C2 Domains

The initial model building was done in EM maps with best resolution for the 30S, mRNA or tRNA and for specific IFs. Then this model was used as a reference for model building in EM maps with lower resolution. The head and the body of the atomic model of 30S of T. thermophilus (PDB: 1HR0) (Carter et al., 2001 (link)) were placed independently into density of each class by rigid-body fitting using Chimera (Pettersen et al., 2004 (link)). Next, the crystal structures of IF1 (PDB: 1HR0) (Carter et al., 2001 (link)), the N and C-terminal domains of Geobacillus stearothermophilus IF3 (PDB: 1TIF and PDB: 1TIG) (Biou et al., 1995 (link)), T. thermophilus IF2 (PDBs: 3J4J and PDB: 4KJZ) (Simonetti et al., 2013 (link)) (Eiler et al., 2013 (link)) and tRNA (PDB: 4WZO) (Rozov et al., 2015 (link)) were docked into density using Chimera. Then, each chain of the model (including ribosomal proteins, rRNA segments, protein factors and tRNA and mRNA) was rigid-body fitted in Coot (Emsley et al., 2010 (link)) and further model building was also done in Coot v0.8.
The availability of crystal structures of N and C-terminal domains of IF3 (PDB: 1TIF and PDB: 1TIG) helped in the model building almost complete IF3 (residue 3 to 170) with the helical linker joining the two domains. Special attention was devoted toward modeling of domain C2 of IF2. Rigid body fitting the NMR structure of C2 of IF2 from Bacillus stearothermophilus (PDB: 1D1N) (Meunier et al., 2000 (link)) was carried out into the density. Orientation of the C2 domain agrees with previous biochemical data (Guenneugues et al., 2000 (link)). It is also in agreement with EM data of its eukaryotic homolog eIF5B (Yamamoto et al., 2014 (link)) and with the orientation of C2 resulting from the superimposition on domain C1 of the crystal structure of its archaeal counterpart (PDB: 1G7T) (Meunier et al., 2000 (link)). In PIC-III, CCA of tRNA and fMet were taken from (PDB: 1ZO1) (Sprink et al., 2016 (link)). Refinement for all but PICs-I and II was carried out in Refmac v5.8 optimized for electron microscopy (Brown et al., 2015 (link)), using external restraints generated by ProSMART and LIBG (Brown et al., 2015 (link)). Average FSC was monitored during refinement. Final model was validated using MolProbity (Chen et al., 2010 (link)). Cross-validation against overfitting was calculated as previously described (Brown et al., 2015 (link), Amunts et al., 2014 (link)). Refinement statistics are given in Table S1. All figures were generated using PyMOL (DeLano, 2006 ) or Chimera.

All simulations and analysis were carried out in MATLAB 2020b, The MathWorks, Inc. The predicted responses of the release of inhibition models and the benchmark allosteric model to input [Ca²⁺](t) traces were generated by Monte Carlo estimation from multiple stochastic simulations. Specifically, we used the direct Gillespie algorithm^{49 (link)} which proceeds by iteratively generating a randomised time at which the system next changes its state and then randomly selecting the identity of the new state. The Ca²⁺ binding and SV fusion dynamics of the allosteric model were described by a six-state kinetic scheme with a single occupied state which is updated at each step of the algorithm. The release of inhibition models consisted of either six (in the case of Syt1^P) or twelve (in the cases of Syt1^P/Syt1^T and Syt1^P/Syt7^T) four-state kinetic schemes (Fig. 1b), one for each Syt C2 domain. Rather than updating a unique state in the resultant macroscopic Markov chain which had either 4⁶ or 4¹² states, we monitored each synaptotagmin C2 domain concurrently and updated one of their states according to the algorithm. We assumed that at the start of each simulation both the Ca²⁺ sensor in the benchmark allosteric model and all SNARE-associated synaptotagmin C2 domains in the release of inhibition models were in the Ca²⁺ unbound state.
In simulations shown in Figs. 2–4 we did not include the mechanism for vesicle replenishment and individual stochastic simulations terminated when vesicle fusion occurred. In simulations describing vesicular release in response to a 10 × 100 Hz AP train we included a mechanism for SV replenishment. Upon vesicle fusion the release site remained unoccupied for a fixed refractory time of 2.5 ms after which a SV was replenished in the initial state with the rate of $k_{\mathrm{rep}}=$  0.02 ms⁻¹, as was estimated in our previous work^{28 (link)}.
For each scenario, the collection of stochastic simulations yielded a set of times at which SV fusion occurred. We used the cumulative count of these vesicle fusion times, normalised to the total number of stochastic simulations, as an estimate for the expected cumulative number of vesicles exocytosed at a single release site by time $t$ ( $n_T\left(t\right)$ ). In the absence of vesicle replenishment $n_T\left(t\right)$ corresponds to the cumulative vesicular release probability $p_v\left(t\right)$ . In production, $n_T\left(t\right)$ was calculated by gathering release event times into a histogram with an adaptive bin width to capture the features of release kinetics at different temporal scales. The release rate was estimated as $\frac{d{n}_T\left(t\right)}{dt}$ with a moving average smoothing applied to limit the sensitivity of peaks to stochastic variation.