{"id":"a550421a-53e6-4566-8614-0713ddca5cda","arxiv_id":"2411.13263","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A cryo-EM analysis method computes the 'softest modes' of the likelihood Hessian, revealing which map deformations are least constrained by the images.","lead":"This paper derives a way to estimate the softest (least constrained) deformation modes of a cryo-EM reconstruction by computing the smallest eigenvalues of the Hessian of the log-likelihood. The authors show on synthetic and real datasets that these modes reveal which parts of a reconstructed molecule are uncertain, depending on viewing-angle coverage and defocus.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The computed soft modes are soft modes of a low-temperature surrogate objective; whether they match the exact marginalized likelihood at realistic SNR is never tested, so the central uncertainty claim is not yet established.","rationale":"Good-faith reading: the 3D-integral rewrite and the multipole Hessian-vector products are internally consistent, and the derivation is plausible within the stated low-noise regime; the Lanczos convergence example in Fig. 23 provides some independent support. The concern is not an internal algebraic contradiction but a validation gap in the domain of validity: the abstract promises that the interpretation holds in 'many practical scenarios', yet only one empirical affine fit supports the noise-marginalized extension, and no experiment compares the approximate Hessian with the exact marginalized likelihood at realistic SNR. This is the same weakest assumption the reader identified, so I agree. The conditional verdict remains appropriate: the method is promising, but its central uncertainty quantifications need a direct numerical check before being applied to real reconstructions.","tokens_in":50284,"tokens_out":6943,"duration_ms":81301,"concrete_test":"At low resolution (e.g., 2*pi*K = 8 or 16) with J = 512 images from EMPIAR-10005 and their empirical CTFs and noise level, compute the exact Hessian of -log P({A_j} | F) from Eq. 75 by discretizing the alignment integral on a dense SO(3) grid and using numerical or automatic differentiation of the log-sum-exp. Compare the five smallest eigenvalues and eigenvectors with those obtained from the low-temperature H (Eqs. 77, 102, with Section 21 corrections) and from the affine noise-marginalized version (Eq. 176 with Fig. 26 coefficients). If the relative eigenvalue differences exceed 20% or the subspace overlap falls below 0.9 at the empirical SNR, the approximate soft modes are not the exact soft modes, and the uncertainty interpretation would need to be restricted to the low-temperature limit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the smallest eigenvectors of the Hessian computed from the 3D-integral representation are the 'softest modes' of the actual cryo-EM log-likelihood. The derivation in Sections 13 and 20 computes H from the low-temperature limit (Eq. 77), replacing the alignment integral in Eqs. 47/75 by a saddle point at tau_opt(F). This is exact only as noise sigma_hat tends to zero. The case studies use empirical noise, CTFs, and finite image numbers, where the exact Hessian contains additional second-order contributions from the width of the alignment posterior. Section 23 attempts a noise-marginalized version, but Eqs. 168-169 again use a large-dof saddle point over alignments, and then approximate log(l) by an affine function fit to one FIB-milled lamellae dataset (Fig. 26, a approximately 0.537, b approximately -0.378). No comparison is made between the approximate H and the exact Hessian of log P({A_j} | F) (Eq. 75) at realistic SNR. If those differ, the reported soft modes and their use for identifying trustworthy or untrustworthy regions are properties of a different objective, not of the reconstruction problem the paper claims to analyze. Since Section 19 concedes that the exact smallest-eigenvalue problem has no known efficient solution, this approximation is not a minor technicality; it is the basis of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a computational strategy for estimating the smallest eigenvalues and associated eigenvectors (the 'softest modes') of the Hessian of the log-likelihood in ab-initio single-particle cryo-EM reconstruction. The central device is a rewriting of the low-temperature negative log-likelihood as a three-dimensional volumetric integral (Eqs. 54, 73, 77), which makes Hessian-vector products tractable and permits iterative Lanczos estimation of the spectral tail. The authors argue that the softest modes quantify which volumetric deformations and image-alignment perturbations are most confusable with a given reconstruction, and they use this to identify trustworthy and untrustworthy regions of maps. The main text presents case studies on TRPV1 (EMPIAR-10005), on synthetic viewing-angle distributions, on defocus variation, and on free-energy paths; the appendix derives the low-temperature and noise-marginalized representations, the multipole treatment of implicit alignments, and several ill-posedness examples.","tokens_in":50568,"tokens_out":3237,"duration_ms":43342,"significance":"If the central approximation is valid at realistic noise levels, this is a useful and potentially important contribution to cryo-EM uncertainty quantification. The derivation from the Gaussian likelihood to the volumetric form is self-contained, elegant, and largely parameter-free in its low-temperature version; the Hessian-vector-product formulation and the Lanczos procedure are natural and computationally plausible. The synthetic-image case studies are well controlled and illustrate that the method can recover intuitive effects of viewing-angle coverage, defocus, and noise. The paper also connects the approach to free-energy estimation and multi-particle ill-posedness, broadening its potential impact. However, the method's practical value depends on whether the computed soft modes describe the Hessian of the actual cryo-EM log-likelihood at realistic signal-to-noise ratios, and this is not tested in the manuscript. The noise-marginalized generalization also relies on an empirical affine approximation fitted to a single dataset, which needs independent validation.","major_comments":[{"comment":"The Hessian whose soft modes are estimated is the Hessian of the low-temperature surrogate objective, not of the exact Gaussian posterior marginalized over alignments. Equation (54) and the subsequent volumetric forms (73), (77) are obtained by replacing the alignment integral in Eq. (75) with a saddle point at the optimal alignments; at realistic noise the exact Hessian of Eq. (75) contains additional second-order contributions arising from the width of the alignment posterior. Section 19 explicitly states that the exact smallest-eigenvalue problem has no known efficient solution, so this approximation is not a minor technicality. The manuscript uses empirical noise levels, CTFs, and finite image numbers in the case studies, but provides no comparison between the approximate Hessian and the exact Hessian of Eq. (75) (or the noise-marginalized version) at those noise levels. I recommend adding a small-scale numerical comparison—for example, a system with few images and low resolution where the exact Hessian can be formed by brute-force integration over a discretized alignment grid—and reporting the resulting eigenvalue differences. Without such a check, the claimed link between the computed soft modes and the actual uncertainty of the reconstruction is not established.","section":"Sections 13, 20; Eqs. (54), (73), (75), (77)"},{"comment":"The noise-marginalized version of the method rests on an empirical affine approximation log(ℓ) ≈ aℓ + b, with a ≈ 0.537 and b ≈ −0.378, fitted to a single FIB-milled lamellae dataset. The text claims these constants are 'fixed and independent of the image-index j', but Figure 26 shows variations of the same underlying micrograph set under different added noise levels; it does not show independent image pools, different specimens, or different imaging conditions. Since Eq. (176) and the subsequent volumetric representation depend on this approximation being universal, the manuscript needs either a derivation of the affine relation from the noise-marginalized integral or a demonstration across multiple independent datasets, together with a sensitivity analysis of the resulting soft modes to plausible variations in a and b. Without this, the noise-marginalized scenario—which the abstract identifies as a practical route for realistic data—remains unvalidated.","section":"Section 23, Fig. 26, Eq. (176)"},{"comment":"The case studies validate internal consistency rather than predictive accuracy. The synthetic images in Section 3.1 are generated from the same volume that is later treated as the optimal reconstruction, and the text explicitly assumes this volume is a global optimum. Real ab-initio reconstructions often converge to local minima or contain systematic errors, and the Hessian and its soft modes will generally be evaluated at such imperfect reconstructions. The manuscript does not test whether the softest-mode structure is stable with respect to moderate reconstruction error or whether it correctly predicts localization of errors when the reconstruction is not the truth. A numerical experiment where the input volume is perturbed away from the global optimum, or where a known suboptimal local minimum is used, would strengthen the practical relevance of the trustworthiness interpretation.","section":"Section 3.1 and Section 18"}],"minor_comments":[{"comment":"In the first paragraph of Section 3.1, 'quallity' should be 'quality'.","section":"Section 3.1"},{"comment":"In the sentence introducing the 3-dimensional representation, 'emphazised' should be 'emphasized'; similarly, 'trival' in Section 20.1 should be 'trivial'.","section":"Section 20"},{"comment":"In the membrane-protein example, 'lammelae' should be 'lamellae'.","section":"Discussion"},{"comment":"The figure shows Rayleigh quotients for the softest seven modes at three defocus values, but no error bars, iteration counts, or convergence criteria are reported; a brief description of how many Lanczos iterations were used and how convergence was assessed would be helpful.","section":"Fig. 17"},{"comment":"After offset marginalization, the text states images and templates can be assumed centered; it would be clearer to state explicitly whether the Hessian calculation is performed on the centered data and whether this changes the definition of M11 in Eq. (80).","section":"Section 23.1, Eq. (174)"}],"recommendation":"major_revision","confidential_remarks":"I see this as a methods paper with a clean central derivation and potentially high practical value for cryo-EM uncertainty analysis. The main gap is validation of the low-temperature and noise-marginalized surrogates against the exact likelihood at realistic noise levels; this is fixable with small-scale numerical experiments. I would also encourage the authors to provide code or pseudocode for the Hessian-vector product and Lanczos procedure, since reproducibility will be important for the community to adopt the method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is concrete: a way to write the cryo-EM log-likelihood as a 3D volumetric integral (Eq. 77) so that Hessian-vector products become cheap, and then to use Lanczos iteration to pull out the low end of the spectrum. That combination is not in the cited literature, and it is a genuinely useful tool for local uncertainty analysis. The derivation is careful and mostly self-contained; the case studies show the expected physics (polar-concentrated views make the poles soft, defocus changes the soft subspace, and so on). The ill-posedness examples in the appendix are also good, and the paper is honest about its own limits, including the explicit statement in Section 19 that no efficient general solution to the exact smallest-eigenvalue problem is known.\n\nThe soft spot is the one the stress-test flags: the Hessian whose tails you compute is the Hessian of a low-temperature surrogate objective, not the exact Gaussian posterior, and the paper never checks how much those differ at realistic SNR. The noise-marginalized version in Section 23 replaces the exact marginal with an affine fit to one lamellae dataset (Fig. 26). That is a thin basis for the central claim that the computed soft modes describe the true reconstruction uncertainty. The quadratic approximation is also extrapolated to finite perturbations without checking higher-order terms. These are validation gaps rather than obvious errors; the derivation itself holds together. But the paper as written would benefit from a direct comparison of the approximate and exact Hessians on a small, computationally tractable case, or at least a sensitivity check across noise levels. No code is provided, and there is no quantitative comparison against FSC or bootstrap, which the authors themselves position as complementary.\n\nWho is this for? People building cryo-EM methods and possibly structural biologists who want a principled way to think about where a map is locally reliable. A serious referee should engage with it, but the referee should push hard on the validation question, because the uncertainty claim is only as good as the surrogate. I would not desk-reject this, but I would send it back for real work on that front.","headline":"Cryo-EM uncertainty via Hessian soft modes; the derivation is clean and the idea is useful, but the gap between the surrogate likelihood and the real one at empirical noise levels keeps the central claim provisional.","tokens_in":51140,"tokens_out":995,"would_cite":true,"duration_ms":15164,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65F15","65R30","92C55"],"pacs":[],"model":"deepseek-v4-flash","headline":"A cryo-EM reconstruction can now say where it is uncertain: in the low-noise limit its log-likelihood collapses to a single 3D integral, and the smallest eigenvalues of the resulting second-derivative matrix mark the deformations the data…","keywords":["cryo-electron microscopy","single-particle reconstruction","Hessian spectrum","softest modes","log-likelihood","Lanczos iteration","uncertainty quantification","viewing-angle distribution"],"falsifier":"On a small single-particle problem where the exact Gaussian likelihood can still be evaluated pixel-by-pixel, compute the posterior covariance by direct numerical integration or by extensive sampling at a realistic signal-to-noise ratio, and compare its leading eigenvectors with the soft modes obtained from the 3D-integral Hessian. A large subspace angle between the two, or a mismatch between the inverse Hessian eigenvalues and the sampled variances along those directions, would show that the low-temperature approximation breaks down precisely where uncertainty quantification is needed.","tokens_in":50040,"feed_emoji":"🔬","tokens_out":17872,"duration_ms":170762,"temperature":0.7,"pith_summary":"At stake is whether a cryo-EM reconstruction can tell you where it is wrong. This paper argues that it can, locally and cheaply: when the image noise is small, or after marginalizing over the noise level, the log-likelihood of an image stack against a reconstructed volume collapses from a sum of two-dimensional projection integrals into a single three-dimensional integral over frequency space. With that form, the Hessian (second-derivative matrix) of the log-likelihood can be applied to any perturbation quickly enough for Lanczos iteration, yielding the smallest eigenvalues and their eigenvectors—the 'softest modes' of the reconstruction problem. These modes encode the volumetric deformations and the accompanying per-image alignment shifts that the data most readily confuses with the current map, so soft-mode amplitude marks which regions of a map are trustworthy or unreliable, and the associated eigenvalues give a local measure of Fisher information. If the claim holds, cryo-EM sensitivity analysis gains a quantitative complement to Fourier-shell correlation, with consequences for data collection, validation, and free-energy estimation.","feed_headline":"One integral finds the shaky regions in a cryo-EM map","feed_subtitle":"The Hessian's softest modes, which flag unreliable map regions, become computationally accessible.","key_machinery":"The central object is the Hessian $H$ of the scaled negative-log-likelihood $L$ in the reduced model where each image's optimal viewing angle is implicitly determined by the volume. The key move is to hold two representations of $L$ side by side: the template-wise 2D sum, which makes per-image alignment derivatives easy, and the volumetric 3D integral, which makes volume derivatives diagonal in Fourier space. The moments $M_{20} = \\sum_j |\\mathrm{CTF}_j(k)|^2$, $M_{11} = \\sum_j \\mathrm{CTF}_j(k)\\hat T(-\\delta_j,k)\\hat A_j(k)$, and $M_{02} = \\sum_j |\\hat A_j(k)|^2$ over the back-propagated shells carry all of the data; $H\\cdot \\Delta\\hat F$ is built from the blocks $H_{FF}, H_{F\\tau}, H_{\\tau F}, H_{\\tau\\tau}$ with the implicit alignment response $\\Delta \\tau_j$ computed to first or second order from Euler-angle expansions of the template. Lanczos iteration (an iterative method for the extreme eigenvalues of a symmetric matrix) then targets the smallest eigenvalues while the three rigid-rotation directions are projected away, and the same machinery yields the Rayleigh quotient $R(s) = \\partial_s \\hat F^\\dagger H(s)\\partial_s \\hat F$ for paths in volume space.","core_discovery":"The paper's central discovery is that the Hessian of the low-temperature log-likelihood has a usable spectral structure. It rewrites $L = -\\hat{\\sigma}^2 \\log P$ as $\\frac{1}{2}\\int \\bigl(M_{20}(\\kappa)|\\hat F(\\kappa)|^2 - 2\\Re(M_{11}^\\dagger(\\kappa)\\hat F(\\kappa)) + M_{02}(\\kappa)\\bigr)\\,d\\kappa/|\\kappa|$, where $M_{20}, M_{11}, M_{02}$ are accumulated Fourier-space moments of the CTF-weighted, alignment-corrected image data. From the template-wise 2D form the alignment block of the Hessian is block-diagonal per image and the volumetric block is diagonal in Fourier space; the cross-blocks $H_{F\\tau}$ and $H_{\\tau F}$ are evaluated through the 3D form, using back-propagated monopole and dipole impulses. When each image's optimal alignment is treated implicitly as a function of the volume, the Hessian of the reduced likelihood is assembled from these blocks, a first- and second-order expansion of the implicit alignment shift gives the correction, Lanczos iteration extracts the extremal spectrum after rigid rotations are projected out, and the small eigenvalues are the soft modes. On TRPV1 with realistic image pools the softest modes localize uncertainty to the bottom of the map under polar-biased views, shrink dramatically for the empirical viewing-angle distribution, move with viewing-angle geometry for other molecules, and stiffen as defocus increases; along an RNA conformational path the overlap of the path tangent with the soft subspace flags locations where free-energy estimates will be poorly resolved.","pith_inferences":["A natural extension the paper leaves implicit is to use the soft subspace as a local covariance model: the inverse Hessian eigenvalues provide per-direction variances, so a per-voxel error map could be produced at the cost of a few Lanczos iterations and compared against bootstrap or split-map FSC uncertainty.","The Hessian spectrum suggests an experimental-design rule that the authors only gesture at: new images should be chosen to enlarge the smallest eigenvalues of the Hessian (for instance by adding viewing angles or defocus values that couple to the soft modes), rather than only increasing the total image count.","The constructed degeneracies point to a practical diagnostic threshold: if the softest eigenvalues are of order the noise variance divided by the number of images, the local likelihood surface is nearly flat; flagging such reconstructions could catch the spurious heterogeneity the paper demonstrates, since the fake sub-volumes show enhanced softness in exactly the regions where they were under-con","For the noise-marginalized variant, the paper's validity check is the scatter around a single empirical affine fit; monitoring that scatter on each new dataset would give a cheap, dataset-specific test of whether the computed modes still describe the true posterior uncertainty."],"forward_implications":["The softest eigenvalue converts into a hypothesis-testing bound: a deformation along the softest mode can be scaled until the likelihood ratio drops by a chosen factor (for example 20, as in the paper's figures), bounding the largest volumetric change that a standard test could not reject, together with its linked alignment shifts.","Soft-mode amplitude per map region acts as a per-voxel trust map: regions with large soft-mode component are the least constrained by the image pool, which is exactly the information needed when interpreting flexible or low-confidence domains.","Changing the image pool shifts the soft subspace in predictable ways: polar-biased viewing angles soften the caps, equatorial bias softens a complementary set, and broader viewing-angle coverage raises the smallest eigenvalues, quantifying which additional viewing angles are needed.","Increasing defocus at fixed resolution raises the softest eigenvalues and shrinks the soft subspace, so imaging conditions can be ranked by information content rather than by CTF appearance alone.","Along a conformational path, locations where the path tangent has small Rayleigh quotient $R(s)$ with the Hessian are exactly where neighbouring conformations are projectionally confusable; a collective variable whose tangent follows stiff modes avoids this source of poor conditioning in free-energy estimates."],"supporting_citations":[{"why":"Supplies the volumetric back-propagation formulation and low-resolution solver from which the 3D integral moments are built.","marker":"[42]"},{"why":"Provides the fast rigid image-alignment factorization used to evaluate the 2D template inner products and their angle derivatives.","marker":"[47]"},{"why":"Supplies the radial-recombination scheme that projects per-shell alignment information into the volume representation.","marker":"[48]"},{"why":"Defines the Bayesian cryo-EM framework whose posterior likelihood is the object being differentiated.","marker":"[16]"},{"why":"Provides the noise-marginalized likelihood model whose empirical affine fit is used for realistic noise levels.","marker":"[44]"},{"why":"Sets up the cryo-EM free-energy estimation problem that motivates the path Rayleigh quotient case study.","marker":"[43]"},{"why":"Gives the rotationally invariant representation used to argue that uniformly distributed viewing angles typically determine a unique reconstruction.","marker":"[52]"},{"why":"Supplies the bispectrum-inversion uniqueness result against which the paper's constructed degenerate examples are contrasted.","marker":"[53]"},{"why":"Supplies the TRPV1 volume and image pool used in the case studies demonstrating soft-mode localization.","marker":"[41]"},{"why":"Supplies the lamellae micrographs whose per-image mismatch values justify the affine log-likelihood fit used for the noise-marginalized Hessian.","marker":"[46]"}],"fun_headline_variants":["Integral trick finds cryo-EM map's shakiest zones","Softest Hessian modes expose unreliable cryo-EM regions","One integral maps out weak spots in cryo-EM volumes","Hessian spectrum flags which cryo-EM map parts to trust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole calculation is done on a simplified likelihood: the low-noise limit, or a noise-marginalized form in which $\\log \\ell$ is replaced by an affine fit $a\\ell + b$ from a single dataset. The soft modes describe the true posterior uncertainty only if that approximation faithfully represents real, noisy cryo-EM data; if it does not, they describe an approximation instead.","fun_headline_variants_meta":{"raw":{"variants":["Integral trick finds cryo-EM map's shakiest zones","Softest Hessian modes expose unreliable cryo-EM regions","One integral maps out weak spots in cryo-EM volumes","Hessian spectrum flags which cryo-EM map parts to trust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000914,"raw_usage":{"total_tokens":4078,"prompt_tokens":1252,"completion_tokens":2826,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":868,"completion_tokens_details":{"reasoning_tokens":2751}},"tokens_in":868,"tokens_out":2826,"duration_ms":21406,"temperature":1.0,"reasoning_tokens":2751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:39:00.823945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small single-particle problem where the exact Gaussian likelihood can still be evaluated pixel-by-pixel, compute the posterior covariance by direct numerical integration or by extensive sampling at a realistic signal-to-noise ratio, and compare its leading eigenvectors with the soft modes obtained from the 3D-integral Hessian. A large subspace angle between the two, or a mismatch between the inverse Hessian eigenvalues and the sampled variances along those directions, would show that the low-temperature approximation breaks down precisely where uncertainty quantification is needed.","supporting_citations":[{"cited_title":"Rangan and Leslie Greengard","cited_arxiv_id":null,"evidence_quote":"Supplies the volumetric back-propagation formulation and low-resolution solver from which the 3D integral moments are built."},{"cited_title":"Factorization of the translation kernel for fast rigid image alignment","cited_arxiv_id":null,"evidence_quote":"Provides the fast rigid image-alignment factorization used to evaluate the 2D template inner products and their angle derivatives."},{"cited_title":"Radial-recombination for rigid rotational alignment of images and volumes","cited_arxiv_id":null,"evidence_quote":"Supplies the radial-recombination scheme that projects per-shell alignment information into the volume representation."},{"cited_title":"A Bayesian view on cryo-EM structure determination","cited_arxiv_id":null,"evidence_quote":"Defines the Bayesian cryo-EM framework whose posterior likelihood is the object being differentiated."},{"cited_title":"Cryolike: A python package for cryo-electron microscopy image-to-structure likelihood calculations","cited_arxiv_id":null,"evidence_quote":"Provides the noise-marginalized likelihood model whose empirical affine fit is used for realistic noise levels."},{"cited_title":"A bayesian approach to extracting free-energy profiles from cryo-electron microscopy experiments","cited_arxiv_id":null,"evidence_quote":"Sets up the cryo-EM free-energy estimation problem that motivates the path Rayleigh quotient case study."},{"cited_title":"Zhao and A","cited_arxiv_id":null,"evidence_quote":"Gives the rotationally invariant representation used to argue that uniformly distributed viewing angles typically determine a unique reconstruction."},{"cited_title":"Bendory, N","cited_arxiv_id":null,"evidence_quote":"Supplies the bispectrum-inversion uniqueness result against which the paper's constructed degenerate examples are contrasted."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TRPV1 volume and image pool used in the case studies demonstrating soft-mode localization."},{"cited_title":"New statistical metric for robust target detection in cryo-em using 2dtm","cited_arxiv_id":null,"evidence_quote":"Supplies the lamellae micrographs whose per-image mismatch values justify the affine log-likelihood fit used for the noise-marginalized Hessian."}],"review_version":1}