PaperlayerView exact revision on arXiv

Source-linked AI summary

Retrieval-guided Twin Fusion with Similarity-aware Contrast for Molecule-Text Alignment

Shunshun Gu, Shengqi Qiu, Hang Zhou, Xiao Luo

arXiv:2608.16005v1cs.LG

TL;DR

Molecule–text alignment methods often overlook fine-grained substructure–text semantics. RISEN uses retrieval-guided twin prototype fusion and similarity-aware soft contrastive targets, achieving consistent overall gains over strong baselines on retrieval and molecular property prediction benchmarks.

  • Problem

    Existing molecule–text alignment methods operate on global pairs and overlook fine-grained semantic relationships between substructures and texts.

  • Method

    RISEN retrieves related texts and molecules, fuses attention-pooled twin prototypes with substructure embeddings, and uses similarity-based soft contrastive targets.

  • Results

    RISEN shows consistent overall gains over strong baselines on zero-shot retrieval and molecular property prediction benchmarks.

  • Takeaways & Limitations

    The results support semantic retrieval and soft contrastive supervision as effective components for cross-modal molecular representation learning.

  • Takeaways & Limitations

    RISEN’s soft-label mechanism relies on frozen MoLFormer similarities, which may not capture all task-relevant similarities.

Abstract

from arXiv · show

This paper studies the problem of molecule-text alignment, which aims to project molecules and their textual descriptions into a joint latent space for downstream tasks including molecule search and molecular property prediction. Previous approaches typically combine graph structure mining with contrastive learning to enhance joint representation learning. However, they typically neglect fine-grained semantic relationships between substructures and texts, leading to suboptimal performance on downstream tasks. Towards this end, we propose a novel approach named Retrieval-guided Twin Fusion with Similarity-aware Contrast (RISEN) for molecule-text alignment. The core idea of RISEN is to construct a latent twin molecule for each substructure with cross-modal retrieval for semantic enhancement. In particular, for each substructure query, we retrieve relevant textual descriptions and sample several molecules that share similar descriptions of substructures. Then, we aggregate their representations via attention pooling for a twin latent representation, which would be further fused with the original substructure for representation enrichment. In addition, we measure the similarity across substructures and texts, which would further guide cross-modal contrastive learning with soft thresholding. Extensive experiments on benchmark datasets validate the superiority of the proposed RISEN in comparison with existing baselines.

1 Introduction

Molecule–text alignment jointly embeds molecular structures and textual descriptions for cross-modal reasoning, but fine-grained matching can introduce noisy alignments and hard contrastive learning can create false negatives. RISEN addresses these limitations through retrieval-guided twin fusion and similarity-aware soft targets, achieving consistent gains on retrieval and property-prediction benchmarks.

  • Background: Molecule–text alignment embeds molecular structures and textual descriptions into a shared semantic space for cross-modal reasoning.
  • Limitations: Fine-grained substructure–phrase supervision can introduce weak alignments, while one-hot contrastive targets treat chemically similar in-batch molecules as false negatives.These issues can dilute useful signals, degrade generalization, and distort semantic neighborhoods.
  • RISEN: Retrieval-guided Twin Fusion retrieves semantically related mirror molecules, attention-pools them into a twin prototype, and fuses it with each original substructure embedding.This enables local motifs to inherit broader structural context from complete molecules.
  • RISEN: Similarity-aware Contrast replaces one-hot targets with soft targets derived from MoLFormer [18] embedding similarities to preserve semantic neighborhoods.
  • Results: RISEN achieves consistent overall gains over strong baselines on molecule–text retrieval and molecular property prediction benchmarks.The experiments evaluate both downstream tasks.

2 Related Work

Related work spans global and fine-grained molecule–text alignment, soft-target contrastive learning, and retrieval augmentation. RISEN combines these directions by enriching substructures with retrieved molecular context and similarity-aware contrastive targets without explicit fragment annotations or inference changes.

  • Molecule–Text Multimodal Modeling: Global methods align complete molecules with descriptions, whereas fine-grained methods model correspondences between molecular fragments and chemical phrases.Global approaches include contrastive pre-training, cross-modal projection, and instruction-tuned multi-task learning; fine-grained approaches use modality blending, optimal transport, and hierarchical alignment.
  • Molecule–Text Multimodal Modeling: RISEN enriches substructure representations with semantically related complete molecules, providing broader chemical context without explicit fragment-level annotations.Retrieved mirror molecules are aggregated into an attention-pooled prototype and fused with the original substructure embedding.
  • Contrastive Learning with Soft Targets: Standard contrastive objectives can create false negatives by treating semantically similar in-batch samples as non-matching negatives.Prior work addressed noisy correspondences with momentum-distilled soft labels, embedding-similarity targets, and fragment contrast for structurally similar molecules.
  • Contrastive Learning with Soft Targets: RISEN’s Similarity-aware Contrast constructs soft targets from pairwise MoLFormer embedding similarities to reduce penalties on chemically similar in-batch samples.This extends prior efforts to preserve semantic neighborhoods and mitigate harmful false negatives in molecular contrastive learning.
  • Retrieval-Augmented Learning: Unlike prior molecular retrieval augmentation used mainly for input-level generation, RISEN applies retrieval at representation level during training and leaves inference unchanged.The method aggregates retrieved mirror molecules into an attention-pooled prototype fused with the original substructure embedding.

3 Methodology

RISEN enriches substructure representations by retrieving semantically related molecules and fusing their aggregated embeddings, while replacing hard contrastive targets with similarity-aware soft supervision. Its objectives jointly align molecule-text representations, enforce prototype consistency, and filter unreliable correspondences.

  • Retrieval-guided twin fusion: RISEN retrieves semantically related complete molecules for each substructure and aggregates their embeddings into an attention-pooled twin prototype.The mirror map is precomputed from aligned textual descriptions, and up to five mirror molecules are sampled uniformly during training.
  • Retrieval-guided twin fusion: The prototype is fused with the original substructure representation to provide broader chemical context.The fusion uses embedding dimension d and a fusion coefficient α.
  • Similarity-aware contrast: RISEN constructs soft contrastive targets from pairwise MoLFormer similarities, reducing penalties on chemically similar in-batch molecules.A threshold τ filters weak similarities, and soft labels are mixed with hard labels using weight β.
  • Optimization objectives: The training objective combines bidirectional contrastive learning on soft-hard mixed targets with prototype contrastive loss for fused-prototype alignment consistency.The total objective also includes a classification loss that follows MolBridge’s self-refinement protocol to filter unreliable alignment pairs.
  • Optimization objectives: Together, these objectives target representations that are semantically aligned, structurally enriched, and robust to noisy correspondences.The contrastive loss averages both retrieval directions and is optimized on the fused substructure representation and mirror-molecule prototype.

4 Experiments

RISEN is evaluated on zero-shot molecule–text retrieval and molecular property prediction, where it achieves strong performance across benchmark datasets. Ablations, sensitivity tests, and qualitative analysis attribute these gains to its retrieval-enriched representations and component design.

  • Experimental Setup: Experiments cover zero-shot retrieval on PCDes and PubChem324kV2, plus ROC-AUC property prediction on eight MoleculeNet datasets.The training data comprise 431,877 molecule–description pairs augmented to 2M through substructure-level alignment.
  • Zero-shot Molecule-Text Retrieval: RISEN achieves the best performance on most metrics across both retrieval benchmarks and directions, with particularly strong gains in M2T.The results support retrieval-enriched substructure representations over isolated fragment learning.
  • Molecular Property Prediction: RISEN achieves the best average ROC-AUC across eight scaffold-split MoleculeNet classification datasets, with gains on most tasks.This indicates effective transfer to downstream property-relevant chemical semantics.
  • Ablation Study: Removing soft labels hurts T2M R@1 and MRR, while attention pooling improves both retrieval directions.These findings support soft labels for mitigating false negatives and attention pooling for strengthening retrieval.
  • Ablation Study: Removing mirror loss or prototype fusion yields small T2M gains but degrades M2T, suggesting these components stabilize M2T alignment.The ablation is conducted on PubChem324kV2.
  • Sensitivity Analysis: α=0.1 provides the best balance across directions and β=0.3 yields the strongest results, whereas the sharp drop at β=0.8 shows soft labels should supplement the original objective.Sensitivity analysis varies prototype fusion weight α and soft-label mixing weight β on PCDes.

5 Conclusion

RISEN is a retrieval-guided framework for molecule–text alignment that enriches substructure representations through twin prototype fusion and reduces false negatives with similarity-aware soft targets. Experiments on zero-shot retrieval and molecular property prediction benchmarks show consistent overall gains over strong baselines.

  • Method: RISEN enriches substructure representations through retrieval-guided twin prototype fusion.The framework uses semantic retrieval to support cross-modal molecular representation learning.
  • Method: Similarity-aware soft targets alleviate false negatives during cross-modal contrastive learning.The approach uses soft contrastive supervision to account for similarity across molecular and textual representations.
  • Results: RISEN shows consistent overall gains over strong baselines on zero-shot retrieval and molecular property prediction benchmarks.These results highlight the effectiveness of semantic retrieval and soft contrastive supervision in cross-modal molecular learning.

6 Limitations

RISEN has bounded but nontrivial memory and scalability costs from encoding mirror molecules, while its retrieval and soft-label mechanisms may face challenges at larger scales and under domain shift. Its evaluation is limited to retrieval and property prediction, leaving generative applications for future study.

  • Computational and scaling costs: Encoding up to k=5 mirror molecules per substructure sample moderately increases memory usage, with larger batches or longer retrieval neighborhoods potentially challenging to scale.The additional mirror molecules enable structurally enriched representations that are empirically beneficial.
  • Soft-label limitations: Soft labels rely on similarities from frozen MoLFormer embeddings, which may miss task-relevant chemical semantics and reduce quality for underrepresented scaffolds or domain-shifted distributions.The limitation concerns the semantic fidelity of similarity-based supervision.
  • Evaluation scope: Evaluation focuses on molecule–text retrieval and molecular property prediction, so RISEN’s performance on other task types remains unassessed.The paper identifies generative applications as an important extension for evaluating broader utility.
  • Future evaluation: Extending RISEN to molecule captioning and text-guided molecule generation would provide a more complete picture of its general utility.These generative tasks are proposed as future evaluation settings.
Loading 2608.16005v1…