acceptodds
Under review as a conference paper at ICLR 2027

When Does the Auxiliary Encoder Help in Multimodal Molecular Pretraining?

Abstract

Multimodal molecular pretraining often improves graph representations by aligning them with a second view of the same molecule, such as a SMILES string or 3D conformation. These gains are usually attributed to information learned by the auxiliary encoder. We show that this explanation is often incomplete: in several settings, a frozen, randomly initialized auxiliary network recovers much of the benefit of a trained one. We test this in controlled experiments that keep the graph model and training procedure fixed while replacing the trained auxiliary encoder with an untrained network of the same architecture. Markedly, on QM9 and Alchemy, random SMILES targets reduce mean raw-feature linear-probe error by 5.0% and 3.0% relative to graph-only training, respectively. The effect varies by settings: trained targets matter more for 3D supervision on PCQM4Mv2, for longer training on QMugs, and under GraphMVP fine-tuning. Additional controls show that random graph encoders and fixed random codes can also help, while simple representation expansion does not explain the gains. Together, these results show that cross-modal improvements can arise even when the second encoder has learned no molecular information. A frozen random target should therefore be included as a baseline before attributing such gains to transferred cross-modal content.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
molecular representation learning, graph neural networks, multimodal pretraining, cross-modal contrastive alignment, molecular property prediction, random features

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_V4tsOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper challenges a common view in multimodal molecular learning: specifically, it asks whether aligning multiple learned views of the same molecule (such as 3D and SMILES) truly improves representation learning and predictive performance. The authors show that in certain scenarios, cross-modal improvements can occur even when the auxiliary encoder has learned no molecular information, and that some of these gains persist even without a second molecular view (e.g., with random graph encoders or fixed random codes). Based on these findings, they recommend including a frozen random auxiliary encoder as a baseline before attributing performance gains to learned cross-modal information.

Critical strengths
  • The paper is well written and clear.
  • The main idea and message of the paper are very simple, but the observation is relevant and important for multimodal molecular representation learning, especially since the performance gains of multimodal molecular models are usually marginal.
  • The experimental evaluation is sound and comprehensive. The authors test random molecular encoders and fixed random codes, and ablate the second molecular view across various datasets and models.
  • The GraphMVP results, where learned targets outperform random ones, show that in certain scenarios using multiple modalities can be beneficial, further motivating the recommendation to include random targets as baselines when developing multimodal architectures.
Critical weaknesses
  • The connection to self-supervised learning is somewhat limited, as most experiments retain a supervised objective during training. It would be interesting to include more self-supervised methods in the analysis. At the very least, methods such as [1], [2], [3], [4] should be included in the related work discussion.
  • The paper offers limited discussion and guidance on when random targets are likely to be helpful.
  • Apart from QMugs, the datasets contain relatively small molecules. Evaluating on MARCEL [5], which includes larger molecules represented as conformer ensembles, would strengthen the analysis, especially for the 3D view.

I believe that the paper is interesting and offers an important discussion regarding multimodal molecular representation learning. I recommend acceptance.

Questions for authors
  • The regularization controls and geometry diagnostics are informative (Appendix E2), but the mechanism remains unresolved. What is the authors' hypothesis regarding the performance gains when using random features or random encoders as the second view? Could it be just a regularization effect? Do you believe that this is a property specific to molecular representation learning?
  • Can the authors expand the scope of the paper and include more self-supervised methods? This is especially interesting since with GraphMVP, the only fully self-supervised method studied, learned targets outperform random ones on all three datasets.
Reviewer_DpTfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the role of auxiliary encoders in multimodal molecular pretraining, particularly whether their benefits depend on learned representations. The authors show that randomly initialized and frozen auxiliary encoders can recover a substantial portion of the improvements obtained with trained encoders in certain settings. They further demonstrate that even fixed random targets without auxiliary molecular information can provide performance gains. Based on these observations, the paper suggests including random-target baselines when evaluating multimodal molecular pretraining methods.

Critical strengths

The paper investigates a potentially overlooked issue in the evaluation of multimodal molecular pretraining methods. The empirical findings provide useful insights into the interpretation of performance improvements attributed to auxiliary encoders.

Critical weaknesses

The paper is framed around when an auxiliary encoder is beneficial, whereas its experiments primarily investigate the sources of improvement associated with auxiliary alignment, particularly the role of auxiliary encoder training. The framing could better reflect this distinction.

Questions for authors

The main experiments use supervised pretraining with molecular property labels in addition to the alignment loss. Could the authors clarify the motivation for this design and discuss whether similar conclusions would hold in a purely self-supervised setting without property supervision?

Reviewer_d9qdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether auxiliary encoder can improve model performance without being training in multimodal molecular pretraining. The paper proposes random-target control, which replace a trained auxiliary encoder with a random frozen encoder in same structure. Experiments are conducted in several datasets with SMILES and 3D two molecule modal. The results show that random encoder can achieve close improve performance.

Critical strengths

S1. This paper studies an interesting problem: random encoder can sometimes improve the performance.

S2. The experimental design is clear and easy to understand.

S3. The experiments are extensive. The authors evaluate several datasets under different training settings and conduct many control experiments.

Critical weaknesses

W1. Whether the trained auxiliary encoder has learned enough useful information remains unclear. The authors use a fixed number of training steps without showing that the encoder has convergence. Futhermore, auxiliary encoder is learned on the same five attribution, which might mean that auxiliary encoder doesn't learn enough new informantion.

W2. The theoretical analysis doesn't fully explain the experiment results. Prop. 1 assumes the student model can preserve its original predictive ability and learn the useful informantion from random target. However, these important assumptions are not verified in training process. The paper doesn't explictly explain why random encoders are effective on some tasks, but perform much worse than trained encoders on others.

W3. The main results are sensitive to the evaluation settings. As shown in Table 11, the recovered fraction changes substantially simply by normalizing the graph representations. This makes me question how much the results depend on the specific evaluation method.

Questions for authors

See W1~3

Comments on the paper

1 comment

Dear Reviewers,

The author-reviewer discussion phase is ending soon. If you have not done so yet, please read the author rebuttals and respond.

Please recall that, as a reviewer, you have agreed to engage in discussion with the authors to clarify and improve the paper. Failure to do so may result in penalties on your own submissions, as per the official ICLR policy. Given the high volume of submissions, your participation is crucial.

Thank you for your time and help.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.