The brain is made up of dense networks of interconnected neurons. Mapping the anatomy of these dense networks is one of the biggest challenges in neuroscience. Electron microscopy provides the highest resolution and is used as a gold standard in connectomics; however, its data size hampers large-scale circuit reconstruction at the millimetre scale. Light microscopy combined with tissue clearing is a new emerging approach for mesoscopic circuit mapping. However, the reconstruction of densely labelled circuits is challenging as its limited resolution hinders the discrimination of different neurons. Stochastic multicolour labelling strategies, such as Brainbow, utilise a combination of 3 fluorescent proteins (XFPs) to create different colour hues. Allowing for the reconstruction of densely labelled circuits. However, these tools only produce ~20 colour hues, which is not enough to reconstruct neuronal circuits at sufficient density. Moreover, manual circuit tracing based on the colour hue is a rate limiting step in this strategy.
We aimed to solve these issues by increasing the number of colour hues available, then use machine learning to automatically reconstruct neurons based on their colour hue alone. Firstly, we increased the number of colour hues by stochastically expressing a combination of 7 different fluorescent proteins, then separating the spectral overlap through linear unmixing. Our modelling suggests that this can generate ~1,200 different colour hues. Secondly, as our eyes are limited to trichromatic vision, we developed a pipeline to automatically recognise the combination of >3 colours. This pipeline includes a newly developed unsupervised clustering algorithm, named the “Euclidean Crawler”, which classifies data points in N-dimensional space purely based on the threshold Euclidean distance. It holds an advantage over other distance-based clustering algorithms as we do not need to specify the number of clusters (like K-means), nor the density of clusters (like mean-shift clustering).
As proof of concept, first, we successfully reconstructed densely labelled layer 2/3 neurons in S1 (~300 neurons). Secondly, we automatically reconstructed long range (>2 mm) axonal terminals of mitral and tufted cell axons in the olfactory cortex. Finally, we used our automatic circuit reconstruction pipeline to register neighbouring brain sections: this was done by identifying neurites that go between sections by their colour hue, then performing piece-wise linear mapping for registration. Thus, super multi-colour labelling is a powerful tool for highly multiplexed circuit reconstruction on the mesoscopic scale.
Join in person or through zoom online
Webinar ID
670 8203 9447
Webinar Passcode
359683
