Automated analysis of a diverse synapse population

Abstract

Synapses of the mammalian central nervous system are highly diverse in function and molecular composition. Synapse diversity per se may be critical to brain function, since memory and homeostatic mechanisms are thought to be rooted primarily in activity-dependent plastic changes in specific subsets of individual synapses. Unfortunately, the measurement of synapse diversity has been restricted by the limitations of methods capable of measuring synapse properties at the level of individual synapses. Array tomography is a new high-resolution, high-throughput proteomic imaging method that has the potential to advance the measurement of unit-level synapse diversity across large and diverse synapse populations. Here we present an automated feature extraction and classification algorithm designed to quantify synapses from high-dimensional array tomographic data too voluminous for manual analysis. We demonstrate the use of this method to quantify laminar distributions of synapses in mouse somatosensory cortex and validate the classification process by detecting the presence of known but uncommon proteomic profiles. Such classification and quantification will be highly useful in identifying specific subpopulations of synapses exhibiting plasticity in response to perturbations from the environment or the sensory periphery.

Figure 1. The synaptogram as a tool for high-dimensional proteomic visualization.

(A) A maximum projected volume of Synapsin I labeling. 41 slices, 70 nm per slice, total thickness of 2.87 . (B) Randomly-colored segmentation of individual synapsin puncta. (C) Rendering of a single punctum from the volume showing synapsin (white), imaged together with VGluT1 (red), PSD95 (green), GluR2 (blue), GAD (magenta) and VGAT (magenta). From top to bottom: all proteomic markers, glutamatergic presynaptic labels, glutamatergic postsynaptic labels, GABAergic labels. This appears to be a glutamatergic synapse. (D) The synaptogram derived from the same synapse. Synapsin, top row, is repeated in red for the rest to provide spatial context. Not shown, sixteen other colors and two redundant labels (synapsin and VGluT1). Scale bar: 5 , size of synaptogram/render volume, 1100 nm × 1100 nm × 630 nm.

Figure 2. Clustering of synapsin I imaged with array tomography.

When the first and third principal components of the local brightness feature eq 1 are plotted against each other, they form clusters identifiable as known synaptic subtypes, and explain 50.4% of the variance in the data.

Figure 3. Relative feature importance for different molecular labels.

When all classes were averaged (top left), our local brightness feature (ii) saw the most use, followed by integrated brightness (i), center of mass (iii) and moment of inertia (iv). GAD, VGAT, PV, VGluT3, VGluT2, VGluT1, PSD95, VAChT, and TH each make slightly different use of the feature set. VGluT3, VGluT2, and VAChT are notable in that they rely most heavily on features other than local brightness.

Figure 4. Comparison of human rating to machine learning.

(A) Accuracy rates. i-vi - When compared against the average decisions of their peers in a VGluT1 synapse discrimination task, humans performed at different accuracy levels based on their stringency of classification. vii - The random forest ensemble, (VGluT1 PSD95), trained by human rater i, performed comparably to the humans. (B) Rater agreement histogram. 100 individually-classified synaptic loci are scored according to the number of “yes” votes received among the five humans composing the gold standard consensus. Situations with unanimous agreement (0,5) make up half of the set (49 loci), with an additional 37 examples having only one dissenting opinion (1,4). (C) Receiver operating characteristics (ROC) curve, for VGluT1 and PSD95 classifications on human-rated data. The ROC curve describes the tradeoff between reducing false positives (left side of the curve) and maximizing true positives (right side of the curve). The displayed diagonal line represents chance, with better classifiers occupying large areas between the diagonal and their own curves.

Figure 5. Density and size of synapse classes as a function of depth through the cortex.

(A) Synapse density through the cortex. * - VGluT2 synapse density peaks in layer IV. PV-positive GABAergic synapse density is slightly decreased in layer I, and significantly lacking in layer VI. (B) Synapse size estimated using the synapsin local brightness measurement. ** - VGluT1 size peaks in layer Va (p0.05).

Figure 6. Positive and negative pairwise channel copresence.

Symbols denote interesting comparisons with statistical significance of . Red squares represent label pairs which are copresent more than expected, blue squares less than expected by chance. * - GABAergic markers are copresent with each other, but avoid glutamatergic and TH markers. ? - VGluT1/2 are copresent with PSD95, but not with each other. # - VGluT3 is present with all three GABAergic markers, but avoids VGluT1 and PSD95. & - VGluT2 shows some presence with TH. ∼ - TH tends to avoid VAChT.