Towards a connectomics-based predictive model of the inner retina

视觉处理从视网膜开始，在那里提取至少40种不同的特征，并通过平行通道发送到大脑的高级视觉中心。视网膜研究中最大的挑战之一是了解这些不同的表征是如何在视网膜回路中产生的。到目前为止，其中只有少数被很好地理解；在每种情况下，理解一种特定类型的中间神经元的作用是揭示计算机制的关键。中间神经元是主要抑制性视网膜细胞类的一部分，统称为无突细胞（ACs）。尽管ACs在视网膜计算中起着关键作用，但令人惊讶的是，人们对ACs的60多种遗传类型及其在视网膜内的复杂网络知之甚少。在优先计划2041“计算连接组学”的第一阶段，我们旨在进一步剖析交流电路在图像处理中的功能作用。我们开发了一种新的双光子成像范式，并将记录的功能数据与广泛的视觉刺激和基于接触的连接数据相结合，开发了一个性能良好的预测模型，用于视网膜内的时间处理，包括交流电路。在下一阶段，我们将扩展我们的模型，包括空间处理，并基于新的电子显微镜重建，具有突触级连接和双成像功能数据，使我们能够以前所未有的方式约束扩展模型。首先，我们将生成一个具有突触分辨率的新的功能注释的小鼠视网膜连接组数据集。我们将使用一种新型的多束扫描电子显微镜，它可以在几周而不是几个月内收集大量的组织。然后，我们将开发一个自动管道，将重建的单元格分类为类型，并将它们与功能数据对齐。我们将使用双光子成像和轴向扫描同时测量双极细胞（BC）轴突末端的谷氨酸释放-视网膜内部的兴奋性输入-以及整个内丛状层中交流树突的活动。接下来，我们将基于BC和AC类型之间的连通性数据建立时空处理模型，并根据功能数据推断其参数，从而整合两种数据模式。通过迭代，我们将不断添加更多关于连通性的细节，并将其与基于前几轮建模结果设计的实验中获得的功能数据进行拟合，从而改进模型。总之，这将使我们能够进一步了解ACs在视网膜计算中的作用。

Visual processing starts in the retina, where at least 40 distinct features are extracted and sent through parallel channels to higher visual centres in the brain. One of the biggest remaining challenges in retinal research is to understand how these diverse representations arise within the retinal circuits. So far, only a few of these are well understood; in each case, understanding the role of a specific type of interneuron, part of the largely inhibitory retinal cell class collectively called amacrine cells (ACs), was key to revealing the computational mechanisms. Despite the key role of ACs in retinal computations, surprisingly little is known about the great majority of the 60+ genetic types of ACs and their intricate networks in the inner retina. In the first phase of the Priority Program 2041 “Computational Connectomics”, we aimed to further dissect the functional roles of AC circuits for image processing. We developed a new two-photon imaging paradigm and combined functional data recorded with a wide range of visual stimuli with contact-based connectivity data to develop a well-performing predictive model for temporal processing in the inner retina, including AC circuits. In the next phase, we will extend our model to include spatial processing and base it on new electron microscopy reconstructions with synapse-level connectivity and dual-imaging functional data, allowing us to constrain the extended model in unprecedented ways. First, we will generate a new, functionally-annotated connectomics dataset of the mouse retina with synapse resolution. We will use a novel multibeam scanning electron microscope that enables large tissue volumes to be collected in weeks instead of months. Then, we will develop an automatic pipeline to classify the reconstructed cells into types and align them with functional data. We will use 2-photon imaging with axial scans to simultaneously measure glutamate release from bipolar cell (BC) axon terminals – the excitatory input to the inner retina – and activity in AC dendrites through the entire depth of the inner plexiform layer. Next, we will integrate both data modalities by setting up a model for spatio-temporal processing based on the connectivity data between BC and AC types and inferring its parameters based on the functional data. Through iteration, we will improve the model by successively adding more details on connectivity and by fitting it with functional data gained from experiments designed based on modelling results from previous rounds. Together, this will allow us to advance our knowledge of the ACs’ roles in retinal computations.