Large-language models facilitate discovery of the molecular signatures regulating sleep and activity

Di Peng, Liubin Zheng, Dan Liu, Cheng Han, Xin Wang, Yan Yang, Song Li, Miaoying Zhao, Yanfeng Wei, Jiayi Li, Xiaoxue Ye, Yuxiang Wei, Zihao Feng, Xinhe Huang, Miaomiao Chen, Yujie Gou, Yu Xue, Luoying ZhangNature Communications 2024 · DOI

Abstract

Sleep, locomotor and social activities are essential animal behaviors, but their reciprocal relationships and underlying mechanisms remain poorly understood. Here, we elicit information from a cutting-edge large-language model (LLM), generative pre-trained transformer (GPT) 3.5, which interprets 10.2-13.8% of Drosophila genes known to regulate the 3 behaviors. We develop an instrument for simultaneous video tracking of multiple moving objects, and conduct a genome-wide screen. We have identified 758 fly genes that regulate sleep and activities, including mre11 which regulates sleep only in the presence of conspecifics, and NELF-B which regulates sleep regardless of whether conspecifics are present. Based on LLM-reasoning, an educated signal web is modeled for understanding of potential relationships between its components, presenting comprehensive molecular signatures that control sleep, locomotor and social activities. This LLM-aided strategy may also be helpful for addressing other complex scientific questions.

Introduction

Understanding the genetic basis of behavior is a "needle in a haystack" problem. While sleep, locomotion, and social behavior are known to be interlinked, the specific molecular signatures coordinating them—especially in social settings—remain poorly understood. Traditional research faces two primary bottlenecks: the technical difficulty of tracking nuanced behaviors of multiple interacting animals simultaneously, and the immense human effort required to synthesize knowledge from tens of thousands of genes and millions of publications.

Methods: LLM-Aided Discovery and High-Throughput Tracking

The researchers employed a dual-track strategy to overcome these bottlenecks. First, they used GPT-3.5 to conduct a "silico interpretation" of 42,794 fly protein isoforms. By utilizing Chain-of-Thought (CoT) prompting, they elicited complex reasoning regarding gene functions and potential regulatory relationships.

Second, to handle behavioral data acquisition, they developed the Video-to-Location (VTL) instrument. This system transforms massive raw video files (over 100 TB) into compact 2D coordinate text files, enabling high-accuracy tracking of multiple flies in social environments. The VTL system maintained >98% accuracy for groups of up to 5 flies, significantly outperforming standard automated systems.

Overview of the LLM-aided strategy and the VTL tracking system.

Finally, they conducted a genome-wide RNAi screen, knocking down 5,588 genes in the fly nervous system to provide an experimental dataset to validate the LLM’s predictions.

Key Insights: The "Social Sleep" Gene and LLM Performance

A striking discovery was the identification of mre11 as a regulator of "social sleep." Traditionally known for DNA repair, mre11 was found to regulate sleep only when flies are in a group. In isolated flies, mre11 deficiency had no effect, suggesting the gene processes social-related sensory cues that influence rest.

The study also benchmarked the LLM's performance. GPT-3.5 acted as a "conservative" expert, showing high specificity (~93%) but low sensitivity (~20%). While it was highly reliable when predicting a gene's involvement (low false positives), it missed newer discoveries due to its training data cutoff. However, its efficiency was remarkable: the LLM performed 3,655 pairwise gene-relationship comparisons in just 3 hours—a task estimated to take a human researcher over a week.

Results: Mapping the Molecular Landscape

The genome-wide screen identified 758 genes regulating sleep and activity: 285 for sleep, 310 for locomotion, and 359 for social activity. Notably, 30 "master regulators" were found to influence all three behaviors simultaneously.

Genome-wide screening results and overlap of genes regulating sleep, locomotor, and social behaviors.

Using LLM reasoning, the team modeled an "educated signal web" for 86 candidate genes. Validation showed that 74.1% of the LLM-generated relationships were supported by existing literature, while the remainder represent novel hypotheses for future experimental testing.

Discussion and Future Directions

While powerful, the LLM-aided approach is limited by the model's knowledge cutoff (e.g., 2022 for GPT-3.5), leading to Type II errors where recent discoveries are missed. Furthermore, LLM-proposed relationships require experimental validation. Future work will involve utilizing more advanced models (like GPT-4) and expanding the VTL system to track larger, more complex social groups across different species. This "human-AI interactive" framework represents a significant leap in accelerating the synthesis of complex biological networks.

Summary updated

Abstract

Sleep, locomotor and social activities are essential animal behaviors, but their reciprocal relationships and underlying mechanisms remain poorly understood. Here, we elicit information from a cutting-edge large-language model (LLM), generative pre-trained transformer (GPT) 3.5, which interprets 10.2-13.8% of Drosophila genes known to regulate the 3 behaviors. We develop an instrument for simultaneous video tracking of multiple moving objects, and conduct a genome-wide screen. We have identified 758 fly genes that regulate sleep and activities, including mre11 which regulates sleep only in the presence of conspecifics, and NELF-B which regulates sleep regardless of whether conspecifics are present. Based on LLM-reasoning, an educated signal web is modeled for understanding of potential relationships between its components, presenting comprehensive molecular signatures that control sleep, locomotor and social activities. This LLM-aided strategy may also be helpful for addressing other complex scientific questions.
BibTeX
@inproceedings{peng2024large,
  author = {Di Peng and Liubin Zheng and Dan Liu and Cheng Han and Xin Wang and Yan Yang and Song Li and Miaoying Zhao and Yanfeng Wei and Jiayi Li and Xiaoxue Ye and Yuxiang Wei and Zihao Feng and Xinhe Huang and Miaomiao Chen and Yujie Gou and Yu Xue and Luoying Zhang},
  title = {{Large-language models facilitate discovery of the molecular signatures regulating sleep and activity}},
  booktitle = {Nature Communications},
  year = {2024},
  doi = {10.1038/s41467-024-48005-w},
}
Figure