acceptodds
Under review as a conference paper at ICLR 2027

SOURCE-ANCHORED FLOW MATCHING FOR LOCAL MOLECULAR EDITING

Abstract

Lead optimisation requires local edits that improve a property while preserving the parent scaffold. Property-conditioned generators steer samples toward a desired label without anchoring them to a given parent, so scaffold retention can only be enforced after sampling. We introduce a recipe for local property editing in the frozen latent space of a pretrained SMILES encoder. Sparse Cross-Improvement Neighbours (SCIN) mines directed source-to-target pairs whose observed property gain exceeds a margin within a latent-similarity neighbourhood, and Source-Anchored Flow Matching (SAFM) learns a velocity field conditioned on the source embedding and the property labels. Classifier-free guidance and an inference-time edit budget control the fidelity–improvement trade-off without retraining. We evaluate on ZINC250K and MOSES for QED maximisation, then transfer the same recipe to cyclic-peptide permeability on CycPeptMPDB across the PAMPA and Caco-2 assays. SAFM leads faithful editing on both modalities, at 36.3% on CycPeptMPDB, and at 9.8% on ZINC250K. Descriptor analysis of the peptide edits recovers a consistent aromatic-to-aliphatic exchange that medicinal chemistry associates with improved passive permeation. The recipe transfers across chemical modalities with no domain-specific machinery beyond a frozen encoder and modality-appropriate validity gates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular editing, cyclic peptides, permeability, flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rkXJOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles the lead optimization pipeline and aims to improve desired properties while maintaining structural similarity to the lead compound. The method uses a pre-trained SMILES encoder with frozen embeddings, whose latents are fed into SCIN network used to construct supervised local improvement pairs. A source-anchored flow matching (SAFM) model then trains a velocity network to transport from source latents towards these mined targets; it has an edit budget which truncates the integration path. SAFM is conditioned on source embedding and property labels, and guided using classifier-free guidance. At inference time a trained decoder is used to map latent representations back into SMILES, generating eight candidates of which the first passing domain validity gates is kept. A regressor is trained as a property oracle to predict target property.

Critical strengths

This paper tackles the lead optimization pipeline and aims to improve desired properties while maintaining structural similarity to the lead compound. The method uses a pre-trained SMILES encoder with frozen embeddings, whose latents are fed into SCIN network used to construct supervised local improvement pairs. A source-anchored flow matching (SAFM) model then trains a velocity network to transport from source latents towards these mined targets; it has an edit budget which truncates the integration path. SAFM is conditioned on source embedding and property labels, and guided using classifier-free guidance. At inference time a trained decoder is used to map latent representations back into SMILES, generating eight candidates of which the first passing domain validity gates is kept. A regressor is trained as a property oracle to predict target property.

Critical weaknesses

How come the actual mined supervision targets fails the evaluation criteria? The paper describes mining "real improvement neighbours" and "supervised local improvement trajectories," which implies these targets should be acceptable edits, but on ZINC they are not: table 1 SCIN target retrieval (M3) scores 0. This suggests the model is trained to move toward targets its own evaluation would fail which, in the worst case, means that the quoted SAFM result is gained by the model 'not' reaching the desired training target.  Section 4.2.1. Lays out ‘non-negative’ (Ln 293) improvement as a criteria for faithful-improve rate. A paper with a central claim to improve molecular properties should not count zero change as improvement.

Figure 4 case study shows the deletion of a functionally important group which the authors also concede is a plausible pharmacophore loss. Morgan-Tanomoto similarity score is indifferent to which substructures are preserved so edits that might remove functionally critical groups can still pass faithfulness criteria. As detailed in the introduction, lead optimization requires edits that improve a property while also preserving functionality/affinity.

No limitations mentioned:  A section that promises limitations and delivers none reads as either unaware or evasive. If the decoder construction quality is around 73% for ZINC250K (Table 9), how come Table 3 and Table 5 recover 1.00 identity on tanimoto and constraint passing criteria? Please clarify. The big difference in constraint passing (0.730 vs 0.616) for In-house SAFM with and without conditioning in Table 3, makes the Table 1’s exactly same (0.363) faithfulness scores look implausible.  I’m also not fully bought on the faithful improve rate as a metric. ‘The paper’s central question: whether an editor can improve a molecule locally, rather than generate a different high-scoring molecule.’ Appendix H counters this as the rearrangements described there gives a much better Tanimoto similarity (0.93 vs 0.84) compared to single-residue substitutions, despite affecting every residue in the macrocycle. Classifier free guidance should be referenced to Ho and Salimans (2022). The paper contains inconsistencies on important terms e.g. the method is called SAFM throughout in the paper yet Figure 5 describes it as “SC-CFM (ours)”. The primary metric of evaluation is called “fidelity-improvement trade off” or ‘faithful-improvement rate’. Table 5 calls ‘de novo filter’ what Table 1 calls ‘generate-and-filter’.

Questions for authors

Please Report Faithful-Improve on ZINC under similarity floors of 0.6 -  0.8, and statistical paired test over the 15 cells for SAFM vs. latent interpolation. Please also give the per-criterion failure breakdown for SAFM, SCIN retrieval and latent interpolation on ZINC, and state exactly what the small-molecule constraint gate

Correct the Faithful-Improve metric. Re-report Faithful-Improve under strict \Deltau>0 and provide the distribution of  \Delta u among passing edits.

Please write the limitations section.

Please provide an analysis of nearest-neighbor in the training set for accepted edits including the distribution of molecules tanimoto similarity to the train set.

4 comments

We thank the reviewer for recognising how SCIN, SAFM, budgeted inference, and decoding form an integrated source-anchored editing pipeline.

R1.1: SCIN targets versus evaluation criteria\textbf{R1.1: SCIN targets versus evaluation criteria}

We thank the reviewer for this observation. The key distinction is that SCIN targets provide directional supervision for the velocity field (they define where property improves relative to the source in latent space) and are not intended to be individually recoverable edit endpoints.

SCIN pairs are mined using latent cosine similarity (τ≥0.50\tau \geq 0.50) and a property margin (δmin⁡\delta_{\min}), whereas the evaluation gate is applied in Morgan-fingerprint space (Tanimoto ≥0.70\geq 0.70 to the source). These similarity measures are not equivalent: a higher-QED latent neighbour can remain distant from the source in fingerprint space. This mismatch is largest on ZINC, where gg imposes no additional structural constraint and decoded constraint-passing SCIN retrieved candidates have mean source Tanimoto 0.090.09; on peptides, the equal-monomer-length restriction narrows the training-pair space and decoded constraint-passing SCIN retrieved candidates reach mean Tanimoto 0.560.56.

SAFM uses SCIN targets as directional supervision rather than endpoints to reproduce. At inference, it integrates a learned velocity field from the held-out source embedding without access to a target molecule, allowing partial transport to balance improvement and source retention. The random-pair ablation supports the contribution of SCIN neighbourhood structure on peptides: replacing SCIN with globally sampled improving pairs raises mean Δ\Delta from 0.330.33 to 0.480.48 among the constraint-passing edits, but reduces mean source Tanimoto from 0.850.85 to 0.600.60. Thus, broader supervision produces larger property gains at the cost of source retention, while SAFM need not retrieve an individual SCIN target for that supervision to be useful.

R1.2: Strict positive improvement criterion\textbf{R1.2: Strict positive improvement criterion}

Thank you for this valuable comment. We have substantially revised the manuscript to address this concern and enforce a strict improvement criterion (Δu>0\Delta u > 0). The updated numbers are reported in the revised Tables 1 and 3--5 and summarised above.

R2: Functional-group preservation and affinity\textbf{R2: Functional-group preservation and affinity}

We agree. Morgan Tanimoto is indifferent to which fragments survive, so an edit can clear the 0.70.7 Tanimoto floor while removing a pharmacologically relevant group, as in Figure 4. The updated analysis adds a further case, since 156 of the 367 accepted edits invert a stereocentre and score a Tanimoto of exactly 1.00.

The new Limitations section states that the recipe improves an intrinsic property while leaving others unguarded, so lead optimisation in the full sense requires a target-aware objective. It also records the blind spots of the fingerprint floor that motivate reporting the residue-level classification alongside it.

R3.1: Missing limitations\textbf{R3.1: Missing limitations}

Section 6 now carries a Limitations subsection, covering five points.

The recipe optimises one intrinsic property, so properties outside the objective are left unguarded, which the Figure 4 edit illustrates by removing a plausible pharmacophoric element. We set out the multi-property extension and the co-labelled data it would require.

Fidelity is measured by Morgan Tanimoto, which is global, indifferent to which fragments survive, and insensitive to configuration, so the 156 stereocentre inversions among the accepted peptide edits score exactly 1.00. The residue-level classification is reported alongside it for this reason.

SAFM acts on a pooled sequence embedding and decodes the whole molecule, so there is no site-level control, and the edit budget acts globally as the only inference-time dial.

For peptides, all reported property changes are predictions from learned proxies, and the cross-assay analysis in the appendix compares predictors rather than measurements. QED on ZINC is computed analytically via RDKit.

The chemical characterisation covers one assay on one dataset, and its conformational component is computed in vacuum with a classical force field.

R3.2: Identity baseline and decoder quality\textbf{R3.2: Identity baseline and decoder quality}

We thank the reviewer for the keen observation. The Identity row (M0) is a sanity check: it returns the source molecule directly without passing through the decoder, so it trivially scores 1.001.00 on Tanimoto. It is not a claim about decoder quality. We agree that including it alongside decoder-dependent methods invites exactly this confusion, so we have removed the Identity row from the revised tables.

R3.3: Composite metric and constraint pass rate\textbf{R3.3: Composite metric and constraint pass rate}

We appreciate this scrutiny and have acted on both parts of it. The old Faithful-Improve score was the intersection of three gates (Δ\Delta, Tanimoto ≥0.7\geq0.7, and constraint passage), not a direct function of the marginal constraint-pass rate. The source-conditioned and no-source variants passed the individual gates on different subsets of molecules, and their three-way intersections happened to yield the same score of 0.3630.363. Thus, the identical composite scores were possible despite the different marginal constraint-pass rates, but we agree that the aggregate obscured how each variant succeeded and failed.

The revised main text therefore no longer uses the composite as the primary result. Table 1 reports the strict %Δ>0\Delta>0 rate, mean Δu\Delta u, and mean Tanimoto separately, while the expanded appendix tables report the constraint-pass rate. This makes every dimension of the fidelity--improvement trade-off visible. Consistent with Table 1, among constraint-passing peptide edits, SAFM with source conditioning achieves a 50.050.0% strict improvement rate, mean Δu\Delta u of 0.330.33, and mean Tanimoto of 0.850.85; without source conditioning, the corresponding values are 41.141.1%, 0.280.28, and 0.890.89. The constraint-pass rates over all decoded edits, reported separately in the expanded appendix tables, are 55.055.0% and 60.560.5%, respectively.

R3.4: Tanimoto blind spots and edit locality\textbf{R3.4: Tanimoto blind spots and edit locality}

We agree. Morgan Tanimoto is a global measure, so an edit that conserves the residue multiset can score above a single-residue substitution while moving every residue, and the fingerprint floor alone does not establish locality.

In the re-run analysis, 156 of the 367 accepted edits invert a stereocentre and score exactly 1.00, the fingerprint being insensitive to configuration. Rearrangements reach 0.87 and 0.88 against 0.84 for single-residue substitutions, so the ordering holds with a narrower margin than the figures quoted.

Appendix C therefore classifies edits by residue multiset rather than by fingerprint. Under that classification the accepted edits are dominated by single-site changes, 241 of 367, with 51 rearrangements and 75 multi-residue changes.

We have moved the classification into Section 5.2 so that the locality claim rests on edit composition.

R3.5: Classifier-free guidance citation\textbf{R3.5: Classifier-free guidance citation}

We have added the Ho & Salimans (2022) citation where classifier-free guidance is introduced in Sections 2 and 3.3. During this revision, we also aligned the sampler with this standard classifier-free-guidance formulation and re-ran the affected experiments under the unified protocol and strict Δu>0\Delta u>0 criterion. The revised results are reported in Section 5.1 and summarised in the general response.

R3.6: Terminology consistency\textbf{R3.6: Terminology consistency}

We thank the reviewer for identifying these inconsistencies. We have replaced "SC-CFM'' with "SAFM'' in the figure labels and legends, and standardised "de novo filter'' to "generate-and-filter'' throughout. We now reserve "fidelity-improvement trade-off'' for the relationship between property improvement and source similarity; the revised evaluation reports its components separately through the strict %Δ>0\Delta>0 rate, mean Δ\Delta, mean Tanimoto, and constraint-pass rate rather than using Faithful-Improve as the primary composite metric. These revisions also address the metric concern raised in R3.3.

Q1: Similarity-floor sweep and ZINC constraint breakdown\textbf{Q1: Similarity-floor sweep and ZINC constraint breakdown}

On ZINC the constraint gate requires only that the decoded SMILES is valid (RDKit-parseable) and distinct from the source by canonical-SMILES inequality. No additional structural or drug-likeness filter is applied. Figure 7 in Appendix B now reports the per-gate pass rates at b=1.0b=1.0 across the 15 seed--fold cells.

Similarity-floor sweep.\textit{Similarity-floor sweep.} We re-evaluated all methods on ZINC at Tanimoto floors from 0.10.1 to 1.01.0 across the 15 cells. SAFM without source conditioning leads latent interpolation at every floor up to 0.70.7: at floor 0.30.3 it scores 45.745.7% against 15.015.0%; at 0.50.5, 24.324.3% against 12.712.7%; and at 0.70.7, 7.87.8% against 6.66.6%. At floor 0.80.8, the two converge at 2.62.6% and 2.72.7%, respectively. Generate-and-filter, SCIN retrieval, and SAFM with random targets fall below 11% by floor 0.30.3 and reach 0.00.0% by floor 0.50.5.

Paired statistical test.\textit{Paired statistical test.} A two-sided Wilcoxon signed-rank test over the 15 matched seed--fold cells confirms that SAFM (no source conditioning) significantly outperforms latent interpolation at all floors ≤0.6\leq 0.6 (p<0.001p < 0.001 at each). At floor 0.70.7 the difference remains significant (p=0.033p = 0.033). At floor 0.80.8 the two methods are statistically indistinguishable (p=0.72p = 0.72).

Q2: Strict Faithful-Improve and improvement distribution\textbf{Q2: Strict Faithful-Improve and improvement distribution}

We thank the reviewer for pressing this point. The revised manuscript enforces strict Δu>0\Delta u > 0 everywhere: in all result tables, and in the general response summary above. All methods were re-run under this tightened criterion with the same CV folds and seeds.

The revised headline results are reported in Table 1, with constraint-pass and per-gate results reported separately in Appendix B. This reporting exposes strict improvement, improvement magnitude, source similarity, and validity without collapsing them into a single primary score.

As an additional distributional view, we report Δu\Delta u among best Δ\Delta per source at b=1.0b=1.0. On CycPeptMPDB (PAMPA), SAFM with source conditioning has n=2,359n=2{,}359, median Δy^=0.18\Delta\hat{y}=0.18, and IQR [0.01,0.67][0.01,0.67]; without source conditioning, n=2,370n=2{,}370, median 0.140.14, and IQR [0.01,0.64][0.01,0.64]. On ZINC under analytic QED, SAFM with source conditioning has n=3,855n=3{,}855, median ΔQED=0.13\Delta\mathrm{QED}=0.13, and IQR [0.04,0.28][0.04,0.28]; without source conditioning, n=3,511n=3{,}511, median 0.100.10, and IQR [0.01,0.22][0.01,0.22].

Q3: Limitations section\textbf{Q3: Limitations section}

Done. Section 6 of the revised manuscript now contains a dedicated "Discussion and Limitations'' subsection covering five points, as detailed in our response to R3.1 above: single-property optimisation, Tanimoto blind spots, absence of site-level control, proxy-based evaluation (with analytic QED on ZINC), and scope of the chemical characterisation.

Q4: Nearest-neighbour analysis\textbf{Q4: Nearest-neighbour analysis}

For this analysis we selected the constraint-passing edits with positive improvement (Δ>0\Delta>0), one per source at the best Δ\Delta and budget b=1.0b=1.0, and computed the maximum Morgan-Tanimoto similarity (radius 2, 2048 bits, chirality not encoded) to any molecule in the corresponding training fold. On CycPeptMPDB (PAMPA, in-house oracle), SAFM with source conditioning (n=1,843n=1{,}843) reaches mean nearest-neighbour Tanimoto 0.9850.985 and median 1.001.00, and without source conditioning (n=1,886n=1{,}886) 0.9860.986 and 1.001.00.

A Tanimoto of 1.001.00 here does not establish molecular identity, because the Morgan fingerprint is computed without stereochemistry. An edit that inverts a single stereocentre yields a diastereomer, an epimer with identical 2D connectivity and therefore an identical fingerprint, even though it is a distinct molecule. Section 5.2 attributes 156 of these unit-Tanimoto cases to stereocentre inversions. For cyclic peptides this distinction is chemically load-bearing, since backbone stereochemistry (L versus D residues) is among the primary determinants of passive membrane permeability, so an epimerization that the fingerprint cannot see is exactly the kind of permeability-relevant edit the model is meant to produce.

On ZINC, SAFM with and without source conditioning (n=3,235n=3{,}235 and 2,7152{,}715) reaches mean nearest-neighbour Tanimoto 0.4520.452 and 0.4770.477, median 0.440.44 and 0.470.47, with only 0.060.06% and 0.110.11% of edits above 0.90.9, so training-fold overlap is low. The contrast with the peptides partly reflects the size sensitivity of the fingerprint, since a comparable local modification changes a larger fraction of the substructural features in the smaller ZINC molecules than in the larger peptides. We discuss these limitations on Morgan Tanimoto under the Section 6.

Thank you for your detailed scrutiny of the supervision, evaluation and presentation.

Your central question concerns why mined improvement targets can fail the structural evaluation. SCIN mines pairs using latent cosine similarity and a property margin, whereas source retention is evaluated using Morgan-Tanimoto similarity; these criteria are distinct, particularly on ZINC, where mining imposes no additional structural constraint (Section 3.2). SCIN supplies observed improvement directions, and SAFM learns source-initialised transport with an edit budget that balances improvement and source retention rather than requiring reproduction of an individual mined target (Section 3.3). The revised evaluation requires strict positive improvement (Δu>0\Delta u > 0): Table 1 reports its rate, improvement magnitude and source similarity separately, with constraint-pass rates in Appendix B. We also aligned the classifier-free-guidance sampler with Ho and Salimans (2022), added the citation and re-ran the affected comparisons under matched folds, seeds and evaluation conditions (Sections 2–3.3 and 5.1).

On your requested analyses: (1) at each ZINC similarity floor, reported percentages reflect distinctness from source, constraint pass, strict positive QED improvement and source similarity at or above the specific floor. At floor 0.5, SAFM without source conditioning succeeds for 24.3% of held-out sources versus 12.7% for latent interpolation. It significantly outperforms interpolation at floors 0.1–0.6 (p < 0.001 at each) and 0.7 (7.8% versus 6.6%; p = 0.033); at 0.8, the methods converge (2.62% versus 2.69%; p = 0.72) (added to Appendix B.3). (2) Appendix B now reports per-gate constraint breakdown for ZINC in addition to peptides, and Section 3.4 defines the ZINC gate as RDKit validity and canonical-SMILES distinctness. (3) Appendix B.4 now reports improvement distributions among constraint-passing edits, with positive medians for both SAFM variants on both datasets. (4) Appendix B.2 quantifies training-fold overlap through nearest-neighbour means and medians and explains why unit Tanimoto does not imply molecular identity when stereochemistry is omitted.

We have also resolved the reporting distinctions behind your questions: the Identity reference returned the source directly rather than reconstructing it through the decoder and has been removed. The earlier identical composite scores were intersections of gates on different subsets, not consequences of the marginal constraint-pass rates; the revised reporting exposes those components separately. On rearrangements, Morgan Tanimoto can rank a whole-ring residue rearrangement above a single-residue substitution, so we assess locality using complementary residue-level evidence rather than the fingerprint alone. This analysis identifies 241 single-site changes among 367 modifications in the analysed seed-0/fold-0 set across budgets (Section 5.2; Appendix C). On functionality, SAFM demonstrates controllable intrinsic-property improvement with source retention, providing an editing component for lead optimisation. Preserving binding affinity and biological function is explicitly discussed in Section 6 (Discussion and Limitations) as a direction for future target-aware, multi-property optimisation. We have elaborated on the limitations in Section 6, distinguishing source similarity from pharmacophore preservation and setting out the objective needed for full lead optimisation, alongside fingerprint blind spots, absence of site-level control, proxy-based evaluation and the scope of the chemical characterisation. Method and baseline terminology has been standardised.

Together, these corrections and analyses strengthen the evidence for controllable source-anchored editing while defining its scope precisely. We respectfully ask you to reconsider your assessment in light of the revised manuscript. As the discussion period draws to a close, please let us know if any point remains unclear or requires further clarification so that we can respond before it ends.

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

Author proposed a method for local molecular editing in SMILES latent space, evaluated on QED optimization and predicted PAMPA permeability.

Critical strengths
  1. Local modification of existing molecules could make property optimization more efficient. It is a valid direction to explore.
Critical weaknesses
  1. For the organization of the paper, consider including more results in the main text and moving detailed evaluation procedures to the appendix.
Questions for authors
  1. The rules defined are too absolute. N-methylation for example, is not always an "incresed" favor, so as aromatic rings, etc. Both are sites and context dependent (see citations below). How are these effects selected and considered?

White, Tina R., et al. "On-resin N-methylation of cyclic peptides for discovery of orally bioavailable scaffolds." Nature chemical biology 7.11 (2011): 810-817. Kelly, Colin N., et al. "Geometrically diverse lariat peptide scaffolds reveal an untapped chemical space of high membrane permeability." Journal of the American Chemical Society 143.2 (2021): 705-714.

  1. Figure 4 gives the impression of targeted functional-group editing, although SAFM modifies a pooled molecular embedding and then decodes the whole molecule. Please clarify that the highlighted substitution is an observed output, and whether the method can actually control which functional group or residue is edited.
  2. Does pairing each source with only a few nearby, higher-scoring molecules limit diversity and OOD exploration? How is this evaluated?
  3. With the current framework, how can other properties, such as binding affinity and stability, be optimized alongside permeability?
7 comments

We thank the reviewer for recognising our local molecular editing approach as a valid and potentially more efficient way to address property optimisation.

R1: Organization of the results\textbf{R1: Organization of the results}

We thank the reviewer for this suggestion. In the revised manuscript, Table 1 now reports the strict improvement rate (%Δ>0\Delta>0), mean property improvement (Δu±\Delta u \pm SD), and mean Morgan-Tanimoto similarity to source (±\pm SD), making the fidelity--improvement trade-off directly visible. We have also expanded the explanations of the quantitative results in Section 5.1 and the chemical characterisation in Section 5.2. Constraint-pass rates are reported in the expanded per-method tables in Appendix B, while detailed chemical-characterisation procedures are provided in Appendix C.

Q1: Definition of medicinal chemistry rules\textbf{Q1: Definition of medicinal chemistry rules}

We agree that these effects are site- and context-dependent, and we have revised the text to reflect that.

Table 2 is a post-hoc characterisation of the accepted edits. It plays no part in training, in SCIN mining or in the editing objective, and SAFM is fitted on the permeability label alone.

The sign test counts violations explicitly, so context dependence appears in the unfavourable column. In the updated analysis sp3 fraction moves favourably in 73 pairs against 42, aromatic ring count in 39 against 21, and rotatable bonds in 70 against 38.

On the two descriptors raised, TPSA is non-directional in this population, at 24 against 18 with a two-sided sign test p=0.44p = 0.44, and is excluded from the composite. N-methyl amide count is preserved in 91.8% of pairs, and in the minority where it changes it moves favourably, at 25 against 5. Following White et al. (2011), where the benefit depends on position, we report it outside the composite. Kelly et al. (2021) report high permeability outside the classical descriptor envelope, which is consistent with reading Table 2 as a description of this population.

We have revised Section 4.2.2 and Appendix C to state that these are population-level descriptors applied post hoc, and both references are now cited. We have also replaced the earlier description of an aromatic-to-aliphatic exchange. The re-run ring inventory gives 65 ring systems removed against 42 introduced, with saturated systems in near balance at 17 against 16, so the text now reports a net removal of 25 aromatic ring systems.

Q2: Site-level control and Figure 4\textbf{Q2: Site-level control and Figure 4}

SAFM acts on a pooled sequence embedding and decodes the whole molecule, so it has no site-level control. The substitution in Figure 4 is an observed output of that decoding, selected post hoc for illustration under a filter fixed before inspection. The edit budget is the only inference-time control and acts globally by truncating the transport path.

Locality is therefore an empirical property of the accepted edits. Of the 367 accepted peptide edits, 241 change a single site, comprising 85 single-residue substitutions and 156 single-stereocentre inversions.

We have revised the Figure 4 caption and the case study paragraph to state that the substitution is an observed output and to describe the selection filter. The new Limitations section records the absence of site-level control.

Q3: Diversity and out-of-distribution exploration\textbf{Q3: Diversity and out-of-distribution exploration}

SCIN retains at most three improving targets per source from a k=50k{=}50 nearest-neighbour pool after applying the property-margin, cosine-similarity and applicable structural filters. This limits redundant supervision while retaining several observed improvement directions. The three-target cap applies to training supervision, not to the number of possible outputs. The velocity network is trained jointly over SCIN pairs from all training sources (Section 3.3), and inference integrates the learned field from a held-out source embedding rather than selecting one of its training targets. The decoder then produces one greedy decode and seven stochastic samples at temperature 0.80.8. These mechanisms allow candidate variation.

Does the locality constraint limit OOD exploration?\textit{Does the locality constraint limit OOD exploration?} SCIN favours local property improvement rather than unrestricted exploration. The random-pair ablation (A1) examines broader pairing by drawing improving targets from a global pool without the cosine threshold, while retaining the applicable structural filter. Among constraint-passing peptide edits, random pairing achieves higher mean PeptiVerse improvement (0.480.48 versus 0.330.33) but lower mean source Tanimoto (0.600.60 versus 0.850.85) than SAFM (Table 1). This illustrates the trade-off between property improvement and source retention. The edit budget provides a complementary inference-time control: smaller budgets prioritise source retention, while larger budgets permit greater displacement along the learned path without retraining.

Q4: Multi-property optimisation\textbf{Q4: Multi-property optimisation}

The framework conditions on a property label, so additional objectives enter as a vector label, with SCIN retaining the pairs that improve all of them jointly and each property receiving its own guidance scale at inference. The barrier is data, since the properties must be labelled on the same molecules, which is scarce for cyclic peptides. Affinity is a further case, since it is defined relative to a chosen target and requires an oracle that scores against it. The edit in Figure 4 shows the risk, improving predicted permeability while removing a naphthylalanine side chain that is a plausible pharmacophoric element. The new Limitations section states that the recipe optimises a single intrinsic property and sets out this extension with its data and target requirements.

Thank you for the clarifications. The revisions address my concern about Figure 4 and clarify the scope of the method.

However, the sign test counts descriptor changes in predefined directions. It does not establish whether those directions improve permeability in a particular molecule. The "unfavourable" counts therefore do not demonstrate that context dependence is accounted for. In addition, stochastic decoding allows different outputs, but how many distinct valid edits are actually produced for the same source, and how many are absent from the training set?

I appreciate the revisions, but these evaluation concerns remain, so I maintain my score.

Reviewer_qmjXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a new framework for local molecular editing which consists of two main components: 1) a data mining framework named SCIN, which can extract source-to-target pairs of molecules from any molecular dataset with property labels based on molecular similarity of their SMILES embeddings; 2) a flow matching model mapping the paired molecules conditioned on source molecule and property labels. Experimental results show that the proposed method yields the best results in the small molecules and cyclic peptide use cases within the different variants tested.

Critical strengths
  • Novel idea: both components of the proposed idea are quite interesting. Mining the paired molecules for editing from existing dataset with property appears to be a direction presenting great potential. While the high quality molecular editing dataset is very limited today, there are more and more drug / chemistry / materials datasets with properties available, which could potentially yield high quality molecular editing dataset as well. Using the flow matching model to map the paired molecules could also enable larger scale training / inference for the molecular editing model.
Critical weaknesses
  • Semantic similarity of SMILES does not guarantee chemical transformability: This is my main concern about this work. The entire framework is built on an implicit assumption that a pair of molecules with high similarity in SMILES is reachable from one to the other through local molecular editing. However, considering that SMILES-based language models are pretrained based on SMILES tokens, this is not actually true. Two molecules with high SMILES similarity could sit in different scaffolds or even be completely distinct from each other. In fact, this can also be supported by results presented in Table 1 and Figure 3. Table 1 shows that even at an edit budget of 1, the proposed method can only achieve a faithful improve rate below 10%. Figure 3 shows that when the edit budget is below 0.4, the gain in property is also minor. In other words, while the molecular pairs mined from existing datasets may present some similarity in the latent space, going from one molecule to another could very likely need non-local edit, contradicting the main idea of this paper.
  • Lack of real benchmark: while the authors constructed a series of models for comparison, they are more like a detailed ablation rather than the real benchmarks. Considering this is a new proposed framework for the molecular editing problem, it is important to compare it with existing benchmarks (like Graph GA).
Questions for authors
  1. In section 4.2.1, the authors said that comparing the target property improvement at different bins can analyze if the edits are applied on the critical positions. Can authors elaborate more on how does this work?
  2. In Table 1, peptide edits task scores a much higher Faithful-improve rate compared to the small molecules editing. Is it because of the additional constraint added for peptide editing ( g(xi, xj) = 1) or something else?
  3. In all the successful edits involved, how many of them are real local edits like the case study presented in Figure 4?

Minor point:

  • My understanding is that fidelity and molecular similarity are exchangeable in the context of this work. If so, I would suggest the authors make it clear at the first time mentioning fidelity.
8 comments

We thank the reviewer for recognising the potential of both SCIN and SAFM. SCIN converts property-labelled datasets into editing supervision, while SAFM learns scalable transport over the mined pairs.

R1: Semantic similarity of SMILES does not guarantee chemical transformability\textbf{R1: Semantic similarity of SMILES does not guarantee chemical transformability}

We thank the reviewer for raising this point. We agree that the reported low Faithful-Improve rate raises a substantive question about how often property improvement and structural retention are achieved together. In the revision, we require strict positive improvement (Δu>0\Delta u>0), re-run all methods under a unified protocol, and report the fidelity-improvement trade-off through the improvement rate, mean improvement, mean Tanimoto to source, and constraint-pass rate rather than a single composite score. These components make the trade-off more transparent. The updated results are provided in the revised manuscript.

On the implicit assumption.\textit{On the implicit assumption.} We agree that proximity in a SMILES encoder's latent space does not guarantee a local structural edit or chemical transformability. Our intended claim is narrower: SCIN provides candidate improvement trajectories, and the locality of the decoded outputs must be assessed empirically. SCIN selects pairs using a latent cosine-similarity threshold (τ=0.50\tau=0.50), a property-margin gate (δmin⁡\delta_{\min}), and an optional structural constraint g(xi,xj)=1g(x_i,x_j)=1; it does not use SMILES-string similarity. As stated in the Introduction, we learn transport initialised with the source molecule itself. This source initialisation and the mining filters are modelling choices, not guarantees that every target is structurally close or reachable through a feasible chemical transformation.

On pretrained encoders and latent structure.\textit{On pretrained encoders and latent structure.} We agree that similarity learned from SMILES need not imply a shared scaffold. ChemBERTa-77M and PeptideCLM-2 MLM-large provide frozen representations learned through masked-language pretraining, but that training objective alone does not establish that their cosine neighbourhoods correspond to local chemical transformations. SCIN uses these representations to propose supervision rather than to certify locality. For peptides, the additional equal-monomer-length constraint (g(xi,xj)=1g(x_i,x_j)=1) restricts the target pool, but matching monomer length likewise does not guarantee a local edit. The relevant evidence is therefore the structure of the decoded edits, not the encoder's pretraining objective.

On ZINC, where SCIN imposes no additional structural constraint beyond the latent-similarity threshold, SAFM variant without the optional source conditioning, SAFM (no source)\textit{SAFM (no source)}, leads the fidelity-improvement trade-off: it improves QED in 77.377.3% of its constraint-passing edits while keeping a mean Tanimoto of 0.510.51 to the source, the highest fidelity of any method with a comparable constraint-passing improvement rate. Generate-and-filter, and Latent gradient methods achieve high %Δ>0\Delta>0 (92.292.2% and 80.280.2%) at the cost of similarity to source, at mean Tanimoto of 0.120.12 and 0.260.26. For cyclic peptides, where the equal-monomer-length constraint already anchors the SCIN pairs, SAFM with source conditioning leads the fidelity-improvement trade-off: it achieves 50.050.0% Δ>0\Delta>0 at a mean Tanimoto of 0.850.85 on the external PeptiVerse scorer, whereas generate-and-filter reaches 91.091.0% Δ>0\Delta>0 only at 0.420.42. ZINC tests the performance of SAFM under weak structural constraints in the latent space, and CycPeptMPDB tests the performance under harder domain-specific structural and validity constraints. Across both modalities the SAFM variants pair a substantial improvement rate with source similarity while generating distinct valid molecules; every other baseline sacrifices at least one of improvement, fidelity and validity (Table 1).

On the role of the mined targets and source initialisation.\textit{On the role of the mined targets and source initialisation.} Our locality argument does not rest on latent similarity alone. SCIN targets provide training supervision for the direction of property improvement, while SAFM learns the transport itself from source embeddings rather than from generic noise. For each training pair, the path starts at the source latent, and the velocity field vθ(zt,t,z0,y0,ytarget)v_\theta(z_t,t,z_0,y_0,y_{\mathrm{target}}) (Eq. 2) learns transport along that path. At inference, there is no target molecule: integration starts from the test source embedding, is steered by the property label, and is truncated by the edit budget bb. This source initialisation remains present even in SAFM (no source)\textit{SAFM (no source)}, which removes only the optional source-conditioning input. Thus, even on ZINC, where SCIN imposes no additional structural constraint, SAFM retains a source-specific starting point that property-conditioned generate-and-filter lacks. The observed comparison supports the value of this design: among constraint-passing edits, SAFM (no source) retains mean source Tanimoto 0.510.51 while improving QED in 77.377.3% of edits, compared with 0.120.12 and 92.292.2% for generate-and-filter.

On Figure 3 and the low-budget regime.\textit{On Figure 3 and the low-budget regime.} The y-axis in Figure 3 reports predicted changes in PAMPA permeability in log units. The smaller gains at low budgets reflect the intended fidelity-improvement trade-off: low budgets prioritise source retention, while higher budgets permit greater structural displacement for larger property gains. This is the intended design, not a limitation; the edit budget gives the practitioner an explicit dial to navigate the trade-off for their use case.

R2: Lack of real benchmark\textbf{R2: Lack of real benchmark}

We thank the reviewer for this suggestion. We first clarify the design rationale for the main-text baselines and then report the requested external comparison.

On the controlled baselines and existing external comparisons.\textit{On the controlled baselines and existing external comparisons.} We clarify that the submitted manuscript already included external comparisons in the appendix: a MoltenFlow-style optimisation recipe on ZINC250K and fold-retrained C2PO on CycPeptMPDB. These external comparisons complement the controlled main-text evaluation, which compares latent gradient ascent (M1) and property-conditioned generate-and-filter (M2), together with SCIN-target retrieval (M3) and latent interpolation toward a SCIN target (M4) as diagnostic references. The controlled methods use the same frozen encoder, decoder, held-out sources, candidate count, and feasibility gates to isolate differences in the editing rule\textit{editing rule}. M3 and M4 additionally use training-derived improvement targets at inference, whereas SAFM does not require a target molecule at deployment. The revised appendix retains the MoltenFlow-style comparison (Appendix B.1, Table 5) and C2PO comparison (Table 7). Following the reviewer's specific suggestion, we have now added Graph GA on ZINC250K, as detailed below.

Comparison with Graph GA (MolGA).\textit{Comparison with Graph GA (MolGA).} Following the reviewer's suggestion, we have run the Graph GA Jensen (2019) baseline from mol_ga (https://github.com/AustinT/mol_ga) on the same 15 ZINC250K seed--fold cells and held-out sources. MolGA uses 100 generations with offspring size 100 and a candidate pool of 256, seeded with the held-out source molecules and scored by analytic QED. The best candidate per source is evaluated with the same Morgan-Tanimoto, constraint, and Δ\DeltaQED criteria as all other methods. Across 15 cells, MolGA achieves a mean Δ\DeltaQED of 0.371±0.0030.371 \pm 0.003, a constraint-pass rate of 100100%, but a mean Tanimoto to source of 0.107±0.0240.107 \pm 0.024. MolGA is highly effective at property maximisation and produces exclusively valid molecules, but because it searches over a broad chemical space the resulting candidates diverge from the source structure. This is consistent with its strength as a general-purpose molecular optimiser; the comparison highlights that property maximisation and source-anchored local editing are complementary rather than competing objectives. For comparison, SAFM (no source) achieves a mean Tanimoto of 0.527±0.0190.527 \pm 0.019 with Δ\DeltaQED of 0.139±0.0090.139 \pm 0.009 and a constraint-pass rate of 78.078.0%; SAFM (with source conditioning) achieves 0.255±0.0180.255 \pm 0.018 with Δ\DeltaQED of 0.184±0.0240.184 \pm 0.024 and a constraint-pass rate of 85.785.7%. On ZINC the constraint gate requires a valid, RDKit-parseable SMILES that is distinct from the source by canonical-SMILES inequality.

The Graph GA comparison reinforces the central distinction tested in this work: higher property scores do not by themselves establish successful source-retaining editing. Under the same source-relative evaluation criteria, SAFM retains greater similarity to the specified parent than MolGA, while MolGA achieves greater QED improvement and a higher constraint-pass rate. Thus, the evidence for SAFM is not limited to in-stack comparisons: the submitted manuscript already included MoltenFlow-style optimisation and fold-retrained C2PO, and the revision adds the specifically requested Graph GA benchmark (Appendix B.1, Table 6). Together, these comparisons evaluate SAFM against both controlled editing strategies and established external methods.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of the revisions

We made substantive revisions addressing the reviewers' central concerns: (i) aligned the classifier-free-guidance sampler with Ho & Salimans (2022); (ii) replaced non-negative improvement with strict Δu>0\Delta u>0 throughout; and (iii) re-ran SAFM and directly comparable baselines under matched CV folds and seeds, edit budget, candidate count, and evaluation criteria. The revised results sharpen the central finding: SAFM provides a controllable fidelity-improvement trade-off while remaining source-anchored.

On ZINC, SAFM without optional source conditioning improves 77.3% of constraint-passing edits at mean source Tanimoto 0.510.51; on CycPeptMPDB, source-conditioned SAFM achieves 50.0% at 0.850.85. We report strict improvement rate, improvement magnitude, source Tanimoto, and constraint pass rate separately rather than as a single primary score (Table 1 and Appendix B).

The chemical characterisation of the peptide edits was re-run on the updated candidate set and now covers 367 accepted modifications from 409 unique pairs. The locality and backbone-preservation conclusions are unchanged. Two descriptions were revised. The ring inventory shows a net removal of aromatic ring systems with saturated systems in near balance, so the earlier description of an aromatic-to-aliphatic exchange has been replaced by aromatic removal. N-methyl amide count now moves favourably in the minority of pairs where it changes, and the surrounding discussion has been updated while the descriptor remains outside the composite.

Dear reviewers, thank you again for your detailed feedback. We have posted responses and revised the manuscript to address the principal concerns raised by Reviewer qmjX\textbf{Reviewer qmjX} (chemical transformability, locality, and the requested Graph GA comparison), Reviewer hTAu\textbf{Reviewer hTAu} (descriptor context dependence, site-level control, output diversity, and training-set novelty), and Reviewer rkXJ\textbf{Reviewer rkXJ} (strict improvement criteria, similarity-floor and constraint breakdown, statistical testing, training-set proximity, and limitations). As the discussion period closes tomorrow, we would be grateful to know whether these revisions address your main concerns. If any specific point remains unclear, or if there is a focused additional analysis, we would be happy to clarify it or provide additional empirical evidence. Thank you for your time and consideration.

We ask the Area Chair to assess the revised paper on the full post-discussion record. Reviewers qmjX and hTAu noted the novelty and promise of mining supervision from labelled corpora (SCIN) and learning source-anchored flow transport (SAFM). We addressed every concrete soundness, benchmark, and discussion follow-up with matched re-runs and new analyses in the revised PDF.

Core contribution. SAFM performs local molecular editing: improving a target property while keeping the molecule structurally close to a given parent (the lead-optimisation setting), where a chemist must retain a known scaffold rather than find an unrelated high-scoring molecule. SAFM is thus a source-anchored editor, not a de novo generator or global optimiser: from a parent embedding it integrates a property-guided velocity field up to an inference budget b∈(0,1]b \in (0,1], trading property gain against scaffold retention without retraining, across small molecules and cyclic peptides.

Summary of the revisions: (i) enforced strict Δu>0\Delta u > 0 everywhere and aligned classifier-free guidance with Ho & Salimans (2022); (ii) replaced the single composite score with decomposed, transparent metrics (improvement rate, magnitude, source similarity, and constraint pass) plus a ZINC similarity-floor sweep with paired tests; (iii) added the requested Graph GA external benchmark alongside the existing MoltenFlow and C2PO; (iv) moved the residue-level peptide locality analysis into the main text (Sec 5.2), added a training-fold novelty audit, and expanded the Limitations section (Sec 6). Exact numbers and pointers are in the table below.

Clarifications for AC evaluation

  • Targeted editing vs. de novo / global optimisation: The task is to improve a property and stay close to the parent, so source similarity is a task requirement, not a metric we favour ourselves on. Global optimisers (Graph GA, MoltenFlow, generate-and-filter) reach higher raw improvement only by abandoning the scaffold, which is not desirable for retaining a known lead; SAFM instead improves while staying close to the parent, and the edit budget dials retention explicitly.
  • Directional supervision, not retrieval: SCIN supplies directions of improvement during training; inference uses no target molecule. The large majority of accepted edits are absent from the training fold, so SAFM generates rather than memorises.
  • Both SAFM and SAFM (no source) are our variants: Both start from the parent embedding; no source removes only the optional source-conditioning input.

Summary of Reviewer Concerns and Evidence

Reviewer & main concernsContext & framingHow addressedPaper changes & evidence
R1 (qmjX): Latent proximity ≠\neq local editability; low-budget Δ\Delta; add Graph GA; bin lifts, domain gap, fidelity, local-edit rate.SCIN gives directional supervision, not scaffold guarantees; small Δ\Delta at b<0.4b<0.4 is the conservative end of the budget dial.Re-ran strict Δu>0\Delta u>0; ran Graph GA on 15 ZINC cells; cleaved amides to count 1-site edits.Table 1, App B.1 (Table 6): Graph GA 0.107 Tanimoto vs. SAFM (no source) 0.527. Sec 5.2: 241/367 (65.7%) 1-site edits. Sec 1, 5.1, 6: Fidelity, bin lifts & size scaling.
R2 (hTAu): Rule context-dependence (White 2011; Kelly 2021) & causality; Fig. 4 control; novelty/diversity under 3-target cap; multi-obj.Descriptors are post-hoc summaries (SAFM trains on labels alone); the 3-target cap restricts training pairs, not outputs.Bounded sign-test scope; cited both papers; clarified pooled decoding in Fig. 4; ran 4,500-case novelty audit.Sec 5.2, App C: Population scope + 3D ensembles (-6% amide-H; -25 aromatic rings). Fig. 4, Sec 6: Observed-output note & vector labels. App B.2 (Nov 16 follow-up): 56.3% (peptide) / 99.97% (ZINC) edits absent from train fold; 42 rings added.
R3 (rkXJ): SCIN retrieval (M3) ZINC floor; Δu≥0\Delta u\ge 0 ties; Tanimoto blind spots; Limitations, Identity row, 0.363 tie, CFG citation, naming; floor/Wilcoxon/gate/IQR/NN tests.One conceptual point (latent proximity ≠\neq transformability) plus metric/reporting items; tightening every metric left SAFM's family-level lead intact, and M3 failing the ZINC floor is positive evidence that SAFM does not copy SCIN targets.Enforced strict Δu>0\Delta u>0 & Ho–Salimans CFG; dropped Identity row; unified naming; wrote Limitations & all 4 analyses.Sec 2, 3.3, 6: Standard CFG + 5-point Limitations. App B (Fig. 7): ZINC gates. App B.3: SAFM (no source) >> interp. to floor 0.7 (p<0.001p<0.001 to 0.6, p=0.033p=0.033 at 0.7). App B.2, B.4: Medians 0.18/0.13; NN Tanimoto (0.45 ZINC; 0.99 peptide, chirality-unaware 2D FP).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.