Model-to-crop conserved NUE Regulons enhance machine learning predictions of nitrogen use efficiency

In this data-rich era, the promise of systems biology is to learn gene regulatory networks controlling key agricultural traits. However, validating these networks in crops remains challenging. By integrating gene regulatory network and machine learning, we functionally validated network regulons predicting nitrogen use efficiency (NUE) in Arabidopsis and maize. Our time-course nitrogen response transcriptome analysis uncovered a conserved N-response cascade between maize and Arabidopsis. Using Dynamic Factor Graph, we inferred N-regulated gene regulatory networks (N-GRNs) in maize and validated TF-target interactions for 23 maize TFs with the TARGET, a cell-based TF-perturbation assay. We pruned the N-GRNs by Precision-Recall analysis. Combining these data, we uncovered a previously unknown role for KNOTTED1 in the dynamic N-signaling network. We learned gene-to-NUE trait models across 16 maize varieties using XGBoost trained on N-response genes conserved model-to-crop. Integrating NUE importance scores within our GRN, we ranked maize TFs by their NUENet scores. In a model-to-crop approach, we validated orthologous N-regulated TF-targets for the top-ranked maize NUENet TFs (MYB34/R4?24 targets) and the orthologous Arabidopsis TF (AtDIV1?23 targets) using the cell-based TARGET assay. The genes in this orthologous model-to-crop NUENet regulons were superior at predicting NUE traits in XGBoost models learned in both maize and Arabidopsis. Thus, our model-to-crop approach combining GRNs, machine learning, and orthologous network modules offers a strategic framework for crop trait improvement. Overall design: We conducted the fine time-course nitrogen treatement experiment in maize B73 seedlings.On the day of the sampling (day ten), two hours after the start of the light period (10 am), seedlings were transferred to the fresh hydroponic medium containing either the Nitrogen in MS medium (20 mM KNO3 + 20 mM NH4NO3) or 20 mM KCl and harvested at time intervals 0, 5, 10, 15, 20, 30, 45, 60, 90, and 120 minutes.

在这个数据富集的时代，系统生物学（systems biology）的核心愿景在于解析调控关键农业性状的基因调控网络（gene regulatory networks, GRNs）。然而，在作物中验证这类网络仍颇具挑战。
我们通过整合基因调控网络与机器学习技术，对拟南芥（Arabidopsis）与玉米（maize）中预测氮素利用效率（nitrogen use efficiency, NUE）的网络调控子（regulons）进行了功能验证。本研究的时序氮响应转录组分析揭示了玉米与拟南芥之间保守的氮响应级联通路。
我们借助动态因子图（Dynamic Factor Graph）推断了玉米中受氮素调控的基因调控网络（N-regulated gene regulatory networks, N-GRNs），并利用基于细胞的转录因子扰动检测技术（TARGET）验证了23个玉米转录因子（transcription factor, TF）的靶基因互作关系。随后通过精确召回（Precision-Recall）分析对N-GRNs进行了剪枝优化。
整合上述数据后，我们发现KNOTTED1在动态氮信号网络中存在此前未被报道的功能。我们基于跨物种保守的氮响应基因训练XGBoost模型，进而在16个玉米品种中构建了基因与氮素利用效率性状的关联模型。将氮素利用效率重要性评分整合至我们的GRN后，我们基于NUENet评分对玉米转录因子进行了排序。
我们采用跨物种模型-作物转化的研究策略，利用基于细胞的TARGET检测技术，验证了排名靠前的玉米NUENet转录因子（MYB34/R4，共24个靶基因）及其拟南芥同源转录因子AtDIV1（共23个靶基因）的受氮素调控的靶基因互作关系。在同时基于玉米与拟南芥构建的XGBoost模型中，该跨物种同源NUENet调控子所包含的基因对氮素利用效率性状的预测性能更为优异。
综上，我们将基因调控网络、机器学习与同源网络模块相结合的跨物种模型-作物转化研究策略，为作物性状改良提供了一套系统性的研究框架。
实验整体设计：我们以玉米B73幼苗为材料，开展了精细时序氮素处理实验。在采样当日（第10天），光照周期开始后2小时（上午10时），将幼苗转移至新鲜的水培培养基中，培养基分别添加MS培养基中的氮源（20 mM KNO3 + 20 mM NH4NO3）或20 mM KCl，并分别于0、5、10、15、20、30、45、60、90、120分钟时采集样本。