[论文解读] InstructMol: Multi-Modal Integration for Building a Versatile and Reliable Molecular Assistant in Drug Discovery

He Cao, Zijing Liu+3 more|arXiv (Cornell University)|Nov 27, 2023

Machine Learning in Materials Science被引用 14

一句话总结

tldr: InstructMol 是一种多模态大语言模型，通过两阶段指令微调管线将分子图和序列与自然语言对齐，从而提升分子任务性能并缩小与专门模型之间的差距。

ABSTRACT

The rapid evolution of artificial intelligence in drug discovery encounters challenges with generalization and extensive training, yet Large Language Models (LLMs) offer promise in reshaping interactions with complex molecular data. Our novel contribution, InstructMol, a multi-modal LLM, effectively aligns molecular structures with natural language via an instruction-tuning approach, utilizing a two-stage training strategy that adeptly combines limited domain-specific data with molecular and textual information. InstructMol showcases substantial performance improvements in drug discovery-related molecular tasks, surpassing leading LLMs and significantly reducing the gap with specialized models, thereby establishing a robust foundation for a versatile and dependable drug discovery assistant.

研究动机与目标

• Bridge molecular representations (graphs and sequences) with natural language to enable a versatile drug-discovery assistant.
• Leverage limited domain data via a two-stage training strategy to align modalities and adapt to downstream tasks.
• Demonstrate improvements over generalist LLMs and reduce gaps to specialized models across molecular tasks.

提出的方法

• Encode molecular structure as a graph and sequence, align via a lightweight projection to text space.
• Two-stage training: (i) alignment pretraining with molecule-text pairs to project graph representations into the textual space, (ii) task-specific instruction tuning using LoRA on top of a frozen graph encoder and LLM.
• Utilize Vicuna-7B as the base LLM and a GIN-based molecule graph encoder pre-aligned with text (MoleculeSTM).
• Train with molecule-description data and domain-specific task instructions to adapt the LLM to property prediction, description generation, and reaction analysis.

Figure 2 : Overview of InstructMol model architecture design and two-stage training paradigm. The example molecule in the figure is Terephthalaldehyde [ 62 ] (CID 12173).

实验结果

研究问题

• RQ1 Can a multimodal instruction-tuned LLM effectively align molecular graphs and sequences with natural language for drug-discovery tasks?
• RQ2 Does a two-stage pipeline (alignment pretraining plus task-specific instruction tuning) improve performance over generalist LLMs and single-modality baselines?
• RQ3 How does InstructMol perform on compound property prediction, molecule description generation, and chemical reaction tasks relative to expert models and other LLM-based baselines?

主要发现

• InstructMol consistently improves LLM performance on molecular understanding and design tasks compared to state-of-the-art generalist LLMs.
• The two-stage training with a lightweight alignment projector yields better modality alignment than more complex cross-attention methods under data scarcity.
• Molecule description generation and property prediction show gains over single-modality LLM baselines, though expert models still hold an edge on some metrics.
• Ablation suggests graph-text pre-alignment and the use of LoRA for LLM adaptation are important for convergence and task performance.

Figure 3 : Comparison of biomolecule-domain molecule-text dataset scale with existing general domain vision-language datasets [ 4 , 81 , 13 , 87 , 71 ] .