[Paper Review] DrugAssist: A Large Language Model for Molecule Optimization

Geyan Ye, Xibao Cai+6 more|arXiv (Cornell University)|Dec 28, 2023

Machine Learning in Materials Science4 citations

TL;DR

DrugAssist is an interactive large language model fine-tuned on Llama2-7B-Chat for molecule optimization through human-machine dialogue, enabling iterative, expert-guided refinement. It achieves state-of-the-art performance in single- and multi-property optimization, including range-constrained objectives, and demonstrates strong zero-shot and few-shot transferability.

ABSTRACT

Recently, the impressive performance of large language models (LLMs) on a wide range of tasks has attracted an increasing number of attempts to apply LLMs in drug discovery. However, molecule optimization, a critical task in the drug discovery pipeline, is currently an area that has seen little involvement from LLMs. Most of existing approaches focus solely on capturing the underlying patterns in chemical structures provided by the data, without taking advantage of expert feedback. These non-interactive approaches overlook the fact that the drug discovery process is actually one that requires the integration of expert experience and iterative refinement. To address this gap, we propose DrugAssist, an interactive molecule optimization model which performs optimization through human-machine dialogue by leveraging LLM's strong interactivity and generalizability. DrugAssist has achieved leading results in both single and multiple property optimization, simultaneously showcasing immense potential in transferability and iterative optimization. In addition, we publicly release a large instruction-based dataset called MolOpt-Instructions for fine-tuning language models on molecule optimization tasks. We have made our code and data publicly available at https://github.com/blazerye/DrugAssist, which we hope to pave the way for future research in LLMs' application for drug discovery.

Motivation & Objective

• Address the lack of interactive, expert-feedback-driven molecule optimization in existing LLM-based approaches.
• Enable real-time, multi-turn dialogue between domain experts and an LLM to iteratively refine molecular properties.
• Develop a model capable of optimizing molecules for multiple properties simultaneously, including within specified value ranges.
• Improve transferability by enabling zero-shot and few-shot optimization of previously unseen properties.
• Facilitate future research by releasing a large instruction-based dataset, MolOpt-Instructions, for fine-tuning LLMs in drug discovery.

Proposed method

• Fine-tune Llama2-7B-Chat on a large instruction-based dataset, MolOpt-Instructions, for molecule optimization tasks.
• Represent molecules using SMILES strings and frame optimization as a conditional generation task guided by natural language instructions.
• Support multi-turn dialogues where users provide feedback and new optimization goals, enabling iterative refinement.
• Incorporate expert feedback by allowing users to correct model outputs and guide subsequent generations toward desired molecular properties.
• Enable zero-shot transfer by generalizing to new property combinations not seen during training.
• Support few-shot transfer by leveraging a few examples of successful optimizations to guide the model on new, unseen properties.

Experimental results

Research questions

• RQ1 Can an LLM-based model effectively support interactive, iterative molecule optimization through natural language dialogue with domain experts?
• RQ2 How well does the model generalize to multi-property optimization tasks, especially when optimizing within specific value ranges?
• RQ3 To what extent can the model transfer knowledge to optimize properties not included in its training data?
• RQ4 How effective is the model in correcting its own errors through human-provided feedback in a multi-turn interaction setting?
• RQ5 Can a fine-tuned LLM achieve state-of-the-art performance in molecule optimization while maintaining strong transferability and interactivity?

Key findings

• DrugAssist achieves state-of-the-art performance in both single- and multi-property molecule optimization, including challenging range-constrained objectives.
• The model demonstrates strong zero-shot transferability, successfully optimizing previously unseen property combinations without retraining.
• In few-shot settings, DrugAssist can optimize new properties (e.g., logP) not present in training data by leveraging a few demonstration examples.
• The model exhibits robust iterative optimization capability: when initial outputs fail to meet criteria, it corrects its approach based on human-provided examples.
• DrugAssist maintains high validity and success rates across diverse optimization tasks, with results consistently outperforming prior LLM-based and traditional methods.
• The publicly released MolOpt-Instructions dataset contains sufficient data diversity to support fine-tuning of LLMs for molecule optimization with both similarity and property difference constraints.