{"id":"bbfba32b-5e66-4360-b9b9-95badef7483c","arxiv_id":"2412.11082","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"EquiFlow applies equivariant conditional flow matching with optimal transport to predict 3D molecular conformations, reporting improved RMSD and coverage on QM9 datasets.","lead":"EquiFlow trains an equivariant neural network to move random 3D point clouds into molecular conformations using optimal transport and flow matching. On QM9 and GEOM-QM9 benchmarks it reports better average conformer accuracy than several prior deep learning and cheminformatics methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper asserts rotational invariance from Kabsch alignment but never proves the OT coupling or the conditional target u_t is equivariant; under the natural simultaneous-rotation action it likely is, but the missing proof leaves the training objective's consistency unverified.","rationale":"The reader's weakest assumption correctly identifies a missing proof: the paper never demonstrates that the Kabsch-based OT pairing yields an equivariant conditional target. This is a genuine soft spot because the entire method relies on training an equivariant network to regress u_t. However, analysis shows that the property holds if the group action is taken to be simultaneous rotation of x0 and x1: the Kabsch-aligned RMSD cost is invariant under such rotations, the OT assignment commutes with the action, and the resulting target rotates covariantly. So the concern is not that the method is wrong, but that the paper has not established it and the surrounding text is ambiguous enough to invite an incorrect implementation (e.g., rotating x1 in the target rather than in the cost). A direct numerical verification would settle the matter. Other issues the reader raised — no code, no error bars, no single-conformation baseline — remain valid but are secondary to whether the training objective is internally consistent. The verdict should stay CONDITIONAL: the authors should supply the equivariance proof or the numerical check, along with the missing artifacts, before the central accuracy claim can be fully accepted.","tokens_in":13346,"tokens_out":20196,"duration_ms":183080,"concrete_test":"Write a small numerical check using the paper's own procedure: for a molecule with K atoms, sample x0 ~ N(0,1) and a set of conformations x1; apply zero CoM to both, compute the Kabsch-aligned RMSD cost matrix, solve the OT assignment M, and form u_t = x1_j - x0_i and x_t = x0_i + t u_t for t in {0, 0.5, 1}. Then apply a random rotation R to both x0 and x1 (retaining the same pairing indices), recompute the cost matrix and OT assignment, and verify that (i) the assignment M is identical and (ii) the new target equals R u_t up to numerical tolerance. If either check fails, the training objective is not equivariant under the simultaneous-rotation action and the reported results cannot be attributed to a consistent equivariant flow-matching training procedure.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central accuracy claim depends on the flow-matching training objective being consistent with the equivariant network. In Algorithm 1, the target is u_t = x1_j - x0_i, where the pairing (i,j) is found by solving an OT problem with a Kabsch-aligned RMSD cost. For an equivariant model to fit, the joint distribution of (x_t, u_t) must be invariant under the appropriate group action, and the target must transform covariantly. The paper never states this action explicitly. If the action is taken to be a rotation R applied simultaneously to x0 and x1, then the Kabsch cost is invariant, the pairing M is unchanged, and u_t rotates as R u_t; so the objective is equivariant. However, the text says 'x1 is rotated using the Kabsch algorithm' in the cost computation, while Algorithm 1 then uses the unrotated x1_j in u_t. If an implementation instead uses the Kabsch-rotated x1_j in the target, or computes the cost without internally centering, the target may not be equivariant. Because no proof or code is provided, the consistency of the training signal for an equivariant network — and therefore the reported COV/MAT numbers — remains unverified. This is the load-bearing assumption: if it fails, the equivariant network cannot fit the objective and the central claim collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"EquiFlow applies conditional flow matching with optimal transport (OT-CFM) to 3D molecular conformation prediction. The training procedure pairs Gaussian noise conformations with true conformations using an OT cost based on Kabsch-aligned RMSD, then regresses a modified EquiformerV2 network to the straight-line conditional vector field u_t = x1_j - x0_i. The paper reports state-of-the-art or competitive multi-conformation results on GEOM-QM9 (COV-R 95.9%, MAT-R 0.130 Å, MAT-P 0.164 Å) and a single-conformation QM9 RMSD of 0.17 Å, and argues that the ODE-based sampler is faster than SDE-based diffusion.","tokens_in":13642,"tokens_out":5430,"duration_ms":52890,"significance":"If the reported results are reproducible and the equivariance issue below is resolved, EquiFlow would be a solid contribution: it combines the efficient training of CFM with a high-degree equivariant backbone, and the use of a Kabsch-aligned RMSD cost in the OT coupling is a sensible adaptation to molecular symmetry. The empirical numbers on GEOM-QM9 are competitive with strong published baselines. However, the manuscript currently lacks a proof that the OT-CFM training objective is SE(3)-equivariant, reports no variance estimates or code, provides no runtime measurements despite speed being a central motivation, and presents the single-conformation claim without baselines. These gaps prevent a full verification of the central accuracy and efficiency claims.","major_comments":[{"comment":"The paper asserts that Kabsch alignment ensures rotational invariance, but it never states the group action or proves that the OT pairing M and the target vector field u_t = x1_j - x0_i transform covariantly under SE(3). This is load-bearing: if the pairing or target is orientation-dependent, the equivariant network cannot fit a consistent objective and the reported accuracy numbers are not explained. The natural proof (simultaneously rotating both point clouds leaves the Kabsch cost invariant, so M is unchanged, and u_t rotates as R u_t) is not written down, and Algorithm 1 is ambiguous about whether the Kabsch-rotated x1 or the original x1 is used in Eq. (11). Please add a formal equivariance statement for the OT-CFM coupling and specify exactly which coordinates enter the target. Also reconcile the order of centering: §3.2 says Zero CoM is applied before computing the Kabsch cost, while Algorithm 1 centers only after OT.","section":"§3.2, Algorithm 1 lines 6-9"},{"comment":"The GEOM-QM9 baseline numbers are directly taken from Jing et al. (2022), yet the manuscript modifies the dataset by grouping by SMILES and removing invalid conformations. Unless this filtered test set is provably identical to the one used by Torsional Diffusion, the comparison is not valid; the paper should state the overlap or rerun the baselines on the same filtered set. In addition, the headline single-conformation RMSD of 0.17 Å on QM9 is reported with no comparison table, no description of the exact protocol beyond the split, and no baseline numbers, so this central claim is currently unsupported.","section":"Experiments, Table 1 and single-conformation paragraph"},{"comment":"No error bars, seeds, or confidence intervals are provided for any EquiFlow result, and no code is released. Since EquiFlow's COV-P (91.8%) is below Torsional Diffusion's (92.7%) and several metric differences are small, it is impossible to assess whether the reported improvements are statistically significant. Please report results over multiple seeds with mean and standard deviation, and release code or at least a detailed configuration to make the numbers reproducible.","section":"Experimental Setups / Experimental Results"},{"comment":"The stated motivation is avoiding slow training and enabling faster inference than diffusion models with SDEs, but no runtime measurements are reported for training or sampling, and no ODE-vs-SDE inference-time comparison is given. If speed is one of the two central claimed advantages, it must be measured; otherwise the speed claims in the abstract and introduction should be removed or substantially softened.","section":"Abstract and Introduction"}],"minor_comments":[{"comment":"The caption refers to 'ConfromFlow' instead of 'EquiFlow'.","section":"Figure 1 caption"},{"comment":"The text says 'MES loss function' where it should say 'MSE loss function'.","section":"Algorithm 1 / Training Procedure"},{"comment":"The COV-R and COV-P definitions use xp without an explicit existential quantifier or set-builder binding; as written, xp appears as a free variable. Please clarify that the inner condition is 'there exists xp in Sp' (or equivalently 'min over xp').","section":"Equations (12)-(15)"},{"comment":"The notation C(x0_i, x1_j) does not indicate that x1 is Kabsch-rotated relative to x0 before the RMSD is computed; this is a source of the ambiguity noted in the major comment and should be made explicit in the equation or surrounding text.","section":"§3.2, Equation (7)"},{"comment":"The main text says the straight-line OT approach 'may pose issues in 3D molecular conformation prediction due to spatial symmetry (see details in Appendix A)', but Appendix A is a review of SE(3) equivariance and tensor products; it does not directly discuss why straight-line OT fails under molecular symmetry. Please connect the arguments more explicitly.","section":"Appendix A reference"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the empirical results are promising, but the manuscript is not yet verifiable: the equivariance proof is missing, the experimental comparisons are incomplete, and the speed claims are unmeasured. I do not see a fatal internal contradiction, so major revision rather than rejection seems appropriate. The authors should also confirm that their GEOM-QM9 filtered test set is identical to the one used by the baselines whose numbers they reproduce."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. EquiFlow reports the best COV-R, MAT-R, and MAT-P on GEOM-QM9, with COV-R 95.9 against 92.8 for Torsional Diffusion and MAT-R 0.130 against 0.178. That is a real margin. The second thing is that the result is conditional: the paper never proves the training target is equivariant, and there is no code or error bars behind the table.\n\nWhat is new: the first application of OT-CFM to 3D conformation prediction, built on a modified EquiformerV2 with time embedding, bond features, and a 1-degree prediction head. These are incremental but sensible changes. The evaluation uses the standard Torsional Diffusion split, and the comparison table is the main evidence. The single-conformation RMSD of 0.17 Å on QM9 appears without a comparison table, so it is hard to interpret.\n\nThe soft spots are real but not fatal. The equivariance of the OT target is asserted, not established. Under the natural action where the same rotation is applied to both x0 and x1, the Kabsch cost is invariant, so the pairing is unchanged and ut transforms covariantly. But the paper never states that action, and Algorithm 1 is ambiguous: it says x1 is Kabsch-rotated for the cost, then uses unrotated x1_j in ut. If an implementation feeds the rotated coordinates into the target, the objective is not equivariant. Without code or a proof, the central accuracy claim rests on an unverified assumption. The fix is straightforward: state the group action, give a short invariance argument, and release the code. The other issues are minor: no error bars, no runtime measurements despite the speed claims, and a few typos (MES loss, ConfromFlow in Figure 1).\n\nWho should read this: people working on generative models for small-molecule conformations. It is a legitimate new application with strong benchmark numbers. The paper deserves a serious referee, but should not be accepted as is. The referee should demand a proof or clear derivation of the equivariance, error bars, and a code release. If those check out, the results are likely solid.","headline":"Plausible new SOTA on GEOM-QM9 from OT-CFM, but missing equivariance proof and missing code keep the result conditional.","tokens_in":14160,"tokens_out":3710,"would_cite":false,"duration_ms":28568,"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":"EquiFlow predicts 3D molecular conformations more accurately than prior models by pairing conditional flow matching with optimal transport.","keywords":["molecular conformation prediction","conditional flow matching","optimal transport","equivariant graph neural network","SE(3) equivariance","Equiformer","QM9","GEOM-QM9"],"falsifier":"Train EquiFlow on GEOM-QM9 twice: once with conformations in their native orientations and once with every molecule randomly rotated before both training and evaluation. If the OT pairing is not rotation-covariant, the rotated run should show materially worse COV-R and MAT-R; additionally, with a fixed-orientation test, a random global rotation of a test molecule should not change the RMSD of the predicted conformation to the reference after re-alignment, and any such change would indicate broken equivariance.","tokens_in":13167,"feed_emoji":"🧪","tokens_out":7142,"duration_ms":55889,"temperature":0.7,"pith_summary":"The paper proposes EquiFlow, a model that predicts 3D molecular conformations by learning a conditional flow from Gaussian noise on atomic coordinates to real conformations, with the flow's target vector field chosen by optimal transport. It claims this is the first application of conditional flow matching to molecular conformation prediction, and that the simulation-free training plus an ODE solver makes it faster to train and sample than diffusion models. On the GEOM-QM9 benchmark, EquiFlow reports higher coverage (COV-R 95.9 vs 92.8) and lower matching error (MAT-R 0.130 Å vs 0.178 Å) than the previous best deep-learning method, and 0.17 Å RMSD on single-conformation QM9 prediction. A careful reader would care because molecular geometry governs binding and reactivity, and a fast, accurate predictor would be directly useful in drug and materials design.","feed_headline":"Flow matching predicts molecular 3D shapes more accurately","feed_subtitle":"EquiFlow pairs equivariant flow matching with optimal transport to beat prior models on QM9 and GEOM-QM9.","key_machinery":"The load-bearing object is the equivariant conditional flow matching objective with optimal transport. A Gaussian point cloud $x_0$ is paired with a real conformation $x_1$ by solving a minibatch optimal transport problem with RMSD cost $C(x_{0i}, x_{1j}) = \\sqrt{\\frac{1}{K}\\sum_k \\|x_{0ik}-x_{1jk}\\|^2}$ after zeroing center of mass and rotationally aligning the structures; the conditional target is the straight-line vector field $u_t = x_1 - x_0$ along $x_t = x_0 + t u_t$. A modified Equiformer, a Transformer-style equivariant graph neural network built on SO(3) irreducible representations, regresses this vector field from atom types, bond types, relative positions, and time, using a degree-1 (vector) prediction head. The paper's argument is that this pairing keeps the training target compatible with SE(3) equivariance while giving straighter, easier-to-learn probability paths than diffusion.","core_discovery":"The central discovery claimed is that conditional flow matching with an optimal-transport pairing is a better training objective for 3D conformation prediction than the diffusion and torsion-space objectives used previously. EquiFlow pairs each sampled Gaussian coordinate set $x_0$ with a true conformation $x_1$ by solving a transport problem whose cost is the rotation-aligned RMSD between conformations, then regresses the straight-line velocity $u_t = x_1 - x_0$ with a modified Equiformer network that represents atomic positions as degree-1 irreducible vectors and takes atom types, bond types, and time as inputs. In the reported experiments this reaches COV-R 95.9 and MAT-R 0.130 Å on GEOM-QM9, with COV-P 91.8 and MAT-P 0.164 Å, and 0.17 Å RMSD on QM9 single-conformation prediction. The paper presents these numbers as evidence that combining optimal transport with equivariant flow matching improves both diversity and accuracy over prior state-of-the-art models.","pith_inferences":["If the equivariance of the OT pairing holds, the same recipe should transfer to larger molecules and to conformer ensembles of protein fragments, since the architecture does not depend on molecule size; the paper only demonstrates up to QM9-sized heavy-atom counts.","The reported COV-P (91.8) is slightly below Torsional Diffusion (92.7), so the improvement is not uniform across all precision metrics; a practitioner interested in precision at the 0.5 Å threshold should look at per-molecule distributions before switching methods.","Because the model needs bond types, its applicability to conformer generation of molecules with unknown bond orders or reaction intermediates is limited unless bond assignment is provided.","A direct stress test would be to train EquiFlow on randomly rotated versus fixed-orientation conformations; if the Kabsch-based OT pairing is not truly rotation-covariant, the rotated training run should underperform."],"forward_implications":["Single-conformation prediction at 0.17 Å RMSD on QM9 implies the model is accurate enough for geometry-sensitive downstream tasks on small molecules, if the number reproduces under independent evaluation.","GEOM-QM9 COV-R of 95.9 and MAT-R of 0.130 Å would mean the generated ensembles cover essentially all reference conformations with lower error than torsion-based diffusion (92.8, 0.178 Å).","Because training is simulation-free and sampling uses an ODE, the same accuracy would come with faster training and inference than score-based diffusion; the paper states this speed advantage explicitly.","Using bond types and higher-degree equivariant features (up to degree 6) would show that chemistry-informed edge features and high-degree tensors help conformation prediction, not just property prediction.","The method is end-to-end for conformations given a 2D molecular graph, avoiding the precomputed rigid substructures required by torsion-based approaches."],"supporting_citations":[{"why":"Defines the flow-matching objective and straight-line conditional paths that EquiFlow adapts to conformations.","marker":"Lipman et al. 2023"},{"why":"Extends flow matching to conditional flow matching with minibatch optimal transport, the training scheme EquiFlow uses.","marker":"Tong et al. 2024"},{"why":"Provides the Torsional Diffusion baseline, GEOM-QM9 split, and evaluation metrics used in the comparison table.","marker":"Jing et al. 2022"},{"why":"Introduces GeoDiff baseline and the RMSD coverage/matching metrics plus zero center-of-mass convention that EquiFlow follows.","marker":"Xu et al. 2022"},{"why":"Supplies EquiformerV2, the equivariant Transformer backbone that EquiFlow modifies with time and bond embeddings and a degree-1 head.","marker":"Liao et al. 2024"},{"why":"Gives the optimal rotation alignment used inside the RMSD cost for the optimal-transport pairing.","marker":"Kabsch 1976"},{"why":"Provides the QM9 and GEOM-QM9 datasets used for training and evaluation.","marker":"Ramakrishnan et al. 2014"},{"why":"Equivariant flow matching for 3D molecule generation; the closest flow-matching baseline for 3D molecular geometry.","marker":"Song et al. 2024"}],"fun_headline_variants":["EquiFlow: equivariant flow matching with optimal transport for accurate conformers","Conditional flow matching with optimal transport improves 3D conformation prediction","EquiFlow beats diffusion models for 3D molecular conformations","Simulation-free training makes 3D conformation prediction faster and more precise","Flow matching + optimal transport: new state of the art for 3D conformers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the optimal-transport pairing computed with rotation-aligned RMSD cost remains rotation- and translation-covariant, so the straight-line target field $u_t = x_1 - x_0$ can be learned by an SE(3)-equivariant network; the paper asserts this but does not prove it.","fun_headline_variants_meta":{"raw":{"variants":["EquiFlow: equivariant flow matching with optimal transport for accurate conformers","Conditional flow matching with optimal transport improves 3D conformation prediction","EquiFlow beats diffusion models for 3D molecular conformations","Simulation-free training makes 3D conformation prediction faster and more precise","Flow matching + optimal transport: new state of the art for 3D conformers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2671,"prompt_tokens":903,"completion_tokens":1768,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":1684}},"tokens_in":519,"tokens_out":1768,"duration_ms":12319,"temperature":1.0,"reasoning_tokens":1684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:18:35.371451+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EquiFlow on GEOM-QM9 twice: once with conformations in their native orientations and once with every molecule randomly rotated before both training and evaluation. If the OT pairing is not rotation-covariant, the rotated run should show materially worse COV-R and MAT-R; additionally, with a fixed-orientation test, a random global rotation of a test molecule should not change the RMSD of the predicted conformation to the reference after re-alignment, and any such change would indicate broken equivariance.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends flow matching to conditional flow matching with minibatch optimal transport, the training scheme EquiFlow uses."},{"cited_title":"M.; Das, A.; and Smidt, T","cited_arxiv_id":null,"evidence_quote":"Supplies EquiformerV2, the equivariant Transformer backbone that EquiFlow modifies with time and bond embeddings and a degree-1 head."},{"cited_title":"O.; Rupp, M.; and Von Lilienfeld, O","cited_arxiv_id":null,"evidence_quote":"Provides the QM9 and GEOM-QM9 datasets used for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Equivariant flow matching for 3D molecule generation; the closest flow-matching baseline for 3D molecular geometry."}],"review_version":1}