acceptodds
Under review as a conference paper at ICLR 2027

SkeleMol: Molecular Property Prediction with Skeletal Images

Abstract

Molecular property prediction is central to drug discovery and chemistry, as it lets candidate molecules be screened before synthesis and testing. It typically relies on specialized representations such as molecular graphs, SMILES strings, or 3D conformers, together with models pre-trained specifically on molecular data. Models for each representation are growing increasingly complex, and the field is fragmenting along input modalities. We investigate a different approach: representing molecules as skeletal images and adapting general-purpose vision encoders to molecular prediction. We introduce , an image-based framework that uses only a molecular drawing at inference time. It enables pre-trained vision encoders for molecular property prediction, replacing the molecule-specific architecture. During pre-training, we progressively inject chemical and geometric information through image-SMILES alignment, distillation from a 3D molecular model, and a chemistry-informed curriculum. Across 10 vision architectures and seven pre-training strategies, we find that the choice of molecular supervision has a substantially larger effect than the choice of vision backbone, with consistent strategy rankings across architectures. In particular, transferring 3D knowledge improves an image encoder despite the absence of conformers at inference, while curriculum learning further improves data efficiency. Across 10 benchmarks spanning physical, biological, and quantum properties, achieves state-of-the-art results on five tasks using 2 M pre-training molecules, while retaining a standard vision architecture and image-only inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular property prediction, molecular images, vision encoders, cross-modal distillation, curriculum learning, contrastive pre-training, drug discovery

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_TCBAOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces SkeleMol, which adapts general-purpose vision encoders for molecular property prediction using skeletal images. It combines multiple stages in pretraining, including image–SMILES alignment, distillation from a 3D molecular model, and chemistry-informed curriculum learning. It outperforms molecule-specific baselines on physical properties (ESOL, LD50), and a couple of drug discovery-related properties including BBBP and HIV.

Critical strengths
  • The framework makes good use of molecular data in diverse forms, including skeletal images, SMILES, and representations from a 3D molecular model.
  • Chemistry-informed curriculum learning seems an interesting way to help the encoder learn basic chemical structure patterns before introducing more complex molecules. Further, the curriculum ablations provide useful evidence of its downstream benefits.
Critical weaknesses
  • The authors do not include a performance comparison with image-based property prediction models such as ImageMol [1], which are their most direct baselines. Further comparison with these baselines would help demonstrate the efficacy of the proposed training recipe over existing image-based approaches.
  • Table 13 reports large error shifts when images are rendered with PIKAchU rather than RDKit. This raises concerns about its generalization to images drawn from other renderers or published papers, and limits its broader applicability.

References

[1] Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework, Nature Machine Intelligence (2022).

Questions for authors
  • What motivates using teacher distillation to transfer 3D information? PCQM4Mv2 also provides 3D coordinates for training molecules, so a straightforward alternative would be to add an auxiliary coordinate prediction head and supervise it directly with the reference structures, which does not necessitate employing a teacher model.
1 comment

Summary. ImageMol is now a row of Table 5. It pre-trains on 10 M molecules and SkeleMol on 2 M. Table 14 measures sensitivity to the renderer. We distil from Uni-Mol because its embedding is already invariant to frame and atom order.

We thank the reviewer for the positive assessment and for two concrete requests, both now in the paper.

ImageMol as the most direct baseline

ImageMol is now a row of Table 5 with the scaffold-split values its paper reports (its Supplementary Table S5), and App. C.4 records the split behind the row. Like every row of Table 5, it is a comparison of reported values. ImageMol trains a molecule-specific image encoder on 10 M molecules, and SkeleMol adapts an off-the-shelf vision encoder with 2 M. Section 2 also places MaskMol and VideoMol.

Sensitivity to the renderer

The depiction table (Table 14) quantifies it by scoring the same encoder and heads under six RDKit variants and under PIKAchU on 1,000 held-out molecules. Within RDKit, predictions are stable under layout and typography. Bond width, rotation, type size and explicit methyls each remain within 10% of the default MAE (for example rotation 35° +2.2% on Δε, explicit methyls −0.1%). The element colours and the training resolution carry signal the encoder uses, and Table 14 gives each value. PIKAchU changes bond layout, label sizing and ring rendering at once and lies outside the training depictions, and its Δε error is 47.7% higher on the 992 molecules it drew. The encoder is specific to the drawing style it trained on, and we validate it on drawings from its training pipeline (App. D.3). Depiction-augmented pre-training (sampling renderers, palettes and resolutions) is the natural extension and changes only the rendering step.

Why distil from a teacher and not regress coordinates directly?

  1. Coordinates have no canonical frame or atom order from an image. A coordinate head must output N × 3 values in some frame and atom ordering, and the image fixes neither (PCQM4Mv2's geometries come in arbitrary frames, and the drawing does not say which pixel is atom 1). Making the target invariant means predicting a distance matrix or a set of invariants, which amounts to designing a geometry representation. The Uni-Mol embedding is already an SE(3)- and permutation-invariant summary of the conformer, so distillation avoids this step.
  2. The teacher transfers geometric priors learned from a large conformer corpus. Uni-Mol is pre-trained on 209 M conformers, and its embedding encodes what it learned about plausible geometries and their relation to properties beyond the one structure in hand. Table 23 shows the transfer. Distances in the student space correlate with the teacher's at ρ = 0.65, against 0.24 for the released encoder and 0.01 for a random one.
  3. The gap head already covers direct supervision. The β term in Eq. 2 is direct supervision from a PCQM4Mv2 label (the HOMO–LUMO gap). Eq. 3 shows why a scalar target constrains one embedding direction per molecule while the teacher constrains all of them. Direct coordinate supervision would lie between the two and would need the invariant target of (1).
Reviewer_V6YmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SkeleMol for molecular property prediction from 2D skeletal images, combining cross-modal pretraining, 3D knowledge distillation, and curriculum learning. It evaluates different pretraining strategies and visual backbones on ten downstream tasks, while requiring only images at inference.

Critical strengths
  1. Evaluation across multiple training strategies and downstream tasks.
  2. The paper highlights that pre-training strategies have a greater impact on model performance than the choice of network architecture.
Critical weaknesses
  1. The method primarily combines established components: contrastive visual pretraining, cross-modal alignment, teacher-student distillation, and an easy-to-hard curriculum. I do not identify a substantive methodological innovation in either the pretraining objectives or the curriculum mechanism, and the paper does not establish a sufficiently new insight arising from their combination.
  2. Despite its focus on molecular images, the paper omits relevant methods such as VideoMol [1], MaskMol [2], and IEM [3] from its discussion and comparisons. It also overlooks prior work on curriculum learning for molecular representation learning, including CurrMG [4] and SemiMol [5]. These omissions lead to an overestimation of the innovativeness of the method proposed in this paper.
  3. Methods such as IEM [3] already use 3D-derived supervision during training while requiring only graphs at inference. Since SkeleMol’s images are rendered from available molecular structures, the advantage of an image-only student remains unclear without a controlled comparison against graph-only alternatives in predictive performance and computational cost.
  4. Figure 2 shows drawing-style variations, but the main text does not clarify their use in augmentation or provide corresponding quantitative results. Given the fixed RDKit rendering pipeline, their relevance to the evaluated setting is unclear; I suggest moving the figure to the appendix.
  5. Table 2 defines T0 as hydrocarbons containing no heteroatoms; however, the T0 molecules for PCQM4Mv2 shown in Figure 4 explicitly contain B, S, and O.

[1] VideoMol. A molecular video-derived foundation model for scientific drug discovery. Nature Communications 15, 9696 (2024). [2] MaskMol. MaskMol: knowledge-guided molecular image pre-training framework for activity cliffs with pixel masking. BMC Biology 23, 279 (2025). [3] IEM. An Image-enhanced Molecular Graph Representation Learning Framework. IJCAI, 6107–6115 (2024). [4] CurrMG. An efficient curriculum learning-based strategy for molecular graph learning. Briefings in Bioinformatics 23(3), bbac099 (2022). [5] SemiMol. A Semi-supervised Molecular Learning Framework for Activity Cliff Estimation. IJCAI, 6080–6088 (2024).

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response (1/2)

Summary. We compare ten vision backbones under seven pre-training strategies with everything else fixed. On quantum tasks we observe that the strategy matters more than the backbone. The revision discusses the five suggested references, adds a graph baseline and a cost comparison, and revises Fig. 2 and Fig. 4.

We thank the reviewer for the five references, all now discussed, and for the note on Fig. 4.

What the paper contributes

We train ten vision backbones under seven pre-training strategies and test them on ten tasks. The renderer, head, schedule, seeds and 10-epoch budget are the same for every run. We make five observations.

  1. Every backbone ranks the strategies in the same order, and every strategy ranks the backbones in the same order (Fig. 4a). On the quantum tasks the strategy matters more than the backbone. From S3 to S6 the PCQM error falls several-fold for both backbones and Δε falls further than any change of backbone achieves. The backbone effect is at most ÷1.9 and −10% (Table 3). The full matrix makes this visible.
  2. Image–SMILES alignment beats image–description alignment on most tasks (Table 3, Δε 0.0102 vs 0.0122, BBBP 93.6 vs 91.0), even though a description has more words. Canonical SMILES encodes the whole graph, and a description encodes part of it.
  3. Ordering the corpus by chemistry tiers improves the result with 43% fewer molecule-views (5.74 M vs 10 M, with ESOL −10.7%, Δε −12.9% and PCQM −19.1% in Table 4), and the gains concentrate on the tasks that depend on fine structural detail.
  4. A student that reads only a drawing beats its 3D-conformer teacher on the properties governed by 2D structure (ESOL 0.627 vs 0.788, and BBBP, BACE and HIV) and remains within 1.6× of the best 3D model on PCQM4Mv2 without computing a conformer (Table 6).
  5. An off-the-shelf natural-image encoder with 2 M molecules of alignment matches or exceeds the best reported values on five of ten tasks, next to molecule-specific models pre-trained on up to 1.1 B molecules (Table 5).

ImageMol pre-trains on about 10 M molecules, and VideoMol and MaskMol on about 2 M, each with molecule-specific visual objectives. SkeleMol also uses 2 M molecules and starts from an off-the-shelf natural-image encoder.

Related work

Section 2 now discusses VideoMol (a 60-frame video of a rotating conformer, also needed at inference), MaskMol (pixel masking of atoms, bonds and motifs) and IEM (an image-enhanced graph learner), and Sec. 4 discusses CurrMG (difficulty measures for graph curricula) and SemiMol (curriculum for pseudo-label admission). SkeleMol differs from all five in evaluating general-purpose vision encoders under a fixed-budget matrix.

Why an image-only student (IEM)

IEM runs the other way. It distils a teacher trained on rendered conformer images into a graph student, so its inference input is a graph. Our input is a drawing, which is available wherever a structure exists only as a picture. The revision adds a controlled graph comparison with ChemProp's D-MPNN (v2.3, default architecture, 50 epochs, three seeds) on our exact files, validation-selected, with test scored once.

task (our split)ChemProp D-MPNNSkeleMol S6
ESOL RMSE ↓0.850 ± 0.0160.627
BACE AUC ↑87.6 ± 1.689.0
ClinTox AUC ↑86.6 ± 0.786.8
Tox21 AUC ↑73.4 ± 0.377.6
HIV AUC ↑81.4 ± 0.383.2
QM9-Δε MAE ↓ (Ha)0.00430 ± 0.000050.00697
QM9-U0 MAE ↓ (Ha)0.0206 ± 0.00050.499

S6 reads a single image and scores higher on ESOL, BACE, Tox21 and HIV, and it matches the D-MPNN on ClinTox. The D-MPNN reads the graph and is stronger on the two QM9 targets.

On a strict scaffold re-split of HIV and BACE, same checkpoint and files for both, S6 leads the D-MPNN on both tasks (BACE 83.2 ± 2.2 vs 79.8 ± 0.5 and HIV 76.0 ± 0.8 vs 75.3 ± 2.0) and leads the same encoder without pre-training by 11 points on BACE (72.4 ± 1.3).

We timed the routes on 10,000 PCQM4Mv2 molecules from SMILES. The conformer route needs 34 ms per molecule for conformer generation alone, fails on 10.4% and costs at least 45 ms end to end. Our ViT-B/16 renders in 6.65 ms with no failures and predicts in 9.25 ms end to end. Table 6 of the revision lists the cost of all three routes.

AuthorsOpenReview ↗

Response (2/2)

Summary. Fig. 2 moves to the appendix as Fig. 10, and Table 14 backs it with numbers. A script draws the tier figure from the curriculum's tier maps, and every T0 panel is a hydrocarbon.

Figure 2

Fig. 2 moves to the appendix as Fig. 10, and the depiction table (Table 14) backs it with numbers. It scores the S6 encoder under six drawing variants and a second renderer on 1,000 held-out molecules. Predictions are stable under layout and typography (bond width, rotation, font size and explicit methyls each remain within 10% of the default MAE), and the element colours and the training resolution carry signal the encoder uses.

Figure 4

Every T0 panel is a hydrocarbon. A script draws the figure from the tier maps the curriculum trains on, one molecule per tier and corpus, and each panel matches the map it comes from. The T0 examples are a methylnaphthalene from MolTextNet and an alkenyl-cycloheptane from PCQM4Mv2. In the revision the figure is Fig. 3.

I appreciate the authors' substantial efforts during the rebuttal, including the additional experiments and comparisons. I have also carefully considered the other reviewers' comments. However, my main concern remains unresolved: the five observations are largely expected, and the paper reports empirical findings without sufficiently explaining the underlying reasons (e.g., why image–SMILES alignment outperforms image–description alignment, or why pre-training strategies matter more than backbones for quantum tasks). The lack of deeper analysis and generalizable insights limits the scientific contribution. Therefore, I have decided to maintain my original rating.

We thank the reviewer for clarifying what was required and we would like to appropriately answer it.

Both answers come from one point. A structure drawing is a plain line image, few colours, even strokes, no texture, so what makes one vision encoder differ from another barely shows up in it.

Why the training signal matters more than the backbone. Eight of our ten encoders share the same image tower, so what varies is only what they were pre-trained on, and a line drawing gives that little to act on. Table 17 shows it. DINOv2-base never saw a molecule and still beats our distilled encoder on both scaffold-split QM9 targets. On the quantum tasks the drawing cannot show 3D structure, so the training signal is the only way in, and changing it is the only thing that moves those numbers. The teacher gives the student 512 numbers per molecule while the gap head gives one, which is why the teacher stage carries most of that gain, and the match with Uni-Mol's 3D space rises from 0.24 to 0.65. We will add a direct check, comparing all ten encoders on drawings and on ordinary photos.

Why SMILES beats descriptions. Every molecule has its own SMILES string. Descriptions repeat, because two molecules with the same functional groups get the same text, and the loss then treats a correct image and text pair as a wrong one. We will report how often descriptions repeat, how much structure each kind of text carries, and what changes once the repeats are removed.

On the findings being expected. Two are not. A general encoder with no molecular training beats ours on QM9, where labels are plentiful, so molecular pre-training stops paying once data is dense. And freezing the encoder instead of fine-tuning grows the pre-training gain three to five times, so fine-tuning hides most of what pre-training does. The revision leads with these.

We hope that all our analyses and empirical evidences are sufficient to positively convince the reviewer of the merits of our work.

Reviewer_nf7aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether a molecule's skeletal image, interpreted by a standard vision encoder, suffices for property prediction, and what pre-training supervision such an encoder requires. With renderer, head, schedule and seeds held fixed, the authors vary the pre-training strategy across ten backbones, ranging from image–image and image–SMILES alignment to 3D-teacher distillation with a HOMO–LUMO auxiliary head and a five-tier curriculum. Two million-molecule corpora are used and ten tasks evaluated. The authors contend that supervision matters more than backbone, and that a SigLIP2 encoder trained under the curriculum attains state-of-the-art results on five of ten benchmarks.

Critical strengths

1.The paper innovatively proposes a framework that reads only a skeletal image at inference while injecting the knowledge the image lacks during pre-training.

2.The proposed method achieves improvements across the strategy matrix, the low-label controls and the efficiency measurements.

Critical weaknesses

1.§5 states only that "each benchmark's standard split" is followed, yet the splits for BBBP, BACE, HIV, Tox21, ESOL, and LD50 are never specified; the 25-point spread in the BBBP column (68.5–93.7) requires explanation.

2.S5 simultaneously introduces a teacher stage, an auxiliary head, and an embedding term, so their individual contributions cannot be disentangled; an ablation is needed.

3.No stop-gradient is annotated on L_coord, so it cannot be ruled out that the loss is reduced by moving the projection rather than the student; the ρ = 0.65 in App. G.2 is compatible with both readings, and the DINOv2-base ρ control is missing.

4.The curriculum is also a 43% data-subset intervention; App. C.6 reports only contrastive loss and explicitly excludes those runs from the main tables, while the mixed schedule reaches the same terminal loss (0.0004). Fine-tuning the three existing checkpoints would settle the ordering question in a single sweep, with a fixed held-out set and runs extended to 30/100 epochs.

5.In Table 16, DINOv2-base wins on both metrics, with a Δε separation of roughly six combined standard deviations, yet the text does not acknowledge this direction; the defensible contribution is a small low-label improvement on two tasks, not the "state-of-the-art on five tasks" advertised in the abstract.

6.Both the kekulised and "no stereo" variants in Table 13 are no-ops (kekulisation is the default, and the flag does not affect wedge/hash bonds), so the only experiment targeting the "stereo annotations" claim tests nothing; PIKAchU, the sole genuinely out-of-distribution renderer, is reported only as "large," without a number.

7.The reporting of seeds and dispersion is internally inconsistent (no standard deviations in the main tables; Table 14 covers only two experiments), and several headline contrasts fall within the authors' own reported seed spread; mean ± sd should be given for the S3/S5/S6 decomposition, single-run cells marked, and the "+5.8" ClinTox jump not attributed to a column the authors concede is too noisy to rank.

8.Bundling multiple changes in S3→S6 makes the "5.8×" unattributable; on BACE the backbone spread is 0.072 versus a pre-training gain of 0.016, so the magnitude claim runs contrary to the paper's own controlled comparison.

Questions for authors

1.Which split does SkeleMol use for each of the ten tasks? Please add a split column to Table 10. If the answer is "the official MoleculeNet split", say so explicitly.

2.What are the numbers for α = 0, β = 0, a 2D teacher, a shuffled teacher, and an equal-size extra-corpus run without distillation?

3.Is the Linear projection in Alg. 2 frozen, or does ẑ_g carry a stop-gradient? And what is ρ (App. G.2) for DINOv2-base?

4.Will you fine-tune the mixed, sequential and anti-curriculum checkpoints from App. C.6 and report downstream results? Is Fig. 10 recomputable on a fixed held-out set spanning all tiers, and does the S5–S6 gap survive at 30/100 epochs?

5.Will you report DINOv2-base on all ten tasks? And how would you state the contribution in one sentence that remains true of Table 17's +0.004 on BBBP?

6.What is the PIKAchU MAE, numerically? And why is the no-stereo variant pixel-identical, given that addStereoAnnotation=False is the RDKit default?

4 comments
AuthorsOpenReview ↗

Response (1/4)

Summary. Table 10 now states the split of every task. BBBP numbers differ across papers mainly because the papers use different scaffold splitters. S6 scores higher than a graph baseline on ESOL, BACE, Tox21 and HIV. New runs test the parts of the S5 loss and the curriculum order, and DINOv2 now covers all ten tasks.

We thank the reviewer for a careful reading. The revision addresses the points the review rates as fixable.

  1. W1 / Q1a. Which splits, and which scaffold splitter
  2. W1 / Q1b. A graph baseline on our exact files
  3. W2 / Q2. What each S5 term contributes
  4. W3 / Q3. Stop-gradient and the projection; ρ for DINOv2
  5. W4 / Q4. Curriculum as a data-subset intervention; App. C.6; Fig. 10; 30/100 epochs
  6. W5 / Q5. DINOv2 on the label-rich tasks and the one-sentence contribution
  7. W6 / Q6. Depiction variants
  8. W7. Seeds and dispersion
  9. W8. The 5.8× and the BACE spread

W1 / Q1a. Which splits, and which scaffold splitter

Each task uses the following split.

tasksplittest ntasksplittest n
ESOLDeepChem scaffold 80/10/10113Tox21DeepChem scaffold, 12 tasks783
LD50TDC scaffold1,478HIVstratified random 60/20/208,224
BBBPDeepChem scaffold 80/10/10204QM9random 80/10/10 (scaffold in App. D.1)13,389
BACEsource split 1,433/48/3232PCQM4Mv2official OGB-LSC73,545 (valid)
ClinToxDeepChem scaffold 80/10/10148

Four tasks use the DeepChem scaffold split, LD50 the TDC scaffold split, HIV a stratified random split and BACE its source split. Table 10 of the revision states this for every task. Our BBBP split is the DeepChem scaffold split (1,631/204/204, no scaffold shared between partitions, verified on the released file).

The 25-point spread in the BBBP column comes from the splitter. "Scaffold split" is a family of partitions. MoLFormer-XL uses the MoleculeNet split, SELFormer builds its own with ChemProp's splitter, and the graph methods use the DeepChem splitter. These assign different molecules to the test set. ChemBERTa-3 (Singh et al., Digital Discovery 5:662–685, 2026) measured the gap by running one model suite under both splitters.

BBBP ROC-AUCMoLFormer splitsDeepChem splitter
Random forest on fingerprints0.926 ± 0.0020.694 ± 0.013
D-MPNN0.930 ± 0.0020.661 ± 0.001
GCN0.898 ± 0.0050.642 ± 0.011
ChemBERTa-MLM-100M0.961 ± 0.0030.700 ± 0.027
MoLFormer (its paper)0.937not reported
MoLFormer (their re-run)0.900 ± 0.0150.735 ± 0.019

Every model drops 16 to 27 points between the two conventions, which is the size of the spread in our BBBP column. A random forest scores 0.926 under one and 0.694 under the other. A BBBP number is comparable only within one splitter.

We confirmed this on our own files, fine-tuning the public MoLFormer checkpoint (ibm-research/MoLFormer-XL-both-10pct) on our BBBP split under our protocol (train, select on validation, score test once) for 30 epochs and three seeds.

MoLFormer on our BBBP filesROC-AUC
validation, selected epoch97.9 ± 0.4
test, same epoch74.2 ± 0.8
test, best of 30 epochs (invalid protocol, for reference)76.4 ± 0.4

A model that reports 93.7 scores 74.2 on our partition and at most 76.8 at any epoch. This matches ChemBERTa-3's 0.735 for their MoLFormer under this splitter. Together with the ChemBERTa-3 comparison, this shows that the splitter explains most of the gap between the reported 93.7 and the score on our files.

In the revision, App. C.4 gives the two-splitter comparison and our MoLFormer run, Table 10 gives every split and metric, and Sec. 5 points to the splitter behind the BBBP rows of Table 5. The paper's claims compare numbers within one splitter.

W8. The 5.8× and the BACE spread

Table 4 attributes the S3 → S6 gain step by step at fixed backbone and budget. PCQM error falls 0.549 → 0.117 → 0.095 for S3 → S5 → S6, a several-fold drop from the teacher stage and a further 1.24× from the curriculum. The BACE figures you quote (0.072 vs 0.016) are from Table 18. The first is the range across five encoders with no molecular pre-training, and the second is S6's margin over the best of them. Two of those encoders have a seed sd of 0.025 and 0.052 on a 48-molecule validation set. In the matrix (Fig. 4a, Table 3) every backbone ranks the strategies in the same order, and every strategy ranks the backbones in the same order. On the quantum columns the strategy matters more than the backbone. From S3 to S6 the PCQM error drops several-fold for both backbones and the Δε error drops by more than any backbone change gives, and the backbone effect is at most 1.9× and 10%. On the scaffold re-split of BACE (152 test molecules), S6 leads the same encoder without pre-training by 10.8 points and the D-MPNN by 3.4.

AuthorsOpenReview ↗

Response (2/4)

Summary. We trained ChemProp's D-MPNN on our files. S6 scores higher on ESOL, BACE, Tox21 and HIV and on scaffold re-splits of HIV and BACE, and the D-MPNN is stronger on QM9. Three ablations of the S5 loss show that the coordinate term, the gap head and the teacher each help the quantum tasks.

W1 / Q1b. A graph baseline on our exact files

We trained ChemProp's D-MPNN (v2.3, default architecture, 50 epochs, three seeds) on our exact train/val/test files, selected on validation and scored test once.

task (our split)ChemProp D-MPNNSkeleMol S6 (Table 5)
ESOL RMSE ↓0.850 ± 0.0160.627
BACE AUC ↑87.6 ± 1.689.0
ClinTox AUC ↑86.6 ± 0.786.8
Tox21 AUC ↑73.4 ± 0.377.6
HIV AUC ↑81.4 ± 0.383.2
QM9-Δε MAE ↓ (Ha)0.00430 ± 0.000050.00697
QM9-U0 MAE ↓ (Ha)0.0206 ± 0.00050.499

S6 reads a single image and scores higher on ESOL, BACE, Tox21 and HIV, and it matches the D-MPNN on ClinTox. The D-MPNN reads the graph and is stronger on the two QM9 targets.

We also re-split HIV and BACE with DeepChem's scaffold splitter (BACE 1,210/151/152 and HIV 32,896/4,112/4,112, with no test scaffold in training) and re-ran the App. E checkpoint, with the D-MPNN on the same files.

scaffold split, test AUCS6released SigLIP2, no pre-trainingChemProp D-MPNN
BACE (n = 152)83.2 ± 2.272.4 ± 1.379.8 ± 0.5
HIV (n = 4,112)76.0 ± 0.877.4 ± 0.775.3 ± 2.0

On BACE under this strict scaffold split, molecular pre-training adds 11 points over the same encoder without it, and S6 is 3.4 points ahead of the D-MPNN. On HIV S6 is also ahead of the D-MPNN. In the revision the ChemProp result is a row of Tables 5 and 6, and the scaffold re-split is App. D.2.

W2 / Q2. What each S5 term contributes

We ran three ablations of the S5 stage. The first two repeat the S5 distillation from the same S3 checkpoint with one weight set to 0. In α = 0 this is the coordinate weight, and in β = 0 the gap-head weight. The third run trains the gap head alone, with no teacher. We fine-tuned each checkpoint with the Table 4 code under three seeds (42, 1337, 2026). Values are means, with ± sd for the new runs.

S5 (α = 10, β = 1)α = 0β = 0gap head only
ESOL RMSE ↓0.7180.713 ± 0.0190.715 ± 0.0120.731 ± 0.014
QM9-Δε MAE ↓ (Ha)0.007860.00819 ± 0.000130.00804 ± 0.000070.00837 ± 0.00005
PCQM4Mv2 MAE ↓ (eV)0.1280.174 ± 0.0230.274 ± 0.0170.196 ± 0.013
BBBP AUC ↑91.992.1 ± 0.692.4 ± 0.691.4 ± 0.4

The full S5 loss has the lowest error on both quantum tasks, and every seed of every ablation lies above the S5 mean on both. In these runs the coordinate term, the gap head and the teacher each contribute. Without the coordinate term the PCQM4Mv2 error rises 36% and the QM9-Δε error 4%. Without the gap head they rise 114% and 2%. With the gap head alone they are 53% and 6% above S5. ESOL and BBBP stay within 0.013 RMSE and 0.5 points of S5 in every run.

On PCQM4Mv2 the gap head contributes more than the teacher, and its training labels are PCQM4Mv2 gaps. On QM9-Δε, the same quantity on a different dataset, the teacher contributes more. There the teacher without the gap head reaches 0.00804 and the gap head without the teacher 0.00837.

AuthorsOpenReview ↗

Response (3/4)

Summary. The projection has no stop-gradient, and the student's space moves toward Uni-Mol's, with ρ rising from 0.24 to 0.65. We ran DINOv2-base on the remaining tasks, and S6 scores higher on LD50, HIV and BACE. Table 14 tests six drawing variants and a second renderer, and predictions are stable when layout and typography change.

W3 / Q3. Stop-gradient and the projection; ρ for DINOv2

φ trains jointly with the student and carries no stop-gradient, as the revised Sec. 3 states. The coordinate gradient on the student is 2α(hᵢ − qᵢ) whatever φ does (Eq. 3), so the student moves whenever the term is non-zero. App. G.2 measures ρ between the student's own space and Uni-Mol's raw space, with φ outside the measurement, so the rise from 0.24 to 0.65 reflects a change in the student's representation.

We ran the DINOv2 control on the same 2,000 molecules and 49,975 pairs (Table 23). A random SigLIP2 gives ρ = 0.01, DINOv2-base 0.10 (mean) and 0.15 (CLS), the released SigLIP2 0.24 and S5 0.65. Natural-image pre-training carries a little of Uni-Mol's structure, and the teacher stage supplies the rest.

W5 / Q5. DINOv2 on the label-rich tasks and the one-sentence contribution

Table 17 is the label-rich case. With 107k labels and the last two blocks fine-tuned, the drawing itself carries the information and a general-purpose encoder recovers it. On scaffold-split QM9 DINOv2-base is ahead on both targets, with Δε 0.01017 ± 0.00019 and U0 2.25 ± 0.32 against 0.01140 ± 0.00006 and 2.47 ± 0.28 for S6, and the revised text says so. Pre-training shows its effect at fixed backbone and budget (Tables 3–4) and at low label counts.

The +0.004 for BBBP in Table 18 comes from a validation comparison, and the BBBP validation partition is saturated for any model that encodes polarity (TPSA alone scores 0.938 and a random forest on ECFP4 0.978). Six encoders at 0.958–0.965 are at its ceiling, and the revision annotates the column. The low-label evidence is the BACE column (0.932 ± 0.007 vs 0.916 ± 0.025 for the best encoder without pre-training) and the frozen-feature probe (+0.092 on BACE).

Our contribution in one sentence is the observation that at a fixed budget every backbone ranks the pre-training strategies in the same order, and on the quantum targets the strategy matters more than the architecture. Molecular pre-training helps most when labels are scarce (Table 18's BACE column), and with plentiful labels a general-purpose encoder does as well or better (Table 17). "State of the art on five of ten tasks" is a second statement, based on the reported values in Table 5.

DINOv2-base on all ten tasks is now complete. We ran it on every remaining task under the Table 18 protocol (same split and head, 30 epochs, last two blocks unfrozen at epoch 6, three seeds), against S6 measured the same way.

validation, 3 seedsS6 SigLIP2DINOv2-base
LD50 MAE ↓0.534 ± 0.0060.564 ± 0.004
HIV AUC ↑0.785 ± 0.0070.731 ± 0.009
HIV, scaffold re-split ↑0.802 ± 0.0070.743 ± 0.011
BACE, scaffold re-split ↑0.786 ± 0.0090.762 ± 0.025
ESOL RMSE ↓0.961 ± 0.0460.961 ± 0.031
ClinTox AUC ↑0.707 ± 0.0350.769 ± 0.005

The pre-training gain extends beyond two tasks. S6 leads DINOv2-base on LD50, on HIV under both splits, and on BACE under both protocols (0.932 vs at most 0.916 in Table 18).

W6 / Q6. Depiction variants

The "kekulised" and "no stereo" settings reproduce the default render pixel for pixel under RDKit's defaults (prepareMolsBeforeDrawing=True, and addStereoAnnotation=False toggles only the R/S and E/Z text labels), so they coincide with the baseline row. The depiction table (Table 14) scores six drawing variants and PIKAchU with the same heads on the same 1,000 held-out molecules, as the change in MAE against the default render.

variantΔεU0variantΔεU0
thick bonds (width 4)+7.9%−2.6%explicit methyl−0.1%+0.2%
rotated 35°+2.2%−10.0%no element colours+39.7%+107.5%
larger type (1.4×)+4.4%+5.3%96-px render+15.0%+39.0%
PIKAchU (n = 992)+47.7%+103.4%

Predictions are stable under layout and typography, within 10% of the default MAE. The element colours and the training resolution carry signal the encoder uses. Without them its predictions contract toward the training mean (sd 0.0072 and 0.0088 vs 0.0179 Hartree on Δε). PIKAchU now has its numbers. It drew 992 of the 1,000 molecules and is scored against the default render of the same 992, and as a different drawing engine it lies outside the training depictions. We validate the encoder on drawings from its own rendering pipeline (App. D.3).

AuthorsOpenReview ↗

Response (4/4)

Summary. We ran the count-matched control. With the same number of molecules per epoch, the curriculum order is better than uniform sampling on all four tasks. The tier-unlock curve (Fig. 9) has the same shape on a fixed set of molecules from all tiers. Every SkeleMol value in the main tables is a three-seed mean.

W4 / Q4. Curriculum as a data-subset intervention; App. C.6; Fig. 10; 30/100 epochs

Data volume. The curriculum stage sees 5.74 M molecule-views over 10 epochs against 10 M for the uniform stage, so S6 improves on S5 with 43% less data. The subset removes data and improves the result. The count-matched control (uniform sampling matched to the curriculum's per-epoch counts) separates "ordering" from "fewer, simpler examples early". We ran it. Two runs start from the same checkpoint with identical settings. In each epoch the uniform run draws a random sample from all five tiers of exactly the size of the curriculum's active-tier set, so the two see the same number of molecules in every epoch and differ only in which molecules come when. We fine-tuned each checkpoint with the Table 4 code under three seeds (42, 1337, 2026). Values are means, with ± sd for the new run.

equal per-epoch countstier order (curriculum)uniform sampling
ESOL RMSE ↓0.6190.657 ± 0.018
QM9-Δε MAE ↓ (Ha)0.006740.00714 ± 0.00023
PCQM4Mv2 MAE ↓ (eV)0.1040.119 ± 0.009
BBBP AUC ↑93.392.9 ± 0.4

At equal data volume the curriculum is better on all four tasks, by 6% on ESOL, 6% on QM9-Δε, 13% on PCQM4Mv2 and 0.4 points on BBBP. On PCQM4Mv2 every seed of the uniform run is behind the curriculum mean. The data volume is matched in these runs, so what differs is the order in which the molecules arrive.

App. C.6. The mixed, sequential and anti-curriculum runs are stage-1 loss curves. Our evidence for the curriculum claim is the count-matched downstream control, with data volume held equal. The mixed and additive schedules reach the same terminal loss because both see every tier by the end, and this contrastive loss saturates near 4×10⁻⁴ for any schedule that converges.

The tier-unlock curve (Fig. 9). We re-scored the same ten checkpoints, with the trainer's own loss, on one fixed set of 4,096 held-out molecules from all tiers (453 T1, 637 T2, 2,872 T3, 134 T4). The printed curve uses the active tiers of each epoch, and the revision adds the fixed-set curve as a dashed line.

epoch0123456789
active tiers0–10–10–10–20–20–30–30–30–40–4
validation, active tiers (Fig. 9)0.1060.2400.9350.0810.0230.00130.00070.00050.00040.0004
validation, fixed all-tier set0.1700.4530.6000.1080.0370.00170.00100.00080.00070.0006
of which T3 (inactive until epoch 5)0.1720.4730.6050.1130.0360.00140.00080.00060.00050.0005
of which T4 (inactive until epoch 8)0.1580.4030.1580.1010.0170.00180.00190.00110.00200.0009

The fixed-set curve has the same shape as the printed one. The drop when T2 unlocks (epoch 3) appears on every tier, including the T3 and T4 molecules that enter training later, so the drop comes from the model itself. The fixed set ends at 6×10⁻⁴ against 4×10⁻⁴ for the active-tier set.

30/100 epochs. Our claims are made at the 10-epoch budget that every strategy in the matrix shares, which makes it a controlled comparison. Within that budget S6's loss plateaus by epoch 8 (Fig. 9).

W7. Seeds and dispersion

Every SkeleMol value in the main tables is a mean over seeds 42, 1337 and 2026. The tables print the mean for space, and the revised Sec. 5 and reproducibility statement say so, with Table 15 holding per-seed values for the controls that report a spread. The three-seed spread on the decomposition is small. The S6 row gives ESOL 0.627 ± 0.006, LD50 0.369 ± 0.006 and BBBP 93.7 ± 0.2, and the S3 and S5 rows carry comparable standard deviations (ESOL ≤ 0.014, LD50 ≤ 0.012, BBBP ≤ 0.3). ClinTox is unranked in every table, and every gain the paper claims is on another task.

Comments on the paper

2 comments

Dear reviewers,

The rebuttal period is ending soon. Please read the authors rebuttal and check if they addressed your concerns. Please acknowledge the rebuttal and adjust your scores accordingly.

Thanks.

-AC

Summary. The revision states every split, adds a graph baseline on our files and scaffold re-splits of HIV and BACE, and runs DINOv2 on all ten tasks. It also puts numbers on depiction sensitivity, revises Fig. 2, Fig. 4 and Fig. 10, and discusses the suggested related work.

We thank the three reviewers. A standard vision encoder that reads only a molecular drawing is competitive with strong molecule-specific graph, 3D and string models, with the image as its only input at inference. In our controlled study of ten backbones and seven pre-training strategies, every backbone ranks the strategies in the same order, and on the quantum tasks the strategy matters more than the backbone.

We have uploaded a revised PDF (OpenReview shows a pdfdiff). It makes the following changes.

  1. Splits stated (nf7a-W1). Table 10 lists the split, test size and metric of every task, and App. C.4 states the selection rule and the source of every Table 5 row. The 25-point BBBP spread across methods comes from the splitter. App. C.4 shows three models moving 23 to 27 points between the MoLFormer-style and DeepChem scaffold splitters. ESOL, BBBP, ClinTox and Tox21 use the DeepChem scaffold split, LD50 the TDC scaffold split, QM9 a random split with a scaffold split in App. D.1, PCQM4Mv2 the OGB-LSC split, HIV a stratified random split and BACE its source split.
  2. Graph baseline on our exact splits (nf7a-W1, V6Ym). S6 scores higher than ChemProp's D-MPNN (three seeds, validation-selected, test scored once) on ESOL, BACE, Tox21 and HIV and matches it on ClinTox, and the D-MPNN is stronger on the two QM9 targets. The D-MPNN is now a row of Tables 5 and 6.
  3. Scaffold re-split of HIV and BACE (nf7a-W1). S6 retains an 11-point margin over the same encoder without pre-training on BACE and leads the D-MPNN on both tasks.
  4. Representation-transfer controls (nf7a-W3). Random SigLIP2 ρ = 0.01, DINOv2-base 0.10/0.15, released SigLIP2 0.24, S5 0.65, on the same 2,000 molecules. Section 3 now states that the projection trains jointly with the student and carries no stop-gradient.
  5. Depiction sensitivity with numbers (nf7a-W6, TCBA). The depiction table (Table 14) scores six drawing variants and gives the numbers for a second renderer. Within the training renderer, predictions are stable under layout and typography. Fig. 2 moves to the appendix as Fig. 10 (V6Ym).
  6. The tier figure (Fig. 3) is drawn by a script from the tier maps (V6Ym), one molecule per tier and corpus, and every T0 panel is a hydrocarbon.
  7. Related work (V6Ym, TCBA). VideoMol, MaskMol, IEM (Sec. 2), CurrMG and SemiMol (Sec. 4), and an ImageMol row in Table 5 with its scaffold-split values.
  8. DINOv2-base on every task (nf7a-W5). The cross-architecture column is complete. S6 leads on LD50, on HIV under both splits and on BACE under both protocols. The numbers are in our reply to nf7a-W5.
  9. Wording (nf7a-W5, W7). Section 5 and the reproducibility statement state that every SkeleMol value is a three-seed mean, and the sentence on Table 17 matches the table.
  10. The tier-unlock curve (Fig. 9) on a fixed set (nf7a-Q4). We add a curve on one fixed set of 4,096 molecules spanning all tiers. It has the same shape as the printed one, and the drop at the T2 unlock also appears on the tiers that enter training later.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.