{"id":"a69e15ea-b1e6-4258-8012-794f30ef877d","arxiv_id":"2506.19820","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"ProxelGen generates protein backbones by latent diffusion over a VAE-compressed voxel density representation, and uses that representation for spatial inpainting, shape conditioning, and motif scaffolding.","lead":"ProxelGen represents proteins as multi-channel 3D voxel densities ('proxels') instead of atom coordinates, then generates new proxels with a VAE plus latent flow model and decodes them using a fine-tuned Proteina model. The density representation lets designers condition directly on shapes and spatial masks, which atom-coordinate protein models cannot do naturally.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Designability and FID are reported after a 200M Proteina renoise-denoise refinement that is not applied to baselines; the headline may characterize the Proteina-assisted pipeline, not the proxel representation.","rationale":"The paper's central scientific claim is representational: a voxel density with Gaussian backbone channels and chain flow is sufficient for de novo protein structure generation. For that claim to hold, the final structures in Tables 1-2 must be traceable to information carried by the generated proxels. The current pipeline makes that traceability insecure. Section 3.4 fine-tunes a 60M Proteina on latent proxel tokens, and App. B.1 then renoises and denoises the decoded backbone with a 200M unconditional Proteina at t=0.8. This second model is a strong generative prior trained on the same AFDB distribution; the paper reports it 'vastly improved' designability and moved structures by ~3.35 Å RMSD. Applying such a refiner to any roughly-plausible backbone can increase ProteinMPNN/ESMFold self-consistency, so the 53.13 designability in Table 1 may not measure the quality of the generated density. The Native Proxels row is telling: proxelizing a native structure and running the same decoder/refinement already yields 50.39 designability, only 2.7 points below ProxelGen. That leaves little room for the generated density to be the cause of the headline. The comparison is also asymmetric: Proteina baselines are not refined at t=0.8, so part of ProxelGen's apparent parity is a pipeline artifact. The admitted chain-flow limitation (Sec. 6) reinforces this reading: if many generated proxels do not form a single connected chain, the decoder must be using its own prior to thread a chain, which is exactly the kind of information that should come from the chain-flow channel if the representation is doing its job. I do not see this as internal inconsistency; the paper is candid about the refinement and the limitation. The issue is attribution. A set of control rows using the same refinement on baselines and on scrambled proxels would settle it. The motif-scaffolding and shape-conditioning demonstrations remain interesting even if the unconditional designability claim is weakened, which is why I retain the CONDITIONAL verdict rather than moving to REJECT. The reader identified the same weakest assumption; my stress-test pass agrees. No code/data and a self-defined FID are secondary concerns, but the decoder attribution is the single point on which the central claim depends.","tokens_in":15143,"tokens_out":5613,"duration_ms":59544,"concrete_test":"Add control rows to Table 1: (1) ProxelGen samples scored before the 200M t=0.8 refinement; (2) Proteina 200M/H and 400M/H samples passed through the identical t=0.8 renoise-denoise refinement; (3) decoded outputs from scrambled or noise proxels (random latent tokens or permuted chain-flow channels) after the same refinement. If (2) or (3) reach designability near 53 without resembling the input proxels, or if (1) drops substantially, the headline metric is attributable to the Proteina refiner and the comparison must be re-leveled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim that ProxelGen's density representation matches AFDB designability is evaluated only after the generated proxels are decoded by a 60M Proteina fine-tuned on proxel tokens and then renoised/denoised by an unconditional 200M Proteina at t=0.8 (Sec. 3.4, App. B.1). The paper states this refinement 'vastly improved' designability at ~3.35 Å RMSD movement. Because this large atomistic prior is not conditioned on the proxel and is not applied to the Proteina baselines, the Table 1 comparison is partially circular: high designability may be supplied by the refiner rather than by the proxel representation. The Native Proxels row (50.39 designability) is within 3 points of ProxelGen (53.13), suggesting the decoder/refiner alone nearly matches the headline number. The admitted failure of chain flow to form connected chains (Sec. 6) points the same way: the decoder is compensating for missing ordering information. Without a control showing that random or scrambled proxels refined by the same pipeline do not reach comparable designability, the central claim conflates the density model with the Proteina-assisted decoding pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ProxelGen, a generative model of protein structure built on a voxelized 3D density representation called proxels. A protein is encoded as a multi-channel voxel grid containing backbone-atom Gaussian densities, a bond channel, and a three-channel chain-flow vector field. ProxelGen consists of a 3D CNN VAE that compresses proxels into a fixed-size latent space, a latent flow model trained to generate these latents, and a fine-tuned Proteina-based coordinate decoder that maps generated proxels back to atomistic backbones. The paper claims that on unconditional generation ProxelGen achieves higher novelty, better FID, and designability comparable to the training set, and that its spatial conditioning enables competitive motif scaffolding and shape-conditional generation. The central claim is that density-based protein generation is a viable alternative to atomistic point-cloud representations.","tokens_in":15469,"tokens_out":4695,"duration_ms":47332,"significance":"If the central claim holds, the paper opens a useful new axis in protein structure generation: representing proteins as voxelized densities rather than atomistic coordinates or frames. This representation naturally supports fixed-size latents, convolutional architectures, inpainting by spatial masking, and shape conditioning, and it connects protein structure generation to the mature literature on voxel-based generative modeling. The paper also provides useful validation of a self-supervised proxel embedding (ProxCLR) for FID-style evaluation, with sanity checks against perturbations and cluster removal. The use of external oracles (TM-align, ProteinMPNN/ESMFold self-consistency, FoldSeek diversity) is a strength, since the headline quantities are not defined purely by the model itself. However, the main empirical claims are currently not anchored tightly enough: the designability comparison is confounded by a large pretrained refinement model, the FID metric is ambiguously defined, and the reported differences lack statistical error bars. These issues are fixable and do not invalidate the core idea, but they must be addressed before the claims can be accepted.","major_comments":[{"comment":"The central claim that ProxelGen achieves 'the same level of designability as the training set' is evaluated after a renoise-denoise refinement step using an unconditional 200M Proteina model at t=0.8, which is not applied to the Proteina baselines. The paper states that this refinement 'vastly improved' designability while moving structures by about 3.35 Å RMSD. Because the decoder and refiner are large pretrained atomistic priors that are not conditioned on the generated proxels, the reported designability may characterize the Proteina-assisted pipeline rather than the density representation. The Native Proxels row (designability 50.39) being close to ProxelGen's 53.13 strengthens this concern. A control in which random or scrambled proxels are passed through the same decoding and refinement pipeline is needed to attribute the observed designability to the generative model rather than to the decoder/refiner.","section":"Section 3.4, Appendix B.1, Table 1"},{"comment":"All unconditional metrics are computed on 256 samples with no error bars, no multiple seeds, and no statistical significance testing. Several headline differences are small (FID 6.05 vs 7.25 for Proteina 400M (H); novelty 0.73 vs 0.69; designability 53.13 vs 45.70), and the reader cannot tell whether these gaps are meaningful. Please report confidence intervals, multiple seeds, or a statistical comparison for the key metrics in Table 1.","section":"Section 4.2, Table 1"},{"comment":"The FID reported in Table 1 is not clearly defined. The text states that 'the FID of the generated proxels themselves, before decoding back to atomic coordinates, is 6.81,' but Table 1 reports FID=6.05 for ProxelGen. Since the refinement step affects both designability and FID, it is unclear whether Table 1's FID is computed on raw generated proxels via ProxCLR, on decoded structures via ProteinFID, or on proxelized versions of decoded and refined structures. The comparison with Proteina baselines is only meaningful if the same embedding and the same pre/post-processing are used for all methods. Please define the FID variant precisely and report pre-refinement and post-refinement values separately.","section":"Section 3.5, Section 4.2, Appendix B.1"},{"comment":"The paper admits that many generated proxels do not form a single connected chain, that the chain flow may split or merge, and that the coordinate decoder often introduces unnatural kinks when threading a chain through such proxels. This is a load-bearing limitation for the claim that the density representation carries sufficient ordering information for full-chain generation. Please quantify how often chain connectivity fails in the generated proxels, and report how many of the designable samples in Table 1 and the successful scaffolding designs in Table 2 required decoder compensation for disconnected chain flow. Without this quantification, the reported results describe the combined proxel-plus-decoder pipeline rather than the density representation alone.","section":"Section 6"}],"minor_comments":[{"comment":"In the sentence describing the contrastive learning details, 'stochasic' should be 'stochastic'.","section":"Appendix A.2"},{"comment":"The chain flow vector field is defined using vectors v_i for i=1,...,N-1, but the summation in Equation (6) runs over i=1,...,N; the indexing for the terminal C-alpha position should be clarified.","section":"Section 3.1, Equation (6)"},{"comment":"The acronym 'FID' is used for both the ProxCLR-based proxel FID and the more general Fréchet Inception Distance naming convention; using a distinct term such as 'ProxFID' would avoid confusion with FID computed on atomistic structures.","section":"Section 3.5 and elsewhere"},{"comment":"The merged length rows in the upper half of Table 2 are difficult to parse, especially because ProxelGen reports a single number per task while prespecified methods report one number per length range; the caption should explain the row structure more explicitly.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising and timely idea, and the external benchmarks and validation experiments are valuable. The main issue is that the headline designability claim is not yet cleanly attributable to the proxel representation because of the Proteina refinement confound. I would encourage the editor to request the random-proxel control and statistical anchoring described in the major comments; these are feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is the representation, not the headline numbers. ProxelGen is the first protein backbone generator I know of that works directly on multi-channel voxel densities, and the chain-flow ordering field is a sensible way to make that representation decodable. That opens up capabilities coordinate models don't have: you can condition on arbitrary 3D masks or shapes, and mask out spatial regions rather than sequence positions. The motif scaffolding numbers on multi-segment tasks are the strongest evidence. On 1BCF, a four-segment motif, ProxelGen finds 12 unique successes where every baseline gets one; on 1QJG it gets 51 vs 3-18. That is a real gain, not a length artifact, and the paper deserves credit for showing it.\n\nWhere I part ways with the abstract is the unconditional comparison. The designability and FID numbers in Table 1 are for the full pipeline: generated proxels -> 60M Proteina decoder -> 200M Proteina renoise-denoise refinement at t=0.8. The baselines don't get that refinement. The paper admits the refinement 'vastly improved' designability, and the Native Proxels row shows that decoding and refining a native proxel gives 50.39 designability, within three points of ProxelGen's 53.13. That means the headline claim 'same level of designability as the training set' could mostly be coming from the off-the-shelf Proteina prior, not from the generated density. The admitted chain-flow failure points the same way: if the decoder is threading chains through proxels that often split or merge, it is doing the structural heavy lifting. I would want a control where random or scrambled proxels go through the same decode-refine pipeline. If those also reach high designability, the unconditional comparison is circular. That is a clean, addressable experiment, and it is currently missing.\n\nThe other issues are real but minor in proportion. No error bars on 256 samples, so the FID gap 6.05 vs 7.25 is unanchored. The FID metric is self-defined and they explicitly tune toward it; the ProxCLR validation in Appendix C is careful, but it is still a new metric. No code or data, which makes all of this hard to check. The limitations section is candid about chain connectivity, which I respect.\n\nBottom line: the representation and the conditioning results are a genuine contribution, and the paper is worth serious refereeing. But the unconditional designability claim should not be taken at face value until the refinement is controlled for. If I were refereeing, I'd ask for the scrambled-proxel control, pre-refinement designability, and error bars before accepting the headline. I'd still bring this to a reading group and cite it for the spatial conditioning work.","headline":"The proxel representation and its spatial conditioning are the real contributions; the unconditional designability claim is currently entangled with a Proteina refinement step not applied to baselines.","tokens_in":15988,"tokens_out":3225,"would_cite":true,"duration_ms":31829,"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 claims protein structures can be generated directly as multi-channel 3D density grids, producing samples more novel and better matched to the training distribution than atomistic point-cloud models at comparable designability.","keywords":["protein structure generation","3D density representation","voxel grids","latent diffusion","flow matching","motif scaffolding","shape conditioning","proxels"],"falsifier":"Decode the generated proxels with the small coordinate decoder and score designability without the large refinement pass; if designability falls to near zero, the density representation itself is not carrying the reported result. Separately, count the connected components of the generated chain-flow field before decoding: if most samples split or merge, the ordering channel is not doing the threading job the method requires.","tokens_in":14956,"feed_emoji":"🧬","tokens_out":13180,"duration_ms":124093,"temperature":0.7,"pith_summary":"This paper sets out to show that the representation used for protein structure generation need not be atomistic coordinates: a protein can be generated as a 3D density grid while matching or beating point-cloud-based models. The authors build ProxelGen, a latent diffusion model over a voxelized density representation with Gaussian-smoothed backbone channels and a vector field that encodes chain ordering from the N- to the C-terminus. On unconditional generation, ProxelGen's samples are more novel and closer to the training distribution than a leading atomistic baseline, with roughly native-level designability on the metrics they report. The same representation makes spatial conditioning natural: in motif scaffolding it solves multi-segment tasks where atomistic baselines find one success and ProxelGen finds twelve. If the density representation is what is delivering these results, protein generation gains a new axis—spatial inpainting and shape conditioning without fixing protein length—rather than a marginal tweak to existing point-cloud models.","feed_headline":"Proteins as 3D densities match or beat point-cloud models","feed_subtitle":"Voxelized densities with a chain-flow channel add spatial conditioning and scaffold multi-segment motifs no other method solves.","key_machinery":"The load-bearing object is the 'proxel' representation: a 7-channel 3D voxel grid in which three channels sample Gaussian-smoothed densities around backbone C, Cα, and N atoms, one channel encodes bond midpoints, and three channels form a 'chain flow' vector field pointing along successive Cα–Cα vectors from the N to the C terminus. That chain-flow channel is the mechanism that carries ordering information, making it possible to thread a single amino-acid chain through the generated density. On top of this representation, ProxelGen compresses 32×32×32 proxel arrays with a 3D CNN VAE into a latent space (512× spatial compression), trains a stochastic-interpolant flow model to generate latents, and fine-tunes an atomistic flow decoder to convert generated latents back to backbone coordinates. Because conditioning inputs are also voxel grids, spatial constraints can be injected by channel-wise concatenation in latent space.","core_discovery":"The paper's central claim is that a protein structure can be generated as a 3D density rather than as an atomistic point cloud, and that this alternative representation is not merely viable but advantageous. ProxelGen encodes proteins as 'proxels'—multi-channel voxel arrays sampled from Gaussian-smoothed densities around the backbone atoms, plus a bond channel and a vector field that traces the chain from N- to C-terminus—and learns a 3D-convolutional VAE whose latent space is generated by a flow model. On unconditional generation, ProxelGen reports higher novelty, better FID against the training distribution, and designability at roughly native levels when compared with a leading atomistic flow model, and in motif scaffolding it reports 12 unique successful designs for a four-segment motif where every tested baseline finds one. The paper also demonstrates that spatial conditioning follows for free from the representation: masked regions and arbitrary voxelized shapes can be concatenated as inputs, enabling inpainting and shape-conditioned generation without prescribing protein length or the placement of motif segments.","pith_inferences":["A decisive attribution test would be to score designability of decoded structures before the 200M-parameter refinement pass; if designability collapses without refinement, the headline result belongs to the decoder/refinement pipeline rather than to the density representation.","If chain-flow connectivity is enforced as a training objective, density-based models should extend to multi-chain complexes and to conditioning on experimental density maps, where a single connected chain is not the right prior; the paper's current chain flow assumes one chain and the authors note it often splits or merges.","The shape-conditioning setup suggests a direct application the paper leaves untested: conditioning on low-resolution experimental envelopes rather than shapes derived from known structures, using the same shape-adherence metrics.","The fixed-resolution proxel grid also points toward a coarse-to-fine generation scheme—generate a low-resolution shape, then sample higher-resolution refinements—which atomistic representations cannot express naturally."],"forward_implications":["Protein length no longer has to be fixed before sampling: the same voxel grid can represent different numbers of residues, so generation and inpainting can produce whatever size the conditioning shape supports.","Spatial tasks that require awkward bookkeeping in atomistic models—motif scaffolding, masked inpainting, shape-conditioned design—become native operations of concatenating or masking voxel channels.","If the 1BCF and 1QJG results generalize, density-based sampling will be the preferred tool for scaffolding multi-segment motifs, where sequence-anchored methods effectively fail.","The fixed grid lets protein generation borrow mature 3D CNN and latent diffusion infrastructure from image generation, and the representation itself scales with grid resolution rather than chain length."],"supporting_citations":[{"why":"Supplies the atomistic flow-model baseline whose unconditional results ProxelGen is compared against, and the pretrained model that is fine-tuned into the coordinate decoder.","marker":"[Geffner et al., 2025]"},{"why":"Supplies the latent-diffusion design (3D CNN VAE plus latent generative model) that ProxelGen adapts to proxel grids.","marker":"[Rombach et al., 2022]"},{"why":"Supplies the stochastic-interpolant flow objective used to train the latent generative model.","marker":"[Albergo et al., 2023]"},{"why":"Defines the RFdiffusion motif-scaffolding benchmark, success thresholds, and baseline results that Table 2 is built on.","marker":"[Watson et al., 2023]"},{"why":"Provides the Genie2 baseline and the clustered structural database and random-crop motif strategy used for training and finetuning.","marker":"[Lin et al., 2024]"},{"why":"Provides the SimCLR contrastive objective used to train the ProxCLR embedding that defines the paper's FID metric.","marker":"[Chen et al., 2020]"},{"why":"Supplies the FID evaluation methodology and the observation that temperature-tuned designability may not transfer, motivating ProxelGen's FID-tuned sampling.","marker":"[Faltings et al., 2025]"},{"why":"Supplies TM-score, used to measure novelty as the highest structural similarity to the training set.","marker":"[Zhang and Skolnick, 2005]"}],"fun_headline_variants":["Proteins as densities beat point-cloud models","Density-based protein generation tops point clouds","ProxelGen turns proteins into 3D densities","3D densities generate proteins with higher novelty","Density representation enables protein shape control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the generated proxels—not the pretrained coordinate decoder and refinement model—are what make the sampled structures designable, and that the chain-flow channel reliably encodes a single connected chain; if either fails, the headline comparisons describe the whole pipeline rather than the density representation.","fun_headline_variants_meta":{"raw":{"variants":["Proteins as densities beat point-cloud models","Density-based protein generation tops point clouds","ProxelGen turns proteins into 3D densities","3D densities generate proteins with higher novelty","Density representation enables protein shape control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1409,"prompt_tokens":881,"completion_tokens":528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":460}},"tokens_in":497,"tokens_out":528,"duration_ms":5729,"temperature":1.0,"reasoning_tokens":460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:25:28.683721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Decode the generated proxels with the small coordinate decoder and score designability without the large refinement pass; if designability falls to near zero, the density representation itself is not carrying the reported result. Separately, count the connected components of the generated chain-flow field before decoding: if most samples split or merge, the ordering channel is not doing the threading job the method requires.","supporting_citations":[],"review_version":2}