acceptodds
Under review as a conference paper at ICLR 2027

MolLIFT: Lifting SMILES Priors into Local Frames for 3D Molecule Generation

11 views

Abstract

Pocket-conditioned 3D molecule generation requires capturing both protein–ligand geometry and the chemical regularities underlying molecular structure. While 3D representations naturally describe conformations and spatial interactions, SMILES provides complementary information about molecular connectivity and chemical organization. We propose MolLIFT, a bidirectional 1D–3D Bayesian flow framework that incorporates evolving SMILES priors into pocket-conditioned 3D generation while preserving SE(3)-equivariance. MolLIFT constructs geometry-aware atom representations from local protein–ligand environments and uses atom-centered local frames to lift invariant SMILES-derived messages into explicit equivariant coordinate updates. The evolving 3D geometry is then projected back into the same local frames to provide orientation-invariant feedback for refining the SMILES belief, forming a closed bidirectional interaction throughout generation. On the CrossDocked benchmark, MolLIFT improves binding affinity, molecular quality, diversity, and success rate over the underlying property-agnostic generator. Ablation studies show that these gains stem from geometry-aware cross-modal interaction and explicit coordinate lifting rather than increased model capacity. The joint 1D–3D state further enables gradient-guided multi-objective optimization, yielding additional improvements in affinity and molecular quality. Our results establish local-frame lifting as an effective mechanism for transferring sequence-level chemical priors into equivariant 3D molecular generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-based drug design, 3D molecule generation, Bayesian flow networks, Cross-modal learning.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZxHMOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2(−1)
Read review
Paper summary

This paper presents MolLIFT, a joint 1D–3D Bayesian Flow framework for pocket-conditioned molecular generation. The proposed method introduces a local-frame lifting mechanism that transforms SMILES-derived features into explicit SE(3)-equivariant coordinate corrections. It further incorporates a 3D-to-SMILES feedback pathway to enable bidirectional interaction during generation. Experiments on CrossDocked demonstrate competitive improvements in binding affinity, molecular properties, and diversity.

Critical strengths
  1. Geometrically motivated cross-modal architecture. The proposed Local-Frame Lifting (LFL) mechanism provides an interesting approach to incorporating 1D chemical information into 3D molecular generation. Unlike conventional scalar feature fusion, it introduces an explicit equivariant vector pathway for coordinate correction. The mathematical analysis of SE(3)-equivariance is also a strength of the paper.

  2. Competitive experimental performance. MolLIFT achieves promising results across several molecular generation metrics. The matched-capacity comparison and existing ablation studies provide useful evidence for the benefits of geometry-aware interaction and explicit coordinate lifting.

Critical weaknesses
  1. Insufficient ablation of bidirectional interaction. The Invariant 3D-to-SMILES Feedback (IFB) module is an important component of the proposed bidirectional framework. However, the paper does not independently evaluate its contribution. Without a dedicated w/o IFB ablation, it remains unclear how much the reverse feedback pathway contributes beyond the SMILES-to-3D interaction.

  2. Limited evidence for meaningful chemical–geometric correspondence. Although the proposed method improves generation performance, the evidence that LFL learns meaningful relationships between SMILES information and 3D coordinate corrections remains incomplete. Appendix F describes shuffled-SMILES and token-permutation controls, but Figure 5 only reports correct-versus-null comparisons. These additional results would help establish whether the improvements depend on chemically relevant sequence information rather than simply the presence of an additional conditioning pathway.

Questions for authors

Q1. Contribution of IFB Could the authors provide a dedicated w/o IFB ablation to clarify whether bidirectional interaction provides measurable improvements over one-way SMILES-to-3D guidance?

Q2. Chemical–geometric correspondence Could the authors report the shuffled-SMILES and token-permutation experiments described in Appendix F? In particular, it would be valuable to demonstrate that chemically correct SMILES information leads to more accurate and directionally meaningful coordinate corrections than mismatched or permuted sequences. Additionally, could the authors clarify whether similar benefits are observed when using the model's own generated SMILES beliefs rather than ground-truth SMILES?

Q3(Minor). Methodological distinction Appendix E acknowledges that scalar cross-attention can preserve SE(3)-equivariance. Could the authors clarify the additional representational or empirical benefits of explicit local-frame vector lifting over scalar-conditioned equivariant generation? A more focused discussion of related approaches, including LiFT, NExT-Mol, and MolSculpt, would help position the contribution.

4 comments
AuthorsOpenReview ↗

Response to Reviewer ZxHM (Part 1/2)

We sincerely thank the reviewer for their time, careful assessment, and valuable feedback. We appreciate the recognition of MolLIFT's geometrically motivated design, equivariance analysis, and competitive performance. Please see our point-by-point responses below; revisions are marked in blue.

Q1: Could the authors provide a dedicated w/o IFB ablation to clarify whether the bidirectional interaction provides a measurable improvement over one-way SMILES-to-3D guidance?

A1: We added w/o IFB in Table 2 and Section 4.2(v). It disables only 3D-to-SMILES feedback while retaining local-frame-aware scalar interaction and the explicit SMILES-to-coordinate vector message.

MethodVina Score Avg./Med.Vina Min Avg./Med.QEDSA
w/o IFB-6.92 / -7.03-7.26 / -7.200.530.67
MolLIFT-7.01 / -7.12-7.37 / -7.290.540.68

Restoring IFB improves every reported metric. This isolated comparison shows that the forward pathway provides the larger performance gain, while invariant reverse feedback contributes an additional measurable improvement by refining the SMILES state used in subsequent flow steps.

Q2-1: Could the authors report the shuffled-SMILES and token-permutation experiments described in the appendix, and examine whether chemically meaningful sequence information produces more accurate and directionally meaningful coordinate corrections than mismatched or permuted sequences?

A2-1: Appendix H.1 and Figures 6–7 now report the random token-permutation and within-batch reassignment controls. In every paired evaluation, we hold the pocket, noisy 3D state, denoising time, coordinate and atom-type beliefs, and stochastic realization fixed. For a control condition bb, coordinate improvement is defined as Δb=Eb−EtrueΔ^b=E^b-E^{true}, while negative transfer is defined as NTRb=Pr(Etrue>Eb)NTR^b=Pr(E^{true}>E^b). Positive coordinate improvement and directional alignment, together with lower NTR, indicate more useful and reliable sequence guidance.

Randomly permuting the correct token order preserves sequence length and composition while disrupting molecular serialization. Its result is consistent with the null control: during most intermediate and late stages, MolLIFT produces larger coordinate improvements, stronger directional alignment, and lower NTR than w/o LF, while w/o Vec is generally intermediate. These results indicate that the learned coordinate guidance is sensitive to meaningful SMILES organization, rather than simply benefiting from an additional input branch.

Within-batch reassignment replaces the condition with another complete molecular sequence. It produces coordinate differences close to zero over most time bins, weak and non-monotonic alignment, and NTR near 0.5. This control changes molecular identity and composition simultaneously and may supply another chemically valid sequence containing transferable composition-level cues. It is therefore an instance-correspondence test rather than a guaranteed negative ligand. The token-permutation experiment provides the cleaner controlled evidence that the learned coordinate guidance depends on meaningful sequence organization. We therefore explicitly delimit the scope of our claim: these results support the utility of informative, correctly organized sequence guidance, but the current training objective does not explicitly enforce one-to-one ligand identity discrimination.

Q2-2: Could the authors clarify whether similar benefits are observed when the model uses its own generated SMILES belief states rather than ground-truth SMILES?

A2-2: Appendix H.2 and Figure 8 remove access to the reference SMILES. For 100 test molecules, the SMILES receiver starts from a uniform prior. At each flow step, the model predicts a soft clean-token distribution, which is incorporated into the current SMILES belief by the categorical BFN receiver. The updated belief is then passed to the next step. Consequently, the belief evaluated at time tkt_k contains the complete model-generated history from t=0t=0 to tkt_k. No reference token or discretized argmax SMILES is supplied, and the accumulated soft belief enters the 3D branch directly.

Against a null sequence on the same controlled reference-anchored 3D trajectory, MolLIFT gives positive coordinate improvement from t=0.25t=0.25 onward. Directional alignment rises from 0.065 at t=0.25t=0.25 to 0.818 at t=0.85t=0.85, while NTR falls from 0.46 to 0.00 and remains 0.02 in the final bin. W/o LF instead becomes negatively aligned after t=0.6t=0.6 and reaches 0.90 NTR in the final bin. These results demonstrate that path-consistent, model-generated SMILES beliefs provide useful directional guidance without ground-truth SMILES once sufficient geometric information has accumulated. We use the reference-anchored trajectory only to provide a clean coordinate target for this controlled measurement.

AuthorsOpenReview ↗

Response to Reviewer ZxHM (Part 2/2)

Q3-1: Could the authors clarify the additional representational and empirical benefits of explicit local-frame vector lifting over scalar-conditioned equivariant generation, given that scalar cross-attention can also preserve SE(3) equivariance?

A3-1: We agree that scalar cross-attention can preserve SE(3) symmetry; our distinction concerns how sequence information reaches coordinate space. In w/o LF, the SMILES message is incorporated only through invariant scalar atom features and influences coordinates indirectly through later equivariant layers. W/o Vec retains the local-frame-aware scalar interaction but removes the explicit coordinate message. Full MolLIFT additionally predicts invariant coefficients and maps them through an atom-centered equivariant frame using Δxt,iS→X=κFt,ict,iS\Delta\mathbf{x}_{t,i}^{S\rightarrow X}=\kappa\mathbf{F}_{t,i}\mathbf{c}_{t,i}^{S}, where ct,iS\mathbf{c}_{t,i}^{S} contains the invariant coefficients predicted from the SMILES message, Ft,i\mathbf{F}_{t,i} is the atom-centered local frame, and κ\kappa is a learnable scale.

Under a rigid transformation (R,b)(\mathbf{R},\mathbf{b}), the coefficients remain invariant and the local frame transforms as Ft,i′=RFt,i\mathbf{F}'_{t,i}=\mathbf{R}\mathbf{F}_{t,i}. Consequently, the complete coordinate message satisfies (Δxt,iS→X)′=RΔxt,iS→X(\Delta\mathbf{x}_{t,i}^{S\rightarrow X})'=\mathbf{R}\Delta\mathbf{x}_{t,i}^{S\rightarrow X}. Translation does not appear in this transformation because the coordinate message is a displacement vector constructed from relative geometry. The module therefore converts invariant sequence information into a direct SE(3)-equivariant displacement without introducing a global reference frame.

MethodCross-modal pathwayVina Score Avg.Vina Min Avg.QEDSA
w/o LFVanilla scalar attention-6.68-7.130.510.67
w/o VecLocal-frame-aware scalar fusion-6.85-7.250.520.68
MolLIFTScalar fusion + explicit vector lifting-7.01-7.370.540.68

The progression from w/o LF to w/o Vec to MolLIFT separates the two effects. Introducing local-frame-aware scalar fusion improves average Vina Score from -6.68 to -6.85 and Vina Min from -7.13 to -7.25. Restoring the explicit vector message further improves them to -7.01 and -7.37, respectively, while QED increases from 0.52 to 0.54. Thus, geometry-aware scalar fusion is useful on its own, and explicit vector lifting provides a further empirical benefit beyond scalar-conditioned equivariant generation. Appendix F now clarifies this distinction and provides the equivariance argument, including the degenerate-frame fallback.

Q3-2: Could the authors provide a more focused comparison with related 1D–3D approaches, including LiFT, NExT-Mol, and MolSculpt, to clarify the methodological positioning of MolLIFT?

A3-2: We expanded both the Introduction and Appendix A. NExT-Mol integrates a pretrained SELFIES language model with diffusion-based 3D conformation generation from molecular sequences. MolSculpt extracts chemical knowledge from a frozen molecular foundation model through learnable queries and injects it into a 3D diffusion model. LiFT incorporates language-derived SMILES priors into pocket-conditioned flow matching through semantic conditioning and adaptive velocity-field modulation. These methods establish the value of sequence priors for geometric generation.

MolLIFT focuses on a different cross-modal interface: the SMILES belief and 3D ligand states evolve jointly in one Bayesian flow, and information is exchanged bidirectionally at every generation step. The forward pathway uses the local protein–ligand environment to retrieve token information and converts invariant sequence-derived coefficients into explicit equivariant coordinate updates. The reverse pathway projects geometric progress into invariant local-frame features that refine the next SMILES belief. The combination of co-evolving states, direct local-frame coordinate lifting, and invariant reverse feedback is the methodological distinction now emphasized in the revision.

The authors have addressed most of my concerns through the additional experiments. Some limitations remain, particularly the lack of guaranteed consistency between the generated SMILES and the final 3D structures. Nevertheless, the rebuttal has strengthened the paper, and I find the improvements sufficient to raise my rating to 3.

AuthorsOpenReview ↗

Response to Reviewer ZxHM

We thank the reviewer for their thoughtful evaluation and recognition of our additional experiments. We acknowledge the remaining limitation regarding consistency between the generated SMILES and the final 3D structures, and will clarify this limitation in the revised manuscript. We appreciate the constructive feedback, which has helped us improve the paper.

Reviewer_CsSmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolLIFT, a bidirectional 1D–3D Bayesian flow framework for pocket-conditioned molecular generation. It uses atom-centered local frames to integrate SMILES priors into SE(3)-equivariant coordinate updates, while geometric feedback refines the SMILES state. CrossDocked experiments show improvements over the underlying generator in affinity, molecular quality, diversity, and success rate. The joint representation also enables gradient-guided multi-objective optimization for further improvements.

Critical strengths
  1. The motivation experiments are interesting, and based on this observation, utilizing sequence information is reasonable.

  2. The proposed method outperforms other baseline methods.

  3. The proposed method is complex enough, and the authors have a significant workload.

Critical weaknesses
  1. Although the motivating experiment is interesting, comparing only two models provides limited support for the claimed complementarity. Figure 1 should include a broader range of 3D- and sequence-based molecular models to establish whether the observed pattern generalizes.

  2. The SMILES-to-3D attention module is insufficiently explained in the main text. The authors should clarify how ht,jSh^S_{t,j} is obtained from the evolving SMILES belief, how sequence information is initialized and used during inference, and how these features enter the key/value projections in Eqs. (10)–(11).

  3. A simpler alternative is to first generate molecules as SMILES and then obtain their binding poses through docking. This pipeline could also combine sequence-level chemical priors with pocket-specific geometric evaluation, potentially addressing the issue highlighted in Fig. 1. A direct comparison is needed to clarify the advantages of the proposed bidirectional framework over this sequential baseline.

Questions for authors

Too many symbols make the paper difficult to read, and some definitions are missing. The subscripts and superscripts on the symbols in the paper are overly complex; you could include a table in the appendix to explain them.

6 comments
AuthorsOpenReview ↗

Response to Reviewer CsSm (Part 1/2)

We sincerely thank the reviewer for their time, careful assessment, and valuable feedback. We appreciate the recognition of the motivation and empirical performance of MolLIFT. Please see our point-by-point responses below; revisions are marked in blue.

W1-1: Could the authors broaden Figure 1 beyond the current two-model comparison to determine whether the observed complementarity generalizes across a wider range of 3D- and sequence-based molecular representations?

A1-1: Figure 1 has been expanded from one encoder per modality to six pretrained molecular encoders:

ModalityPretrained encoders
3DUni-Mol, IHGCN-PLA, DrugCLIP
SMILESChemBFN, MoLFormer, MolEncoder

For every method, the pretrained encoder is frozen and only a two-layer MLP with the same parameter budget is trained. For pocket-dependent Vina prediction, the MLP receives both ligand and pocket representations. Each 3D method uses its corresponding pretrained ligand and protein/pocket encoders, whereas each SMILES method combines its own molecular encoder with the same pretrained Uni-Mol pocket encoder. For ligand-intrinsic SA and QED, only the molecular representation is used.

We report MAE and RMSE for every encoder and the mean of the three methods in each modality. The modality-level means show a consistent pattern: 3D representations produce lower Vina MAE and RMSE, whereas SMILES representations produce lower SA and QED MAE and RMSE. The motivation is therefore no longer based on one selected model pair. It is supported by three independently pretrained encoders from each modality and indicates that sequence-level chemical priors and explicit 3D geometry capture complementary information. The complete protocol is provided in Appendix I.

W2-1: Could the authors clarify how the relevant sequence features are obtained from the evolving SMILES belief, and how the sequence information is initialized and used during inference?

A2-1: We expanded Section 3.3 and Appendix C.2 and added Algorithms 1–2. MolLIFT maintains the joint Bayesian-flow state consisting of a coordinate belief mean, an atom-type categorical belief, and a soft categorical belief over SMILES tokens. At inference, the coordinate mean is initialized to zero, while the atom-type and SMILES beliefs are initialized as uninformative uniform categorical distributions. Thus, ordinary generation does not require a ground-truth or pre-generated SMILES condition.

At every Bayesian-flow step, the current soft token belief is contextualized as

HtS=DiTSinit(θts,t)\mathbf{H}_{t}^{S}=\mathrm{DiT}_{S}^{\mathrm{init}}(\boldsymbol{\theta}_{t}^{s},t).

The model predicts a clean token distribution, and the categorical BFN receiver incorporates this prediction into the next belief. The updated soft belief is then re-encoded at the next step. Consequently, SMILES information evolves jointly with coordinates and atom types throughout generation; it is not a fixed sequence condition and is not reduced to an argmax sequence between steps.

W2-2: Could the authors clarify how the features obtained from the evolving SMILES belief enter the key and value projections in Eqs. (10)–(11)?

A2-2: The local protein–ligand environment of atom ii first produces the geometry-aware query qt,i\mathbf{q}_{t,i}. For each attention head, the projections are

Q=WQqt,i\mathbf{Q}=\mathbf{W}_{Q}\mathbf{q}_{t,i},

K=WKsg⁡(HtS)\mathbf{K}=\mathbf{W}_{K}\operatorname{sg}(\mathbf{H}_{t}^{S}),

and

V=WVsg⁡(HtS)\mathbf{V}=\mathbf{W}_{V}\operatorname{sg}(\mathbf{H}_{t}^{S}).

Thus, the atom representation supplies the query, while the contextual features extracted from the evolving soft SMILES belief supply both keys and values. Padding positions are excluded by the valid-token mask. The resulting atom-specific SMILES message first updates the invariant scalar atom representation and then predicts invariant coefficients in the atom-centered local frame, producing an explicit equivariant coordinate message.

The stop-gradient operator detaches the source token features but does not freeze the key/value projections, attention module, or cross-modal mapping. In the reverse direction, invariant atom memory is read by the SMILES branch to refine the clean-token prediction used by the next BFN update. Section 3.3, Appendix C.2, and Algorithms 1–2 now present this complete step-by-step data flow.

AuthorsOpenReview ↗

Response to Reviewer CsSm (Part 2/2)

W3-1: Could the authors directly compare MolLIFT with a simpler sequential pipeline that first generates pocket-conditioned molecules as SMILES and then obtains their binding poses through molecular docking?

A3-1: We added TamGen to Table 1 as a sequential baseline. TamGen first generates a pocket-conditioned SMILES and subsequently obtains a binding pose through conformer generation and molecular docking. MolLIFT instead jointly evolves the SMILES belief, ligand coordinates, and atom types during generation.

MethodVina Dock Avg. ↓Vina Dock Med. ↓QED ↑SA ↑Diversity ↑Success ↑
TamGen-6.95-6.840.570.770.7921.0%
MolLIFT-7.95-8.100.540.690.8528.2%

Compared with the sequential pipeline, MolLIFT improves average and median Vina Dock by 1.00 and 1.26 kcal/mol, increases diversity from 0.79 to 0.85, and raises the joint success rate from 21.0% to 28.2%, an improvement of 7.2 percentage points. The main distinction is not simply whether docking is eventually performed. In the sequential pipeline, pocket-dependent geometric evaluation is applied only after the SMILES has been completed. In MolLIFT, the evolving pocket-conditioned geometry influences the sequence state, and the evolving sequence state simultaneously contributes to coordinate denoising at every generation step. The substantially stronger Vina Dock and success results clarify the advantage of incorporating 3D information during generation rather than obtaining the pose only after sequence generation.

Q1: Could the authors provide a notation table to clarify the numerous symbols and definitions, particularly the relatively complex subscripts and superscripts used throughout the method?

A-Q1: We added a dedicated notation section and Table 3 in Appendix B, together with an explicit pointer from the main method. The table defines the principal data variables and joint BFN states, the cross-modal representations and directional messages, the local-frame quantities, the feedback memories, and the principal modules and operators.

The main method is now organized into three explicitly named stages:

  1. Geometric Context Encoding (GCE): constructs a geometry-aware atom query from the local protein–ligand environment.
  2. Local-Frame SMILES-to-3D Lifting (LFL): retrieves information from the evolving SMILES belief and converts invariant coefficients into an equivariant coordinate message.
  3. Invariant 3D-to-SMILES Feedback (IFB): projects geometric changes into invariant local-frame features that refine the next SMILES prediction.

We additionally expanded the explanation surrounding Eqs. (10)–(11) and included separate training and generation algorithms. These revisions retain the necessary time, modality, and direction indices while making every principal symbol and each stage of the forward information flow directly traceable.

Thank you for your response. Regarding Weakness 3-1, I notice that TamGen achieves better QED and SAS scores. Does this suggest that TamGen generates molecules with more favorable chemical properties, whereas your method generates better 3D poses? As illustrated in Fig. 1, your motivation highlights the limitations of 3D methods in generating molecular sequences compared with 1D methods. However, TamGen’s superior performance on these metrics raises the question of whether your proposed method adequately addresses this limitation.

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

MolLIFT jointly generates ligand coordinates, atom types, and SMILES tokens with a Bayesian flow model conditioned on a protein pocket. Its main mechanism converts invariant sequence messages into directional coordinate corrections using local geometric frames. A reverse pathway feeds geometric information into the sequence branch. Experiments on CrossDocked compare ordinary generation and property-guided generation, together with capacity and component ablations.

Critical strengths
  • The local-frame construction gives sequence messages a direct geometric role while preserving equivariance to rotations and translations. This is a concrete architectural contribution, and the treatment of degenerate frames makes the argument more convincing.

  • The comparison with a larger MolCRAFT model and the removal of the explicit vector message are useful controls. The full model improves affinity and drug-likeness beyond both alternatives, so there is evidence that the gains involve the proposed interaction rather than parameter count alone.

Critical weaknesses
  • The strongest evidence concerns the forward sequence-to-coordinate pathway, while the claimed benefit of a closed bidirectional interaction remains untested. There is no isolated ablation of geometric feedback into SMILES. Moreover, the coordinate analysis supplies the correct ligand SMILES, whereas ordinary generation must use an evolving sequence that may describe a different molecular graph. Better corrections under the correct sequence do not establish that the sampled sequence carries equally useful information during generation. The shuffled and permuted controls are described but their results are not reported. Keeping the forward pathway intact while removing reverse feedback would directly test the central bidirectional claim; evaluating sequence perturbations along generated trajectories would clarify whether the benefit depends on meaningful chemical correspondence.

  • The molecular-quality evaluation does not establish whether the generated 3D structures are physically more reliable. QED and SA concern the reconstructed molecular graph, while minimization and docking can alter the generated geometry. The direct Vina Score is useful, but it does not replace checks of molecular validity, internal strain, and steric clashes. The paper also does not make clear how invalid or disconnected candidates enter the reported averages and success denominator. This matters because the unguided success gain over MolCRAFT is only 1.4 percentage points. Reporting valid and physically plausible structures per attempted generation would show whether the improved averages translate into more usable ligands.

Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer rsmb (Part 1/2)

We sincerely thank the reviewer for their time, careful assessment, and valuable feedback. We appreciate the recognition of our local-frame construction and the evidence supporting the proposed interaction mechanism. Please see our point-by-point responses below; revisions are marked in blue.

W1-1: Could the authors provide an isolated ablation of the 3D-to-SMILES feedback pathway to clarify whether the claimed closed bidirectional interaction provides measurable benefits beyond the forward sequence-to-coordinate pathway?

A1-1: We added w/o IFB in Table 2 and Section 4.2(v). This variant disables only the invariant 3D-to-SMILES feedback while retaining local-frame-aware scalar interaction and the explicit SMILES-to-coordinate vector message. It therefore compares bidirectional MolLIFT with a one-way SMILES-to-3D model without altering the forward pathway.

MethodVina Score Avg./Med.Vina Min Avg./Med.QEDSA
w/o IFB-6.92 / -7.03-7.26 / -7.200.530.67
MolLIFT-7.01 / -7.12-7.37 / -7.290.540.68

Restoring IFB improves every reported metric. The average/median Vina Score improves from -6.92/-7.03 to -7.01/-7.12, while average/median Vina Min improves from -7.26/-7.20 to -7.37/-7.29. QED and SA also increase from 0.53/0.67 to 0.54/0.68. These results support a more precise conclusion: the forward pathway supplies the primary gain, while reverse geometric feedback provides an additional measurable benefit by refining the SMILES state used in subsequent Bayesian-flow steps.

W1-2: Could the authors report the shuffled-SMILES and token-permutation controls described in the appendix, so that it can be determined whether the coordinate improvements depend on meaningful chemical correspondence rather than only the presence of an additional conditioning pathway?

A1-2: Appendix H.1 and Figures 6–7 now report all controls described in the original submission. For every paired evaluation, we hold the protein pocket, noisy 3D state, denoising time, coordinate and atom-type beliefs, and stochastic realization fixed.

SMILES conditionControlled questionMain observation
Null inputDoes sequence guidance provide useful information?MolLIFT improves coordinate error and directional alignment over most intermediate and late stages.
Randomly permuted token orderDoes the organization of the sequence matter when token composition is preserved?The trend is consistent with the null control; MolLIFT generally gives larger improvement, stronger alignment, and lower negative transfer than w/o LF.
Within-batch reassignmentIs the effect specific to the paired ligand sequence?Coordinate differences are small, alignment is weak and non-monotonic, and negative transfer remains near 0.5.

The permutation result is particularly informative because it preserves sequence length and token composition while disrupting molecular serialization. It therefore shows that organized SMILES information, rather than merely the presence of another conditioning branch, contributes to coordinate denoising. Within-batch reassignment presents a more challenging setting, as the replacement is another chemically valid molecule rather than a confirmed negative ligand for the pocket. We therefore explicitly delimit the scope of our claim: the experiments demonstrate the utility of informative, correctly organized sequence guidance, but the current training objective does not explicitly enforce one-to-one ligand identity discrimination.

AuthorsOpenReview ↗

Response to Reviewer rsmb (Part 2/2)

W1-3: Could the authors clarify whether similarly useful coordinate corrections are obtained when ordinary generation relies on the model’s evolving SMILES belief, rather than being supplied with the correct ligand SMILES?

A1-3: Appendix H.2 and Figure 8 evaluate 100 test molecules without supplying any reference SMILES token. The SMILES receiver starts from a uniform prior and is updated recursively using only the model-predicted soft token distributions. Thus, the belief at tkt_k contains the complete prediction history from t=0t=0 to tkt_k; it is not independently reconstructed at each time, and no discretized argmax SMILES is fed back to the model.

Along the controlled reference-anchored 3D trajectory, MolLIFT achieves positive coordinate improvement from t=0.25t=0.25 onward. Directional alignment rises from 0.065 at t=0.25t=0.25 to 0.818 at t=0.85t=0.85, while negative transfer falls from 0.46 to 0.00 over the same interval and remains 0.02 in the final bin. In contrast, w/o LF becomes negatively aligned after t=0.6t=0.6 and reaches 0.90 negative transfer in the final bin. These results show that accumulated model-generated beliefs can provide useful and directionally consistent guidance once sufficient geometric information has emerged. We explicitly identify this as a controlled reference-anchored reconstruction analysis, because that setting provides a known clean structure against which coordinate improvement can be measured.

W2-1: Could the authors directly evaluate molecular validity, internal strain, steric clashes, and pose consistency to establish whether the originally generated 3D structures are physically reliable?

A2-1: All generated-side quantities are computed from MolLIFT's original generated coordinates. Conformer relaxation is used only to construct the strain-energy reference, while redocking is used only to construct the comparison pose for symmetry-corrected RMSD. Neither operation replaces the generated geometry being evaluated.

The reference and baseline values are presented in Table 4. The direct comparison with the underlying MolCRAFT generator is:

MethodEnergy Med. ↓Clash Avg. ↓RMSD < 2 Å ↑Connected ↑
MolCRAFT1966.9142.4%96.7%
MolLIFT1575.5244.0%99.1%

Relative to the reported MolCRAFT results, MolLIFT reduces median strain energy from 196 to 157, a 19.9% reduction, and mean clash count from 6.91 to 5.52, a 20.1% reduction. It also increases the proportion of evaluable poses with RMSD below 2 Å from 42.4% to 44.05% and connectivity from 96.7% to 99.14%. These results show that MolLIFT's improvements are accompanied by better conformational stability, steric compatibility, pose consistency, and molecular integrity.

W2-2: Could the authors clarify how invalid, disconnected, or evaluation-failed candidates are included in the reported averages and in the denominator of the success rate?

A2-2: The analysis uses a fixed set of 10,000 selected candidates (100 for each of 100 pockets), with no failed molecule replaced. The original success criterion gives 2,824/10,000 (28.24%); additionally requiring connectivity gives 2,820/10,000 (28.20%), so disconnected candidates are not hidden from the denominator. Among 10,000 MolLIFT candidates, 9,914 are connected, giving a connectivity rate of 99.14%. Among 9,444 successfully evaluated poses, 4,160 (44.05%) have RMSD below 2 Å. More details are shown in Appendix G and Table 4.

Dear Reviewer,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know. If you are satisfied with our responses, we would greatly appreciate your consideration in adjusting the evaluation scores accordingly. We sincerely look forward to your feedback.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful assessments and constructive feedback. We appreciate the recognition of our geometrically motivated local-frame design and equivariance analysis (Reviewers rsmb and ZxHM), the motivation of incorporating sequence information (Reviewer CsSm), and the competitive experimental performance (all reviewers). We have revised the manuscript according to the reviewers’ suggestions, with revisions marked in blue.

We summarize the main additions and clarifications below:

  1. Additional evidence for the bidirectional interaction (Reviewers rsmb and ZxHM). We added an isolated w/o IFB ablation, which preserves the forward SMILES-to-3D pathway while removing reverse geometric feedback. The results show that reverse feedback provides an additional measurable improvement across all reported metrics. We also clarified the benefits of explicit local-frame vector lifting beyond scalar-conditioned equivariant generation (Table 2; Section 4.2; Appendix F).

  2. Further analysis of SMILES-guided coordinate updates (Reviewers rsmb and ZxHM). We now report the token-permutation and within-batch reassignment controls. We additionally evaluate recursively accumulated, model-generated SMILES beliefs without supplying reference SMILES tokens. In the controlled reference-anchored reconstruction analysis, these beliefs provide positive coordinate improvements from t = 0.25 onward, supporting the usefulness of the evolving sequence information (Appendix H; Figures 6–8).

  3. Direct evaluation of the generated 3D structures (Reviewer rsmb). We added evaluations of strain energy, steric clashes, pose consistency, and molecular connectivity. Compared with the reported MolCRAFT results, MolLIFT reduces median strain energy from 196 to 157 and average clash count from 6.91 to 5.52, while increasing connectivity from 96.7% to 99.1%. We also clarified evaluation coverage, failure handling, and success-rate denominators (Appendix G; Table 4).

  4. Broader motivation experiments and a sequential-generation baseline (Reviewer CsSm). We expanded the motivation experiment to six pretrained encoders, with three encoders per modality, and added TamGen as a SMILES-generation-then-docking baseline. MolLIFT achieves a better average Vina Dock score (−7.95 vs. −6.95), higher diversity (0.85 vs. 0.79), and a higher joint success rate (28.2% vs. 21.0%) (Figure 1; Table 1; Appendix I).

  5. Clearer methodological explanations and positioning (Reviewers CsSm and ZxHM). We expanded the descriptions of SMILES-belief initialization, inference-time updates, and cross-modal attention; added training and generation algorithms and a notation table; and provided a more focused discussion of the distinctions from LiFT, NExT-Mol, and MolSculpt (Section 3.3; Appendices A–C and F).

Detailed point-by-point responses are provided under each review. We hope this summary helps the reviewers and Area Chair navigate the additional evidence and revisions. Thank you for your time and consideration, and we welcome any further questions or clarification requests.

Thanks for the additions. Reference-anchored tests leave the coordinate benefits of generated SMILES beliefs during free generation unclear. I maintain my original assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.