Weng Guangzheng, Martin Patrick, Kim Hyobin, Won Kyoung Jae
Biotech Research and Innovation Centre (BRIC), University of Copenhagen, Ole Maaløes Vej 5, Copenhagen, 2200, Denmark.
Department of Computational Biomedicine, Cedars-Sinai Medical Center, Los Angeles, CA, 90069, USA.
Adv Sci (Weinh). 2025 Jan;12(3):e2409990. doi: 10.1002/advs.202409990. Epub 2024 Nov 28.
Gene regulatory network (GRN) inference, a process of reconstructing gene regulatory rules from experimental data, has the potential to discover new regulatory rules. However, existing methods often struggle to generalize across diverse cell types and account for unseen regulators. Here, this work presents GRNPT, a novel Transformer-based framework that integrates large language model (LLM) embeddings from publicly accessible biological data and a temporal convolutional network (TCN) autoencoder to capture regulatory patterns from single-cell RNA sequencing (scRNA-seq) trajectories. GRNPT significantly outperforms both supervised and unsupervised methods in inferring GRNs, particularly when training data is limited. Notably, GRNPT exhibits exceptional generalizability, accurately predicting regulatory relationships in previously unseen cell types and even regulators. By combining LLMs ability to distillate biological knowledge from text and deep learning methodologies capturing complex patterns in gene expression data, GRNPT overcomes the limitations of traditional GRN inference methods and enables more accurate and comprehensive understanding of gene regulatory dynamics.
基因调控网络（GRN）推理是一个从实验数据中重建基因调控规则的过程，它有潜力发现新的调控规则。然而，现有方法往往难以在不同细胞类型中进行泛化，也难以考虑到未见过的调控因子。在此，这项工作提出了GRNPT，这是一个基于Transformer的新颖框架，它整合了来自公开可用生物数据的大语言模型（LLM）嵌入以及一个时间卷积网络（TCN）自动编码器，以从单细胞RNA测序（scRNA-seq）轨迹中捕捉调控模式。在推断基因调控网络方面，GRNPT显著优于有监督和无监督方法，特别是在训练数据有限的情况下。值得注意的是，GRNPT表现出卓越的泛化能力，能够准确预测之前未见过的细胞类型甚至调控因子中的调控关系。通过结合大语言模型从文本中提炼生物学知识的能力和深度学习方法捕捉基因表达数据中复杂模式的能力，GRNPT克服了传统基因调控网络推理方法的局限性，能够更准确、全面地理解基因调控动态。
