{"id":"738522dd-ca50-4c5d-b09a-d236a2f8c591","arxiv_id":"2411.10817","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ConfFlow uses a transformer-based continuous normalizing flow to directly sample 3D molecular conformations and reports large accuracy gains on the GEOM-Drugs benchmark.","lead":"ConfFlow is a machine learning model that predicts 3D shapes of molecules from their 2D chemical structure, using a continuous normalizing flow built with transformer layers. It reports up to 40% better accuracy on large drug-like molecules than earlier learning-based methods, which could matter for drug discovery and virtual screening.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'up to 40% over SOTA' claim is not settled because Table 1 omits GeoDiff and Torsional Diffusion, contemporary learning-based conformer generators with strong GEOM-Drugs results; adding them could change the ranking.","rationale":"The reader's weakest assumption is that the translation-invariance normalization layer may be non-invertible, which would invalidate the exact-likelihood objective in Eq. (2). That concern is legitimate but secondary: even if the flow's likelihood is formally wrong, the reported accuracy numbers could in principle be reproduced, because the model is trained by optimizing that objective and the empirical evaluation is based on RMSD after alignment. Conversely, if the baseline set omits genuine state-of-the-art methods, the central claim 'ConfFlow improves accuracy by up to 40% relative to state-of-the-art learning-based methods' fails regardless of whether the flow is correctly formulated. The paper's own completion date and 2025 revision make the omission of GeoDiff and Torsional Diffusion particularly consequential, as both are contemporary learning-based methods with strong GEOM-Drugs results. The missing comparisons are therefore the single most load-bearing condition for the headline claim. The reader's verdict of CONDITIONAL already captures the need for more rigorous comparison, so my stress test does not change the verdict; it sharpens the reason why the condition matters.","tokens_in":12402,"tokens_out":12180,"duration_ms":140380,"concrete_test":"Reproduce Table 1 on the same GEOM-Drugs test split (200 molecules, 2x sampling per graph) using the official implementations and evaluation protocol of GeoDiff and Torsional Diffusion. If either method attains COV mean/median or MAT mean/median within or better than ConfFlow (currently 88.3/97.6 and 0.895/0.874), then the 'up to 40% over SOTA' statement is unsupported. Additionally, rerun ConfFlow with at least three random seeds to check whether the margins over ConfGF (and over any newly added baseline) exceed seed variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim depends on Table 1, which compares ConfFlow only to CVGAE, GraphDG, CGCF, ConfVAE, ConfGF, GeoMol, and RDKit. The paper states it was completed in December 2022 and was revised in February 2025, yet it omits GeoDiff (ICLR 2022) and Torsional Diffusion (NeurIPS 2022), both of which are learning-based conformer-generation methods that report competitive or better GEOM-Drugs accuracy than the older baselines listed. The headline improvement of 'up to 40% relative to state-of-the-art learning-based methods' is computed against ConfGF (2021), not against the actual state of the art at the time. If GeoDiff or Torsional Diffusion matches or exceeds ConfFlow on COV/MAT for GEOM-Drugs, the central claim is misleading. The paper also gives no error bars or significance tests for any table entry, so the margins shown are not established as robust. This is a correctness risk for the empirical claim itself, not merely a disagreement with current consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ConfFlow, a continuous normalizing flow for molecular conformation generation. The model is conditioned on a molecular graph and directly samples atomic coordinates, using a point-transformer-based message passing network as the flow dynamics, without enforcing explicit geometric constraints. Training maximizes the exact log-likelihood via the change-of-variables formula, with kinetic and Jacobian regularization. Experiments on GEOM-QM9 and GEOM-Drugs compare ConfFlow against seven learning-based baselines plus RDKit, evaluating coverage, matching, mismatch, and ensemble property prediction. The central claim is that ConfFlow achieves state-of-the-art accuracy, especially on large molecules in GEOM-Drugs, improving by up to 40% over prior learning-based methods.","tokens_in":12680,"tokens_out":6568,"duration_ms":63839,"significance":"If the results hold, the paper makes a useful empirical contribution: it shows that a non-equivariant transformer-based CNF can outperform dedicated distance- and score-based generative models and the rule-based RDKit on conformation generation for drug-like molecules, and that direct coordinate sampling can scale to molecules with up to 181 atoms. The open-source release of the code is a strength, as are the ablations over architecture size, depth, and regularization. However, the two main concerns identified below (an underspecified normalization layer and an incomplete baseline set) must be resolved before the state-of-the-art claim is credible.","major_comments":[{"comment":"The 'normalization layer' used to obtain translation invariance is never defined. The Introduction says 'we deal with translation invariance by adding a normalization layer at the end of each CNF block,' while Section 2.1 mentions 'graph-independent reversible batch normalization layers.' If this layer subtracts the center of mass, it is not a bijection on R^{3M} and the change-of-variables formula in Eq. (2) is invalid; if it is a reversible batch norm, the mechanism by which it enforces translation invariance is not explained. Please define the transformation explicitly, state its Jacobian log-determinant, and describe how the base distribution and the coordinate space are modified so that the exact-likelihood training objective is mathematically correct.","section":"Section 2; Eq. (2); Eq. (14)"},{"comment":"The comparison omits GeoDiff (Xu et al., ICLR 2022) and Torsional Diffusion (Jing et al., NeurIPS 2022), which are contemporary learning-based conformer-generation methods with reported results on GEOM-Drugs. Without these baselines, the abstract's claim of 'up to 40% relative to state-of-the-art learning-based methods' is not supported, because the comparison is made only against older methods such as ConfGF. Please include these methods in Table 1, either by running their released code or by citing their reported numbers on the same split, or restrict the claim to the methods actually compared.","section":"Table 1; Section 3.2"},{"comment":"No error bars or statistical significance tests are reported, and several margins are small. For example, on GEOM-QM9 ConfFlow's mean MAT of 0.278 Å is worse than ConfGF's 0.269 Å, while the COV mean differs by only 0.53 percentage points. Please report the variance over multiple training runs or perform a significance test to establish that the observed differences are not due to random seed.","section":"Tables 1, 2, and 3"},{"comment":"The property-prediction evaluation uses only 30 molecular graphs randomly drawn from the GEOM-Drugs test set. With 30 samples, the reported median absolute errors are unlikely to be stable, and the claim that ConfFlow 'is the only learning-based method that achieves lower MedAE than RDKit in four of the six properties' may be sensitive to this small sample. Please provide confidence intervals, use a larger random sample, or report the individual results for all 30 molecules.","section":"Section 3.4"}],"minor_comments":[{"comment":"The phrase 'ConfFlow improve accuracy' is a subject-verb agreement error; it should be 'ConfFlow improves accuracy.'","section":"Abstract"},{"comment":"The notation L=1,3,... and L=2,4,... is ambiguous. Please clarify how the discrete reversible batch normalization layers and the continuous GCF blocks are indexed and alternated, and state how the normalization layers' log-determinants are included in the first sum.","section":"Eq. (14)"},{"comment":"The 'Reversible BN' layer is shown in the architecture diagram but not labeled with its exact operation. A one-sentence definition in the caption or text would remove the ambiguity discussed in the first major comment.","section":"Figure 2"},{"comment":"The statement that ConfFlow 'outperforms baselines by up to 40%' should specify the metric and the baseline (for example, MAT on GEOM-Drugs vs ConfGF), since the 40% figure is not apparent from Table 1 without this detail.","section":"Section 3.3"},{"comment":"GeoDiff and Torsional Diffusion are not cited in the reference list. If they are deliberately excluded, please explain why; otherwise, cite them in the appropriate places.","section":"References"},{"comment":"The set-builder notation for COV, MAT, and MIS is slightly informal; using an Iverson bracket or explicitly writing 'the number of elements in the set' would improve clarity.","section":"Eqs. (16)-(18)"}],"recommendation":"major_revision","confidential_remarks":"The paper was completed in December 2022 and revised in February 2025; the revision must address the 2022 state-of-the-art baselines, which are absent from the comparison. The undefined normalization layer is a more serious concern: if the authors cannot provide an explicit invertible transformation with a tractable Jacobian, the exact-likelihood framing of the method would need to be substantially revised. I recommend major revision rather than rejection, because both issues appear addressable in the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ConfFlow is a genuine new architecture: a graph-conditioned continuous normalizing flow built on point transformers that generates atomic coordinates directly, with no distance-geometry postprocessing. That combination isn't in the cited prior work, and the paper is honest about what it can't do—their own MMD results show distance-based baselines do a better job matching the true distance distribution. The GEOM-Drugs results on large molecules are impressive, and they release code, which is a real plus.\n\nTwo things worry me, and they are in different parts of the paper. First, the translation-invariance \"normalization layer\" is never defined. If it subtracts the center of mass, that map is not bijective on R^(3M), so the change-of-variables in Eq. (2) is invalid. Maybe they have a different mechanism, but the paper doesn't say. That is a load-bearing gap, not a style issue. Second, Table 1 omits GeoDiff and Torsional Diffusion, both published well before the February 2025 revision. These are not obscure baselines; they are the strongest learned conformer generators on GEOM-Drugs. So the \"up to 40% over SOTA\" claim is computed against ConfGF, which is not the actual state of the art. No error bars or significance tests appear anywhere, so the margins reported are not established as robust. That makes the central empirical claim unsettled.\n\nStill, I would send this to serious peer review. The architecture is novel enough, the code is released, and the ablations are thorough. The authors need to add the missing baselines, run multiple seeds, and specify the normalization layer explicitly. If the translation-invariance issue turns out to be real, the likelihood objective is wrong and the model needs to be rethought; if it is a matter of presentation, that can be fixed. Either way, this deserves referee time rather than a desk rejection.","headline":"ConfFlow is a promising new architecture with strong empirical results, but the central SOTA claim is not established until missing contemporaneous baselines and the undefined translation-invariance layer are addressed.","tokens_in":13130,"tokens_out":2070,"would_cite":true,"duration_ms":21588,"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":"ConfFlow shows that molecular conformations can be sampled directly in coordinate space with a transformer-based continuous normalizing flow, cutting error by up to 40% on large molecules versus learned baselines.","keywords":["molecular conformation generation","continuous normalizing flows","graph neural networks","point transformer","3D coordinates","GEOM-Drugs benchmark","drug-like molecules","translation invariance"],"falsifier":"For a two-atom molecule, inspect the normalization layer in the released code and compute its Jacobian: if the layer subtracts the center of mass, the Jacobian is rank-deficient on $\\mathbb{R}^{3M}$, so the change-of-variables log-determinant in Eq. (2) is not exact and the reported likelihood-based training becomes an approximation.","tokens_in":12230,"feed_emoji":"🧪","tokens_out":11914,"duration_ms":108365,"temperature":0.7,"pith_summary":"The paper tries to establish that molecular conformation generation does not need to go through predicted distances or explicit physical constraints: a continuous normalizing flow can evolve atomic coordinates directly from a simple prior into accurate 3D structures. Building on transformer-style message passing over the molecular graph, the proposed model ConfFlow treats each refinement step like a force-field update in molecular dynamics, which the authors argue makes the generative process interpretable. On the large-molecule GEOM-Drugs benchmark, the authors report accuracy improvements of up to 40% relative to leading learned methods, with ConfFlow the only learned model to consistently beat the rule-based RDKit on coverage and matching scores. This matters for drug discovery and molecular modeling, where fast, accurate conformer ensembles are needed for property prediction, docking, and screening.","feed_headline":"Transformer flow boosts large-molecule conformation accuracy by 40%","feed_subtitle":"ConfFlow samples coordinates directly, skips physical constraints, and beats learned baselines on drug-sized molecules.","key_machinery":"The machinery is a graph continuous flow (GCF), a continuous normalizing flow in which each block's time derivative is defined by a graph-conditional point transformer (GCPT). GCPT computes attention-weighted messages from neighboring atom and edge embeddings using coordinate differences, updates the embeddings, and then applies a per-atom multilayer perceptron to produce coordinate updates; integrating those updates defines the flow. The flow's invertibility comes from solving the same ODE forward for sampling and backward for likelihood, with the trace of the Jacobian estimated by Hutchinson's estimator. Two regularization integrals, one penalizing the kinetic energy of the flow and one penalizing the Jacobian Frobenius norm, stabilize training on large molecules. Translation invariance is handled by a reversible batch-normalization layer appended to each CNF block, which is the paper's mechanism for respecting physical symmetries without imposing explicit constraints.","core_discovery":"ConfFlow's central claim is that coordinate-space generation with an invertible, graph-conditioned continuous normalizing flow can outperform distance-based learned generators. Instead of predicting interatomic distances or their gradients and converting them to coordinates, ConfFlow samples atom positions from a Gaussian prior and integrates an ODE whose vector field is produced by point-transformer message passing over the molecular graph. The flow supports exact likelihood training because each layer is invertible, and translation invariance is delegated to a normalization layer at the end of each block rather than to equivariant network structure. On GEOM-Drugs the paper reports a mean coverage of 88.3% versus 61.5% for the strongest learning-based baseline ConfGF, a mean matching score of 0.895 Å versus 1.170 Å, and the best average rank across datasets and metrics, with improvements of up to 40%.","pith_inferences":["Not investigated in the paper, the same coordinate-space flow could be applied to protein side-chain or coarse-grained conformation generation, where nonlocal attention over coordinates may capture interactions that local torsion samplers miss.","The paper's own distance-distribution results show ConfFlow trails distance-explicit models on MMD, which suggests a testable extension: adding a soft distance-consistency auxiliary loss to the coordinate-space flow might improve distribution fidelity without sacrificing the large-molecule accuracy gains; the paper does not explore this.","A direct check of the normalization layer's invertibility, such as computing its Jacobian on a two-atom molecule, would clarify whether the reported exact likelihood is literally exact or an approximation; the paper describes the layer only as a normalization layer and gives no explicit Jacobian formula."],"forward_implications":["If ConfFlow's reported results hold, large-molecule conformation ensembles become much cheaper to produce: sampling is a single forward ODE integration per graph after a one-time training cost.","Distance-based pipelines lose accuracy when converting predicted distances into 3D coordinates; ConfFlow's direct coordinate optimization removes that conversion step, so the advantage should grow with molecule size, matching where the paper reports its best margins.","Better conformations feed directly into downstream property prediction: on the paper's GEOM-Drugs property task, ConfFlow has the lowest median absolute error among learned methods on four of six ensemble properties.","Transformer-based message passing gives the flow a nonlocal receptive field per block, allowing long-range interactions in large molecules to be captured without explicitly adding long-range edges.","Because the iterative updates resemble force-field relaxation steps, the intermediate states of the flow can be read as a trajectory of how a conformer refines, not only as a final sample."],"supporting_citations":[{"why":"Supplies the GEOM dataset of curated drug and QM9 conformations on which ConfFlow is trained and benchmarked.","marker":"Axelrod & Gomez-Bombarelli (2020)"},{"why":"Introduces neural ODEs and the adjoint method that let GCF blocks be integrated, inverted, and backpropagated through.","marker":"Chen et al. (2018)"},{"why":"Provides free-form continuous dynamics and the Hutchinson trace estimator that justify tractable likelihoods for unrestricted vector fields.","marker":"Grathwohl et al. (2019)"},{"why":"Defines the Point Transformer architecture whose self-attention over 3D points becomes the GCPT message-passing layer.","marker":"Zhao et al. (2021)"},{"why":"Supplies the message-passing neural network framework for node and edge updates that the graph-conditioned flow extends.","marker":"Gilmer et al. (2017)"},{"why":"Introduces kinetic-energy and Jacobian-Frobenius regularization, which ConfFlow uses to stabilize ODE training on large molecules.","marker":"Finlay et al. (2020)"},{"why":"Supplies the ConfGF baseline, the strongest learning-based comparator on GEOM-QM9 and the reference for gradient-field distance methods.","marker":"Shi et al. (2021)"},{"why":"Supplies CGCF as a distance-based CNF baseline and defines the COV, MAT, and MIS evaluation protocol used in the main results.","marker":"Xu* et al. (2021)"},{"why":"Provides the GraphDG baseline, the auxiliary-edge strategy for fixing bond angles and dihedrals, and the distance-distribution evaluation task.","marker":"Simm & Hernández-Lobato (2020)"}],"fun_headline_variants":["Transformer flow boosts conformer accuracy up to 40% on large molecules","ConfFlow samples coordinates directly, no constraints, 40% better accuracy","Flow-based conformer generation: coordinate sampling yields 40% gain","Transformer flow for conformers: 40% accuracy jump, no explicit physics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact-likelihood machinery assumes the normalization layer that makes the model translation-invariant is a genuinely invertible transformation; if it only subtracts the molecule's center of mass, it is not one-to-one on the full coordinate space and the computed likelihoods are not exact.","fun_headline_variants_meta":{"raw":{"variants":["Transformer flow boosts conformer accuracy up to 40% on large molecules","ConfFlow samples coordinates directly, no constraints, 40% better accuracy","Flow-based conformer generation: coordinate sampling yields 40% gain","Transformer flow for conformers: 40% accuracy jump, no explicit physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3413,"prompt_tokens":853,"completion_tokens":2560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":2481}},"tokens_in":469,"tokens_out":2560,"duration_ms":17299,"temperature":1.0,"reasoning_tokens":2481,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:16:16.565173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a two-atom molecule, inspect the normalization layer in the released code and compute its Jacobian: if the layer subtracts the center of mass, the Jacobian is rank-deficient on $\\mathbb{R}^{3M}$, so the change-of-variables log-determinant in Eq. (2) is not exact and the reported likelihood-based training becomes an approximation.","supporting_citations":[{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"Introduces neural ODEs and the adjoint method that let GCF blocks be integrated, inverted, and backpropagated through."},{"cited_title":"Ffjord: Free-form continuous dynamics for scalable reversible generative models","cited_arxiv_id":null,"evidence_quote":"Provides free-form continuous dynamics and the Hutchinson trace estimator that justify tractable likelihoods for unrestricted vector fields."},{"cited_title":"Point transformer","cited_arxiv_id":null,"evidence_quote":"Defines the Point Transformer architecture whose self-attention over 3D points becomes the GCPT message-passing layer."},{"cited_title":"Neural message passing for quantum chemistry","cited_arxiv_id":null,"evidence_quote":"Supplies the message-passing neural network framework for node and edge updates that the graph-conditioned flow extends."},{"cited_title":"How to train your neural ode: the world of jacobian and kinetic regularization","cited_arxiv_id":null,"evidence_quote":"Introduces kinetic-energy and Jacobian-Frobenius regularization, which ConfFlow uses to stabilize ODE training on large molecules."}],"review_version":1}