KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge

TL;DR

This work tackles the limited molecule understanding in Mol-LLMs caused by imprecise textual descriptions and suboptimal representations. It introduces KnowMol-100K, a 100K multi-level, fine-grained molecule-description dataset, and chemically-informative representations including SELFIES-based 1D tokenization and hierarchical 2D graph encoding. Leveraging these, KnowMol achieves state-of-the-art performance on molecular understanding and generation tasks across multiple benchmarks, underscored by extensive ablations and qualitative analyses. The results suggest that tailored multi-level annotations and representation strategies are key to bridging molecular science with large-language-model capabilities, with potential impact on diverse chemoinformatics tasks and AI-assisted chemistry.

Abstract

The molecular large language models have garnered widespread attention due to their promising potential on molecular applications. However, current molecular large language models face significant limitations in understanding molecules due to inadequate textual descriptions and suboptimal molecular representation strategies during pretraining. To address these challenges, we introduce KnowMol-100K, a large-scale dataset with 100K fine-grained molecular annotations across multiple levels, bridging the gap between molecules and textual descriptions. Additionally, we propose chemically-informative molecular representation, effectively addressing limitations in existing molecular representation strategies. Building upon these innovations, we develop KnowMol, a state-of-the-art multi-modal molecular large language model. Extensive experiments demonstrate that KnowMol achieves superior performance across molecular understanding and generation tasks. GitHub: https://github.com/yzf-code/KnowMol Huggingface: https://hf.co/datasets/yzf1102/KnowMol-100K

Figures (10)

• Figure 1: (a) Demonstration of InstructMol (baseline) and KnowMol (ours) on four fundamental molecular understanding factors: (1) atoms, (2) functional groups, (3) structure, and (4) properties. Error/hallucination parts are marked in red, while correct parts in green. InstructMol is the state-of-the-art among open-sourced Mol-LLMs. (b) The comparison between the caption in our KnowMol-100K and the widely used caption in PubChem database.
• Figure 2: Our proposed KnowMol, a Mol-LLM with state-of-the-art performance on 7 molecule understanding and generation tasks.
• Figure 3: (a) The pipeline of building the KnowMol-100K. We use a combination of basic data from PubChem databases, an open-source toolkit for cheminformatics, RDKit, and the powerful multi-modal large language model GPT-4o. (b) Building upon the KnowMol-100K, we design two instruction-following pre-train tasks: (1) Multi-round Question Answering, and (2) Description Guided Molecule Generation.
• Figure 4: Chemically-informative model architecture. Three levels of representation are used for molecular features: atomic level, functional group level, and molecule level. We mark baseline model in blue, while improved parts in green.
• Figure 5: Prompt for generating structural construction and physicochemical property descriptions to construct KnowMol-100K.