Fifty shades of spines

Russian neurobiologists from Peter the Great St. Petersburg Polytechnic University have created computer software that can automatically analyze and classify the shape of dendritic spines. The program is based on machine learning techniques. At the same time the primary outcome turned out to be the fact that, apparently, it’s time to refrain from the traditional division of spines into groups: in reality it’s much more complicated. Scientists have told about their achievements at the Volga Neuroscience conference. The results have been published in Scientific Reports.

The dendritic spines of neurons is the site of contact between dendrite and the axon, one part of the synapses. Their form is related to its function, but it’s only partially known in what exact way. Historically they were divided into groups according to the structure: fungiform — shaped like a mushroom; thin — shaped like the previous, but a bit smaller; and hempen, that look like a tree stump. The scientists can’t visually separate them when dealing with numerous spines, but it’s extremely important to analyze the information because spiny pathologies may possibly be underlying some neuropsychiatric and neurodegenerative diseases.

«The very idea of creating software arose from the needs of the laboratory. And all those who were involved in the analysis of shapes os spines synapses agreed on the fact that we lack such algorithm for work, when discussing the topic with colleagues at conferences. What we had at hand was a program that classifies spines and makes multiple mistakes that had to be corrected manually, further increasing the human factor («I am an artist, I see so»). Moreover, it was impossible to export parameter values depicting the form (Neuronstudio, Rodriguez).

After that the program was removed from free access because its most part was incorporated into  commercial program called Neurolucida. We have engaged as many as eight experts into the analysis, but it wasn’t possible to reach an agreement regarding the type of shape of more than 20% of the spines. The accuracy of experts itself also left much to be desired — on average it was 77% on consensus. Nevertheless we have succeeded in educating the classifier to cope with this task at an expert level, and that can significantly save time of data analysis if the researcher had adopted such an approach» —  explains Elena Pchitskaya, the first author of the study, a research associate of the Laboratory of Molecular neurodegeneration at Institute of Biomedical Systems and Biotechnology SPbPU.

According to the scientists, both intravital microscopy and other methods are undermining the traditional classification of spines nowadays (mushroom-shaped — thin — hempcrete), there is mounting evidence to suggest that the spines are constituted on a continuum of forms, gradually transforming from one class into another, and everything is far more complex than written textbooks.

«The one script that didn’t classify, but clustered them (i.e. searched for similar group forms based on specific data) has defined five spine groups. In addition to that, we have suggested a new way of describing the form, something much more interesting than just the resulting parameter in the form of a number, for example, length of volume. A large number of chords have been conducted inside the spine’s volume and a bar chart was drawn afterwards. It turned out that just this one parameter is enough to efficiently cluster them. Our proof-of-concept has been successful and now we are carrying on these studies on finding more complicated and sensitive descriptors of forms that would correspond to the level of complexity of our object» — Pchitskaya continues with the story.

The authors have exported аll of the parameters, shared the dataset and made а comprehensive tutorial on how to use it (the program code and manual may be downloaded here).

«We have approached the software as if it was a product from scientists to scientists, so to say, considering the fact that oftentimes it’s impossible to get the source code for free science software, and if it is available, it only works with data which it was written for, so only the group that created it can fully use this program» — she added.

If you look at the essence of the project from the perspective of artificial intelligence, the point  was to substitute the empirical and manual classification of spines with the one based on data. The AI model may be trained to take significantly more detail into account than a human. Also the idea of the researchers was to thereby discover new types of the spines and also confirm or reject the existing classification.

«The most difficult part of our work was representing the 3-D form of spines in a way that eventually showed a more sustainable classification. We already have several segmentation tools of the spine, a couple of coding algorithms for the 3-D geometry of the spine and several AI algorithms. Such an approach gives us game-changing results.

In the long term, we must create a new open source software for scientists that would not only replace the paid foreign analogs, but give the scientists from all over the country the opportunity to use contemporary AI methods even when there are no funds to create and support search software.

The material was prepared with the financial support of Ministry of Education of Russia within the federal project «Popularization of science and technology».