Source-linked AI summary
DrugAssist: A Large Language Model for Molecule Optimization
Geyan Ye, Xibao Cai, Houtim Lai, Xing Wang, Junhong Huang, Longyue Wang, Wei Liu, Xiangxiang Zeng
TL;DR
Molecule optimization has seen little LLM involvement, and existing approaches do not incorporate expert feedback through interactive refinement. DrugAssist uses human-machine dialogue for molecule optimization and reports leading results in single- and multi-property tasks, including specified-range objectives.
Problem
Molecule optimization has seen little LLM involvement, while existing methods generally learn chemical-structure patterns without incorporating expert feedback through interaction.
Source: §1 · p. 2, §2.1 · p. 3Method
DrugAssist is an interactive molecule-optimization model that uses human-machine dialogue and multi-turn expert feedback.
Source: §1 · p. 2Results
DrugAssist reports leading results in single- and multi-property optimization, including specified-range objectives, and outperforms other LLMs in success and validity rates across tasks.
Source: §1 · p. 2, §4.3 · p. 9Takeaways & Limitations
The authors present DrugAssist as showing potential for transferability and iterative optimization during interaction, and release MolOpt-Instructions for future fine-tuning work.
Source: §5 · p. 12The model has a relatively lower success rate on tasks specifying a range of post-optimization values, including the “esol+ strict” task.
Source: §4.3 · p. 10
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
LLMs have drawn growing interest in drug discovery, but molecule optimization has seen little LLM involvement. DrugAssist addresses this gap with interactive dialogue-based optimization and reports leading results across several optimization settings.
1 INTRODUCTION
- Existing molecule-optimization methods learn from chemical structures but lack interaction for incorporating expert experience and feedback during iterative refinement.The drug-discovery pipeline involves iterative refinement through conversations with domain experts.Source: §1 · p. 2
- DrugAssist achieves leading results in multi-property optimization and performs well when optimized properties must remain within specified ranges.The authors also report leading single-property results and high validity and similarity to the original molecules.Source: §1 · p. 2
- The paper introduces DrugAssist, an interactive molecule-optimization model, and releases MolOpt-Instructions for fine-tuning language models on these tasks.DrugAssist uses multi-turn dialogue so experts can guide further optimization of initially generated molecules.Source: §1 · p. 2
2 RELATED WORK
Prior molecule-optimization work uses sequence or graph representations, while biomedical LLM research has focused more on question answering and related tasks. The section identifies gaps in multi-property and range-based optimization, task transfer, and human interaction.
2.1 TRADITIONAL APPROACHES IN MOLECULE OPTIMIZATION
- Traditional molecule-optimization methods use sequence representations, often treating SMILES conversion as translation, or graph representations that generate molecule graphs.The sequence-based translation framing is described by He et al. (2021).Source: §2.1 · pp. 2–3
- Existing methods rarely optimize multiple properties or target a specified property range, and task changes can require retraining because of catastrophic forgetting.Retraining can add costs, while some properties may lack sufficient experimental training data.Source: §2.1 · p. 3
- The section reports no prior focus on interactivity, although domain experts can provide natural-language feedback and existing methods struggle to use that experience.Source: §2.1 · p. 3
2.2 LLMS IN BIOMEDICAL DOMAIN
- Biomedical LLM work has centered largely on question answering, with ChatMol addressing molecule understanding and generation and ChatDrug supporting drug editing.Source: §2.2 · p. 3
3 METHODS
The methods construct MolOpt-Instructions from similar molecule pairs with property differences, then instruction-tune Llama2-7B-Chat on molecule-optimization and general-knowledge data. The dataset includes direction-, threshold-, and range-based goals, while its molecular and biochemical property distributions are presented as evidence of diversity.
3.1 CONSTRUCTION OF MOLOPT-INSTRUCTIONS DATASET
- MolOpt-Instructions contains over one million molecule pairs covering six properties: solubility, BBBP, hERG inhibition, QED, and hydrogen-bond donor and acceptor counts.The dataset statistics are described as preserving similarity constraints alongside substantial property differences.Source: §3.1 · p. 4, Table 1 · p. 4
- Pairs were selected for similarity greater than 0.65 and logP difference greater than 2.5, then scored with iDrug and converted into manually refined ChatGPT-suggested instructions.The workflow began with one million ZINC molecules and used mmpdb to generate similar pairs; some pairs were reversed to balance increased and decreased target properties.Source: §3.1 · p. 4, Figure 2 · p. 5
- The dataset defines three optimization goals: change a property, meet a threshold for that change, or place the optimized value within a specified range.The authors describe range requirements as making tasks more closely aligned with real-world scenarios.Source: §3.1 · p. 4
- Structural-property distributions cover Bertz complexity, molecular weight, atom count, and ring count, while ADMET-related distributions illustrate biochemical-property diversity.The six named properties are solubility, BBBP, hERG inhibition, QED, and hydrogen-bond donor and acceptor counts.Source: §3.1 · p. 5, Figure 3 · p. 6, Figure 4 · p. 6
3.2 INSTRUCTION TUNING
- Instruction tuning fine-tunes Llama2-7B-Chat on MolOpt-Instructions, mixing general-knowledge and molecule-optimization data to address catastrophic forgetting.The less abundant data category is replicated to reach the chosen mixture ratio.Source: §3.2 · p. 6, Figure 5 · p. 7
4 EXPERIMENTS
The experiments compare DrugAssist with traditional optimization models and other LLMs, then examine transfer and iterative optimization. DrugAssist reports the strongest success rates in the comparisons, while case studies illustrate zero-shot, few-shot, and feedback-guided optimization.
4.1 EXPERIMENTAL SETUP
- DrugAssist was fine-tuned from Llama-2-7B-Chat, and the experiments compared it with sequence-based models, other LLMs, and traditional optimization approaches.Training used MolOpt-Instructions and Stanford Alpaca data, with the Stanford Alpaca data replicated five times before mixing and splitting.Source: §4.1 · p. 7
4.2 EVALUATION METHODS
- The traditional-model comparison evaluated Solubility and BBBP success, validity, and molecular similarity using single-turn prompts.Solubility success required the generated value to fall within the requested interval; BBBP success required the target category.Source: §4.2 · p. 8
- The LLM comparison used 16 tasks and multi-turn prompting; success was assessed with loose and strict criteria, including specified ranges for Solubility.The strict Solubility ranges were [x+0.5, x+1.5] for increases and [x-1.5, x-0.5] for decreases.Source: §4.2 · pp. 8–9
4.3 MAIN RESULTS
- DrugAssist achieves the highest success rates against traditional models and significantly outperforms other LLMs in success and validity across all tasks.The comparisons cover single- and multi-property optimization against traditional models and 16 tasks against LLMs.Source: Table 4 · p. 9, §4.3 · p. 9, Table 5 · p. 10
- DrugAssist had relatively lower success on strict Solubility range tasks, while still achieving competitive multi-property results after training on individual properties.The paper identifies the “esol+ strict” task as a more challenging objective for future exploration.Source: §4.3 · p. 10
4.4 CASE STUDY
- In zero-shot tests, DrugAssist simultaneously increased BBBP and QED by at least 0.1 while generating a molecule structurally similar to the original.The properties were trained individually, but users could combine them at inference; the paper contrasts this with traditional models that often require retraining for multiple properties.Source: §4.4 · p. 11, Figure 7 · p. 11
- With few similar examples in the prompt, DrugAssist increased logP by at least 0.1 despite logP being absent from its training data.The case study demonstrates few-shot optimization of an unseen property.Source: §4.4 · p. 11, Figure 8 · p. 11
- After an initial QED increase failed, a human-provided similar molecule meeting the criterion guided a successful retry that changed the functional groups edited.The final molecule remained structurally similar to the original.Source: §4.4 · p. 11, Figure 9 · p. 12
5 CONCLUSION AND FUTURE WORK
The paper presents DrugAssist as an interactive molecule optimization model that adjusts outputs in response to user feedback. It reports strong single- and multi-property performance and identifies multimodal handling and interaction quality as future-work targets.
5 CONCLUSION AND FUTURE WORK
- DrugAssist uses natural-language interaction to generate optimized molecules and continue adjusting them based on user feedback.The paper reports performance in both single- and multi-property optimization, including tasks specifying property ranges.Source: §5 · p. 12
- MolOpt-Instructions is released as an instruction-based dataset for fine-tuning LLMs in molecule optimization.The paper presents it as a resource for future work in the domain.Source: §5 · p. 12
- Future work targets multimodal data and tasks to reduce hallucinations, alongside improved understanding of users’ needs and feedback.Source: §5 · p. 12
A PROMPT SETTINGS FOR THE COMPARISONS WITH LLMS
The appendix provides prompts for different tasks and identifies the molecule-string placeholder used in them.
A PROMPT SETTINGS FOR THE COMPARISONS WITH LLMS
- Table 6 collects prompts for different tasks and uses “[SMILES]” to represent the molecule’s SMILES string.Source: Table 6 · p. 15