradients. The text now gives this definition explicitly.

4. Definition of ρ\rho and derivation of the particle/density identity. ρ\rho denotes the (voxelized) Coulomb potential, or density, of a single conformation, i.e. the object imaged in the particles of one class. For an atomic model X\mathbf X, its density is ρX\rho_{\mathbf X}, defined in Appendix E.3. The previous notation ρ(H)\rho(\mathbf H) was shorthand for ρX\rho_{\mathbf X} with X∼p(X∣g(H))\mathbf X\sim p(\mathbf X\mid g(\mathbf H)); because the decoder is stochastic, we now write ρX\rho_{\mathbf X} explicitly. The identity holds for every density ρ\rho, hence for every sample X\mathbf X, and therefore also in expectation over X\mathbf X. We have clarified both points in Sec. 2 and added a complete derivation in a new Appendix E.4. It expands both quadratics and shows that Qρ~=q\mathbf Q\tilde\rho=\mathbf q also when Q\mathbf Q is singular, e.g. because of CTF zeros or missing views, so the two sides differ by a constant independent of ρ\rho. The derivation also gives this constant explicitly: it is the particle misfit of the reconstruction ρ~\tilde\rho, 12∑k∥yk−Akρ~∥Σk−12\frac12\sum_k\|\mathbf y_k-\mathbf A_k\tilde\rho\|^2_{\boldsymbol\Sigma_k^{-1}}, which depends only on the data.

5. RELION for ab initio pose estimation (Sec. 3.1). We do use RELION: VDAM, the initialization step in Sec. 3.1, is the ab initio algorithm of RELION-4.0 [Kimanius et al. 2021], run as relion_refine --grad --denovo_3dref (Appendix E.2). We have revised Sec. 3.1 to state this explicitly. The resulting poses are then refined jointly with the structure (Step 2 of Sec. 3.2).

6. Equation numbers. All display equations in the main text and appendix are now numbered.

Particle-domain fitting

This comment addresses the reviewer's remaining question: whether the argument at the end of Sec. 2 calls for tracking a covariance, whether it can be stated as a theorem, and why it favours particle-domain inference over a density augmented with the missing statistics.

When the classes are known, as in the scalar example of Sec. 2, a mean and a second moment are indeed sufficient. Below, we give a proposition and proof sketch showing that this no longer holds in the more general case of joint pose and class uncertainty.

Model

Consider two equally populated states,

ρ1,2=ρˉ±δ,\rho_{1,2}=\bar\rho\pm\delta,

where ρˉ\bar\rho is the consensus (what a homogeneous reconstruction estimates) and δ\delta is the class deviation.

The scalar analogue of the imaging model is

yk=ϕk(ρˉ+ckδ)+εk,y_k=\phi_k(\bar\rho+c_k\delta)+\varepsilon_k,

where

  • ck∈{±1}c_k\in\{\pm1\} is the unobserved class, uniformly distributed;
  • ϕk∼N(1,τ2)\phi_k\sim\mathcal N(1,\tau^2) is the unobserved pose;
  • εk∼N(0,σ2)\varepsilon_k\sim\mathcal N(0,\sigma^2) is the observation noise.

We define rk=yk−ρˉr_k=y_k-\bar\rho as the residual against the consensus, as in B-RECOVAR.

Proposition

For any τ≥0\tau\geq0 and ρˉ\bar\rho, the following statements hold.

(i) The consensus carries no information about the class deviation.

\qquad\text{for every }\delta,$$ and $$\frac1K\sum_{k=1}^K r_k\longrightarrow0 \quad\text{almost surely}.$$ Thus, the consensus cannot recover $\delta$, however many particles are averaged. **(ii) The second moment identifies the magnitude of the deviation.** $$\mathbb E[r_k^2]= \sigma^2+\tau^2 \bar\rho^2+(1+\tau^2)\delta^2.$$ Consequently, a single second moment identifies $|\delta|$. **(iii) No continuous statistic with fewer than $K$ components is sufficient for $\delta$.** In particular, neither a mean together with a covariance nor any fixed number of moments is sufficient when both the class and pose are unobserved. **(iv) With known classes, a fixed-size sufficient statistic exists.** For $\tau=0$, the statistic is $$\sum_k c_k r_k.$$ For $\tau>0$, the four per-class sums of $r_k$ and $r_k^2$ are sufficient. ### Proof sketch **Parts (i) and (ii).** Both follow from direct moment computation, using $$\mathbb E[c_k]=0,\qquad \mathbb E[\phi_k]=1,\qquad \mathbb E[\phi_k^2]=1+\tau^2.$$ **Part (iii).** Marginalizing over the unobserved class gives $$\log p(r\mid\delta)=E(r,\delta)+\log\cosh O(r,\delta),$$ where $E$ and $O$ are the even and odd parts in $\delta$ of the class-conditional Gaussian log-density, each quadratic in $r$. Since every nonconstant even Taylor coefficient of $\log\cosh$ is nonzero (as follows from its Bernoulli-number expansion), we obtain $$\log p(r\mid\delta)=\ell(r;\bar\rho)+\sum_{n\geq1}q_n(r)\delta^{2n},$$ where $q_n$ is a polynomial of exact degree $4n$ when $\tau\bar\rho\neq0$, and $2n$ when $\tau\bar\rho=0$. If two samples $\mathbf r$ and $\mathbf s$ have a likelihood ratio that is constant in $\delta$, then $$\sum_k q_n(r_k)=\sum_k q_n(s_k) \qquad\text{for all }n.$$ Now consider the map $$\Psi(\mathbf r)=\left(\sum_k q_1(r_k),\ldots,\sum_k q_K(r_k)\right).$$ Its Jacobian is $$D\Psi(\mathbf r)=\left[q_n'(r_k)\right]_{n,k=1}^K.$$ Because the polynomials $q_n'$ have distinct degrees, this Jacobian is nonsingular at some point. Hence $\Psi$ is locally injective on an open set. Any continuous sufficient statistic $S:\mathbb R^K\to\mathbb R^M$ must therefore also be injective on that set. For $M<K$, this contradicts invariance of domain. **Part (iv).** With known classes, the likelihood belongs to a per-class Gaussian exponential family, admitting the stated fixed-size sufficient statistics.

We address the reviewer's three questions using the proposition above. Recall that ρˉ\bar\rho denotes the consensus of the two states ρ1,2=ρˉ±δ\rho_{1,2}=\bar\rho\pm\delta, where δ\delta is their deviation from the consensus. We write rk=yk−ρˉr_k=y_k-\bar\rho for the residual of particle kk, and τ2\tau^2 for the variance of the unobserved pose factor. References to (i)–(iv) below are to the corresponding parts of the proposition.

1. Covariance

We agree that tracking a covariance can recover information lost by averaging. In the two-state setting, Proposition (ii) shows that the second moment identifies ∣δ∣|\delta|, provided the noise and pose variances are known. This is the principle underlying RECOVAR and B-RECOVAR, which we use for initialization. However, identifying the magnitude of the deviation is not equivalent to retaining all the information in the particle likelihood. Proposition (iii) shows that, when class assignments are unknown, no continuous statistic of fixed dimension is sufficient as the number of particles grows. Notice that this difficulty is not simply due to pose uncertainty: Proposition (iv) shows that fixed-size sufficient statistics exist when the classes are known, even with pose noise. Hence, it is the uncertainty in class assignments that prevents such a finite-dimensional summary in this model.

More specifically, consider the expansion of the marginal log-likelihood around δ=0\delta=0. We denote by q1(rk)q_1(r_k) the coefficient of δ2\delta^2 in the contribution from particle kk. When τρˉ≠0\tau\bar\rho\ne0, this coefficient is a quartic polynomial in rkr_k. Thus, the leading nonconstant term of the full likelihood, δ2∑kq1(rk)\delta^2\sum_k q_1(r_k), already depends on statistics beyond the mean and covariance. For three or more states, the limitation is more apparent, since the first two moments generally do not determine the individual states.

2. Proposition

We agree with the reviewer that the argument at the end of Sec. 2 would benefit from a formal statement. We have therefore formulated the proposition above, which distinguishes three questions: what information is retained by the consensus (i), what can be recovered from the second moment (ii), and whether these summaries are sufficient for inference when the latent variables are unknown (iii–iv). We will replace the corresponding discussion in Sec. 2 with this proposition and provide the full proof in a new appendix of the camera-ready version.

We have also revised the conclusion to reflect this distinction more precisely: "when poses or classes must be inferred, averaging the particles into a consensus density can discard the information needed to infer them together with the structure."

3. Why particles

Our motivation for working directly with particles is both theoretical and practical. First, Proposition (i) shows that the consensus is independent of the class deviation δ\delta in the two-state model. In particular, the mean residual converges to zero regardless of δ\delta. Notice that collecting more particles does not resolve this limitation: the consensus becomes more accurately estimated, but the deviations of the two states continue to cancel. We observe a corresponding limitation experimentally in Fig. A2. For 8U4E with 500 particles, consensus-map guidance reaches 17.95 Å Cα RMSD, compared with 6.88 Å for Fold'EM.

Second, although additional statistics such as a covariance can preserve information about heterogeneity, they must be estimated using the current pose and class assignments. This creates a difficulty for joint inference: as the atomic structure improves, these assignments may change, and a summary computed using earlier estimates may no longer be appropriate. In Fold'EM, we instead return to the individual particle images and refine the poses and classes against the current atomic model at each iteration (Step 2 of Sec. 3.2). Hence, improvements in the structure can inform the nuisance-variable estimates, which in turn provide better information for the next structure update. This is particularly relevant when few particles are available, since estimating a reliable voxel-wise covariance is itself challenging.

We do not claim that density-based inference is inherently inadequate, or that a density supplemented with sufficiently rich statistics could not support joint inference. Rather, our argument is that a single consensus density is insufficient, while retaining the particles allows the structure and latent variables to be refined together without committing to a fixed summary computed from earlier estimates.

I thank the authors for the detailed answers.

Regarding the last answer, I understand the point now thanks. I wonder now to what extent this is a novel results and what was already understood. The fact consensus alone is not a sufficient statistic seems now almost obvious to me.

I will maintain my positive recommendation as I find the idea interesting and I see value in the proposed approach; still, I strongly suggest the authors to rework the presentation in the manuscript to make it more accessible.

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

The paper introduces a new approach to analyzing cryo-EM data by finetuning Alphafold3 to recapitulate individual particle images. They test this approach in the low data limit, showing that this approach is able to recover information from cryo-EM images at low particle count. To handle the case with heterogeneity, they introduce a modified version of the RECOVAR algorithm titled B-RECOVAR. They use this algorithm to identify latent coordinates for conformational change, which then are used to label individual images for use in the Fold’EM pipeline.

Critical strengths

The paper introduces multiple new directions. Fine-tuning the specific pairformers layer used in the model seems like an elegant approach to modifying protein structure models. Moreover, the proposed B-Recovar algorithm is interesting, and merits further analysis.

The experiments are well presented and the paper is clearly written, making it easy to understand the content.

Also, Fold’em is an excellent name for a cryo EM and protein folding algorithm.

Critical weaknesses

I have some questions about the experimental design for the paper.

  • I am worried that comparing performance using the FSC is a bit like comparing apples to oranges, since a model built according to E.3.1 will have inherently different smoothness properties compared with a model not assuming an atomistic prior. As such, I think comparing in atomic space by first fitting an atomic model to cryosparc / cryodrgn outputs might be a bit more rigorous. This is a minor point, however.
  • As I understand the “continuous” heterogeneity method approach, the authors first run a continuous heterogeneity scheme, then pick specific points which are then used to group images together and train separate PSM parameters. This is perfectly fine, but it should be clear that the proposed approach is inherently discrete, albeit there is also a separate continuous approach (B RECOVAR) that is proposed.
  • The authors say that B-RECOVAR is proposed for “low-sample heterogeneous data.” However, my read of Figure A4A is that B-RECOVAR actually does worse than regular RECOVAR at small data.
Questions for authors
  • How is accuracy defined in Figure A4, and elsewhere? To me a notion of accuracy implies a ground truth assignment is available on a per-particle basis, but I’m not sure how that could be done.
1 comment

We thank the reviewer for their comments

compare in atomic space by fitting an atomic model to cryosparc maps

We thank the reviewer for this suggestion and agree that comparison between map-to-map FSC and model-to-map FSC may be influenced by AlphaFold3. We now provide an atomic-space comparison across three synthetic and four experimental targets. For each dataset, RELION first reconstructs an ab initio consensus ESP while estimating the particle poses. Because particles are limited, the maps reach only about 10 to 16 Å half-map resolution, with median pose errors of 4° to 13°. We then fit an atomic structure to each map using the same optimization framework as Fold'EM, replacing the particle-level objective with an ESP map-fitting loss. Both use identical particle subsets, initial pose estimates, structural initializations, and 570 backward passes through AlphaFold3 denoiser, ensuring equivalent optimization budgets.

PDB ID# particlesMedian pose error (°)Map resolution (Å)Cα RMSD: Density guidanceCα RMSD: Particle guidance# backward passes: Density guidance# backward passes: Particle guidance
Synthetic
9UGB500 (SNR 0.005)4.39.81.951.57570570
9UGC500 (SNR 0.005)4.111.18.513.07570570
8U4B500 (SNR 0.005)13.216.09.273.94570570
Experimental
9GXQ500011.412.25.122.87570570
9GXR500011.011.72.351.54570570
9B7750008.911.12.902.25570570
9B9J50007.313.622.22.33570570

Particle guidance achieves lower Cα RMSD on all seven targets. This demonstrates that Fold'EM's advantage is not solely due to the structural prior, but also arises from directly incorporating individual particle observations during structural refinement. We use RELION as the reconstruction baseline because, like cryoSPARC, it supports ab initio reconstruction without requiring known poses. Detailed results are presented in Table A44, Section C.10.

Clarify continuous heterogeneity approach

We agree with the reviewer's description of the fitting procedure, but would like to clarify how the continuous structural representation is obtained. We first use the continuous latent representations of particles learned by cryoDRGN, RECOVAR, or B-RECOVAR to select clusters of particles along a conformational trajectory. These clusters are then used to optimize a discrete set of anchor states, each with its own continuous structural representation Hk\mathbf{H}_k. Using the barycentric parameterization H(α)=∑kck(α)Hk\mathbf{H}(\alpha)=\sum_k c_k(\alpha)\mathbf{H}_k (Sec. 2), we interpolate between these anchors in H\mathbf{H}-space to decode atomic structures along a continuous conformational trajectory.

Figure A4A fix

Due to incorrect pre-processing of data, RECOVAR received100% classification accuracy in some cases. This has been fixed and Figure A4A has now been updated. Additional details about the pre-processing are described in Section E2 of the manuscript.

classification accuracy

Classification accuracy is defined as the percentage of particles whose predicted conformational state matches the original depositors' reference assignment. These assignments are usually obtained through cryo-EM processing workflows, such as 3D classification in RELION or heterogeneous refinement in cryoSPARC. The depositor assignment is only used for evaluation and are not available to Fold'EM during inference or classification. Also, for each heterogeneous dataset, we sample an equal number of particles from each conformation, ensuring that differences in class abundance do not bias the reported accuracy. Classification accuracy is computed after matching the predicted class labels to the reference labels.

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

This paper introduced Fold'EM, an inference-time framework that infers atomic models directly from cryo-EM particle images by guiding AlphaFold3. It optimizes intermediate Pairformer representations using a particle-level likelihood and alternates structural inference with pose and class assignment. Experiments on synthetic and experimental data show accurate recovery from far fewer particles than conventional density-based methods.

Critical strengths
  • The direct use of experimental particle observations for guidance could, in theory, be beneficial. The paper shows that a consensus density may discard particle specific information especially when poses or class assignments are uncertain.

  • The paper also provides an informative comparison with IT-Opt by distinguishing optimization of the intermediate Pairformer state HH from optimization of the terminal representation ZZ. This design choice is supported by both a theoretical interpretation and a controlled experimental ablation.

Critical weaknesses
  • There is a reason why conventional cryo-EM first reconstructs a density map and then builds an atomic model into the map. Individual particle images have extremely low SNR, and jointly optimizing their poses and other nuisance parameters under a strong structural prior is highly susceptible to "model bias" or the classical “Einstein-from-noise” problem. This is highlighted by the controversy of the reconstruction of the HIV-1 envelope trimer paper [1-3]. In this sense, Fold'EM is better characterized as AF3 structure prediction conditioned on a limited number of particle images, rather than atomic structure inference directly from the particles. Without a known PDB/EMDB reference, the resulting atomic structure cannot be effectively evaluated. It is impossible for the user to determine whether the results are genuinely supported by the experimental data. For example, given a known sequence and a set of pure noise particles to Fold'EM, the strong AF3 prior could still produce a plausible atomic model, even though the input images contain no structural information at all.

  • The reported FSC for Fold'EM compares a density computed from its atomic model with a deposited EMDB map. This is a map-model FSC, not a conventional gold-standard FSC between independently reconstructed half-maps. Therefore, the 0.5 threshold instead of the 0.143 threshold should be used [4]. Moreover, this metric measures agreement of the atomic model with the reference map rather than the intrinsic resolution or reconstruction quality of Fold'EM itself. Actually, most of the metrics used in the paper measure either the agreement of the atomic model with a reference map/model, or the stereochemical quality of the model itself, rather than how well the reconstructed model is supported by the input particles.

  • The core guidance strategy is closely related to IT-Opt: both use experimental likelihood gradients to modify an AF3-like latent during reverse diffusion. The principal distinction is that Fold'EM updates the intermediate representation HH, whereas IT-Opt updates the terminal latent ZZ. Although the paper provides a theoretical motivation and an ablation supporting this choice, and the broader framework contains other components (e.g. particle-level likelihoods, B-RECOVAR), but the central guidance mechanism remains relatively incremental over prior work.

[1] Mao, Y., Wang, L., Gu, C., Herschhorn, A., Xiang, S.-H., Haim, H., Yang, X., and Sodroski, J. "Molecular architecture of the uncleaved HIV-1 envelope glycoprotein trimer." PNAS. https://doi.org/10.1073/pnas.1307382110.

[2] Henderson, R. "Avoiding the pitfalls of single particle cryo-electron microscopy: Einstein from noise." PNAS. https://doi.org/10.1073/pnas.1314449110.

[3] van Heel, M. "Finding trimeric HIV-1 envelope glycoproteins in random noise." PNAS. https://doi.org/10.1073/pnas.1314353110.

[4] Rosenthal, P. B., and Henderson, R. "Optimal determination of particle orientation, absolute hand, and contrast loss in single-particle electron cryomicroscopy." Journal of Molecular Biology. https://doi.org/10.1016/j.jmb.2003.07.013.

Questions for authors
  • In L370, "let DθD_{\theta} denote its current denoised structure estimate". How to compute this DθD_{\theta}?

  • In L96-97, "Fold'EM jointly infers atomic structures and particle-specific latent variables, including pose and conformational assignments". I suppose this means in Eq.(6) in L380, but the paper does not explain how this is performed specifically. For example, at this stage, how are the particle poses refined?

  • The authors discuss the noise model in Appendix E.3.2, but has this noise model be used in anywhere in the method itself?

6 comments

We thank the reviewer for raising the important concern of model bias. We view this as a philosophical difference in how structural inference from cryo-EM particles should be approached, particularly in relation to classical density reconstruction. Fold’EM does not treat AlphaFold3 (AF3) as a prediction to be accepted simply because it is molecularly plausible. Instead, AF3 serves as a generator of plausible candidate structures, and the experimental particles provide the evidence used to discriminate among and refine those candidates. We agree that molecular plausibility alone does not establish experimental validity. Our aim is to determine how the particle observations constrain the inferred structure, without assuming that a high-resolution consensus map must first be reconstructed.

Our theoretical analysis (Appendix H) further shows that averaging particles into a consensus map can obscure information relevant to structural inference when poses or conformational assignments are uncertain, whereas particle-level fitting can retain that information. We therefore first examine whether the inferred structures are supported by held-out particles and whether refinement depends on using the correct experimental particles. We consider map-based and geometric assessments separately, as complementary approaches to evaluate the level of support available for individual structural features.

Support from held-out particles

Show how well the reconstructed model is supported by the input particles.

For each target, we independently sample five random particle subsets and fit Fold’EM structures using known poses. We evaluate the resulting particle-guidance scores on both training particles and held-out validation particles, reporting the mean and deviation across the five runs. We apply the same evaluation to the deposited PDB structure and the AF3 prediction. Validation particles are excluded from pose estimation, class initialization, reconstruction, and structural guidance.

Held-out particle scores across five independently sampled subsets (higher is better; values are mean ± deviation):

PDBParticlesModelTrain ↑Validation ↑Gap (score)
9B775,000Deposited PDB0.1474 ± 0.00030.1472 ± 0.00070.00025 ± 0.00049
Fold’EM0.1494 ± 0.00150.1488 ± 0.00170.00058 ± 0.00080
AlphaFold30.1263 ± 0.00040.1259 ± 0.00070.00040 ± 0.00052
8DQ05,000Deposited PDB0.1754 ± 0.00060.1752 ± 0.00120.00023 ± 0.00153
Fold’EM0.1751 ± 0.00060.1767 ± 0.0012−0.0016 ± 0.00154
AlphaFold30.1310 ± 0.00030.1310 ± 0.00110.00002 ± 0.00123
9B9J5,000Deposited PDB0.1420 ± 0.00130.1425 ± 0.0014−0.00055 ± 0.00259
Fold’EM0.1424 ± 0.00120.1430 ± 0.0016−0.00053 ± 0.00266
AlphaFold30.0359 ± 0.00120.0363 ± 0.0007−0.00043 ± 0.00158
6RAI5,000Deposited PDB0.0948 ± 0.00170.0948 ± 0.0007−0.00002 ± 0.00129
Fold’EM0.0955 ± 0.00170.0954 ± 0.00090.00013 ± 0.00093
AlphaFold30.0443 ± 0.00090.0440 ± 0.00110.00029 ± 0.00171
6RAH5,000Deposited PDB0.0626 ± 0.00130.0617 ± 0.00090.00094 ± 0.00190
Fold’EM0.0658 ± 0.00050.0641 ± 0.00130.00171 ± 0.00138
AlphaFold30.0289 ± 0.00070.0287 ± 0.00080.00022 ± 0.00120

Across these examples, Fold’EM obtains substantially higher held-out scores than the AF3 predictions and reaches scores comparable to those of the deposited structures. Training–validation gaps are small and of a similar scale to those for the deposited models. These results demonstrate that the refinements generalize to unseen particles, providing evidence that they are driven by experimental observations rather than solely by the AlphaFold3 prior. Results for all evaluated structures are provided in Table A42 of the revised manuscript.

Mismatched-particle negative control

We also test whether the same AF3 structure would be refined toward the correct answer in the absence of relevant experimental evidence. In this control, we deliberately provide unrelated particles to Fold’EM while retaining a strong structural prior (Fig. A9). Over three optimization loops, the correct particles drive refinement toward the reference structure, whereas mismatched particles cause the structure to drift away from it. These contrasting outcomes demonstrate that the direction of structural refinement is determined by the experimental particle evidence, rather than solely by the AlphaFold3 prior. The accompanying Q-score analysis provides complementary evidence of agreement with the experimental density.

Structural inference, map support, and validation limits

We agree that structure validation is essential and that the original manuscript did not sufficiently address the question of validating our output structures. We would like to distinguish between two problems: (i) inferring a plausible atomic structure that agrees with the particle observations and approaches the deposited reference; and (ii) determining which features of that structure can be independently validated from the available experimental signal. Our results suggest that the first can be achieved with fewer particles than are needed to resolve and independently validate every atomic detail through conventional reconstruction-based analyses.

The experimental particles and their alignments can be backprojected into a density map for assessing model-map agreement and visual inspection. We use Q-score, which measures agreement between atomic coordinates and map density, as one such measure (Fig. A7). The observed Q-scores are consistent with the apparent resolution of the corresponding backprojected maps (Fig. A6 + below table). When the backprojected density map has near-atomic resolution this metric can be used as an independent validation of the atomic structure.

Geometric quality and map agreement:

SystemParticlesClashscoreRamachandran outliers (%)Rotamer outliers (%)Q-score
UvrA / 9GXQ (PDB)195,2333.8440.0003.9160.6915
UvrA / 9GXQ (Fold’EM)100,0007.0740.2381.9080.5236
9C91 (PDB)185,00010.9220.1852.6490.6484
9C91 (Fold’EM)50,0009.3290.0001.9870.4805

At sufficient particle counts, Fold’EM inferred structures show geometric quality comparable to that of the deposited models across clashscore, Ramachandran outliers, and rotamer outliers, with differences across individual measures (Table above; Tables A7, A8, A13, A14, and A24). These measures assess stereochemical plausibility, but do not by themselves demonstrate that a model is supported by the particle observations. Likewise, the lower Q-scores of the Fold’EM models relative to the deposited models reflect weaker map-level support from smaller particle sets.

We acknowledge that small particle subsets do not allow us to independently validate every atomic detail. In that regime, backprojected maps have limited resolution, and conclusions about individual side chains or local atomic features must remain correspondingly cautious. Fig. A8 illustrates this distinction: Fold’EM can infer a structure close to the deposited reference at a particle count for which classical reconstruction-based measures do not yet provide equally detailed independent validation. Motor/domain placement or other larger-scale features may still receive map-level support even where local atomic validation remains incomplete.

Half-map FSC could in principle be evaluated using separate particle half-sets and final alignments. The present implementation does not retain separate half-sets, because this is not required by the structural-inference procedure. Accordingly, we do not use half-map FSC to claim an independently established resolution for these subsets.

FSC thresholds and interpretation

0.5 threshold instead of the 0.143 threshold should be used.

We now report FSC = 0.143 and FSC = 0.5 thresholds across all original benchmarks (Section C7), and use FSC = 0.5 for the newly added experimental entries (Section C8). All FSC tables and captions explicitly distinguish model-to-reference-map FSC from map-to-map FSC. We interpret these thresholds as measures of model-to-reference-map agreement and distinguish them from independent estimates of the resolution supported by a small particle subset.

Definition of DθD_{\theta}

What is DθD_{\theta}?

DθD_{\theta} denotes AF3’s learned denoising function, parameterized by θ\theta, which estimates the underlying clean structure from its noisy counterpart at a given diffusion timestep.

Noise model

The Gaussian frequency-dependent noise model supplies the probabilistic formulation and the fixed-operator particle/density identity. The implementation described in the current manuscript uses band-limited normalized Fourier correlation as its experimental guidance objective, rather than optimizing the Gaussian log-likelihood literally. Accordingly, the reported empirical image scores should not be read as calibrated log-likelihoods or probabilities. The numerical objective and its normalization are given in the image-formation appendix.

Particle-pose refinement

How are the particle poses refined?

Particle poses are refined using the framework in Eq. (11). Specifically, we use a coarse-to-fine search on a HEALPix grid, which partitions the sphere of viewing directions into equal-area cells. Candidate rotations are initially evaluated on a coarse grid with HEALPix nside = 8 and 48 in-plane angles, retaining the five highest-scoring rotations for each particle. During refinement, these candidates are subdivided into eight finer rotations (four viewing directions and two in-plane angles), evaluated, and again reduced to the five highest-scoring candidates. The highest-scoring candidate is then selected as the refined orientation. This hierarchical search achieves fine angular resolution without exhaustively evaluating the full high-resolution rotation grid. We observe the following reductions in median pose error:

PDB IDInitial median pose error (°)Final median pose error (°)Improvement (°)
6RAH42.827.315.5
6RAI26.514.811.7
9GXQ17.110.66.5
29XL13.78.75.0
9GXR9.26.52.7
Reviewer_V9RmOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

Fold'EM performs inference-time structural optimization using cryo-EM particle images and a pretrained AlphaFold3-style model, implemented with Protenix. Rather than guiding solely against a reconstructed map, it optimizes an intermediate Pairformer state using a particle likelihood and alternates structural inference with particle pose/class refinement. B-RECOVAR supplies low-sample heterogeneity estimates and initial assignments. The paper derives the experiment-induced density metric that would reproduce particle fitting with known poses, and interprets intermediate-state optimization as a local learned preconditioner. Experiments include synthetic homogeneous targets, curated experimental EMPIAR datasets, discrete UvrA–DNA states, and a synthetic continuous IgG trajectory. Several cases improve substantially over unguided predictions using small particle subsets.

Critical strengths
  • Introducing structural priors directly into particle-level inference is a consequential contribution. The method addresses the low-sample regime in which reconstruction alone is difficult. Improvements on examples where the unguided model has the wrong arrangement, such as the integrin–Fab binding site and Pma1 domain organization, are more informative than cases where the prior already matches the target (Figures 1–2). The heterogeneous UvrA example also illustrates experimentally driven corrections beyond an unguided prediction (Figure 3).
  • The paper combines a principled formulation with substantial methodological comparisons. The experiment-induced metric makes clear why ordinary Euclidean map guidance does not reproduce particle fitting. Intermediate-state guidance has an interpretable local geometry and is compared with terminal-state and coordinate guidance (Figure A1; Appendix F.1). The supplementary material includes stereochemical assessment, known-pose and ab initio results, and details of the heterogeneity estimator, making this more than a single demonstration of differentiable image fitting.
Critical weaknesses
  1. The number and diversity of independent experimental cases are insufficient to establish broad generalizability.

Table A1 contains only five PDB entries from four EMPIAR datasets, with 9GXR and 9GXQ representing two conformations of the same UvrA system. Sweeping particle counts or evaluating multiple conformations of the same target does not increase the number of independent biological systems tested. The synthetic benchmarks provide useful controlled evidence, but cannot substitute for diverse experimental targets with realistic noise, compositional heterogeneity, and processing artifacts. The results therefore support a promising proof of concept, but do not yet establish that the method generalizes reliably across molecular sizes, architectures, prior quality, and heterogeneous mixtures. A broader experimental benchmark, with success/failure rates and variability across independently sampled subsets, is needed to support the generality claims.

  1. The method has important dependencies and failure modes that limit its scope as a general atomic structure inference pipeline.

First, it requires curated, centered particles with known CTF parameters, and its ab initio workflow relies on coarse consensus/class reconstructions and pose/frame initialization (Section 3.1; Appendix E.1). Errors in these estimates may compromise subsequent alternating structure/pose/class optimization; robustness to such errors is not sufficiently characterized. The deposited particle count also does not establish the minimum count needed by an optimized alternative. Second, guidance operates through a pretrained structural model, so recovery when the relevant structure is poorly represented by the prior remains insufficiently tested. The UvrA results illustrate the practical consequence: for the harder 9GXQ state, unknown-pose/class inference requires approximately 10,000 particles per state to reach a model-to-reference-map FSC crossing of 4.63 Å (Table A17), which qualifies the broad few-hundred-particle message. Third, components absent from the input/model, including glycans, ligands, and flexible regions, remain unexplained signal, while fine interface-loop geometry is harder to recover than global assembly (Section 5). Thus, improved global agreement does not establish complete local atomic accuracy. The discrete heterogeneity workflow also takes the class count as input, leaving discovery of an unknown number of states and rare-state recovery at unknown abundances unestablished.

These dependencies make independent validation of particle-supported detail particularly important. Appendix E.2.4 reports FSC between a model-derived map and the deposited experimental map at 0.143, rather than a gold-standard FSC between independently reconstructed half maps. A strong structural prior can supply high-frequency detail that the small fitted subset does not independently determine. Deposited-model RMSD/TM-score provide useful evidence of accuracy, but do not establish that a few hundred particles experimentally resolve all atomic detail at the quoted Å value. Held-out-particle likelihood or independent half-set tests are needed to separate prior-supported agreement from experimentally supported detail.

Questions for authors
  1. Can you evaluate structures on particle images excluded from guidance and all pose/class initialization, and report held-out likelihood or an equivalent predictive score? Please explicitly distinguish model-to-reference-map FSC from independent half-map resolution and report whether local/interface agreement follows the particle evidence.

  2. Which exact Protenix weights, structure-training cutoff, MSA/template settings, and template exclusions were used for each target? Can you show reproducibility over independent random particle subsets and pose seeds, particularly where the unguided prior is incorrect?

  3. Can you compare against a reconstruction-plus-the-same-structural-prior pipeline using identical particle subsets, initialization information, and comparable optimization budgets across more than the single map-guidance example in Section 4? Please include total inference time and clarify how the reported fractions change when coarse reconstruction, particle curation, and class initialization are counted.

  4. Can you expand beyond the four experimental datasets and report success/failure rates across targets with different sizes, architectures, and unguided-prior accuracy? On representative cases, please quantify sensitivity to pose initialization and an incorrect class count, and distinguish recovery of global conformation from accurate interface loops or other local atomic details. These tests would help identify when the method can be expected to work and when it fails.

3 comments

We thank the reviewer for their comments.

New examples and reproducibility over independent random particle subsets, poses, and class

Table A1 has been expanded with six additional datasets - three discrete heterogeneity datasets and three homogeneous datasets.

Reproducibility with known poses. For each target, we evaluate Fold'EM across five independently sampled particle subsets using known poses. Fold'EM consistently improves the AF3 predictions in Cα RMSD and TM-score, with small variations across subsets. This shows that the improvements are reproducible across independent particle samples. Detailed results, additional metrics and benchmarks, are in Tables A34-A35, Sec. C8. We used 10k particles for 29xk and 5k for others.

EMPIARPDB IDRMSD ↓ (Fold'EM)RMSD ↓ (AF3)TM-score ↑ (Fold'EM)TM-score ↑ (AF3)
112478dq00.97 ± 0.034.870.994 ± 0.00020.878
104386zbh1.54 ± 0.236.560.986 ± 0.0030.883
104386zbe1.32 ± 0.064.540.989 ± 0.00060.929
103176rai1.31 ± 0.044.060.988 ± 0.0020.856
103176rah1.89 ± 0.185.20.981 ± 0.0080.845
133949gxm1.50 ± 0.267.960.989 ± 0.0020.8252
133949gxn2.10 ± 0.135.330.9932 ± 0.0090.839
133949gxo0.84 ± 0.022.550.9955 ± 0.00020.960
121809c910.63 ± 0.031.10.9967 ± 0.00030.977
1350729xk1.04 ± 0.031.840.9904 ± 0.00050.956

Reproducibility with pose estimation. We extend this to the ab-initio setting, using the same number of particles per run as above. For each subset, initial poses are estimated using RELION and refined in Fold'EM. We again improve upon AF3 predictions across all structures, demonstrating consistent improvements despite changes in particle sampling and pose estimates. But, performance is lower compared to known-pose setting. Detailed results are in Tables A36-A37, Sec. C8.

EMPIARPDBInit pose errorRMSD↓(Fold'EM)RMSD↓(AF3)TM↑(Fold'EM)TM↑(AF3)
112478dq03.4 ± 0.21.25 ± 0.144.870.989 ± 0.0030.878
104386zbh4.2 ± 0.21.75 ± 0.526.560.985 ± 0.0020.883
104386zbe4.3 ± 0.31.66 ± 0.374.540.972 ± 0.0120.929
103176rai15.8 ± 0.51.76 ± 0.324.060.979 ± 0.0070.856
103176rah12.2 ± 3.02.15 ± 0.165.20.970 ± 0.0040.845
133949gxm4.9 ± 0.52.11 ± 0.527.960.980 ± 0.0060.825
133949gxn8.4 ± 2.54.14 ± 0.235.330.895 ± 0.0100.839
133949gxo5.3 ± 0.71.75 ± 0.672.550.965 ± 0.0480.960
121809c914.2 ± 0.21.02 ± 0.071.10.992 ± 0.00080.977
1350729xk10.8 ± 0.61.29 ± 0.071.840.985 ± 0.0010.956

Reproducibility with class. We extend this to heterogeneous experimental datasets with known particle poses, using five independently sampled particle subsets per dataset.

EMPIARPDB IDFold'EM accuracy ↑cryoDRGN accuracy ↑RELION accuracy ↑Fold'EM RMSD ↓Fold'EM TM-score ↑AF3 RMSD ↓AF3 TM-score ↑
104386zbe96.8 ± 0.286.5 ± 2.166.2 ± 21.31.29 ± 0.150.984 ± 0.0044.540.929
104386zbh96.8 ± 0.286.5 ± 2.166.2 ± 21.31.42 ± 0.050.975 ± 0.0066.560.883
103176rai93.1 ± 0.382.3 ± 0.988.6 ± 2.61.51 ± 0.120.981 ± 0.0014.060.856
103176rah93.1 ± 0.382.3 ± 0.988.6 ± 2.61.9 ± 0.250.964 ± 0.0055.20.8448
133969gxq92.4 ± 0.285.4 ± 2.779.8 ± 22.81.92 ± 0.130.976 ± 0.0035.960.734
133969gxr92.4 ± 0.285.4 ± 2.779.8 ± 22.81.68 ± 0.180.980 ± 0.0042.010.886
133949gxm99.4 ± 0.197.9 ± 0.598.6 ± 0.10.87 ± 0.050.982 ± 0.0077.960.8252
133949gxn99.4 ± 0.197.9 ± 0.598.6 ± 0.15.62 ± 0.150.845 ± 0.025.330.8388
133949gxo99.4 ± 0.197.9 ± 0.598.6 ± 0.11.15 ± 0.140.975 ± 0.0112.550.9604

Across the datasets, we achieve higher classification accuracy than cryoDRGN and RELION, with lower spread across the runs. It also improves the AF3 predictions for most states. One exception is 9GXN, where Fold'EM does not improve upon the initial prediction. Lastly, we do not assume the number of classes a priori. We select the optimal number of classes based on a silhouette score, as illustrated in Tab. A23. We extend this evaluation to the ab-initio setting, where particle poses are estimated rather than provided. Detailed results are presented in Tab. A38-A41 (Sec. C8).

Global recovery against local accuracy

We agree that accurate global structural recovery does not necessarily imply accurate local geometry. To investigate this, we revisited three experimental targets from Table A1 where AlphaFold3 incorrectly predicts ligand placement or interface loop geometry. We repeated the inference with additional molecular components, including glycans, ligands, ions, and other small molecules where present.

Metric9b9j (AF3)9b9j (Fold'EM)9c91 (AF3)9c91 (Fold'EM)9gxq (AF3)9gxq (Fold'EM)
lDDT ↑0.8820.9070.89330.89710.8180.903
Loop lDDT ↑0.8580.8940.8780.8790.8050.892
Interface lDDT ↑0.5380.8320.89230.91130.5660.873
Interface DockQ ↑0.4860.8310.9090.9360.5790.896
Fnat ↑0.5060.7690.8690.8850.7030.884
iRMSD ↓2.620.890.460.412.720.72
Ligand RMSD ↓5.291.440.880.791.130.63

Fold'EM improves upon the AlphaFold3 prediction across all evaluated metrics. In particular, we observe substantial improvements in interface geometry and ligand placement, while improvements in local loop accuracy are generally more modest. These results suggest that particle guidance can improve both global conformational accuracy and local structural features, although recovering fine loop geometry remains more challenging than correcting overall assembly and interface arrangements. Detailed results are presented in Table A43 of the revised manuscript.

compare against a reconstruction-plus-the-same-structural-prior pipeline.

We thank the reviewer for this suggestion. We have extended the comparison between particle guidance and ESP guidance to three synthetic and four experimental targets, evaluating both directly in atomic space. For each dataset, RELION first reconstructs an ab initio consensus ESP map while simultaneously estimating poses for the subset of particles that was used to construct the ESP. Due to limited particles, the reconstructed maps achieves half-map resolutions of approximately 10–16 Å, with median pose errors of 4°–13°. We then fit an atomic structure to each reconstructed map using the same optimization framework as Fold'EM, replacing the particle objective with an ESP map-fitting objective. Both approaches use identical particle subsets, RELION-estimated poses, structural initializations, and 570 backward passes through the AlphaFold3 denoiser, ensuring matched structural optimization budgets.

SetPDB ID# particlesMedian pose errorMap resolutionRMSD (unguided)RMSD (ESP guidance)RMSD (particle guidance)# backward passes (ESP)# backward passes (particle)Total time (s, ESP)Total time (s, particle)
Synthetic9ugb500, SNR 0.0054.39.88.961.951.575705702929.82912.7
Synthetic9ugc500, SNR 0.0054.111.111.228.513.075705702975.42935.5
Synthetic8u4b500, SNR 0.00513.2167.789.273.94157057010453.810185.9
Real9gxq500011.412.25.965.122.875705705945.15916.6
Real9gxr50001111.72.012.351.545705706104.75979.3
Real9b7750008.911.13.872.92.2557057054154651.2
Real9b9j50007.313.621.922.22.335705704189.53562.5

Particle guidance achieves lower Cα RMSD on all seven targets, with comparable or shorter inference times (Table A44, Section C.9). These results demonstrate that Fold'EM's advantage is not solely attributable to the structural prior, but also arises from directly incorporating individual particle observations during structural refinement. All reported particle counts include the particles used for coarse reconstruction, pose estimation, and guidance. No additional particles are used at any stage of the pipeline.

explicitly distinguish model-to-reference-map FSC from independent half-map resolution

We agree with the reviewer and have revised all FSC tables in the manuscript to explicitly distinguish between model-to-map FSC and map-to-map FSC. We no longer refer to the reported values as the resolution, but instead as the spatial resolution (in Å) where FSC crosses a specific correlation threshold. We now report both FSC=0.143 and FSC=0.5 thresholds across all benchmarks (Section C.7) [1]. In Section E.2.4, we explicitly clarify that these measurements quantify agreement with the deposited map and should not be interpreted as independent gold-standard half-map resolution.

References: [1] Rosenthal, P. B. et al. "Optimal determination of particle orientation, absolute hand, and contrast loss in single-particle electron cryomicroscopy."

Held out particle likelihood

For each target, we fit Fold'EM on five independent random particle subsets with known poses. Then, we evaluated the training and validation particle likelihood for each structure, over the five independent subset. The validation particles were excluded from all components of the Fold'EM pipeline.

Held-out particle scores across five independently sampled subsets (higher is better; values are mean ± deviation):

PDBParticlesModelTrain ↑Validation ↑Gap (score)
9B775,000Deposited PDB0.1474 ± 0.00030.1472 ± 0.00070.00025 ± 0.00049
Fold’EM0.1494 ± 0.00150.1488 ± 0.00170.00058 ± 0.00080
AlphaFold30.1263 ± 0.00040.1259 ± 0.00070.00040 ± 0.00052
8DQ05,000Deposited PDB0.1754 ± 0.00060.1752 ± 0.00120.00023 ± 0.00153
Fold’EM0.1751 ± 0.00060.1767 ± 0.0012−0.0016 ± 0.00154
AlphaFold30.1310 ± 0.00030.1310 ± 0.00110.00002 ± 0.00123
9B9J5,000Deposited PDB0.1420 ± 0.00130.1425 ± 0.0014−0.00055 ± 0.00259
Fold’EM0.1424 ± 0.00120.1430 ± 0.0016−0.00053 ± 0.00266
AlphaFold30.0359 ± 0.00120.0363 ± 0.0007−0.00043 ± 0.00158
6RAI5,000Deposited PDB0.0948 ± 0.00170.0948 ± 0.0007−0.00002 ± 0.00129
Fold’EM0.0955 ± 0.00170.0954 ± 0.00090.00013 ± 0.00093
AlphaFold30.0443 ± 0.00090.0440 ± 0.00110.00029 ± 0.00171
6RAH5,000Deposited PDB0.0626 ± 0.00130.0617 ± 0.00090.00094 ± 0.00190
Fold’EM0.0658 ± 0.00050.0641 ± 0.00130.00171 ± 0.00138
AlphaFold30.0289 ± 0.00070.0287 ± 0.00080.00022 ± 0.00120

Fold'EM consistently achieves higher validation likelihood than AlphaFold3 and approaches those of the PDB structures. This shows that over five random subsets, Fold'EM consistently closes the gap between the AF3 prediction and the deposited structure without overfitting to the training particles. For sake of brevity, we present five representative examples here. The complete and detailed results are reported in Table A42 of the revised manuscript.

Model weight details

We used the Protenix v1.0.0 base model (protenix_base_default_v1.0.0.pt), with a structure-training cutoff of Sept 30, 2021. For all targets, MSAs were generated using ColabFold, and no structural templates were provided during inference. Hence, no target-specific template exclusions were necessary.

recovery when the relevant structure is poorly represented

One of the main examples in Fig. 1 and 2 addresses cases where AF3 is poorly represented. For 9B9J, AF3 correctly predicts the individual components but incorrectly docks the Fab. Fold'EM recovers the correct binding arrangement, demonstrating that the framework can recover structures where AF3 is poorly represented. Additional cases are shown in Table A34 & A35.

Continuous heterogeneity

We further evaluate Fold’EM on the ALC1–nucleosome complex (EMPIAR-10739), whose ATPase motor exhibits continuous conformational heterogeneity (Figure A10). To characterize this heterogeneity without manually selecting states, we order particles along the first principal component of an ALC1-focused \brecovar{} embedding and reconstruct 50 overlapping sets of 5,000 particles. We show that these reconstructions closely follow the previously reported cryoDRGN trajectory, constructed using 11 manually selected cluster centres (Figure A10B–D). The best-matching cryoDRGN frames progress consistently along the principal component (Pearson r=0.92r = 0.92).

We then use 10 particle sets spanning this component to guide Fold’EM and show that the inferred structures capture the progressive docking of the motor onto nucleosomal DNA (Figure A10A), accompanied by closure of the ATP-binding cleft. Along this trajectory, the motor–DNA distance decreases from 8.8 to 7.7 Å, while the ATP-binding cleft narrows from 15.1 to approximately 13 Å, approaching the deposited closed-state structures (12.3–13.1 Å). We measure these distances without structural superposition and obtain consistent results across two independent Fold’EM runs per particle set. Although the inferred open-end structure remains only partially open relative to deposited open-state references, these results demonstrate that Fold’EM can recover continuous conformational changes from experimental particles ordered along a learned coordinate, without manually selecting discrete states.

Comments on the paper

1 comment
Area_Chair_3K2ROpenReview ↗

discussion and feedback on rebuttal

Dear Reviewers,

Thank you for your efforts so far. The authors have responded to your reviews with detailed feedbacks. Please check them carefully (you may also check other reviewers' comments) and provide your feedbacks.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.