{"id":"264df85b-4a07-4307-9326-9cf730865efb","arxiv_id":"2412.14897","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Diffusion priors with reconstruction-guided posterior sampling reconstruct plausible 3D point clouds from sparse 2D projections, coarse densities, and subunits, outperforming maximum likelihood alone.","lead":"The authors train AI diffusion models on 3D point clouds of chairs, mixed objects, and biomolecular complexes, then use them as Bayesian priors to reconstruct full 3D structures from very sparse inputs such as a few 2D projections, coarse densities, or known subunits. The method beats maximum-likelihood fitting on synthetic benchmarks and targets cryo-electron microscopy, where images are extremely noisy and sparse.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control: the paper never compares guided posterior sampling against sampling from the diffusion prior alone (α=0), so the reported gains over ML may reflect the prior distribution rather than the sparse observations.","rationale":"The paper's strongest evidence is that DPS beats an ML baseline at the same forward-model energy. But the central claim is about posterior sampling from observations, so the relevant baseline is unconditional prior sampling. The ML baseline is deliberately weak (random initialization, 100 Adam steps), so beating it does not establish that the observations are informative. Because both ShapeNet and CryoStruct test sets are in-distribution relative to the prior training data, the prior alone is a strong competitor; the paper's own generation metrics (Table 3) show the priors produce plausible structures. An α=0 ablation is the minimal experiment that separates prior quality from likelihood guidance. This is a different concern from the reader's weakest assumption about the delta approximation in Appendix A.1, but it is more directly load-bearing: if the ablation lands, the reported success is an artifact of the prior distribution, not a demonstration of Bayesian data fusion. I therefore keep the reader's CONDITIONAL verdict; the condition should include this control. I also note that Eq. 9 defines ζ = α(t)/√log p0(y|Dθ), but p0 ∝ exp(-E) makes log p0 negative for the positive energies in Table 1, so the formula as written is not real-valued; this should be clarified, but the missing control is the decisive issue.","tokens_in":20918,"tokens_out":13089,"duration_ms":104177,"concrete_test":"Run Algorithm 1 with α=0 (ζ=0) for all rows of Table 1 and for the CryoStruct settings of Appendices A.4.3 and A.4.6, using identical time steps, β schedules, and the same RMSD/CD/EMD evaluation code. If the α=0 samples achieve errors statistically indistinguishable from or better than the reported DPS values, the central claim that the sparse observations drive the reconstruction fails; if DPS is clearly and consistently better, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that posterior sampling with diffusion priors enables 3D reconstruction from sparse observations requires that the observations shift the samples away from the prior. The paper never tests this. In Eq. 9 the guidance term is ζ∇_{x(t)} log p0(y | Dθ(x(t),t)); setting ζ=0 (equivalently the guidance strength α=0 in Table 4) reduces Algorithm 1 to unconditional prior sampling. Yet Table 1 and the CryoStruct benchmarks in Appendices A.4.1–A.4.7 compare only DPS against maximum-likelihood optimization of the same energies. Without an α=0 control, the lower CD/EMD/RMSD of DPS relative to ML is compatible with two explanations: (i) the likelihood guidance genuinely fuses the observations, or (ii) the diffusion prior alone already generates in-distribution structures close to the test ground truth. Explanation (ii) is plausible because the ShapeNet test instances come from the same categories used for training, and the CryoStruct test inputs are generated from 1024-component Gaussian-mixture point clouds of the same kind the prior was trained on. Table 3 confirms the priors generate plausible, in-distribution samples (1-NNA 54–78%), so the possibility that prior-only samples match or beat the reported DPS errors is not ruled out. The reader's concern about the DPS delta approximation (Appendix A.1, Eq. 17) is valid but secondary: if prior-only samples already achieve the reported errors, the guidance mechanism is not what the experiments demonstrate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Bayesian framework for 3D point-cloud reconstruction in which a score-based diffusion model trained on clean point clouds serves as a prior, and sparse observations (2D projections, coarse-grained structures, or subunits) are incorporated through reconstruction guidance (DPS). The authors train diffusion priors on three datasets (ShapeNet-Chair, ShapeNet-Mixed, CryoStruct), define assignment-based likelihood energies in Eqs. (10)-(14), and compare their approximate posterior sampler against maximum-likelihood optimization with the same forward models. Table 1 reports lower Chamfer and Earth Mover distances for DPS over ML in nine ShapeNet tasks, and Appendices A.4.1-A.4.7 report per-structure RMSD comparisons on 100 held-out biomolecular complexes, with selected examples shown in Figure 2.","tokens_in":21231,"tokens_out":8559,"duration_ms":68913,"significance":"If the empirical claims hold, the paper makes a useful contribution: it demonstrates a flexible, task-agnostic way to condition a generative prior on severely incomplete 3D measurements, and it applies this to cryo-EM-like settings with very few projections. The evaluation has genuine strengths: the prior is trained on a training split and tested on held-out structures, the ML baseline uses the same observational model, and the ShapeNet comparison aggregates 1,000 reconstructions. The manuscript is also honest about runtime costs. However, the central claim that the observations are what drive the reconstructions is not yet supported, because no prior-only (alpha=0) baseline is reported and the guidance strength is manually tuned per task. The paper also relies on the DPS delta approximation for non-smooth assignment energies without a diagnostic. For these reasons the result is promising but not yet established.","major_comments":[{"comment":"Table 1 and the CryoStruct benchmarks compare DPS only against ML; there is no alpha=0 (unconditional prior sampling) control. Setting alpha=0 in the guidance weight zeta=alpha(t)/sqrt(log p0(y|D_theta(x(t),t))) in Eq. (9) reduces Algorithm 1 to sampling from the prior alone, so the reported gains over ML could in principle be an effect of the prior distribution rather than of the likelihood guidance. This is not merely hypothetical: Table 3 shows that the priors generate samples whose 1-NNA values are far from random, and the ShapeNet test instances come from the same categories used in training. The paper should report the CD/EMD (and RMSD, for CryoStruct) of prior-only samples on the same test sets, or otherwise demonstrate that the guidance term shifts the samples toward the measurements (e.g. by showing that the energy of DPS samples is lower than that of prior-only samples). Without this control, the abstract's claim that posterior sampling with diffusion priors allows 3D reconstruction from very sparse observations is not established.","section":"Section 4.2, Eq. (9), Table 1"},{"comment":"Guidance strength alpha is chosen separately for each row of Table 1, with values ranging from 4,000 to 80,000, yet no sensitivity analysis or selection rule is given. Since alpha is the knob that controls the balance between prior and likelihood, the headline advantage of DPS over ML could be sensitive to this tuning. The authors should include a sensitivity sweep (e.g. CD/EMD versus alpha for at least one ShapeNet and one CryoStruct task) and, ideally, a principled way to set alpha before the method can be claimed as generally applicable. The same concern applies to the task-specific choices of beta(t) in Table 4.","section":"Section A.3, Table 4"},{"comment":"Reconstruction guidance replaces p(x(0)|x(t)) with a delta at the denoiser output (Eq. 17), and the guidance gradient is computed through a min-over-permutations energy (Eqs. 10-13). For point clouds with assignment-based likelihoods, this energy is non-smooth and the delta approximation has no error control; it is therefore not clear that the guided samples approximate the intended posterior, as opposed to a prior sample adjusted by a heuristic gradient. The authors should add a diagnostic: for example, compare the measurement energies of DPS samples against prior-only samples and against the energies of the ground-truth test inputs, or run a small experiment with a known posterior to check whether the sampler is measurement-faithful. Section 4.4 lists runtime as the only limitation, but this approximation is at least as important.","section":"Appendix A.1, Eq. (17), Eqs. (10)-(13)"},{"comment":"All experiments are conducted on synthetic measurements generated from the ground-truth point clouds (randomly sampled points projected with known orientations; Eq. (10) and the text in Section 4.2). No real cryo-EM images are used, and the forward model does not include CTF, noise, or unknown orientations. The abstract's statement that the method allows 3D reconstruction from cryo-EM images is therefore stronger than what the experiments support. The authors should either add experiments on realistic simulated cryo-EM images (with noise and unknown orientations) or explicitly scope the central claim to idealized projections with known orientations.","section":"Sections 4.2-4.3"}],"minor_comments":[{"comment":"The atomic-structure RMSD is computed as a one-sided nearest-neighbor distance (argmin over model points for each ground-truth atom), which is not the standard symmetric RMSD; this should be stated and ideally supplemented with a symmetric metric such as CD or EMD.","section":"Section 4.3"},{"comment":"The mapping from the rows of Table 1 to the hyperparameter rows of Table 4 is implicit; adding explicit row identifiers would make the experimental setup reproducible.","section":"Tables 1 and 4"},{"comment":"Neither axes nor scale bars are shown, so the reported RMSD values in angstroms cannot be visually verified; adding a common scale bar or showing the aligned ground truth in the same panels would help.","section":"Figure 2"},{"comment":"After Eq. (17) the notation switches from D(x(t),t) to D_theta(x(t),t) without comment; since the whole point is that the learned denoiser is an approximation, this distinction should be made explicit.","section":"Appendix A.1"},{"comment":"The manuscript does not mention whether code or trained models will be released; given the many training and inference details, a public implementation would substantially aid reproducibility.","section":"General reproducibility"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is a reasonable proof-of-concept, but the headline claim is broader than the evidence. The most pressing fixes are the alpha=0 baseline and the synthetic-only evaluation; both are addressable within the scope of a revision. I would not reject on the basis of the DPS approximation alone, but the authors should provide a diagnostic. If code is not released, the per-task tuning in Table 4 will be hard to evaluate. The fit for a machine learning venue is fine; the novelty over prior diffusion-based inversion is mainly in the 3D point-cloud application and the assignment-based likelihoods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper does something genuinely new—diffusion posterior sampling with assignment-based likelihoods for 3D point-cloud reconstruction from projections, coarse densities, and subunits—and the synthetic results consistently beat maximum likelihood. The stress-test concern is legitimate: there is no alpha=0 control, so the paper never shows what the prior alone produces under the same evaluation. That is a real gap, but I don't think it sinks the paper. Prior-only samples would be random draws from the training categories; they would not match specific held-out test structures well enough to produce the reported CD/EMD/RMSD numbers. Still, a two-line ablation would settle it and should be requested.\n\nWhat's new: the combination of DPS with EDM schedules and a point transformer, plus the min-over-permutations energies for cryo-EM-like forward models. The evaluation is thorough within its synthetic scope: 1000 reconstructions, per-structure plots, EMD and CD, and the appendix is honest about the runtime. The cryo-EM examples are suggestive, though all synthetic.\n\nSoft spots, in order: first, the missing alpha=0 control; second, alpha and beta are tuned per task with no sensitivity analysis, so the robustness of the method to those choices is unknown; third, everything is simulated—no real cryo-EM images—and there is no comparison with existing cryo-EM or shape-completion methods; fourth, no code or data released. The DPS delta approximation is inherited from Chung et al. and the assignment energy is non-smooth, but that is a known trade-off and the results suggest it works in practice; I would call that secondary.\n\nWho this is for: people working on diffusion priors for inverse problems, and structural biologists interested in sparse-view reconstruction. It deserves a serious referee. The central claim is defensible but needs a cleaner experiment isolating the guidance, a sensitivity analysis, and ideally one real-data test.\n\nRecommendation: send to peer review. Ask for the alpha=0 ablation, a small alpha sweep, and at least one non-synthetic demonstration before acceptance.","headline":"Solid proof of concept for diffusion-prior 3D reconstruction, but the missing alpha=0 control means the paper doesn't fully isolate the effect of the guidance.","tokens_in":21779,"tokens_out":3875,"would_cite":false,"duration_ms":31062,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that diffusion models trained on 3D point clouds provide Bayesian priors that make 3D reconstruction from very sparse, low-resolution, partial measurements tractable, producing structures closer to ground truth than…","keywords":["diffusion priors","Bayesian inverse problems","3D point cloud reconstruction","cryo-EM","diffusion posterior sampling","reconstruction guidance","sparse measurements","score-based generative models"],"falsifier":"Generate synthetic projections from a test structure that is clearly outside the distribution the diffusion prior was trained on, for example a random linear polymer, reconstruct with the same pipeline, and measure whether the posterior samples still match the projections or instead snap back to typical training shapes; if the samples ignore the measurements, the guidance is not conditioning on data. A sharper check is to compare DPS samples to samples from an exact conditional diffusion sampler, such as a sequential Monte Carlo approach, on a small point-cloud task; a large gap in RMSD or Chamfer distance would show the delta approximation is doing the work.","tokens_in":20693,"feed_emoji":"🧬","tokens_out":6598,"duration_ms":54753,"temperature":0.7,"pith_summary":"This paper tries to establish that a diffusion model trained on 3D point-cloud structures can act as a Bayesian prior that makes severely ill-posed 3D reconstruction from sparse measurements tractable. The authors integrate the prior with likelihoods for 2D projections, coarse-grained structures, and subunits via approximate diffusion posterior sampling, and compare against maximum-likelihood fitting with the same forward models. Across all tested ShapeNet configurations, the diffusion prior improves Chamfer Distance and Earth Mover's Distance reconstruction errors, even though ML often achieves higher likelihood. On biomolecular complexes with deposited atomic structures, one to five synthetic projections, sometimes plus a low-resolution structure or a known subunit, yield intermediate-resolution models with RMSDs of roughly 2 to 10 angstroms. The upshot is that rich data-driven priors can replace generic regularizers for cryo-EM-like inverse problems.","feed_headline":"A few projections plus a diffusion prior rebuild 3D structures","feed_subtitle":"Learned point-cloud priors beat likelihood-only fitting on sparse cryo-EM-style inputs, giving intermediate-resolution models.","key_machinery":"The machinery is a score-based diffusion model trained on 3D point clouds with perturbation kernel $\\mathcal{N}(x(0), t^2 I)$, combined with reconstruction guidance. Guidance replaces the intractable posterior over clean structures, $p(x(0) \\mid x(t))$, with a point mass at the denoiser output $D_\\theta(x(t), t)$, then backpropagates through assignment-based energies: for projections, the minimum over permutation matrices assigning upsampled projected points to cloud points; for coarse and subunit clouds, analogous minimum-assignment matching. These energies are solved with the Hungarian algorithm, and sampling uses an Euler-Maruyama integrator with second-order correction and stochastic noise injection. The whole chain converts a diffusion prior into a conditional sampler for arbitrary combinations of sparse observations.","core_discovery":"The central claim, on the paper's own terms, is that posterior sampling with a diffusion model prior enables 3D reconstruction from very sparse, low-resolution and partial observations, where generic priors fail. Concretely, the same likelihood that guides the diffusion prior, when optimized alone, produces higher-likelihood but structurally worse reconstructions; adding the prior consistently yields lower EMD and CD errors in every test case. For biomolecular complexes, the method produces intermediate-resolution structures from as few as one to five synthetic 2D projections, with RMSDs of roughly three to ten angstroms against deposited atomic models. The point-cloud representation and assignment-based likelihoods, solved as linear assignment problems, let the same trained prior condition on projections, coarse clouds, and subunits without task-specific retraining.","pith_inferences":["Because the representation is a coarse-grained point cloud, pushing to atomic resolution will likely require hierarchical priors or hybrid density-and-point-cloud models; the paper's stated goal of improving resolution points in that direction.","The minimum-over-permutations energy is non-smooth; replacing it with a soft assignment or entropic optimal transport could give smoother guidance gradients and potentially improve posterior sampling, a direct and testable extension.","If the claims hold on experimental cryo-EM data, this could complement standard reconstruction pipelines by providing models from very few particle images, which is useful for heterogeneous or flexible complexes.","The reported RMSDs depend on radius-of-gyration scaling and kernel-correlation alignment before comparison, so a different alignment convention could change the absolute numbers; cross-study comparisons should be cautious."],"forward_implications":["One trained diffusion prior can be reused for many observation geometries, including 2D projections, coarse-grained structures, subunits, or any weighted combination, without retraining the likelihood or the prior.","Sparse cryo-EM-like data from one to five projections can yield intermediate-resolution models with RMSDs around 2 to 10 angstroms against deposited atomic structures, which could make it feasible to study conformational differences between individual complexes.","The diffusion prior consistently beats maximum likelihood on EMD and CD in every ShapeNet test case, showing that data-driven priors are an effective regularizer for severely ill-posed 3D reconstruction.","Stochasticity in the sampling SDE plus a second-order correction step lowers reconstruction error at fixed network evaluations, so the details of the sampler matter for reconstruction quality.","Combining complementary sparse observations, such as projections plus a low-resolution envelope plus a known subunit, produces the most accurate reconstructions, as demonstrated on the spliceosome and the 26S proteasome."],"supporting_citations":[{"why":"Supplies the reconstruction-guidance approximation used to turn the diffusion prior into a posterior sampler.","marker":"[Chung et al., 2023]"},{"why":"Supplies the EDM diffusion schedule, score parameterization, and the second-order sampler with stochastic noise injection.","marker":"[Karras et al., 2022]"},{"why":"Supplies the point transformer architecture used to parameterize the denoiser.","marker":"[Nichol et al., 2022]"},{"why":"Supplies the chair and mixed-object point-cloud benchmarks used for training and testing the diffusion priors.","marker":"[Chang et al., 2015]"},{"why":"Supplies the curated biomolecular complex dataset used to train and benchmark the CryoStruct prior.","marker":"[Giri et al., 2024]"},{"why":"Supplies the reference atomic structures used as ground truth for biomolecular reconstruction comparisons.","marker":"[Berman et al., 2000]"},{"why":"Supplies the Hungarian algorithm used to solve the linear assignment problems in the likelihood energies.","marker":"[Kuhn, 1955]"},{"why":"Supplies the reverse-time diffusion equation underlying the reverse SDE used in sampling.","marker":"[Anderson, 1982]"}],"fun_headline_variants":["Diffusion priors rebuild 3D from sparse cryo-EM projections","Few projections plus diffusion prior give 3D structures","Bayesian diffusion priors beat likelihood-only fitting for cryo-EM","Posterior sampling with diffusion priors solves sparse 3D reconstruction","Diffusion priors handle sparse, noisy cryo-EM views for 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method relies on the approximation that the denoiser output fully determines the clean structure, treating the posterior over clean data as a point mass; if that approximation is poor for point-cloud likelihoods, samples will be biased toward the prior and the reconstructions will not faithfully reflect the measurements.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion priors rebuild 3D from sparse cryo-EM projections","Few projections plus diffusion prior give 3D structures","Bayesian diffusion priors beat likelihood-only fitting for cryo-EM","Posterior sampling with diffusion priors solves sparse 3D reconstruction","Diffusion priors handle sparse, noisy cryo-EM views for 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001111,"raw_usage":{"total_tokens":4610,"prompt_tokens":906,"completion_tokens":3704,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":3609}},"tokens_in":522,"tokens_out":3704,"duration_ms":15809,"temperature":1.0,"reasoning_tokens":3609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:49:02.885182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic projections from a test structure that is clearly outside the distribution the diffusion prior was trained on, for example a random linear polymer, reconstruct with the same pipeline, and measure whether the posterior samples still match the projections or instead snap back to typical training shapes; if the samples ignore the measurements, the guidance is not conditioning on data. A sharper check is to compare DPS samples to samples from an exact conditional diffusion sampler, such as a sequential Monte Carlo approach, on a small point-cloud task; a large gap in RMSD or Chamfer distance would show the delta approximation is doing the work.","supporting_citations":[{"cited_title":"Diffusion Posterior Sampling for General Noisy Inverse Problems","cited_arxiv_id":null,"evidence_quote":"Supplies the reconstruction-guidance approximation used to turn the diffusion prior into a posterior sampler."},{"cited_title":"Elucidating the Design Space of Diffusion-Based Generative Models","cited_arxiv_id":null,"evidence_quote":"Supplies the EDM diffusion schedule, score parameterization, and the second-order sampler with stochastic noise injection."},{"cited_title":"Point-e: A system for generating 3d point clouds from complex prompts, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the point transformer architecture used to parameterize the denoiser."},{"cited_title":"Cryo2structdata: A large labeled cryo-em density map dataset for ai-based modeling of protein structures","cited_arxiv_id":null,"evidence_quote":"Supplies the curated biomolecular complex dataset used to train and benchmark the CryoStruct prior."},{"cited_title":"The protein data bank","cited_arxiv_id":null,"evidence_quote":"Supplies the reference atomic structures used as ground truth for biomolecular reconstruction comparisons."}],"review_version":1}