acceptodds
Under review as a conference paper at ICLR 2027

SIGMA: Semantic Identifier Grouping for Molecular Autoregression

Abstract

Autoregressive molecular models assign probability to molecular serializations even though chemical identity is invariant to serialization. Equivalent serializations can induce different internal states along a shared molecular continuation. Randomized strings broaden exposure, but do not explicitly supervise correspondence between those states. We introduce SIGMA, a dense suffix-position objective built from chemically certified same-suffix triplets: two equivalent histories, one non-equivalent history, and a shared suffix. SIGMA aligns corresponding pre-token hidden states along the continuation and enforces a finite similarity margin over the non-equivalent history, leaving the language-model objective, decoder, and inference procedure unchanged. We compare SIGMA with canonical training, randomized-serialization training, and last-token alignment across four datasets under SMILES and SELFIES. SIGMA lowers test-reference Frechet ChemNet Distance against every control in all three training seeds on all four SELFIES datasets and on QM9 and ZINC under SMILES. Position-wise analyses show improved state correspondence, chemical discrimination, and top-1 next-token agreement while preserving between-molecule information. Full-corpus ZINC ablations under both representations identify correct state correspondence, rather than extra computation alone, as an effective ingredient. Beyond generation, a whole-molecule SIGMA extension improves mean predictive performance on all six molecular property benchmarks and reduces serialization sensitivity on every task relative to a matched randomized-serialization control.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular generation, autoregressive modeling, molecular representation learning, token-level contrastive learning, chemical equivalence, SMILES, SELFIES

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_EKpTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SIGMA, a fine-grained alignment objective for autoregressive molecular models. It constructs chemically verified same-suffix triplets to align hidden states of equivalent molecular serializations at each suffix position. Experiments on four molecular datasets using SMILES and SELFIES show improvements in generation quality, prediction consistency, and downstream molecular property prediction.

Critical strengths
  1. The chemically verified triplet construction provides a principled way to identify corresponding hidden states across molecular serializations.

  2. The experiments cover multiple datasets and molecular representations, with useful ablations and analyses of prediction consistency.

Critical weaknesses
  1. Insufficient motivation. Canonical SMILES already provides a deterministic molecular representation. The paper does not clearly establish why serialization inconsistency is a critical limitation for molecular generation or why explicit alignment is necessary.

  2. Limited novelty and justification. SIGMA is essentially a fine-grained contrastive alignment objective. It remains unclear why aligning every intermediate hidden state should improve generation, especially since chemically equivalent histories do not necessarily imply identical next-token distributions.

  3. Insufficient baseline comparisons. The randomized-serialization baseline uses fixed alternative serializations rather than online augmentation, and standard whole-molecule contrastive learning baselines are missing. Although the QM9 ablations support dense alignment, broader comparisons are needed to establish its advantage over simpler approaches.

Questions for authors
  1. Given the availability of canonical SMILES, what practical limitation specifically requires fine-grained serialization alignment?

  2. Why should chemically equivalent histories have similar hidden states at every suffix position when their distributions over alternative next tokens may differ?

  3. Can the authors compare against stronger randomized augmentation and whole-molecule contrastive learning baselines under matched training and computational budgets?

3 comments

Thank you for your thoughtful review. We address your questions by testing what the model learns from explicit correspondence between spellings, beyond what it learns from seeing more spellings.

Q1. What does SIGMA add beyond canonicalization and augmentation? Canonicalization chooses one spelling for a completed molecule. Even online augmentation, which gives the model new spellings each epoch, trains it to predict each spelling independently. SIGMA adds an explicit training signal for which intermediate steps correspond. For example, CCO and C(C)O both describe ethanol and can both append C to reach the same ether. SIGMA supervises the correspondence at that prediction step. On SMILES–QM9, given the correct prefixes at verified matching positions, top-1 agreement rose from 72.34% to 80.14% relative to canonical training, and the fraction where both predictions matched the shared next token rose from 49.00% to 54.04%. The predictions became more often correct together, not just more alike.

Q2. Why align intermediate states if their next-token distributions can differ? A chemically verified shared continuation tells us which prediction steps correspond across two spellings. At each supervised step, SIGMA trains the anchor state to identify the corresponding pre-token state in the other spelling. This gives the model a specific target for alignment without requiring the two states to assign identical probabilities to alternative next tokens. Sharing the next token alone is not enough: the chemical check must also verify molecular equivalence along the shared continuation.

Q3. Does the choice of alignment positions improve free generation? With the same auxiliary sequences and forward passes in the ZINC ablation, correct matches lowered validation FCD by 0.051 on SMILES and 0.542 on SELFIES relative to ordinary language-model training. Shuffled matches did not reproduce those gains, supporting the role of correct state pairing beyond exposure to extra spellings.

SELFIES–ZINC test. SIGMA achieved lower test FCD in all three seeds than both whole-molecule contrastive learning and alignment at the final suffix position. Whole-molecule contrastive controls with and without a projection head were tuned on validation. FCD is mean ± SD across three training seeds.

MethodTest FCD ↓
Whole-molecule contrastive learning (EOS; validation-selected)7.120 ± 0.144
Margin-gated alignment at the last suffix position7.225 ± 0.140
SIGMA (position-wise alignment)6.502 ± 0.159

Whole-sequence readout controls. On SELFIES–ZINC validation (training seed 101), we also compared SIGMA with two ways to form a single sequence vector for contrastive learning. Values are mean ± SD over five paired generations.

MethodValidation FCD ↓
Mean pooling over molecular tokens6.028 ± 0.115
Right-padding with zeros, then flattening without chemical alignment5.872 ± 0.075
SIGMA (position-wise alignment)5.643 ± 0.101

Matched GPU training time. On SELFIES–ZINC at 550s, SIGMA with anchor reuse achieved 19.51% lower test FCD than long-run online randomization, winning all three seeds. Each pair used the same physical A100 and the same 111,041 training molecules, with a measured training-time gap under 0.033s (excluding CPU preparation). Online randomization drew new spellings of the same molecules. SIGMA reused the triplet anchor's forward pass for next-token prediction; the alignment loss was unchanged, and the positive and negative views still participated.

FCD ↓SIGMA (anchor reuse)Online
550s test3.253 ± 0.0794.041 ± 0.004
550s validation2.489 ± 0.0433.284 ± 0.027
415s validation3.081 ± 0.0664.094 ± 0.136

Values are mean ± SD across three seeds, with five paired generations per seed. The two 550s rows score the same checkpoints against different references. The separate 415s validation comparison also included the original SIGMA recipe (4.545 ± 0.126); anchor reuse beat both alternatives in all three seeds. See JZ6h (Q1) for the shorter 235s test comparison.

Thank you for your detailed rebuttal. Your responses have addressed my concerns, and I have updated my score to 3.

Thank you for taking another look at our response. Your questions about stronger augmentation and whole-molecule contrastive baselines helped us clarify what aligning intermediate states adds.

Reviewer_JZ6hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SIGMA pretrains an autoregressive molecular model with a dense suffix-position objective built from chemically certified same-suffix triplets. Each triplet contains two equivalent prefixes and a non-equivalent prefix with the same continuation. A position-matching objective aligns corresponding pre-token states, while a gated margin term separates the positive and negative similarities. Experiments evaluate SMILES and SELFIES generation, representation consistency, and molecular property prediction.

Critical strengths
  1. I find the use of chemical identity to supervise intermediate prediction states interesting. The shared suffix gives the model prediction steps to align, making the supervision more targeted than simply encouraging two complete molecular representations to be similar.
  2. The SELFIES results are encouraging: Table 1 shows lower FCD across all four datasets within the certifiable evaluation subsets. I appreciate the three training seeds, five paired generation repetitions per seed, and hierarchical bootstrap intervals.
  3. The method adds no model parameters or autoregressive inference cost. This is a useful practical feature: once training is complete, the model can use the same generation procedure without an additional alignment step.
Critical weaknesses
  1. I would like to see a stronger augmentation baseline. Section 4 uses one fixed alternate serialization per molecule, while SIGMA receives extra views and supervision through its triplets. Resampling serializations during training would provide a more convincing comparison. The shuffled-pair control helps show that correct correspondence matters, but it does not answer whether SIGMA beats ordinary augmentation with comparable molecular exposure.
  2. It is unclear whether the gains justify the extra training time. SIGMA-Margin takes ~4 times as long as the baseline LM in Table A10, while the main six-layer, 256-dimensional models train for only three epochs. Could longer LM training or more augmented examples achieve similar gains at the same cost? Learning curves and a baseline matched for time or FLOPs would help. The SMILES results are mixed: FCD improves by 9.9% on QM9 and 6.7% on ZINC over the strongest controls, but the MOSES gains over canonical and last-token training are inconclusive, and canonical training performs best on ChEMBL.
  3. The split checks need to be clearer. I could not find molecular-identity overlap counts across splits, including auxiliary positives and negatives. Please report these and specify whether the property tasks use random or scaffold splits. I suggest that the authors include the deduplication methodology in the main paper.
  4. I would like a clearer picture of what the FCD gains mean for generation. The appendix includes validity, uniqueness, novelty, connectedness, and NLL. Diversity or scaffold coverage of generated molecules, as well as PMO, would be useful to understand if the proposed approach improves molecular design.
Questions for authors
  1. Does SIGMA outperform repeated randomized-serialization augmentation at comparable molecular exposure, and canonical or augmented LM training at matched compute?
  2. What molecular-identity overlap exists between training, validation, test, and auxiliary examples, and which splitting protocol is used for the property tasks?
  3. Do the FCD improvements preserve molecular diversity and translate into useful goal-directed generation?
4 comments

Thank you for your careful review and constructive feedback. We address your questions below, starting with the comparison against online randomization at matched GPU training time.

Q1. Does SIGMA beat online randomization at the same GPU training time? Yes. In the 550s SELFIES–ZINC comparison, SIGMA with anchor reuse achieved 19.51% lower test FCD than long-run online randomization (3.253 ± 0.079 vs. 4.041 ± 0.004), winning all three seeds. Within each seed, both methods trained on the same physical A100 and the same 111,041 molecules, with a measured training-time gap under 0.033s. Online randomization drew new spellings of those same molecules. We averaged five paired generations per seed against the fixed 2,516-molecule test reference and report mean ± SD across seeds.

Anchor reuse uses the triplet anchor's existing forward pass for next-token prediction instead of a separate pass on the canonical sequence. Both SIGMA variants used the same verified triplets and the same alignment loss; the positive and negative views still contribute to alignment.

550s validation. Against the validation reference, the same six checkpoints gave 2.489 ± 0.043 vs. 3.284 ± 0.027 validation FCD, with anchor reuse ahead in all three seeds.

415s three-way validation. To isolate the effect of reusing the anchor, we also timed the paper's separate canonical prediction pass. Within each seed, all three runs shared the same A100, molecule set, initialization and time-based learning-rate schedule, and differed by less than 0.13s in training time.

SELFIES–ZINC, ~415s GPU trainingValidation FCD ↓Training time (s)Prediction targets (millions)
SIGMA, reusing the anchor forward pass3.081 ± 0.066414.688 ± 0.02722.634
Online randomization4.094 ± 0.136414.674 ± 0.03329.412
SIGMA, separate canonical prediction4.545 ± 0.126414.778 ± 0.02612.377

FCD and time are mean ± SD across three seeds, with five paired generations per seed. Validity was 100% for all three methods. Anchor reuse beat both alternatives in all three seeds: mean validation FCD was 24.76% lower than online randomization and 32.23% lower than separate canonical prediction. Online randomization processed more next-token targets on the same molecules.

Matched-time test confirmation at about 235s. In a separate comparison, anchor reuse and online randomization trained on the same 111,041 molecules and the same physical A100 within each seed, with a training-time difference under 0.4s. Validation FCD was 5.116 ± 0.089 vs. 7.605 ± 0.093 (32.73% lower). Scoring those same six checkpoints on the fixed 2,516-molecule test subset gave 5.975 ± 0.103 vs. 8.557 ± 0.152: 30.17% lower test FCD, with anchor reuse winning all three seeds.

Why reuse the anchor? At a fixed 2,607 updates on the same triplets, it cut training time by 43.67%, though validation FCD was higher than with separate canonical prediction (5.147 ± 0.076 vs. 4.607 ± 0.069). With the 415s budget, anchor reuse fit 4,616 vs. 2,605 updates on average and gave lower FCD. The appendix's 4.05× instead compares the original three-epoch recipe with the full-corpus language model.

CPU preparation. Rebuilding and certifying the shared triplets took 668.961s with 12 CPU workers, once for all three seeds, using an existing serialization cache whose creation is not included. Preparing all 11 online cache epochs took 83, 85, and 85s for seeds 101, 202, and 303 on 126 allocated CPU cores; the timed runs used only part of those caches. Preparation is excluded from the GPU training times; these are not end-to-end comparisons.

Original training recipe. The paper's original SIGMA recipe (with a separate canonical prediction pass) had lower test FCD than three-epoch online randomization in all three settings, but extending online training to match estimated compute rather than measured GPU time gave Online lower test FCD in both ZINC settings and similar FCD on SMILES–ChEMBL.

In the separate ZINC ablation, correct correspondence improved FCD; shuffling did not reproduce the gain at the same auxiliary compute. SIGMA adds no decoding cost.

Q2. Which molecular identities are shared across splits? The triplet construction drew its source molecules from the training pool only. We verified the anchor/positive and negative source identities in all eight representation–dataset settings. The overlaps below were already present in the supplied splits. Each cell gives the counts using two identity rules: canonical isomeric graph signature / Standard InChIKey.

DatasetTrain–validationTrain–testValidation–test
QM917 / 65817 / 6612 / 113
ZINC0 / 10 / 00 / 0
ChEMBL29 / 12365 / 3338 / 24
MOSES0 / 20 / 110 / 1

Triplet source overlap with test references. Counts are distinct Standard InChIKeys shared by training-triplet source molecules and the test scoring references. Each cell lists anchor/positive sources first, then negative sources. These overlaps were inherited from the training pools; the table does not count completed negatives (C + suffix).

DatasetSMILES: anchor/positive / negativeSELFIES: anchor/positive / negative
QM9548 / 332148 / 91
ZINC0 / 00 / 0
ChEMBL21 / 11154 / 99
MOSES1 / 15 / 4

We separately tested whether the QM9 FCD gain depends on exact matches to training molecules. We applied one shared Standard InChIKey exclusion list to both test-reference and generated sets for every method and representation. It includes all next-token training molecules and every completed auxiliary negative (C plus the shared suffix). We deduplicated by the same identity rule and used 2,000 molecules per set. SIGMA beat all three controls in every seed; relative to canonical training, FCD fell from 1.509 ± 0.070 to 1.391 ± 0.064 on SMILES and from 1.323 ± 0.038 to 1.216 ± 0.039 on SELFIES (7.81% and 8.09% lower). In a separate full-corpus check, both SELFIES–QM9 models trained on all 107,059 eligible molecules. SIGMA lowered test FCD from 1.491 ± 0.219 to 1.092 ± 0.134 (26.7%), improving all three seeds.

All six property tasks used scaffold-grouped supervised splits. No molecular identity or nonempty Bemis–Murcko scaffold appeared in both train and test. The shared LM pretraining corpus did contain some downstream test structures, but it did not contain their property labels.

Q3. Does the FCD gain preserve diversity or help goal-directed generation? Across 480 generation runs, mean fingerprint-based internal diversity stayed within 0.002 of canonical training in every setting. At 5,000 proposals, distinct nonempty scaffolds rose from 1,130 to 1,181 on SMILES–QM9 and from 3,407 to 3,445 on SELFIES–ZINC. In a separate warm-start comparison on three Practical Molecular Optimization (PMO) targets, 41 of 72 runs reached the 10,000-call budget. Among full-budget SELFIES runs, performance was similar on GSK3B and JNK3, and canonical training did better on DRD2. The FCD gains did not come at the expense of the measured diversity; they did not consistently translate into better target optimization.

Thank you for the clarifications and additional analyses. The authors addressed my main concerns and I will raise my score accordingly.

Thank you for taking the time to review the additional experiments. Your suggestion to compare against longer online randomization at matched training time made the evaluation stronger.

Reviewer_dVNCOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Autoregressive molecular models assign likelihood to strings while chemical identity is invariant to serialization, so two equivalent histories (CCO, C(C)O) facing the same continuation can induce different internal states. Randomized-SMILES augmentation exposes more traversals but never identifies which intermediate states correspond.

SIGMA makes that correspondence an explicit training signal. The construction is a “chemically certified same-suffix triplet”: an anchor A, an equivalent positive B, and a valid non-equivalent negative C, all sharing an exact token suffix S, with equivalence verified by complete Standard InChIKey at every jointly decodable offset rather than only at the start or the end. At each suffix offset j, suffix-relative indexing rX(j) = bX + j − 2 picks out the unique pre-token state about to predict sⱼ; the positive term is a padding-masked softmax retrieval of rB(j) from A’s state, and the negative term is a hinge max(0, s⁻ − s⁺ + γ) that switches off once the margin is met. The LM loss, head, decoder, and inference cost are unchanged.

Validation spans four datasets (QM9, ZINC, ChEMBL, MOSES-300k) × two representations × three seeds × five paired generations, against canonical, randomized-serialization, and last-token-alignment controls under matched initialization and decoding. Headline: lower test-reference FCD than every control in all three seeds on all four SELFIES datasets and on SMILES QM9/ZINC. Supporting evidence: position-wise mechanism diagnostics (six per domain, all improving), information-preservation probes (effective rank, descriptor R², fingerprint AUROC, scaffold strata), full-corpus compute-matched ablations including a shuffled-pair control and a symmetric-KL output-space variant, a density factorial separating dense-vs-last-token from loss form, four propositions with proofs, an independent re-implementation of the QM9 certification, and an EOS-level extension to six MoleculeNet tasks.

Critical strengths
  1. The ablation design isolates the claimed mechanism rather than merely demonstrating an improvement. The full-corpus study (Table A10, 45 models, 225 paired evaluations) holds LM stream, initialization, schedule, and forward-pass budget fixed, and includes the one control that most such papers omit: a shuffled-pair condition that performs identical auxiliary computation on deliberately corrupted chemical pairings. It yields nothing (SMILES–ZINC Δ = −.0066, CI crossing zero; SELFIES–ZINC +.0026, crossing zero) while correct correspondence yields −.0505 and −.5424 with all three seed directions. Since every auxiliary condition costs 3.3–4.5× LM training wall-clock, this is the comparison that distinguishes “the mechanism works” from “more compute helps,” and it is the reason I believe the central claim. The separate density factorial (Table A12) then isolates the second claim — dense beats last-token under both positive-only (Δ = −.0317, CI [−.0617, −.0068]) and matched positive-plus-negative (Δ = −.0334, CI [−.0567, −.0106]) — and incidentally motivates the margin gate by showing an always-active negative hurts in both regimes (+.0509, +.0492).

  2. The mechanism analysis separates what the loss optimizes from what it does not, and reports both. Retrieval accuracy and ordering violations are directly supervised, so improving them proves little; the paper says so and then measures quantities the loss does not touch. Top-1 next-token agreement between equivalent histories rises .7234 → .8014 on QM9, and suffix-token NLL falls, in all four domains and three seeds. Crucially, the paper anticipates the obvious deflationary explanation — that cosine alignment is just collapsing representations — and tests it four ways: effective rank and descriptor R² increase in every domain, Morgan macro-AUROC moves by at most .0038, and the scaffold ordering (same-molecule > same-scaffold > different-scaffold) holds in every dataset and seed. Having pre-registered what would count as collapse (Section C.4’s “SIGMA is called selective only if...”) and then checked it is what makes the mechanism claim credible rather than decorative.

Critical weaknesses

The method is a net loss on the one representation built for fragment assembly, and the main text doesn’t convey how badly. On SAFE (Table A23), last-token alignment beats SIGMA on all four domains — ZINC 9.5305 vs 9.6728, MOSES 10.2188 vs 10.6665, QM9 .7157 vs .7691, ChEMBL 4.3037 vs 4.3252 — and SIGMA’s connectedness is lower on all four too. Calling this a case where “SAFE exposes a fragment-assembly limitation” reads as a scope note, not as a control beating the method everywhere. What bothers me isn’t the loss so much as what it implies about generality. SAFE connectedness is terrible for everyone (.237 on ZINC, .221 on MOSES), so no objective works in this regime — but it’s also the regime where the denser supervision is consistently worse than the sparse one, which cuts against the density argument that Proposition A3 and Table A12 exist to support. Put that together with SMILES–ChEMBL, where SIGMA beats canonical in 0/3 seeds and no auxiliary objective beats the full-corpus LM (Table A11), and the pattern is that gains show up where strings are short and regular and disappear or reverse where they’re long, repetitive or fragment-structured. The abstract’s framing — four SELFIES datasets plus QM9 and ZINC under SMILES — is accurate, but it’s the favorable half of the grid.

Questions for authors

Does the FCD advantage survive at matched novelty? E.g., report FCD restricted to novel generations, or an FCD–novelty frontier across λ_aux, on QM9 under both representations. This is the single answer most likely to move my score. The matching term is directed A→B and not symmetrized, though chemical equivalence is symmetric. Was the symmetrized version tried? If it was not, is there a reason to expect it would not help, given that the construction already requires both forward passes? The certified-subset SELFIES regime drops LM loss for non-certifiable molecules. Can the SELFIES result be reproduced in the full-corpus regime used for SMILES, at least on QM9, to confirm the gain is not partly a coverage artifact?

4 comments

Thank you for the thoughtful review and for suggesting a direct test on novel molecules. We start with that test, then address the SAFE and alignment ablations.

Q1. Does the FCD gain hold on novel molecules? Yes. SIGMA had the lowest QM9 FCD in every seed under both representations, even after removing molecules seen only as auxiliary negatives. The table compares FCD at a fixed count of 2,000 novel molecules per set, not the fraction of outputs that are novel (reported in Table A6).

QM9 training methodSMILES FCD ↓SELFIES FCD ↓
Canonical1.509 ± 0.0701.323 ± 0.038
Randomized serialization1.626 ± 0.0631.435 ± 0.119
Last-token alignment1.471 ± 0.0781.329 ± 0.036
SIGMA1.391 ± 0.0641.216 ± 0.039

We built one Standard InChIKey exclusion list from the actual next-token training sets for both representations and every completed auxiliary negative (the negative prefix plus its shared suffix). We applied it to every generated set and both test-reference pools, deduplicated by the same identity rule, and took the first 2,000 eligible molecules from each. Every saved generated set had at least 2,210 eligible molecules. The table reports mean ± SD across three training seeds, after averaging five paired generations within each seed.

Relative to canonical training, SIGMA lowered novel-only FCD by 7.81% on SMILES and 8.09% on SELFIES.

Q2. Would within-fragment alignment recover SAFE performance? On SAFE–QM9, restricting alignment to within-fragment positions gave 0.780 ± 0.025 validation FCD, versus 0.773 ± 0.025 when dropping the same number of positions at random. The two masks also produced nearly identical connected yields (809.1 vs. 808.4 molecules per 1,000 proposals). This tests which positions to align while holding their count fixed; it did not identify dot or cross-fragment attachment positions as the cause.

On SAFE–MOSES, the original dense and last-token models reveal a distinction between assembling a connected molecule and the fidelity of the molecules that are assembled. Using the same decoder and proposal budget, we scored both the first 5,000 valid outputs and, separately, exactly 1,000 connected outputs. Values are from validation and report mean ± SD across three training seeds.

SAFE–MOSESLast-token alignmentSIGMA
FCD, 1,000 connected outputs ↓4.417 ± 0.1604.168 ± 0.139
Connected outputs per 1,000 proposals ↑152.1 ± 10.3132.6 ± 12.4
FCD, 5,000 valid outputs ↓10.505 ± 0.28210.995 ± 0.607

SIGMA gave lower connected-only FCD in every seed. Last-token alignment still gave higher connected yield and better all-valid FCD. Dense alignment improved fidelity among connected outputs without solving the assembly bottleneck.

Q3. Does removing repeated-token competitors help? On SMILES–ChEMBL we excluded nonmatching positions that predict the same token from the position-retrieval denominator, keeping the true matching position. The equal-count control removed random competitors instead. Test FCD was 1.104 ± 0.063 with same-token removal and 1.111 ± 0.073 with random removal. The small difference did not close the gap to multi-view prediction (1.069 ± 0.025), so removing those competitors alone was not sufficient.

Q4. Did we try aligning in both directions? Yes. Directed alignment asks the anchor to retrieve its matching position in the positive view. The bidirectional variant also asks the positive view to retrieve its match in the anchor and averages the two losses, using the same correspondences. Test FCD for bidirectional vs. directed alignment was 1.106 ± 0.068 vs. 1.110 ± 0.066 on SMILES–ChEMBL and 6.582 ± 0.100 vs. 6.502 ± 0.159 on SELFIES–ZINC. Reversing the query is valid, but adding that direction did not consistently improve generation.

Continuing our reply to dVNC: below are the chemical-identity audit, the proximity check, and the two tests of whether the results extend beyond the original generation setup.

Q5. How often do stricter identity checks disagree? On QM9, canonical isomeric graph signatures disagreed on fewer than 0.5% of the completed positive pairs in either representation. We checked completed molecules and intermediate positions separately, since the former do not establish equality along the trajectory. The table counts disagreements among Standard InChIKey matches; intermediate denominators include only supervised positions where both prefixes parse.

QM9 audit unitComparedGraph signature differsFixed-H differs
SMILES, completed80,932355 (0.439%)353
SELFIES, completed45,93664 (0.139%)62
SMILES, intermediate188,308555 (0.295%)551
SELFIES, intermediate415,280346 (0.083%)338

For example, two parsed prefixes can place an aromatic hydrogen on different nitrogens and still have the same Standard InChIKey. The audit quantifies that distinction; Standard InChIKey is the training identity rule, not a claim of tautomer-specific graph equality.

Q6. Are generated molecules closer to the training set? Yes, by the nearest-neighbor measure the review suggests. Mean fingerprint Tanimoto similarity to the closest training molecule rose from 0.726 to 0.747 on SMILES–QM9 and from 0.627 to 0.645 on SELFIES–QM9. This is a shift in structural proximity, whereas effective rank measures the spread of internal states. The two quantities need not move together. The novel-only result in Q1 shows that exact matches to next-token training molecules or completed auxiliary negatives are not needed for the FCD gain.

Q7. What transfers to property prediction? We see this as a test of whether alignment makes molecular predictions less sensitive to serialization. The end-of-sequence variant reduced within-molecule prediction variability across paired spellings on all six tasks. For example, on BACE the mean within-molecule prediction SD fell from 0.1039 to 0.0573. Task performance means also favored SIGMA on all six tasks against the matched randomized-serialization baseline (three-seed mean ± SD).

TaskMetricRandomized baselineSIGMA
ESOLRMSE ↓1.013 ± 0.0810.970 ± 0.082
FreeSolvRMSE ↓2.320 ± 0.1552.246 ± 0.108
LipophilicityRMSE ↓0.758 ± 0.0210.722 ± 0.005
BACEAUROC ↑0.823 ± 0.0120.834 ± 0.013
BBBPAUROC ↑0.920 ± 0.0260.943 ± 0.005
ClinToxAUROC ↑0.761 ± 0.1550.854 ± 0.061

The paired-spelling result is the direct evidence for transfer: a prediction changes less when the molecule stays the same and only its serialization changes.

Q8. Does the SELFIES result hold with the full training corpus? Yes. Both SELFIES–QM9 models used all 107,059 eligible training molecules for next-token prediction, including those without a certified triplet. The original raw test pool was filtered only for lossless SELFIES round-trip and context length, not triplet eligibility. This is a separate test from Q1's novel-only scoring.

Test FCD ↓CanonicalSIGMA
Seed 1011.7391.228
Seed 2021.4131.089
Seed 3031.3220.960
Mean ± SD1.491 ± 0.2191.092 ± 0.134

SIGMA lowered test FCD by 26.7% and improved all three seeds. The supervised triplets add an alignment signal without requiring the language model to train on a smaller set of molecules.

Reviewer_dVNCOpenReview ↗

Response

Thanks for the thorough response — the novel-only test addresses my main concern, and I’m raising my score to 4.

My remaining reservation is presentational:

  1. Add the novel-only FCD and full-corpus SELFIES results into the main text.
  2. State the SAFE result in the main text as a control beating SIGMA on all four domains.

Thank you for the follow-up and for suggesting the novel-only evaluation. We’ll report the novel-only FCD and full-corpus SELFIES–QM9 results in the main text and state explicitly that last-token alignment achieved lower FCD than SIGMA on all four SAFE datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.