Graph learning methods for engieering mammalian promoters in bioproduction

新冠肺炎基因疫苗有效地防止了熊猫期间的住院和死亡。在哺乳动物细胞系中可以实现有价值的疫苗和生物治疗的生物生产，但很难开发出健壮、可预测和可持续的表达。因此，设计增强型哺乳动物启动子和基因电路是一个关键的战略性工业目标。这个项目的目的是探索大规模转录数据集的结构特性，并学习如何在哺乳动物细胞中优化工程启动子。作为该项目的一部分，将使用多种公开可用的转录数据集，包括EPD[2]、DEE2统一转录数据库[7]和SRA[5]，用于构建异质图，其中每个节点表示一组DNA序列(即启动子和基因)；边类型由两个节点共享的生物背景决定。首先，利用图神经网络[3，4，8]来学习网络中每个节点的表示，将节点映射到嵌入空间来反映图的结构。这样，细胞和环境的上下文将被嵌入到学习的节点表示中，然后我们可以使用所得到的节点嵌入来预测启动子的活性(即基因表达的速度和数量)。启动子活性预测任务可以与表示学习过程相结合，也可以使用分离预测器来求解。该项目的第二阶段将结合实验的结果，反复改进GNN。经过训练的模型将选择50-100个启动子的基本集合，这些启动子在一系列条件下起作用，并具有期望的性能，作为合成启动子设计的起点。在与伦敦生物铸造厂的合作中，他们的特征将在瞬时转基因中进行测试，以消除由于拷贝数量和物理环境的差异而导致的行为变化。GNNS算法将使用具有验证行为的启动子作为模块构建块，以构建一组具有设计行为的50个合成启动子。构建性能将与GNN产生的预测进行比较，并用于改进模型。总的来说，该项目将利用GNN的表达能力和大规模转录数据集来开发一个广泛适用的框架，用于基因表达问题分析，并应用于预测上下文启动子行为和设计新的哺乳动物启动子。此外，学习模型的可解释性可通过应用符号回归[1]和GNN-解释器[6]来实现。

mRNA COVID-19 vaccines have effectively prevented hospitalization and deathduring the pandamic. Bioproduction of valuable vaccines and biotherapeuticsin mammalian cell lines can be achieved, but it is difficult to develop robust,predictable, and sustainable expression. The design of enhanced mammalianpromoters and genetic circuits is therefore a key strategic industrial target. Thisproject aims to explore the structural properties of large-scale transcriptomicdatasets with graph representation learning to optimise engineered promoters inmammalian cells. Demonstration will be performed in mammalian cell lines us-ing automated DNA assembly and analytics available at the London Biofoundry.As part of this project, multiple publicly available transcriptomic datasetswill be used, including the EPD [2], DEE2 uniform transcriptomic database[7], and SRA [5] for constructing heterogeneous graphs, in which each noderepresents a set of DNA sequences (i.e. promoters and genes); edge types aredetermined by the biological context shared by two nodes. The initial noderepresentation may be generated using the internal DNA sequence structure.With constructed heterogeneous network, there are two objectives: Firstly,we will use Graph neural networks (GNNs) [3, 4, 8] to learn the representationof each node in the network, which maps the nodes into embedding space toreflect the graph structure. In this way, the cellular and environmental con-text will be embedded into the learnt node representations; and then we canuse the resulting node embedding to predict the promoter activity (i.e. thespeed and quantity of gene expression). Promoter activity prediction task maybe integrated with the representation learning process or solved using separatepredictors. The second stage of the project will combine results from exper-iments to iteratively refine the GNNs. The trained model will select a baseset of 50-100 promoters functional over a range of conditions and with desiredperformance to be used as starting point for synthetic promoter design. In col-laboration with the London Biofoundry, their characteristics will be tested intransient transfection to eliminate changes in behaviour due to differences incopy number and physical context. Promoters with verified behaviour will beused by the GNNs algorithm as modular building blocks to build a set of 50synthetic promoters with designed behaviour. Construct performance will becompared against predictions generated by the GNNs and used to refine modelperformance.Overall, the project will leverage the expression power of GNNs and thelarge-scale transcriptomic datasets to develop a broadly applicable frameworkfor gene expression problem analysis, with the application in predicting con-textual promoter behaviours and engineering new mammalian promoters. Inaddition, the interpretability of learnt model can be achieved by applying sym-bolic regression [1] and GNN-explainer [6].