Methodological Improvements

In this Section, our work on method development is described. This work enabled the studies on protein and lipid diffusion in membranes crowded by proteins. Publications I and II consider this work.

5.1.1 Membrane Protein Insertion

As described in Chapter 4, prior to an MD simulation, an initial configuration of the simulated components needs to be set up. Unfortunately, generating such configurations for membrane protein simulations has traditionally been tedious, as lipid chains show a plethora of conformations that need to be fitted to complex protein surfaces. Luckily, many tools have been developed for this purpose [228, 229, 230, 231, 232, 233]. These tools either embed a protein into an existing bilayer or generate the whole system from scratch. Unfortunately, all of them suffer from a few limitations. The tools that insert proteins into existing bilayers require the removal of lipids. With new lipidomics data fostering simulations of multi-component membranes, this might result in unwanted changes in their lipid compositions. Most tools are also compatible only with specific models and simulation software or require the installation of third-party software. Moreover, they are merely suitable for creating a single-protein configuration.

To tackle these issues, we developed a protocol for embedding proteins into pre-existing lipid bilayers. This approach is based on pushing a protein into a membrane from its side by applied pressure. Briefly, the simulation box around the bilayer is increased in the plane of the bilayer so that the protein can be placed next to the bilayer without any overlap of atoms. This configuration is visualized in

54 5. New Insights and Advancements Provided by This Thesis Figs. 4.1A and 4.1B for AA and CG models, respectively. Next, a set of restraints are applied to maintain the integrity of the bilayer and the structure of the protein.

Finally, significant pressure is applied in the plane of the membrane, which causes the fluid membrane to encapsulate the protein. Simultaneously, the lateral size of the box shrinks as the initial vacuum around the membrane disappears. After a brief simulation of approximately one nanosecond, the membrane is again continuous, yet now hosts the protein. This process is visualized in Fig. 5.1.

Figure 5.1 Snapshots (top view) of the embedding process with an AA system. The protein is shown in yellow and the POPC lipids in green. Water and ions are omitted for clarity.

Periodic images of the system are displayed to visualize the process better. The simulation box (unit cell) is highlighted in blue.

This new protocol avoids most of the pitfalls of the earlier methods. It is universal to all models and independent of simulation software. Instead, any molecular dynamics engine can be employed as long as it implements the standard functions required – a barostat and position restraints. Most importantly, our approach does not require the removal of any lipids. Unfortunately, it does not come without limitations. The approach is not compatible with doughnut-shaped proteins that enclose lipids within their structure, and it is limited to work only with planar geometries. Finally, it is worth emphasizing that since making our protocol accessible for the community, two other tools, CHARMM-GUI [213, 214, 215] andinsane [195], have emerged and rendered our approach obsolete in some cases.

5.1.2 Adjustment of Protein–Protein Interactions

During the research presented in this Thesis, it became apparent that the Martini model is not well parametrized for protein–protein interactions. This issue manifests itself in the excessive aggregation of membrane proteins under crowded conditions [189, 234, 235] (see Publication II for more examples). Without exception, simulations

5.1. Methodological Improvements 55 performed using the Martini model result in a superaggregate that contains all the simulated proteins clustered together in a nonspecific manner. What is more, the estimated dimerization free energies of both trans-membrane (TM) peptides and proteins show incredibly high values of up to ⇠160 kJ/mol [235]. Such energies refer to an entirely irreversible association, which is in disagreement with experimental observations and our intuitive picture of biological processes.

To systematically evaluate the level of protein–protein interactions in Martini, we performed umbrella sampling simulations on two TM domain dimers of known structures and dimerization free energies (see Section 4.3). An example conformation is shown in Fig. 4.2A. Additionally, the dimerization free energies of both dimers had been estimated in earlier Martini simulations [236, 237], yet using different simulation parameters and membrane compositions. We mimicked the experimental setups as closely as possible in our simulations. Additionally, we simulated the spontaneous formation of five dimers of TM domains, for which NMR structures are available.

These dimers are shown in the CG scheme in Fig. 4.2B.

As expected and as shown in Table. 5.1, we find that the dimerization free energies of the TM domains extracted from umbrella sampling simulations using Martini are of the order of ⇠35–40 kJ/mol. In contrast, the Förster resonance energy transfer (FRET) experiments provide values of around⇠10–15 kJ/mol (see Table 5.1).

Furthermore, as demonstrated in Table 5.1, using the polarizable Martini model [238]

does not improve this agreement. Our analysis also highlights that the peptides form higher-order oligomers or even a superaggregate in a simulation containing multiple membrane-embedded peptides. This result is in disagreement with the FRET studies that report the lack of such higher-order oligomers [220, 221].

Table 5.1 EphA1 and ErbB1 TM domain dimerization free energies (in kJ/mol).

EphA1 ErbB1

FRET experiments [220, 221] 15.4±0.5 10.5±0.4 Standard Martini 29.9±1.0 39.5±1.0

Polarizable Martini 33.5±1.0 –

Scaled (Publication II) 15.2±1.0 15.3±0.3 Previous studies [236, 237] 60±2 38±3

Considering next the structures of the spontaneously formed dimers, the comparison with NMR data (not shown here) unfortunately reveals that Martini is unable to

56 5. New Insights and Advancements Provided by This Thesis correctly predict any of the dimer structures. Moreover, the 10 replicas simulated for each dimer provide a very diverse set of structures, indicating that any protein–protein contact can lead to irreversible protein aggregation.

Following the approach of Stark et al. applied to water-soluble proteins [197], we scaled down the Lennard-Jones (LJ) interactions between the protein beads by various amounts and repeated the simulations. We applied this scaling to either all protein beads or only to those that resided mostly in water. The former approach turns out to be more successful, and a subtle 10 % decrease in the LJ energy parameter (✏) brings the dimerization free energies to the same ballpark with experiments (see Table 5.1). Also, the oligomer sizes in multi-peptide simulations decrease, and the peptides no longer form a single superaggregate, in more reasonable agreement with experiments.

Unfortunately, the quality of the structures of the spontaneously formed dimers of TM domain peptides does not improve in this process. Instead, we still observe dimerization interfaces and dimer conformations that drastically differ from those resolved by NMR.

Despite the issues that prevent the general adaptation of the scaled Martini protein model, certain types of studies can significantly benefit from it. These include research on membrane dynamics in crowded environments, where excessive protein–

protein interactions lead to abnormal protein aggregation and subsequently to strong confinement effects by these aggregates. In such cases, the scaling down of the protein–

protein interactions can provide more realistic transient aggregation behavior. This is the case in Publication IV, where the proposed scaling approach is applied to crowded membranes.