SciCore-Mol: A Plug-in Architecture for Infusing Molecular Cognition into Large Language Models

The SciCore-Mol project aims to address the lack of molecular cognition capabilities in large language models (LLMs) in fields like chemistry and biology. Through an innovative plug-in architecture, it infuses LLMs with professional capabilities to understand and reason about molecular structures, chemical properties, and biological activities. The core design concept is modular enhancement—no need to retrain specialized models, and it can seamlessly integrate with existing LLMs.

In fields such as drug development and materials science, traditional computational methods are accurate but costly and require professional knowledge; while LLMs are highly versatile, they lack deep molecular understanding—unable to intuitively 'see' 3D conformations, understand functional group interactions, or predict biological activities. Therefore, it is necessary to equip LLMs with specialized molecular cognition modules to enable them to think about molecules like chemists do.

Core Architecture: Develop a set of pluggable molecular cognition modules covering molecular representation (converting SMILES etc. into vectors), structural analysis (identifying skeletons/functional groups/stereochemistry), property prediction (solubility etc.), activity evaluation (target binding affinity), and synthesis planning (synthesizability and routes).

Technical Implementation: Adopt a multimodal fusion strategy to map information from different modalities (e.g., 1D SMILES, 2D structural formulas, 3D conformations) into an LLM-compatible space, enabling seamless interaction between text and molecular structures (e.g., asking about molecular properties in natural language, and the model calls modules to analyze and return results).

Application Scenarios:

• Drug discovery: Rapidly screen compound libraries, evaluate ADMET properties, and shorten R&D cycles;
• Materials science: Predict new material properties and guide experimental design;
• Education: Intuitively explain chemical concepts and lower learning barriers.

Plug-in Advantages: Preserve the general capabilities of LLMs, with optional function activation; easy to expand new modules; convenient to integrate with different base models.

Current Challenges: High cost of molecular data annotation, need to improve model interpretability, and limitations in modeling complex biological systems. Future Directions: Introduce more experimental data to enhance accuracy, develop fine-grained molecular dynamics simulation modules, and explore integration with experimental automation systems.

SciCore-Mol represents an important direction in the AI for Science field. The 'general foundation + professional plug-in' architecture is not only applicable to molecular science but also provides a reference for AI applications in other scientific fields. We look forward to AI accelerating human understanding and transformation of nature in scientific research.