Researchers have developed ChatNT, a groundbreaking multimodal conversational agent designed to interpret and analyze complex biological sequences like DNA, RNA, and proteins. By integrating a DNA encoder with a large language model, this system allows users to solve advanced genomics tasks using simple English instructions rather than complex code. The model achieves state-of-the-art accuracy across dozens of tasks, including predicting promoter activity, splice sites, and protein stability. Its unique architecture utilizes an English-aware projection to extract specific biological features based on the user's unique questions. Additionally, the study introduces a perplexity-based method to verify the model's confidence, ensuring more reliable predictions for scientific research. Ultimately, ChatNT makes sophisticated genomic analysis accessible to a broader range of scientists, bridging the gap between artificial intelligence and biological discovery.

References:

• de Almeida B P, Richard G, Dalla-Torre H, et al. A multimodal conversational agent for DNA, RNA and protein tasks[J]. Nature Machine Intelligence, 2025: 1-14.