# MolSight: A Graph-Aware Vision-Language Model for Unified Chemical Image Understanding

**Authors:** Wenda Wang, Yihan Tong, Yuwei Hu, Zhewei Wei  
**arXiv:** [2607.01982](https://arxiv.org/abs/2607.01982)  
**Date:** 2026-07-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2607.01982v1/

## Key points

- Molecular large language models (LLMs) are widely used in molecular design and drug discovery for understanding molecular structures and functions.
- Existing LLMs struggle to fully capture the visual representation of molecular structures, limiting their effectiveness.
- Molecular vision-language models (VLMs) show promise but face challenges in structural alignment and topological modeling for accurate molecular understanding.
- MolSight is a new graph-aware vision-language model designed to enhance the understanding of molecular images by VLMs through a Molecular Topology Module and a Molecular Grounding Module.
- MolSight significantly outperforms existing VLMs, molecular LLMs, and specialized tools in various chemical visual understanding tasks, achieving higher levels of molecular image reasoning.

## Summary

, , , , 

In the field of molecular design and drug discovery, the use of molecular large language models (LLMs) has become a prominent trend for understanding molecular structures and functions. However, existing LLMs struggle to fully capture the visual representation of molecular structures, limiting their effectiveness. While molecular vision-language models (VLMs) show promise, they still face challenges in structural alignment and lack topological modeling necessary for accurate molecular understanding. To address these limitations, a new framework called MolSight has been proposed. MolSight is a graph-aware vision-language model designed to enhance the understanding of molecular images by VLMs. It integrates a Molecular Topology Module that injects chemical-bond adjacency information into vision tokens, as well as a Molecular Grounding Module that aligns visual features with chemical symbolic semantics. Through experiments, it has been demonstrated that MolSight significantly outperforms existing VLMs, molecular LLMs, and specialized tools across various chemical visual understanding tasks, achieving a higher level of molecular image reasoning. Accurately identifying molecular structures and inferring their physicochemical properties is crucial for advancements in molecular design and drug discovery. This process involves combining various modalities such as molecular structure images, SMILES strings, and natural-language descriptions to identify key structural features and reason about properties and functions. Large-scale textual data have enabled LLMs to learn general chemical knowledge and apply it to tasks like molecular generation and property prediction. Molecular LLMs require the ability to process complex chemical languages containing structural information like canonical SMILES representations. MolSight's innovative approach addresses this need by incorporating graph-aware techniques that improve the alignment between visual features and chemical semantics in order to enhance overall understanding of molecular images. This advancement represents a significant step forward in the development of AI-driven approaches to chemistry research and applications in drug discovery.