MULTIMODAL CELL CONTEXT INSTRUCTION TUNING FOR CONDITIONAL DNA REGULATORY SEQUENCE GENERATION WITH LARGE LANGUAGE...

Area of Science:

• Synthetic biology
• Genomics
• Computational biology

Background:

• Designing effective regulatory DNA sequences like enhancers is crucial for synthetic biology.
• Current methods struggle with cell-type-specific transcription factor (TF) recruitment, leading to suboptimal enhancer designs.

Purpose of the Study:

• To introduce Leonine, a novel framework for designing cell-type-specific enhancers.
• To redefine enhancer design as a multimodal question-answering task integrating DNA sequences and cellular context.

Main Methods:

• Developed Leonine, a framework utilizing large language models (LLMs) trained on DNA sequences.
• Integrated multimodal cellular data, including promoter sequences, gene expression, and cell type information.
• Created a large-scale dataset of over 1.5 million cell-type-specific promoter-enhancer pairs and a benchmark for evaluation.

Main Results:

• Leonine demonstrated superior performance compared to state-of-the-art LLMs across seven cell types.
• Generated biologically and functionally coherent enhancer sequences tailored for specific regulatory environments.
• Outperformed existing methods in optimizing enhancer sequences for effective transcription activation.

Conclusions:

• Leonine establishes a new paradigm for context-aware DNA sequence design.
• Advances the field of synthetic biology by enabling precise control of gene regulation.
• Provides a powerful tool for optimizing enhancer function in specific cellular contexts.