KnowMol

KnowMol is a multimodal molecular large language model (LLM) built to reason about small molecules and their properties using chemically grounded knowledge. Molecular LLMs promise to let researchers query, describe, and design molecules in natural language, but their quality is limited by the training data: most molecule-text datasets provide only coarse, high-level captions that miss the functional groups, substructures, and reaction-relevant details a chemist relies on. As a result, models can produce fluent but chemically shallow answers.

To close this gap, the authors introduce KnowMol-100K, a dataset of 100,000 fine-grained molecular annotations that describe molecules across multiple levels of chemical knowledge — from atoms and functional groups up to whole-molecule properties. Building on this resource, KnowMol pairs a large language model with a molecule graph encoder to align chemically informative molecular representations with text, improving both molecular understanding and generation.

Developed by researchers at the Institute of Computing Technology, Chinese Academy of Sciences, and released as a preprint in October 2025, KnowMol joins a line of molecular LLMs such as MolT5, Mol-Instructions, and MoleculeSTM, distinguished by its emphasis on multi-level chemical annotation as the foundation for grounding.

#Key Features

• Multi-level chemical annotations: The KnowMol-100K dataset captures molecular knowledge from functional groups and substructures to whole-molecule properties, providing richer supervision than coarse caption datasets.
• Multimodal architecture: A molecule graph encoder is combined with a large language model so that structural and textual information are reasoned over jointly.
• Understanding and generation: The same model handles molecule-to-text understanding tasks and text-to-molecule generation within a unified framework.
• Open dataset and code: Both the KnowMol-100K dataset and model code are released publicly, supporting reproducibility and further research.

#Technical Details

KnowMol uses Vicuna-7B as its base language model and incorporates a molecule graph encoder (MoleculeSTM or GraphMVP) with projection layers that map graph embeddings into the language model's token space; LoRA is used for parameter-efficient fine-tuning. Training proceeds in two stages: a pretraining phase using multi-round question answering and molecule-generation tasks over KnowMol-100K, followed by task-specific instruction tuning. The code is released under Apache 2.0 and the dataset under the MIT license, with usage additionally bound by the licenses of LLaMA, Vicuna, LLaVA, Mol-Instructions, and GPT-4. The authors report improved performance over prior molecular LLMs across molecular understanding and generation benchmarks.

#Applications

KnowMol serves cheminformatics and drug-discovery researchers who want to interact with molecules through natural language: describing a molecule's properties and functional groups, answering questions about structure, and generating candidate molecules from textual specifications. Its fine-grained grounding makes it particularly suited to tasks where substructure-level reasoning matters, such as property prediction and molecule captioning.

#Impact

KnowMol demonstrates that the quality of molecular language models is tightly coupled to the granularity of their training annotations, and that multi-level chemical knowledge can meaningfully improve grounding. By releasing KnowMol-100K alongside the model, the work provides a reusable resource for the broader molecular LLM community. As a recent preprint awaiting peer review, its robustness across chemical space and its advantages over the strongest contemporary molecular LLMs remain to be independently confirmed.