SensitiveCancerGPT: Leveraging Generative Large Language Model on Structured Omics Data to Optimize Drug Sensitivity...

Area of Science:

• Pharmacogenomics
• Computational Biology
• Artificial Intelligence in Oncology

Background:

• Vast pharmacogenomics data offers opportunities for drug sensitivity prediction (DSP) in precision oncology.
• Generative Large Language Models (LLMs) show potential but struggle with structured pharmacogenomics data.

Purpose of the Study:

• Adapt prompt engineering for LLMs to optimize DSP performance on structured pharmacogenomics data.
• Evaluate LLM generalization in real-world DSP scenarios.
• Compare LLM DSP performance against state-of-the-science baselines.

Main Methods:

• Systematically investigated Generative Pre-trained Transformer (GPT) on four pharmacogenomics datasets across five cancer tissue types.
• Employed novel prompt engineering with instruction, instruction-prefix, and cloze templates, integrating pharmacogenomics features.
• Assessed GPT via zero-shot, few-shot, fine-tuning, and clustering pretrained embeddings.

Main Results:

• Fine-tuning GPT achieved the best DSP performance (28% F1 increase), outperforming few-shot learning.
• Prompt engineering, particularly instruction-prefix templates, enhanced F1 performance by 22%.
• GPT demonstrated superior mean F1 performance (16% gain) against baselines and comparable cross-tissue generalization.

Conclusions:

• Generative LLMs like GPT are viable in silico tools for guiding precision oncology.
• Optimized prompt engineering is crucial for leveraging LLMs with structured pharmacogenomics data.
• GPT's performance highlights its potential in drug sensitivity prediction and personalized cancer treatment.