{"id":"1e5a9a3b-0bd4-49a8-940d-33ad78c9c4a4","arxiv_id":"2412.13223","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A diffusion model is guided by crystallographic electron density to sample non-independent protein ensembles that recover alternate backbone conformations.","lead":"Proteins change shape, and X-ray crystallography measures the average of many copies. This paper guides a generative AI protein model to reproduce that average, recovering multiple alternative structures that match the data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Forward-model calibration is the load-bearing risk: the uniform PDB-average B-factor and implicit unit noise scale in Eq. 2 are admitted approximations, and a misspecified bandwidth can manufacture or erase the bimodality on which the recovery claim rests.","rationale":"The reader's weakest_assumption identifies the forward model's B-factor treatment as the load-bearing risk, and my reading converges on the same point. The paper explicitly labels the B-factor choice 'naive' and says it should depend on ensemble size, yet no ablation or sensitivity analysis is performed. Since the likelihood in Eq. 2 also has no noise scale, the entire guidance signal in Eq. 4 is an unnormalized L2 penalty whose effective strength is set by eta and gradient clipping. The evaluation further confounds the question: the same density objective used for guidance is used for matching-pursuit filtering and for the reported cosine similarity, and altloc recovery is assessed after that filter. A misspecified kernel can therefore generate spurious bimodality, so the central claim 'accurately recovers complicated multi-modal alternate protein backbone conformations' is not secured against the admitted forward-model approximation. This does not change the reader's CONDITIONAL verdict, but it sharpens the condition: the B-factor/noise-scale sensitivity check should be a required revision, not an optional future improvement. The proposed concrete test directly examines whether the recovered modes are stable under the admitted approximation and whether they correspond to the true altloc coordinates.","tokens_in":8444,"tokens_out":9686,"duration_ms":103278,"concrete_test":"Run the full pipeline on all 13 targets with B in {0, B_PDB/2, B_PDB, 2*B_PDB, c/M} and, separately, with an estimated noise variance in Eq. 2, keeping all other hyperparameters fixed. For each setting, report the altloc proximity scores and bimodality of the selected ensembles, the cosine similarity, and the spread over multiple seeds. If the recovered bimodality and proximity to PDB altlocs are essentially unchanged across the B range and over seeds, the concern does not land. If bimodality appears or disappears, or the scored modes shift by more than a backbone bond length, the headline claim is contingent on an untuned bandwidth and the verdict should require a calibrated forward model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the guidance gradient in Eq. 4 to pull ensembles toward densities that match the true crystallographic measurement. That is only true if the forward model Fc in Eq. 3 is calibrated. The paper fixes a single scalar B for every atom, set to the PDB-wide average of per-atom B-factors, even though each generated sample is already a complete structure and the authors themselves note that B should be inversely proportional to the ensemble size. Eq. 2 also contains no noise variance, so the likelihood scale is arbitrary up to the guidance weight eta=9 and gradient clipping; the density grid spacing and L2 normalization are never specified. A wrong B does not merely blur the output: it changes the geometry of the loss landscape. With B too large, the smoothed mean density of a unimodal ensemble can fit a bimodal Fo; with B too small, the gradient rewards overfitting coordinate noise. Because the claimed 'accurate recovery of altlocs' is measured by proximity to PDB-deposited altloc coordinates after matching-pursuit filtering against the same density target, a miscalibrated kernel can produce a bimodal score distribution that is an artifact of the KDE bandwidth rather than evidence that the true alternate conformations were found. The paper provides no sensitivity analysis, no comparison to standard multiconformer refinement (e.g., qFit), and no synthetic test with known ground-truth modes, so this admitted approximation is the weakest link between the observed map and the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative approach for reconstructing protein structure ensembles from crystallographic electron density maps. The method combines a pre-trained diffusion model (Chroma) as a structural prior with a differentiable forward model that computes electron density from an ensemble of backbone and sidechain coordinates (Eq. 3). Sampling is performed via a non-i.i.d. score-guidance scheme (Eq. 4), where the entire ensemble is guided toward the observed density using a Gaussian likelihood (Eq. 2). The authors evaluate their method on thirteen protein structures with known alternate conformations (altlocs), reporting improved density alignment and recovery of bimodal backbone distributions compared to unconditional sampling.","tokens_in":8801,"tokens_out":12062,"duration_ms":112565,"significance":"The problem is timely and of genuine biological interest: connecting generative structure models to experimental electron density measurements could provide a systematic way to extract ensembles from crystallographic data. The paper contributes a clear inverse-problem formulation, a non-i.i.d. ensemble guidance scheme, an ablation of batched versus i.i.d. sampling, and a unimodal control experiment. These are useful ingredients for the community. However, the quantitative evidence for the central claim is weakened by the selection-based evaluation and by an apparent sign inconsistency in the guidance formula. If these issues are resolved, the method would constitute a valuable step toward data-driven ensemble modeling; in its current form, the paper's conclusions outrun the evidence.","major_comments":[{"comment":"The sign of the guidance term is inconsistent with the log-likelihood. Equation (2) defines log p(Fo|X, a) = -||Fo - (1/|X|) Σ Fc||^2, so the likelihood gradient points in the direction that decreases the squared density error. Equation (4) contains the drift term -PP^T(∇ log p_t + η ∇ log p), which, after using ∇ log p = -∇(squared error), yields a guidance contribution +η PP^T ∇(squared error). Algorithm A1, line 17, explicitly uses xn-1 = μ + λ ∇x [||Fc(...)-Fo||^2], again the positive gradient of the squared error. Both expressions move the ensemble in the direction that increases the density mismatch, i.e., they implement gradient ascent on the error rather than the likelihood guidance described in the text. Standard classifier guidance (Dhariwal and Nichol, 2021) adds ∇ log p(y|x) with the opposite sign. If the implemented code uses the negative sign, then the equations and algorithm as written are wrong; if the positive sign is actually used, the reported improvement in alignment is unexplained. This must be clarified and corrected.","section":"Section 2, Eq. (2), Eq. (4), Algorithm A1 (line 17)"},{"comment":"The primary quantitative evaluation is circular with respect to the filtering procedure. Algorithm A2 selects a subset I of samples that maximizes corr(d_{I∪i}, dt), i.e., the correlation (or related density match) with the observed density Fo. The evaluation then reports cosine similarity between the mean density of the selected subset and Fo (Figures A3 and 2, Table A2). Because the selection criterion is the same metric as the reported outcome, the absolute density-alignment numbers are inflated by construction; they do not independently measure the quality of guidance. The guided-versus-unconditional comparison is less affected because both arms are filtered identically, but the absolute claim that the method \"accurately recovers\" densities is not supported. The authors should report unfiltered ensemble metrics, or use a held-out region or a metric not used by the matching pursuit selector.","section":"Section 3, Evaluation criteria and Algorithm A2"},{"comment":"The forward model is not calibrated, and no sensitivity analysis is provided. The B-factor in Eq. (3) is set to the average PDB B-factor, which the authors themselves describe as \"naive,\" and it acts as a global KDE bandwidth. The Gaussian likelihood in Eq. (2) contains no noise variance, leaving the likelihood scale arbitrary up to the guidance weight η. The bandwidth directly controls the geometry of the guidance landscape: a too-large B-factor can make a unimodal ensemble's average density fit a bimodal Fo, while a too-small B-factor can reward overfitting to coordinate noise. Without a systematic sensitivity analysis over B (and the noise scale, grid spacing, and normalization), and without a synthetic test with known ground-truth modes, the claim that the guidance gradient recovers the true alternate conformations rather than artifacts of the approximate forward model is not established.","section":"Section 2, Modeling the likelihood (Eqs. 2 and 3)"},{"comment":"The structural recovery claim rests on the bimodality scores shown in Figure 3, computed only for samples that survived the matching pursuit filter (at most 5 of 16 initial samples). The figure shows score distributions but provides no error bars, no significance tests, and no distance threshold that would let the reader judge whether the recovered modes are truly close to the PDB altlocs. With such a small filtered sample size, the bimodality may be dominated by the selection procedure rather than by the guidance signal. The authors should either report the full unfiltered ensemble or provide a quantitative metric such as the fraction of samples within 1 Å of each altloc, with uncertainties, before claiming accurate recovery.","section":"Section 3, Bimodality and structural alignment (Figure 3)"}],"minor_comments":[{"comment":"There are several typos: \"heterogenity\" in the abstract, \"dihderal\" and \"backpropagtaion\" in Section A.1, and \"angels\" in Section A.2 (should be \"angles\").","section":"Abstract and Appendix A.1"},{"comment":"In the algorithm's input list, the target density is denoted \"dt\", but the text refers to it as \"Fo\"; the notation should be made consistent.","section":"Algorithm A2"},{"comment":"The description of the density conversion (Lang et al. method) is brief; a reference to the exact equation or software used to convert PDB maps to e-/Å^3 would improve reproducibility.","section":"Section 3, Experimental setup"},{"comment":"The statement that results were obtained \"without any additional tuning\" is difficult to reconcile with the absence of a sensitivity study for the central hyperparameters (guidance scale, gradient clipping, B-factor); a small table varying these would strengthen the robustness claim.","section":"Section A.2"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency in Eq. (4) and Algorithm A1 is the most serious technical issue in the manuscript. It is possible that this is a typographical error in the paper while the implementation uses the correct sign; however, as written, the guidance moves the ensemble in the wrong direction. The editor should ask the authors to state explicitly the implemented sign and to rectify the equations and algorithm accordingly. The circularity in the evaluation and the lack of forward-model calibration also need to be addressed before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, worth a look if you care about generative models for protein structure. The core idea is simple and genuinely new: instead of guiding a diffusion model sample-by-sample, they guide the joint distribution of a whole batch, with a likelihood that couples all samples through the mean calculated density. That non-i.i.d. formulation (Eq. 4) is not in Dhariwal-Nichol or Weiss et al., and it is the right kind of move for an inverse problem where the measurement is an ensemble average. The paper also uses Chroma as a fixed prior rather than training a new model, which is sensible.\n\nWhat they show: on 13 high-res structures with deposited altlocs, guided sampling produces density maps with higher cosine similarity to the observed map than unguided Chroma, and the guided ensembles are bimodal in the altloc regions where unguided sampling is not. The ablation in Table A2 (batch vs i.i.d.) supports the central modeling claim. That is real evidence, and it is more than a lot of method papers ship.\n\nThe soft spots are also real. First, the matching-pursuit filter selects the subset that maximizes the same likelihood that is later used to report density alignment, so the cosine numbers are upper-bounded by design. The structural bimodality score is a separate metric, which helps, but the filter still selects samples by density fit. Second, the forward model is rough: fixed average B-factor, no noise scale in Eq. 2, unspecified grid spacing. The authors admit the B-factor treatment is naive. That matters because the guidance gradient depends entirely on this forward model, and there is no sensitivity analysis or synthetic test with known modes to show bimodality is not a bandwidth artifact. Third, no code or data release, no comparison to qFit or other multiconformer refinement tools, no error bars. The paper's own unimodal control is a nice touch, but only on one protein.\n\nOverall, the central idea holds up; the evaluation is not yet rigorous enough to take the 'accurate recovery' claim at face value. This is a solid conference submission that needs a revision round: add synthetic ground-truth experiments, vary B, compare to a standard multiconformer tool, and release code. I'd send it to review, and I'd cite the non-i.i.d. guidance idea if working on guided diffusion for ensembles. Take the altloc recovery plots as suggestive, not definitive.","headline":"Non-i.i.d. ensemble guidance over a diffusion prior is a genuinely new idea for crystallographic ensemble fitting, but the evaluation leans on filtered self-consistency and an admittedly approximate forward model, so read the altloc claims as promising rather than proven.","tokens_in":9308,"tokens_out":2445,"would_cite":true,"duration_ms":21767,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion model guided by crystallographic electron densities recovers protein conformations that unconditional sampling misses.","keywords":["protein ensembles","alternate conformations","electron density","diffusion models","score guidance","crystallography","inverse problem","protein structure generation"],"falsifier":"Run the guidance on a structure whose true alternate conformations and per-atom B-factors are known, or on a synthetic density generated from a known two-state ensemble, and recompute the ensemble with the B-factor taken inversely proportional to ensemble size instead of the naive average. If the recovered states shift or the bimodality score collapses, the forward-model approximation is the deciding factor; if recovery is unchanged, the method is insensitive to this approximation. Another decisive test: compare the guided ensemble's calculated density with the deposited density at unconstrained regions away from altlocs, since overfitting to noise there would indicate the likelihood model is too simplistic.","tokens_in":8231,"feed_emoji":"🧬","tokens_out":9275,"duration_ms":73486,"temperature":0.7,"pith_summary":"This paper attempts to solve a long-standing inverse problem in structural biology: given a crystallographic electron density map of a protein, produce an ensemble of backbone conformations that explains the map, including the alternate conformations (altlocs) that crystallographers model when the density is bimodal. The authors' central claim is that this can be done by guiding an existing protein-structure diffusion model with a non-independent ('non-i.i.d.') ensemble likelihood: instead of asking each sampled conformer to match the density on its own, the model is guided by how well the whole ensemble's average calculated density fits the observed map. They show that this guidance recovers the two distinct backbone conformations in several high-resolution crystal structures, whereas the same diffusion model sampled unconditionally captures neither the correct states nor the bimodality.","feed_headline":"Density-guided diffusion finds alternate protein states","feed_subtitle":"Coupling an ensemble of conformers to the density map recovers alternate states that unguided generation misses.","key_machinery":"The load-bearing object is the non-i.i.d. ensemble guidance term in the backward SDE of a score-based diffusion model. The variable $X$ concatenates all backbone coordinate sets in the ensemble; the unconditional prior score is block-diagonal (separable), but the guidance score $\\nabla_X \\log p(F_o | X, \\chi(X), a)$ is inseparable, so the ensemble is optimized jointly. This score is computed through the differentiable forward model of Eq. 3 (a kernel-density estimate of atomic Gaussians with tabulated form factors and a uniform B-factor) and the Gaussian likelihood of Eq. 2, with sidechain dihedrals $\\chi(X)$ produced by the model's differentiable sidechain packer. A matching-pursuit algorithm then selects the subset of generated samples that best explains the observed density, which the authors find necessary because some ensemble members overfit noise.","core_discovery":"The paper's central discovery is methodological: a differentiable forward model of electron density can be inserted into the backward diffusion step of a pretrained protein backbone model as a guidance score, and the correct way to use it is to compute the likelihood of the entire sampled batch at once rather than per-sample. The calculated density $F_c$ is a kernel-density estimate of atomic Gaussian spheres with tabulated form factors (Eq. 3), the likelihood is a Gaussian comparing the observed map to the ensemble-averaged calculated map (Eq. 2), and the guidance term $\\nabla_X \\log p(F_o | X, \\chi(X), a)$ couples all backbone coordinates in the joint SDE (Eq. 4). On a set of high-resolution structures with known alternate locations, this non-i.i.d. guidance improves density-map cosine similarity and produces bimodal distributions straddling both deposited altlocs, while unconditional sampling fails to reproduce the experimental states. The authors also show that batched (non-i.i.d.) guidance beats per-sample (i.i.d.) guidance, and that a matching-pursuit filter that greedily selects the best-fitting subset of the ensemble further stabilizes the result.","pith_inferences":["If the B-factor is truly a free parameter, the naive average-B-factor choice could be replaced by an optimized or learned bandwidth during sampling; the authors' own comment that B 'should be inversely proportional to ensemble size' points to a concrete improvement that might further sharpen altloc recovery.","Because the likelihood couples all ensemble members through their average density, the method implicitly enforces a first-moment constraint on the sampled distribution; a natural extension would be to match higher-order statistics, such as the variance of density, to capture broad conformational ensembles with weak bimodality.","The matching-pursuit filter discards samples that overfit noise, which suggests the raw non-i.i.d. posterior still over-disperses; a differentiable subset selection or a sparse ensemble prior could reduce reliance on post-hoc filtering.","The failure of unconditional generation to capture experimentally stable altlocs implies that generative protein models may systematically under-represent subpopulations; density guidance offers a data-anchoring mechanism to recover such subpopulations without retraining."],"forward_implications":["Density-guided generation yields ensembles whose mean calculated density matches the observed map substantially better than unconditional sampling, across all tested high-resolution structures.","In regions with modeled alternate locations, guided sampling produces samples near both altloc states, while unconditional sampling concentrates on a single mode, so the method recovers experimental conformational heterogeneity rather than just the prior's preferred state.","The non-i.i.d. (batched) likelihood is essential: ablating it to per-sample guidance degrades density alignment on the tested proteins, confirming that ensemble coupling is what carries the recovery.","On unimodal control regions, guided sampling stays unimodal while still fitting the density better than unconditional sampling, suggesting the guidance does not invent multimodality where the data do not support it.","The same guidance machinery can in principle be applied to other experimental observables, since only the differentiable forward model and the likelihood need to be swapped."],"supporting_citations":[{"why":"Supplies the pretrained score-based diffusion prior and differentiable sidechain packer that the guidance modifies.","marker":"[5]"},{"why":"Provides the classifier-guidance formulation that the non-i.i.d. ensemble score generalizes in Eq. 4.","marker":"[3]"},{"why":"Supplies the guided-diffusion SDE form used as the joint sampling equation.","marker":"[16]"},{"why":"Documents the catalogue of alternate locations and shows that existing ensemble predictors miss these states, motivating the task and providing the evaluation set.","marker":"[11]"},{"why":"Gives the procedure for converting mean-centered deposited density maps into physical units for the likelihood comparison.","marker":"[7]"},{"why":"Provides the tabulated atomic form factors used in the forward density model of Eq. 3.","marker":"[10]"},{"why":"Underlies the matching-pursuit filter that greedily selects the best-fitting subset of sampled conformers.","marker":"[8]"},{"why":"Provides the backbone-dependent rotamer library used by the sidechain packer.","marker":"[12]"}],"fun_headline_variants":["Non-i.i.d. guidance spots hidden protein states","Diffusion model uses density maps to find alternate folds","Ensemble guidance resolves conformation ambiguity in crystals","Batch-aware diffusion recovers alternate backbones from X-ray data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success hinges on the forward model and likelihood of Eqs. 2-3 being faithful: specifically, that a uniform B-factor set to the average of the deposited per-atom B-factors, and an implicit unit noise scale, correctly describe how the true conformations would appear in electron density.","fun_headline_variants_meta":{"raw":{"variants":["Non-i.i.d. guidance spots hidden protein states","Diffusion model uses density maps to find alternate folds","Ensemble guidance resolves conformation ambiguity in crystals","Batch-aware diffusion recovers alternate backbones from X-ray data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1415,"prompt_tokens":845,"completion_tokens":570,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":507}},"tokens_in":461,"tokens_out":570,"duration_ms":5385,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:05:57.870797+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the guidance on a structure whose true alternate conformations and per-atom B-factors are known, or on a synthetic density generated from a known two-state ensemble, and recompute the ensemble with the B-factor taken inversely proportional to ensemble size instead of the naive average. If the recovered states shift or the bimodality score collapses, the forward-model approximation is the deciding factor; if recovery is unchanged, the method is insensitive to this approximation. Another decisive test: compare the guided ensemble's calculated density with the deposited density at unconstrained regions away from altlocs, since overfitting to noise there would indicate the likelihood model is too simplistic.","supporting_citations":[{"cited_title":"A smoothed backbone-dependent rotamer library for proteins derived from adaptive kernel density estimates and regressions","cited_arxiv_id":null,"evidence_quote":"Provides the backbone-dependent rotamer library used by the sidechain packer."},{"cited_title":"Ingraham, Max Baranov, Zak Costello, Karl W","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained score-based diffusion prior and differentiable sidechain packer that the guidance modifies."},{"cited_title":"Bron- stein, and Renana Gershoni-Poranne","cited_arxiv_id":null,"evidence_quote":"Supplies the guided-diffusion SDE form used as the joint sampling equation."},{"cited_title":"Rosenberg, Sanketh Vedula, Alex M","cited_arxiv_id":null,"evidence_quote":"Documents the catalogue of alternate locations and shows that existing ensemble predictors miss these states, motivating the task and providing the evaluation set."},{"cited_title":"Therese Lang, James M","cited_arxiv_id":null,"evidence_quote":"Gives the procedure for converting mean-centered deposited density maps into physical units for the likelihood comparison."},{"cited_title":"International Tables for Crystallography, Volume C: Mathematical, physical and chemical tables","cited_arxiv_id":null,"evidence_quote":"Provides the tabulated atomic form factors used in the forward density model of Eq. 3."},{"cited_title":"Matching pursuits with time-frequency dictionaries","cited_arxiv_id":null,"evidence_quote":"Underlies the matching-pursuit filter that greedily selects the best-fitting subset of sampled conformers."}],"review_version":1}