Scholarlink

Purpose <p>This work aims to accelerate and enhance pharmaceutical drug dissolution prediction by integrating advanced Large Language Models (LLMs) and AI-diffusion models to reduce reliance on time-consuming, costly empirical experiments. The framework sets a foundation for broader adoption of generative AI in drug development.</p> Methods <p>This work introduces a DeepSeek based LLM framework augmented by prompt engineering (zero-shot, few-shot, chain-of-thought) and adaptive weighted retrieval-augmented generation (RAG) to systematize dissolution profile from basic physical properties. Moreover, a diffusion model synthesizes SEM-derived morphological parameters (e.g., particle size, surface area), circumventing error accumulation from multi-instrument characterization workflows. These parameters feed the RAG database, enabling LLM predictions grounded in structure-performance relationships rather than idealized assumptions.</p> Results <p>Overall, the LLM generated dissolution profile (few-shot chain-of-thought with RAG) provides a good agreement between experimental and the prediction result among others. Sensitivity analysis is investigated to quantify the reliability and stability of the prompt content. Additionally,&#xa0;diffusion-generated structural data&#xa0;from SEM images combined with the LLM's predictive capabilities are tested to connect macro-scale physical properties with microstructural characteristics, achieving a close profile trend with acceptable RMSE and PCC.</p> Conclusions <p>This study demonstrates the potential of the DeepSeek-based LLM framework to describe the dissolution of drug powders. Among the different system prompt strategies, few-shot chain-of-thought with RAG performs the best dissolution profile among others. While it may overcomplicate straightforward tasks in certain scenarios. The combination of diffusion models&#xa0;successfully bridges&#xa0;AI-driven insights&#xa0;(e.g., dissolution predictions) with&#xa0;physical and structural drug properties&#xa0;(e.g., particle geometry from SEM images).</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DeepSeek-LLM with Adaptive RAG for Pharmaceutical Dissolution Prediction

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Leqi Lin,
  • Xingyu Zhou,
  • Kaiyuan Yang,
  • Yang Liu,
  • Xizhong Chen

摘要

Purpose

This work aims to accelerate and enhance pharmaceutical drug dissolution prediction by integrating advanced Large Language Models (LLMs) and AI-diffusion models to reduce reliance on time-consuming, costly empirical experiments. The framework sets a foundation for broader adoption of generative AI in drug development.

Methods

This work introduces a DeepSeek based LLM framework augmented by prompt engineering (zero-shot, few-shot, chain-of-thought) and adaptive weighted retrieval-augmented generation (RAG) to systematize dissolution profile from basic physical properties. Moreover, a diffusion model synthesizes SEM-derived morphological parameters (e.g., particle size, surface area), circumventing error accumulation from multi-instrument characterization workflows. These parameters feed the RAG database, enabling LLM predictions grounded in structure-performance relationships rather than idealized assumptions.

Results

Overall, the LLM generated dissolution profile (few-shot chain-of-thought with RAG) provides a good agreement between experimental and the prediction result among others. Sensitivity analysis is investigated to quantify the reliability and stability of the prompt content. Additionally, diffusion-generated structural data from SEM images combined with the LLM's predictive capabilities are tested to connect macro-scale physical properties with microstructural characteristics, achieving a close profile trend with acceptable RMSE and PCC.

Conclusions

This study demonstrates the potential of the DeepSeek-based LLM framework to describe the dissolution of drug powders. Among the different system prompt strategies, few-shot chain-of-thought with RAG performs the best dissolution profile among others. While it may overcomplicate straightforward tasks in certain scenarios. The combination of diffusion models successfully bridges AI-driven insights (e.g., dissolution predictions) with physical and structural drug properties (e.g., particle geometry from SEM images).

Graphical Abstract