Related Experiment Videos
Speaking the Native Language of LLMs: A Discrete Architecture for Molecular Comprehension
IEEE Journal of Biomedical and Health Informatics
|August 10, 2026
Summary
MolLingua, a new framework, enables Large Language Models (LLMs) to understand molecular structures by converting 2D and 3D chemical data into discrete tokens. This approach bridges the gap between spatial information and LLM language, improving chemical reasoning and generative tasks.
Area of Science:
- Computational chemistry
- Artificial intelligence
- Biomedicine
Background:
- Large Language Models (LLMs) are powerful tools for scientific discovery but struggle with complex molecular structures.
- Current multimodal LLMs use continuous projections, creating a modality gap that limits performance.
- A native understanding of molecular structure is crucial for advancing LLM applications in chemistry and biology.
Purpose of the Study:
- To introduce MolLingua, a novel token-centric, dual-modal framework for native molecular understanding in LLMs.
- To address the modality gap by discretizing molecular features into learnable tokens.
- To enhance LLM capabilities in chemical reasoning and molecular generation.
Main Methods:
- Developed a dual-branch Residual Vector Quantization (RVQ) engine to discretize 2D and 3D molecular features.
- Integrated these discretized features as learnable tokens within the LLM vocabulary.
- Implemented a token-centric framework for seamless integration of spatial and linguistic information.
Main Results:
- MolLingua effectively aligns spatial chemical knowledge with LLMs through a fully discrete approach.
- Achieved state-of-the-art performance in complex chemical reasoning tasks.
- Demonstrated competitive results in molecular generation tasks.
Conclusions:
- MolLingua provides a unified framework for native molecular understanding in LLMs.
- The token-centric, dual-modal approach overcomes limitations of continuous projections.
- This framework has the potential to significantly advance LLM applications in biomedicine and chemistry.