Construction of the dynamic graphical model

The individual protein views were constructed using custom‐written plug‐ins and scripts in the 3D software Autodesk Maya (Autodesk Inc., San Rafael, CA). Protein structure information was derived from the UniProt database. The same individual protein view models were used as in Wilhelm et al (2014), and the references used are presented in the particular paper. When available, we used protein database (PDB) coordinates in order to reconstruct proteins. If not available, we relied on structure information provided by a number of prediction servers. We used the following types of information: secondary structure information (http://bioinf.cs.ucl.ac.uk/psipred/); disorder calculations (http://mbs.cbrc.jp/poodle/poodle-s.html; http://mbs.cbrc.jp/poodle/poodle-w.html; alignment (http://web.expasy.org/sim/); predictions of coiled coil regions (http://toolkit.tuebingen.mpg.de/pcoils; http://mbs.cbrc.jp/poodle/poodle-l.html), information on transmembrane domains (http://www.ch.embnet.org/software/TMPRED_form.html); information on glycosylation domains (http://www.glycosciences.de/modeling/glyprot/php/main.php); domain identification (http://smart.embl-heidelberg.de/index2.cgi); and the presence of homologue proteins (http://web.expasy.org/blast/).

As mentioned in the main text, simulated different particle motion behaviors, with different movement speeds, in order to find the behaviors that most closely reproduced the FRAP results. We transformed the particle motion in artificial FRAP movies by applying fluorophore point‐spread‐functions onto the tracks. We then compared the results to the original FRAP data, in order to find the models that best reproduced the biological time constants in the axons and in the synapses. We then placed the protein structures in the 3D space of the model synapse, relying on the same movement tracks we used in the rest of this work. For each protein type, we used for the graphical models a number of protein tracks from the model that had reproduced best the FRAP behavior of the respective protein. The number of tracks was chosen as equal to the expected protein copy number for the respective protein (Richter et al, 2018). The protein views were then placed on every pixel of the tracks and were also allowed to turn around their own axis. Synaptic vesicles were presented according to their known composition (Takamori et al, 2006), relying on previously generated models (Wilhelm et al, 2014). The vesicles are typically shown in grayscale. To avoid confusing the viewer, no vesicle motion is shown (or only a rotational/vibrational motion); this is a reasonable procedure, since the net diffusive vesicle motion is expected to be extremely limited for the time interval we show.