Scholarlink

Predicting drug response is critical for advancing precision medicine, as many drugs fail in clinical trials due to a limited understanding of molecular mechanisms. We introduce DRPLLM (Drug Response Prediction using Large Language Model), a novel framework that integrates the Llama- 8b large language model (LLM) with a deep neural network (DNN) regression head. DRPLLM uses embeddings from cell line gene expression, mutations, and drug properties formatted into natural language prompts. Trained on the Genomics of Drug Sensitivity in Cancer v2 and Cancer Cell Line Encyclopedia datasets, DRPLLM achieved a Spearman rank correlation coefficient (SCC) of 0.71, improving to 0.74 after hyperparameter optimization. Validation on a patient- derived xenograft (PDX) cohort showed that the embedding-based approach outperformed the traditional prompt-engineering method SCC = 0.45 (p < 1e-16) versus SCC = 0.06 (p > 0.05), demonstrating its potential for accurate, generalizable drug response predictions in precision oncology.

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

DRPLLM: A Large Language Model-Based Framework for Predicting Drug Response in Cancer Using Multi-Omics Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Rohan Gnanaolivu,
  • Priyanka Vasanthakumari,
  • Brenna C. Novotny,
  • Jamie C. Overbeek,
  • Yuanhang Liu,
  • Alexander Partin,
  • Nicholas Chia,
  • Yitan Zhu,
  • Chen Wang

摘要

Predicting drug response is critical for advancing precision medicine, as many drugs fail in clinical trials due to a limited understanding of molecular mechanisms. We introduce DRPLLM (Drug Response Prediction using Large Language Model), a novel framework that integrates the Llama- 8b large language model (LLM) with a deep neural network (DNN) regression head. DRPLLM uses embeddings from cell line gene expression, mutations, and drug properties formatted into natural language prompts. Trained on the Genomics of Drug Sensitivity in Cancer v2 and Cancer Cell Line Encyclopedia datasets, DRPLLM achieved a Spearman rank correlation coefficient (SCC) of 0.71, improving to 0.74 after hyperparameter optimization. Validation on a patient- derived xenograft (PDX) cohort showed that the embedding-based approach outperformed the traditional prompt-engineering method SCC = 0.45 (p < 1e-16) versus SCC = 0.06 (p > 0.05), demonstrating its potential for accurate, generalizable drug response predictions in precision oncology.