acceptodds
Under review as a conference paper at ICLR 2027

TangentMol: Local Tangent Consistency for Few-Step 3D Molecular Generation

Abstract

Diffusion and flow-based models generate 3D molecules by gradually denoising random initial structures over many steps, jointly refining atomic coordinates and chemical structure (atom types, bonds, and charges). Reducing the number of steps lowers sampling cost, but each step must then make a large update, and sample quality degrades when the model's predictions change sharply across noise levels. Consistency training mitigates this for continuous data by encouraging agreement between predictions at nearby noise levels, but it does not naturally extend to chemical structure, which changes through discrete jumps rather than along a continuous path. In this study, we introduce TangentMol, a teacher-free training objective that keeps predictions of both coordinates and chemical structure consistent across nearby noise levels. From the noisy molecule already used in standard training, a single Jacobian–vector product computes how each prediction changes along a shared coordinate–time direction. We use the coordinate tangent to form a regression target and the categorical tangents to form soft probability targets for atom, bond, and charge predictions. These terms are added to the standard denoising loss and leave the network architecture and sampler unchanged. Compared with the same-configuration VEDA-FS baseline, TangentMol reduces Fréchet ChemNet Distance from 4.15 to 0.99 on QM9 (8 steps) and from 28.60 to 13.50 on GEOM-Drugs (16 steps), while raising 4-step QM9 connected validity from 53.3% to 79.3%. At 8 steps on GEOM-Drugs, TangentMol surpasses 16-step VEDA, FlowMol2, and FlowMol3 in connected validity, molecule stability, FCD, and strain energy. In pocket-conditioned generation on CrossDocked, it raises 3D validity from 45.3% to 62.8% at 10 steps over PocketVE.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D Molecular Generation, Few-Step Generation, Consistency Models, Diffusion Models, Structure-Based Drug Design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dx4fOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TangentMol (Local Tangent Consistency) for few-step 3D molecular generation. It addresses the inconsistency between continuous coordinates and discrete chemical structures (atom types, bonds, charges) during large single-step updates in diffusion models. Validation on GEOM-Drugs and CrossDocked shows that TangentMol (teacher-free, solver-free) significantly improves chemical validity, stability, and distribution quality, outperforming baselines with strong ablation support.

Critical strengths

The paper crisply identifies the continuous-discrete consistency gap as the dominant failure mode in few-step 3D molecular generation and resolves it through a unified local tangent objective obtained from a single joint JVP, pairing Euclidean regression for coordinates with a logit-shift softmax target for the three categorical fields. Because the formulation is a pure training-time additive loss that leaves architecture, backbone, and sampler untouched at inference, the reported gains admit a clean attribution absent from most concurrent diffusion-Mol works. This is borne out empirically: TangentMol retains state-of-the-art connected validity, molecular stability, and FCD at the practically relevant extreme operating points (4/8/16 NFE on GEOM-Drugs; 10–30 steps pocket-conditioned on CrossDocked), precisely where VEDA, FlowMol2, and FlowMol3 collapse, while preserving geometric fidelity (strain energy, bond-length alignment) instead of sacrificing it for discrete correctness.

Critical weaknesses

A notable strength of TangentMol is that the tangent loss is presented as a plug-and-play regularizer, which makes the broad framing in the introduction—covering both equivariant diffusion and flow-matching generators—natural and appealing. However, the current experimental suite does not yet substantiate the full scope of that claim: every TangentMol-trained model uses VE-diffusion backbones (VEDA-FS / PocketVE), while FlowMol-family models are only evaluated as unchanged external baselines. The authors acknowledge that porting to flow-matching requires adapting the shared JVP direction and the categorical transition targets. I would encourage minimal series experiments in the revision (or a clear timeline in the rebuttal) that includes the flow-matching instantiation trained with the proposed loss under matched settings.

Questions for authors

Q1. Hyperparameter selection protocol

The paper reports λ=1.0\lambda = 1.0 and a=0.1a = 0.1 (with RMS normalization) as the optimal setting, obtained from a 4×44 \times 4 grid over QM9 4-step. Was this grid evaluated on a proper held-out validation split, or directly on the same test split that produces the main Table 1 numbers?

Q2. Transfer assumption for flow-matching

Does the current TangentMol instantiation rely on VE-specific properties, or is the 'plug-and-play' design expected to reuse the same RMS normalization (a=0.1a = 0.1) for flow-matching backbones?

Q3. Revision suggestion

I encourage including flow-matching training (not just baselines) in the revision.

1 comment

Thank you for the authors' response. My concerns have been addressed, and I will raise the score.

Reviewer_Nq3HOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces TangentMol, a training objective for improving few-step 3D molecular generation. The method uses a shared coordinate–time Jacobian–vector product to construct consistency targets for continuous coordinate predictions and discrete chemical categories through their logits. These objectives supplement standard denoising training without changing the inference architecture or sampler. Experiments demonstrate substantial improvements at small inference budgets in both unconditional and pocket-conditioned molecular generation.

Critical strengths
  • The problem is timely and of broad interest. The approach provides an effective way to improve molecular generation when only a small number of function evaluations is available.
  • The joint treatment of continuous and categorical outputs is an interesting methodological contribution with potential applicability beyond the evaluated models.
  • The evaluation is extensive, with appropriate baselines and informative ablations. The matched-configuration baseline is particularly helpful for attributing improvements to the proposed objective.
  • The appendices provide substantial methodological detail and additional evaluations.
  • The paper is generally clear, and its limitations are appropriately acknowledged.
Critical weaknesses
  • All weaknesses are minor.
  • The intuition behind the tangent-derived targets could be communicated more effectively. A visual illustration showing how the local derivative constructs a target and encourages consistency across noise levels would complement the existing text.
  • Benefits at larger inference budgets are mixed: improvements in distributional fidelity can coexist with worse validity or geometric quality. The practical value when inference cost is not a constraint could be discussed more explicitly.
  • Broader applicability remains to be established, as the experiments focus on variance-exploding diffusion denoisers. This is an acknowledged future direction.
  • A reference for the FlowMol2 baseline appears to be missing.
Questions for authors

When inference cost is not a primary constraint, under what conditions would you recommend TangentMol over the base model?

Could you add a visual explanation of how the tangent-derived target encourages consistency and why this supports larger sampling steps?

Reviewer_SW4bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TangentMol aims to generate 3D molecules with fewer sampling steps. It adds a training penalty that keeps predicted geometry, atom types, and bonds consistent under small changes in noisy coordinates and noise level. The network and sampling procedure remain unchanged.

Critical strengths
  1. The method improves few-step molecular generation by adding a training objective while keeping the architecture and sampler unchanged.
  2. Matched baselines show substantial practical gains. Additional controls suggest that extra training time alone does not explain the improvement.
Critical weaknesses
  1. Limited methodological novelty. The coordinate objective follows sCM, while the categorical loss and certainty regularizer draw on dParallel. The contribution is primarily their adaptation to molecular generation.
  2. The consistency objective misses actual categorical transitions. Atom and bond inputs are held fixed when computing derivatives, although they change during sampling. The paper does not establish that the resulting penalty improves consistency across those changes.
  3. The categorical extension has mixed benefits. Much of the improvement comes from coordinate consistency and certainty regularization. Adding the categorical loss can improve FCD while reducing validity and stability, and missing GEOM-Drugs controls leave its independent value unclear.
Questions for authors
  1. What does the shared coordinate-time direction contribute beyond applying existing coordinate and categorical consistency losses separately?
  2. Does the categorical loss improve prediction consistency when atom and bond types actually change during sampling?
  3. On GEOM-Drugs, can you isolate the benefits of the categorical loss and certainty regularization, report uncertainty, and explain why improved FCD sometimes comes with lower validity and stability?
4 comments

Thank you for the review and constructive feedback.

Q1 (what the shared direction adds). The coordinate sCT loss follows sCM, while the certainty regularizer is an auxiliary term adapted from dParallel. Our proposed logit sCT loss extends tangent supervision to categorical predictions. Specifically, we construct a centered, RMS-normalized logit tangent along the same coordinate–time direction used for geometric supervision and use it to shift the reference logits. This couples geometric and categorical supervision through a single JVP covering both coordinate and categorical heads. A matched time-only categorical control remains untested.

Q2 (consistency at categorical updates). We compared X (VEDA-FS + coordinate sCT) with X+Z (VEDA-FS + coordinate sCT + logit sCT) at entries whose categorical input changed. Each model followed its own trajectory from the same 1,000 initial validation states per dataset. We measured clean-category argmax agreement and Jensen–Shannon divergence (JS, in nats using natural logarithms) for 8/16-step grids from t0∈{0.45,0.65,1.0}t_0\in\{0.45,0.65,1.0\}. Each cell gives the min–max difference (X+Z minus X) across starting times (t0t_0).

DatasetGridAtom agreement Δ ↑Atom JS Δ (×10⁻³ nats) ↓True bond agreement Δ ↑True bond JS Δ (×10⁻³ nats) ↓
QM98+1.89 to +2.37−3.55 to −3.24−0.61 to +0.92−0.88 to +0.17
QM916+1.08 to +1.23−1.88 to −1.67−0.39 to +0.30+0.30 to +0.49
GEOM-Drugs8+2.96 to +2.99−5.20 to −4.03−0.95 to −0.44+1.86 to +2.57
GEOM-Drugs16+1.92 to +2.16−4.42 to −3.41−0.56 to −0.16+1.00 to +1.31

Atom agreement increases and JS decreases in all 12 settings, while bond effects are mixed on QM9 and both bond metrics worsen on GEOM-Drugs. We will ablate the bond logit-sCT weight on QM9 within the discussion period.

Full conditionwise estimates and paired molecule-bootstrap 95% intervals will be included in the appendix.

Q3 (component effects). We completed the QM9 factorial with the missing X+C and Z+C configurations. C denotes certainty regularization. The table isolates Z with C held fixed, and Table A5 reports the remaining configurations.

QM9 model4 NFE Conn. validity (%)4 NFE Mol. stability (%)4 NFE FCD8 NFE Conn. validity (%)8 NFE Mol. stability (%)8 NFE FCD
VEDA-FS53.349.16.5492.791.64.15
C66.562.05.9694.493.44.08
Z+C64.060.75.0293.893.04.13
Change from adding Z−2.5−1.3−0.94−0.6−0.4+0.05
X+C77.676.41.3893.094.21.03
X+Z+C (TangentMol)79.378.11.4193.394.10.99
Change from adding Z+1.7+1.7+0.03+0.3−0.1−0.04

On QM9, the factorial study shows complementary effects between Z and coordinate supervision. At 4 NFE, adding Z to X+C improves connected validity and molecular stability by 1.7 pp each. Adding Z to C alone improves FCD (5.96 → 5.02) while reducing connected validity (66.5% → 64.0%). The contribution of Z therefore depends on the accompanying objectives.

On GEOM-Drugs, we isolate Z given X and C given X+Z. At 8 NFE, adding Z improves connected validity (47.6% → 52.8%), molecular stability (47.2% → 54.3%), FCD, and strain. At 16 NFE, Z improves FCD (14.5 → 12.2) while reducing connected validity (81.5% → 76.8%). Adding C then raises connected validity to 85.4%, with FCD at 13.5. Relative to X, the combined model gains 3.9 pp in connected validity and lowers FCD by 1.0. The 16-NFE result shows practical complementarity between Z and C when coordinate supervision is present. Certainty regularization recovers validity while preserving an FCD gain over X.

We have completed the bond logit-sCT ablation on QM9. We set only the bond logit-sCT weight to zero, retaining standard denoising loss, atom and charge logit-sCT, coordinate sCT, and certainty regularization. We evaluated 10000 molecules per NFE using the paper's generation protocol.

NFEFull model Conn. validity (%)Without bond sCT Conn. validity (%)Full model Mol. stability (%)Without bond sCT Mol. stability (%)Full model FCDWithout bond sCT FCD
479.377.5978.175.731.411.253
893.392.3494.193.850.990.969
1696.196.5696.597.231.571.533
10098.899.0398.299.052.302.338

Full-model values are the three-run means reported in Table A5. The new ablation uses one training run.

Bond logit-sCT helps few-step generation: at 4 NFE, it gives higher connected validity and molecular stability, and both remain higher at 8 NFE. FCD is lower without bond sCT at 4 to 16 NFE. We will include this ablation in the revised appendix.

Dear Reviewer SW4b, Thank you for your thoughtful feedback. As the ICLR author–reviewer discussion closes today (November 18, AoE), we wanted to follow up on our responses. We have addressed each of your comments and completed all the analyses you requested.

We would be grateful if you could take these results into account as you finalize your assessment. If any questions or concerns remain, we would be happy to address them before the discussion closes.

Best, The Authors

Reviewer_SW4bOpenReview ↗

Response after author rebuttal

The new sampling diagnostics and model transfer experiment improve its practical contribution, but the novelty remains modest, and the benefit of the categorical loss is still unclear.

Comments on the paper

1 comment

Dear reviewers,

As the authors have posted their responses, I would encourage you to actively parcitipate in the dicussion.

Thank you.

Best,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.