Automated Synapse Detection and Validation by Correlated Array Tomography and Scanning Electron Microscopy

Array tomography (AT) is a method for mapping the expression patterns of many proteins at high resolution in three dimensions. Using ultrathin serial tissue sections, multiple cycles of immunohistochemistry, optical imaging, and 3D reconstruction, AT can map the expression of multiple proteins at sub-micrometer resolution across large tissue volumes. Using antibodies to synaptic proteins, AT has been used to reconstruct synaptic architecture in brain tissue, and we have developed automated routines for counting synapses. To validate our algorithms and immunostaining, we examined AT samples by both light and electron microscopy in order to correlate synaptic protein immunoreactivity and ultrastructure. We cut a 30-section array of 70 nm serial sections of Lowicryl-embedded mouse hippocampus and ran two cycles of immunostaining and light microscopy. It was then stained with heavy metals and a subregion of six serial sections imaged by field-emission scanning electron microscopy (SEM). After deconvolving, stitching, aligning and merging AT and SEM data, we manually scanned the SEM volume and detected 98 synapses, largely by the presence of electron-dense post-synaptic densities. We then overlaid fluorescence signals from anti-PSD-95 and anti-synapsin antibodies (Cell Signaling) and compared them to the ultrastructurally-defined synapses. Finally, we ran light-level data through our automated synapse-detection algorithm which detects the co-localization of PSD-95 and synapsin local maxima. To determine false negatives, we examined how many SEM-identified synapses were detected by the algorithm. An optimized algorithm detected 72% of SEM-identified excitatory synapses. False-negatives were caused by failure of either immunostaining (12%) or the algorithm (16%). To determine false positives, we examined how many algorithm-detected synapses corresponded to SEM-identified synapses. 81% were validated, while 19% were false-positives often due to the close-packing of multiple synapses. We expect to extend these analyses to other markers and algorithms.