{"id":"8a6eb26a-9ae2-4b05-959a-df1fd2e2a3e0","arxiv_id":"2504.17656","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"polyGen generates polymer chain conformations in periodic boxes from repeat-unit SMILES, and reports bond, angle, and dihedral distribution fidelity on a held-out DFT test set.","lead":"polyGen is a machine-learning model that generates three-dimensional atomic structures of polymer chains from only the repeat unit's chemical connectivity, such as a SMILES string. It combines a graph-based encoder, a variational autoencoder, and a latent diffusion transformer, trained on 3,855 DFT-optimized polymers and augmented with small-molecule data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central 'low-energy, realistic' claim is not directly tested: the forward-KL metric targets one DFT structure per polymer, the connectivity filter does not certify 3D accuracy, and no energy validation is reported.","rationale":"The paper is a plausible proof-of-concept with a sensible architecture and useful internal ablations: joint training with QM9 improves local metrics, and relative positional attention bias improves filter success. However, the central claim in the abstract and Figure 1 is that generated conformations are 'realistic' and at potential-energy minima, and that claim is not directly validated anywhere in the manuscript. The evaluation is geometric and distributional only, with a single DFT reference per polymer. That setup cannot certify an ensemble: forward KL to a sparse one-structure histogram does not directly score conformations that are absent from the reference, the connectivity filter explicitly does not guarantee accurate 3D structure, and no energy or relaxation check is reported. The overlap between these evaluation metrics and the VAE loss terms further weakens the inference that the model learned physics rather than the evaluation quantities. I partially agree with the reader's weakest assumption: the single-reference and filter premises are indeed problematic, but the sharper issue is that even if the reference were a perfect ensemble representative, the reported metrics would not test energy validity. A targeted re-optimization and energy comparison would settle whether the generated structures are actually low-energy. Under that condition, the reader's CONDITIONAL verdict remains appropriate; I would not move to REJECT because the architecture, ablations, and exemplar figures provide credible proof-of-concept evidence, but the headline claim should be explicitly conditioned on such validation.","tokens_in":11721,"tokens_out":11872,"duration_ms":121608,"concrete_test":"For 10 test polymers, including the 208-atom system, generate 10 filtered structures per polymer and relax them with the same DFT protocol used for polyChainStructures, or with GFN2-xTB/MMFF for the largest systems. Compare the relaxed energy distribution and backbone RMSD of each generated structure to the reference DFT structure. If median relaxed energies are within roughly 1 kcal/mol/atom of the reference and relaxed structures align to it, the low-energy claim is supported; if they relax to different minima or remain high-energy, the distributional KL matches do not establish realistic generation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 evaluates against empirical histograms of bonds, angles, and dihedrals from a single DFT-optimized structure per polymer via D_KL(Q_DFT || Q_pred) (Eq. 1). This is a geometric overlap criterion, not a physical validity or energy criterion: a generated structure with reference-like bonded local geometry but severe nonbonded strain or incorrect global packing can still score well, and the metric does not certify that the ensemble is low-energy. The Section 4.5 filter checks graph isomorphism and rejects bond lengths below 0.8 Angstrom, but the paper itself concedes that successful connectivity does not guarantee an accurate 3D structure. No energies of generated structures are computed, no re-optimization or relaxation is reported, and no comparison to physics-based baselines (PSP, RDKit ETKDG, PolyConf) is given. The evaluation quantities also overlap with the VAE training losses L_bond, L_angle, L_dihedral (Eq. 8), so the reported distributional agreement partly reflects optimizing the same local-geometry objectives during training. The central claim that polyGen generates realistic, low-energy conformations is therefore stronger than the evidence supports.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces polyGen, a latent diffusion framework that generates atomic-level 3D polymer structures conditioned only on the repeat-unit chemistry (SMILES). The model combines a graph-based conditioning module, a variational autoencoder that reconstructs fractional coordinates, Cartesian coordinates, and the periodic-box z-height, and a diffusion transformer with a relative-position attention bias. Training uses 3,855 DFT-optimized polymer chains from polyChainStructures, augmented with 100K QM9 small molecules. The paper evaluates generated structures by forward KL divergence between predicted and DFT bond-length, angle, and dihedral distributions (Eq. 1), and by a post-generation filter requiring graph isomorphism and a bond-length cutoff. The authors report that joint training with QM9 improves bond and angle KL divergences, that position-biased attention improves generation success and dihedral/angle metrics, and that the model produces diverse conformations and captures z-height trends. The stated contribution is the first generative model for synthetic polymer conformations, with the central claim that polyGen generates realistic and diverse low-energy polymer structures from minimal input.","tokens_in":11974,"tokens_out":5042,"duration_ms":52593,"significance":"If the central claim is established, polyGen would be a useful proof-of-concept for a genuinely under-addressed problem: generating diverse, physically reasonable polymer conformations without external relaxation. The paper has tangible strengths: it introduces a new evaluation protocol adapted to conformational diversity, it ablated the contribution of joint training and position-biased attention, and it reports limitations honestly, including low success rates for large systems and the concession that connectivity filtering does not certify 3D accuracy. However, the evidence as presented does not support the strong claim of generating 'realistic and diverse low-energy conformations.' The evaluation metrics overlap with the training objectives, the reference is a single DFT structure per polymer, no energy validation is reported, and no baselines are included. The paper would be strengthened substantially by adding energy-based validation, non-overlapping evaluation quantities, and comparisons to existing tools such as PSP, RDKit ETKDG, or PolyConf.","major_comments":[{"comment":"The headline claim that polyGen generates 'realistic and diverse low-energy conformations' (Abstract) is not directly tested. The evaluation in Eq. (1) compares bond/angle/dihedral histograms to a single DFT-optimized structure per polymer, and the Section 4.5 filter only checks graph isomorphism and a 0.8 Å bond-length cutoff. Section 4.5 itself states that 'a successful prediction of connectivity doesn't guarantee an accurate 3D structure.' No energies of generated structures are computed, no relaxation or re-optimization is reported, and no comparison to physics-based structure generators (PSP, ETKDG, PolyConf) is given. A generated structure with reference-like local geometry but severe nonbonded strain or incorrect global packing could pass all reported criteria. Please add energy-based validation (e.g., DFT or force-field single-point energies of generated versus relaxed structures) or explicitly narrow the claim to local-geometry distribution matching.","section":"§2.1, §2.3, §4.5"},{"comment":"The evaluation quantities are not independent of the training objective. The VAE loss in Eq. (8) includes L_bond, L_angle, and L_dihedral, which directly supervise bond lengths, angles, and dihedrals; the evaluation metric in Eq. (1) measures exactly those quantities. Therefore, the reported distributional agreement partly reflects optimizing the same structural objectives during training, rather than an emergent ability to generate physically valid conformations. Please report at least one evaluation metric on quantities that are not directly optimized (e.g., nonbonded distances, density or z-height distributions, or energies), or ablate the structural losses and show how the KL metrics change.","section":"Eq. (8) and Eq. (1)"},{"comment":"The quantitative claims lack baselines and uncertainty estimates. Improvements such as success rate increasing from 36.9% to 64.8% with position bias, and KL divergences decreasing by 41.0% (bonds), 29.0% (angles), and 16.1% (dihedrals) with joint training, are reported as point estimates with no standard deviations across seeds or per-polymer confidence intervals. There is also no quantitative comparison against existing polymer or conformer generators, even though PSP [13] and PolyConf [27] are cited in the introduction. Without baselines and error bars, the relative improvements and the claim of being 'the first' model for this task cannot be assessed.","section":"§2.4 and §4.4.1"},{"comment":"The dataset contains only one DFT-optimized conformation per polymer, as the paper acknowledges in Section 2.3. The forward KL divergence in Eq. (1) therefore measures agreement with a single point in what is acknowledged to be a broad conformational ensemble, not with the polymer's true low-energy distribution. The authors should justify that a single relaxed structure is representative of the ensemble, or evaluate against multiple conformers (e.g., from MD trajectories or repeated PSP/DFT relaxations) for at least a subset of test polymers. Without this, the distributional matching results do not establish that the model captures polymer flexibility.","section":"§2.1 and §2.3"}],"minor_comments":[{"comment":"Please specify the binning scheme for the KL divergence computation. The text says the discrete buckets use dz=0.001 over ranges that imply ~1,100 (bond), 180,000 (angle), and 360,000 (dihedral) buckets; it is unclear whether such fine binning is actually used and how empty buckets are handled.","section":"§2.1, Eq. (1)"},{"comment":"The caption and text do not state whether the reported KL divergences are means or medians over polymer test cases, or how the error bars in the figure are defined. Please clarify the aggregation and provide per-polymer distributions or boxplots.","section":"Figure 5"},{"comment":"The symbol h_i^(0) is used both for the initial atom embedding and for the MLP output after concatenating positional encodings. Using a distinct symbol such as h_i^(0,raw) and h_i^(0) would remove ambiguity.","section":"§4.2, Eq. (2)"},{"comment":"The text mixes flow-matching and denoising-diffusion terminology. Equations (9)-(13) define a flow-matching objective, but Section 2 says 'latent diffusion transformer' and the abstract says 'latent diffusion.' Please state the exact formulation, noise schedule, and how the flow-matching ODE corresponds to the diffusion process.","section":"§4.4"},{"comment":"The qualitative statement 'Generation is precise to the order of Å, but not on the scale of picometers' would be more useful with numerical values, e.g., mean absolute errors or RMSDs for the bond-length distributions shown in Figure 3.","section":"§2.2"},{"comment":"The dataset link is present, but code and checkpoints are only promised 'upon publication.' Since the results are central to the paper and no code is available for review, please provide a link to an anonymous repository or describe how to reproduce the key numbers with the dataset alone.","section":"§5 and §6"},{"comment":"Please clarify how the r²=0.854 for z-height is computed: is it a per-polymer mean of generated heights, or based on a single sample per polymer? The current text is ambiguous about which quantity is regressed.","section":"§2.4, Figure 5d"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned as the first generative model for polymer conformations, but the closely related PolyConf work (ref [27]) is cited yet not compared. A head-to-head comparison, or at least a clear technical distinction, would materially improve the contribution. The lack of energy validation is the main correctness risk; if the authors can add even a small DFT or force-field energy study on a subset of test polymers, the core claim would be much better supported. The manuscript is within scope for a computational materials venue, but the current evidence supports only a claim about local geometric distribution matching, not full physical realism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on polyGen. The real contribution is a working pipeline: graph-conditioned VAE, latent diffusion with position-biased attention, joint training with QM9, and a polymer-specific evaluation protocol based on bond/angle/dihedral distributions. That combination, applied to repeat-unit SMILES as the only input, is new, and the ablations are informative: joint training cuts bond/angle KL by substantial margins, and the position bias roughly doubles the pass rate of the connectivity filter. The paper is also honest about its failures—it shows bad samples, admits picometer precision is missing, and concedes that successful connectivity doesn't guarantee accurate 3D structure.\n\nThe soft spots are real but not fatal. The strongest claim in the abstract is \"realistic and diverse\" conformations, and \"low-energy\" appears in Figure 1 and the discussion. But no energy is computed for generated structures. The evaluation is forward KL against a single DFT-optimized structure per polymer, and the structural losses used to train the VAE (bond, angle, dihedral MSE) overlap with the evaluation quantities, so some of the reported agreement is expected. That said, the authors don't hide this; they define the benchmark as distributional matching, not energy validation. The bigger practical issue is that no code or checkpoints are released and no baselines are run against PSP, RDKit ETKDG, or PolyConf, so an independent check isn't possible yet. The \"first generative model\" phrasing is also a bit strong given PolyConf appears in the references as concurrent work.\n\nOverall, this is a solid proof-of-concept that deserves serious peer review. A referee should push for code release, baseline comparisons, and ideally some energy sanity check (e.g., single-point DFT or MM energies on a sample of generated structures). The central claim should be softened from \"low-energy\" to \"geometrically plausible.\" I'd bring it to a reading group and would cite it, conditional on code appearing.","headline":"A useful proof-of-concept for polymer structure generation, honest about its limits, but the 'low-energy' claim is not directly tested.","tokens_in":12518,"tokens_out":1607,"would_cite":true,"duration_ms":15565,"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":"polyGen is a generative model that turns a repeat-unit SMILES string into an ensemble of realistic three-dimensional polymer conformations.","keywords":["polymer structure generation","latent diffusion","flow matching","graph conditioning","position-biased attention","conformational diversity","DFT-optimized polymer structures","structure matching"],"falsifier":"Take a polymer with an independently known multi-conformer ensemble, for example from molecular-dynamics sampling or many DFT minima, generate 100 polyGen samples, and measure the KL divergence of bonds, angles, and dihedrals against that full ensemble. The central claim falls if a populated low-energy basin is systematically missed or if per-bond-type peaks stay displaced by more than about 0.1 Å while the connectivity filter still passes.","tokens_in":11482,"feed_emoji":"🧪","tokens_out":9928,"duration_ms":89307,"temperature":0.7,"pith_summary":"polyGen is a generative model that aims to turn a minimal input—the repeat unit's SMILES string, a text shorthand for atomic connectivity—into a distribution of realistic three-dimensional atomic structures for the polymer chain. If the claim holds, polymer simulations gain an on-demand way to generate initial conformations without heuristic packing steps. The authors train a graph-conditioned latent diffusion transformer on 3,855 DFT-optimized infinite-chain polymer structures, jointly with 100,000 small molecules from QM9, and score the output by comparing bond-length, angle, and dihedral distributions with DFT references. They report that joint training raises the filter-pass rate from 27.4% to 36.1%, that a position-aware attention bias lifts graph-connectivity success from 36.9% to 64.8%, and that predicted repeat-unit lengths track DFT values with $r^2 = 0.854$. The evaluation rests on the premise that each polymer's single stored DFT conformation represents its true low-energy ensemble.","feed_headline":"polyGen generates polymer 3D structures from repeat-unit SMILES alone","feed_subtitle":"A latent diffusion model matches bond, angle, and dihedral distributions to DFT-optimized chains.","key_machinery":"The load-bearing object is a latent diffusion transformer over an atom-wise variational latent space, conditioned by a graph-interaction-network encoding of the repeat unit and modified by a learnable relative-position attention bias. The bias is computed from a one-hot graph-distance tensor that classifies each atom pair as identical, bonded, angle-separated, dihedral-separated, or more than four bonds apart, and it is added directly to the attention logits so local interactions can dominate while global attention remains possible. The same network weights encode both polymers and small molecules, letting the much larger QM9 set share the latent space, and a post-generation filter keeps a sample only if its computed connectivity graph matches the input and every bond exceeds 0.8 Å.","core_discovery":"polyGen's central claim is that polymer conformation generation can be posed as conditional latent diffusion: a molecular graph of the repeat unit is encoded into atom-wise conditioning vectors, a variational autoencoder learns a shared latent space for polymers and small molecules, and a diffusion transformer denoises that latent to decode back into fractional coordinates, Cartesian positions, and the chain-axis box height. Given only the repeat-unit connectivity, the model is claimed to generate a diverse ensemble of linear and branched conformations whose bond-length, angle, and dihedral distributions peak near the DFT-optimized references, rather than a single fixed geometry. The authors further claim that joint training with small-molecule conformers and a learnable relative-position attention bias are what make the approach viable on a dataset of only 3,855 polymers.","pith_inferences":["A stronger validation would compare polyGen ensembles against a genuinely multi-conformer ground truth, such as molecular-dynamics sampling or several DFT minima per polymer; if the distributional match survives that comparison, the realism claim is much firmer.","The connectivity-match plus 0.8 Å filter is a crude physical-validity proxy, so re-ranking accepted samples by a fast force field or a single-point DFT energy could be a cheap way to improve the usable fraction.","Because joint training with QM9 mainly improved local features (bonds and angles) rather than dihedrals and chain-axis heights, further gains in global conformation accuracy will likely require polymer-scale training data rather than more small-molecule data.","The same conditioning and attention-bias recipe could extend to cross-linked or multi-chain amorphous polymer boxes, which share the same local-geometry constraints but also require handling non-bonded interactions that this single-chain proof-of-concept does not address."],"forward_implications":["Given only a repeat-unit SMILES string, the model can produce a pool of diverse polymer conformations suitable as starting points for DFT or molecular-dynamics simulation.","Jointly training on QM9 small molecules cuts bond-length KL divergence by 41.0% and angle KL divergence by 29.0% relative to polymer-only training, showing that molecular local structure transfers to polymers.","The relative-position attention bias is the main feasibility lever: it raises the graph-connectivity pass rate from 36.9% to 64.8%, and the vanilla transformer cannot generate any valid structure for polymers above roughly 150 atoms, whereas the biased version still can.","Filter-passing ensembles are diverse rather than collapsed to one geometry, and the predicted chain-axis repeat length tracks the DFT reference with $r^2 = 0.854$, so the model distinguishes dense from sparse chains.","Current precision is angstrom-scale rather than picometer-scale, so the generator cannot yet reliably separate bond types, and generation success falls off sharply for the largest repeat units."],"supporting_citations":[{"why":"Supplies the 3,855 DFT-optimized infinite-chain polymer structures that form the training and evaluation dataset.","marker":"[12]"},{"why":"Supplies the baseline latent diffusion transformer architecture and self-conditioning flow-matching setup that polyGen adapts.","marker":"[30]"},{"why":"Provides the 100,000-molecule DFT small-molecule training set used for joint augmentation.","marker":"[25]"},{"why":"Supplies the relative-position encoding attention bias and the structural-geometry loss terms used in the autoencoder.","marker":"[21]"},{"why":"Describes the existing physics/heuristic polymer structure predictor that this method targets as the prior approach for initial structure generation.","marker":"[13]"},{"why":"Provides the flow-matching formulation behind the latent-space ordinary differential equation and training objective.","marker":"[15]"}],"fun_headline_variants":["polyGen: first generative model for atomic-level polymer structures","polyGen: polymer 3D structures from repeat units alone","Diffusion model generates realistic polymer conformations","From repeat-unit SMILES to diverse polymer conformations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation's load-bearing premise is that the single DFT-optimized conformation stored for each polymer stands in for that polymer's full ensemble of low-energy conformations, and that the combination of a matching connectivity pattern with a 0.8 Å bond cutoff certifies a physically useful structure.","fun_headline_variants_meta":{"raw":{"variants":["polyGen: first generative model for atomic-level polymer structures","polyGen: polymer 3D structures from repeat units alone","Diffusion model generates realistic polymer conformations","From repeat-unit SMILES to diverse polymer conformations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1327,"prompt_tokens":924,"completion_tokens":403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":339}},"tokens_in":540,"tokens_out":403,"duration_ms":4119,"temperature":1.0,"reasoning_tokens":339,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:34:43.924047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a polymer with an independently known multi-conformer ensemble, for example from molecular-dynamics sampling or many DFT minima, generate 100 polyGen samples, and measure the KL divergence of bonds, angles, and dihedrals against that full ensemble. The central claim falls if a populated low-energy basin is systematically missed or if per-bond-type peaks stay displaced by more than about 0.1 Å while the connectivity filter still passes.","supporting_citations":[{"cited_title":"Polymer structure prediction from first principles","cited_arxiv_id":null,"evidence_quote":"Supplies the 3,855 DFT-optimized infinite-chain polymer structures that form the training and evaluation dataset."},{"cited_title":"Polymer structure predictor (psp): a python toolkit for predicting atomic-level structural models for a range of polymer geometries","cited_arxiv_id":null,"evidence_quote":"Describes the existing physics/heuristic polymer structure predictor that this method targets as the prior approach for initial structure generation."}],"review_version":1}