Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Cross-Modality Controlled Molecule Generation with Diffusion Language Model

Created by
  • Haebom

Author

Yunzhe Zhang, Yifei Wang, Khanh Vinh Nguyen, Pengyu Hong

Outline

To overcome the limitations of the existing SMILES-based molecular generation diffusion model, which only supports single-modal constraints, this paper proposes Cross-Modality Controlled Molecule Generation with Diffusion Language Model (CMCM-DLM), which supports multi-modal constraints and the addition of new constraints. CMCM-DLM applies constraints of various modalities, such as molecular structures and chemical properties, stepwise by adding a Structure Control Module (SCM) and a Property Control Module (PCM) to a pre-trained diffusion model. The SCM establishes the molecular skeleton in the early stage, and the PCM fine-tunes the chemical properties of the generated molecules to target values in the later stage. Experimental results demonstrate the efficiency and adaptability of CMCM-DLM, suggesting a significant advancement in molecule generation in the field of new drug discovery.

Takeaways, Limitations

•
Takeaways:
◦
We overcome the limitations of existing models by proposing a molecular production model that supports multimodal constraints.
◦
We present an efficient method to leverage pre-trained models by adding new constraints without retraining.
◦
Proof of its potential for use in molecular generation in various fields, including new drug development.
◦
Implementation of effective control functions through separate application of structure and characteristic control modules.
•
Limitations:
◦
Further research is needed on the application of constraints to modes other than the two presented (molecular structure, chemical properties).
◦
Generalization performance evaluation is needed for molecules of various sizes and complexities.
◦
Further research is needed on the model's interpretability and explainability.
◦
Performance evaluation and scalability verification on large-scale datasets are required.
👍
Made with Slashpage