ChIP-GPT

ChIP-GPT extracts metadata from biomedical databases, particularly chromatin immunoprecipitation (ChIP) targets and cell lines, through fine-tuning of the Llama generative pre-trained transformer (GPT) model with iterative program prompting and answer text handling.

Key Features:

• Automated LLM interactions: ChIP-GPT automates interactions with the large language model to enable scalable database querying.
• Record pruning and summarization: The tool handles limitations on input length through pre-processing steps that prune or summarize records.
• Few-shot and fine-tuning capabilities: ChIP-GPT uses well-designed short 'few-shot' examples or fine-tuning based on larger sets of well-curated examples to behave reliably.
• Error-tolerant extraction: The model achieves 90-94% accuracy when trained with 100 examples and seamlessly extracts data from records containing typos or absent field labels.
• Adaptability: The method is easily adaptable to customized questions and different databases.

Scientific Applications:

• ChIP target identification: ChIP-GPT identifies chromatin immunoprecipitation targets from biomedical database records.
• Cell line extraction: The model extracts cell line metadata from database entries.
• Sequence Read Archive metadata extraction: ChIP-GPT processes entries from the Sequence Read Archive at scale.
• Large-scale biomedical data characterization: The tool enables large-scale analyses of amassing biomedical data for applications in biology and medicine.

Methodology:

ChIP-GPT is based on fine-tuning of the generative pre-trained transformer (GPT) model Llama and on a program prompting the model iteratively and handling its generation of answer text.