Coding

Part:BBa_K3187028

Designed by: iGEM TU_Darmstadt 2019   Group: iGEM19_TU_Darmstadt   (2019-10-12)
Revision as of 03:32, 21 October 2019 by JonathanFu (Talk | contribs)

ext install esbenp.prettier-vscode

Sortase A7M (Ca2+-independent variant)

Profile

Name Sortase A7M
Base pairs 450
Molecular weight 17.85 kDa
Origin Staphylococcus aureus, synthetic
Properties Ca2+-independent, transpeptidase, linking sorting motif LPXTG to poly-glycine Tag

Usage and Biology

generic filler text

Transpeptidase: Sortase

generic filler text

Reaction

generic filler text

Sortase variants

generic filler text

Sortase A7M

generic filler text

Methods

generic filler text

Cloning

generic filler text

Expression and purification

generic filler text

SDS-Page

generic filler text

Flourescence Resonance Energy Transfer (FRET)

generic filler text

Mass Spectrometry

generic filler text

Results

generic filler text

Characterization of Sortase A7M (and comparison to Sortase A5M)

generic filler text

How do we measure if our purified sortases are active?

generic filler text

how do we measure sortase reaction kinetics

generic filler text

Development of a new FRET pair

generic filler text

Why are enzyme-substrate ratio and duration important parameters of the sortase reaction?

generic filler text

Who wins - Sortase A7M or Sortase A5M

generic filler text

What about other substrates?

generic filler text

Primary Amines

generic filler text

Yield

generic filler text

Is Sortase A7M able to attach cargo to P22 coat protein?

generic filler text

Does methionine affect Sortase linking?

generic filler text

Are there other Sortases that might me useful?

generic filler text

Modeling

Introduction

In synthetic biology, theoretical models are often used to gain insights, predict and improve experiments. In our project we are modifying Virus-like particles (VLPs) by attaching proteins to the surface of the P22 capsid through a linker. The linking is catalyzed using the enzyme Sortase A7M, which is a calcium independent mutant of the wild type Sortase A from Staphylococcus aureus. We performed modeling to predict the unknown structure of the Sortase A7M, to improve the linker between proteins and therefore optimizing the modification efficiency of our platform.
Two different modeling approaches were used to determine the structure of Sortase A7M. We compared machine learning approaches to traditional comparative, Monte-Carlo based modeling methods. The results were evaluated using an energy-scoring function and molecular dynamics (MD) simulations. The most promising Sortase A7M structures were used to perform a docking simulation to screen for optimal linkers.

Structure determination

In silico modeling and simulation of proteins requires a 3D structure, which can be obtained from the RCSB Protein Data Bank. However, if no 3D structures are annotated, as it is the case with sortase A7M, the structure has to be determined by other means. The structure prediction of sortase A7M was done using two different approaches.

RosettaCM

Background

In our second approach we used the RosettaCommons comparative modeling (RosettaCM), which is based on homology modeling. Homology modeling is a protein modeling method, which requires one or more template structures as base the protein to be modeled on. The protein sequences are aligned with the sequence of the target protein. Unaligned sections are modeled using fragment or protein libraries, which leads to creating protein structures based on different sequence homologues of the protein of interest. Ab-initio or de novo modeling on the other hand attempts to find protein structures solely based on physicochemical principles applied to the primary sequence, which can be compared to the refolding of a denaturated protein.

RosettaCM combines ab-initio modeling with homology modeling. The homologus structures for which a resolved 3D structure with sufficiently similar sequence exists are generated using homology modeling. Afterwards the unaligned sequences are modeled de novo. By combining the two methods RosettaCM represents a precise and resource efficient tool for protein structure prediction. Rosetta applications rely on the Monte-Carlo Optimization, which is a probabilistic approach to finding a local minimum in the energy landscape of protein conformations. The underlying equation serving as the fundament of the statistical Monte-Carlo method is the Metropolis acceptance criterion:


where kB is the Boltzmann constant, ΔE the difference in energy of the two states and T the temperature. The term kBT can also be written as a single factor β.

During the statistical protein folding based on the Monte-Carlo method, the initial structure is changed by small random perturbations of the atom locations. Whether the structure is accepted or not is decided by the Metropolis acceptance criterion. If ΔE < 0, the structure is accepted, otherwise the newly proposed structure is accepted with probability p as described in the Metropolis acceptance criterion.

Procedure

The RosettaCM protocol requires evolutionary related structures and sequences, as well as fragment files of the target structure. The fragment files serve as a structure template for the proteins and they consist of peptide fragments of sizes 3 and 9. We gathered five evolutionary related structures from the RCBS PDB with the accession numbers:

  • 1ija
  • 1itw
  • 1itp
  • 1ito
  • 2mlm

The five RCBS entries represent different structures of sortases from Staphylococcus aureus. Fragment files can be created with the Robetta online server or with the Rosetta FragmentPicker application.

The RosettaCM procedure is best described in the following steps:

  1. sequence and structural alignment of templates
  2. fragment insertion in unaligned sections
  3. replacement of random segment with segment from a different template structure
  4. energy minimization
  5. all-atom optimization

The alignment can be performed with various tools. We used MAFFT to generate the multiple sequence alignments. Prior to using the alignments as an input, they were converted to the grishin alignment format as RosettaCM requires the alignments to be in said format. The minimization is performed using the Rosetta controid energy function. For the centroid function to be applied, the protein is converted to the centroid representation. A protein in centroid representation consists of the backbone atoms N, Cα;, OCarbonyl and an atom of varying size representing the side chain. The advantage of using the centroid representation is that the energy landscape can be traversed easier due to the smoother nature of the centroid energy landscape. Finally the generated structure undergoes a second minimization in an all-atom model by means of Monte-Carlo optimization. This is similar to the energy minimization but without the amino acids being represented as centroids of their functional groups. Structures computed through all-atom optimizations can reach atomic resolutions {{Quelle rosetta paper}} which is crucial for a model meant to be used to estimate atomic interactions.

Results

The run yielded 15,000 structures which have been compared using the Rosetta scoring functions (talaris2013). From the 15,000 structures generated, we inspected the ten best scoring structures.

As can be seen in figure 5, the most prominent differences can be found in the regions close to the N- and C-terminus. As fluctuations in those regions are not untypical, we decided to use the best scoring structure, candidate S_14771 (figure 6), as the input for the simulations to follow.

Figure x : The structural alignment of the ten best scoring sortase structures displaying minor differences with the exception of the C- and N-terminal regions. N- and C-terminal regions tend to show strong fluctuations, thus it is unsurprising to find the terminal regions to be unaligned

Figure x : Sortase A7M candidate S_14771 created through RosettaCM.

In order to evaluate the secondary structure of the Sortase A7M candidate S_14771 a Ramachandran plot has been created and compared to the five sortases used as input for the comparitive modeling. Comparisons were also drawn with the Sortase predicted by Deep Learning as well as a database of randomly sampled proteins. Ramachandran plots of dihedral angles (fig x) can be a first indicator whether the structures computed are valid.

Figure x : Caption?

Figure x : Caption?

Figure x : Caption?

Figure 5: The comparison of the ramachandran plot of structure S_14771 and the ramachandran plot found on Protopedia suggests that secondary structures are present. Hence the structure appears to contain α-helices, β-sheets and a small amount of lefthanded α-helices.

Conclusion

We used machine learning methods, as well as monte-carlo simulations to determine the structure of the mutated transpeptidase Sortase A7M. The machine learning approach using AlQuarishi's Deep Neural Network yielded a structure which seemed to not have any secondary structures. To exclude the possibility of an error in the PyMOL visualization software by Schroedinger, a Ramachandran plot (figure xyz) was created. The plot shows that no typical secondary structures are present which is a strong indicator of a failed approach to determine a structure. The approach, using Rosetta Comparative Modeling, yielded 15,000 structures scored with the talaris2013 scoring function. The ten best structures were aligned and exhibited almost identical secondary structures (figure xzy). The greatest structural differences are present in the N- and C-terminal regions. Since terminal regions tend to fluctuate more strongly than non-terminal segments of the protein, we deemed those fluctuations non-relevant for the proteins functionality.
Being the best scoring candidate, structure S_14771 was analyzed structurally using a Ramachandran plot (figure xyz). The plot shows all the relevant and typical structures sortases exhibits and serves as an indicator for a successful structure prediction.
In the steps to follow, a molecular dynamics (MD) simulation will be performed on both structures. Even though structure CASP12 does not seem to be a valid structure, refolding processes during a MD simulation might lead to a relaxation of the protein and allow for a promising prediction of the sortase A7M structure.

Molecular dynamics

Introduction

The structure predictions made so far were based on statistical methods with physical constraints. The Deep Learning algorithm uses a neural network trained to find a function associating the amino acid sequence and the final 3D positions of the atoms within the protein. On the other hand, predictions were made with Rosetta using the Monte Carlo Method. Here random movement of individual atoms occurs, and the energy is estimated after each step.

Even though both methods use physical constraints to find plausible protein structures, neither of them actually simulates the behavior of these molecules within a physical force field. Moreover, both methods do not necessarily output fully relaxed protein structures and simulate water implicitly by preferring hydrophilic parts of the proteins to be on the outside. Thus, we conducted a molecular dynamics (MD) simulation to verify the plausibility of our protein structure and allow equilibration. The molecular dynamics simulation provides the opportunity to simulate water as discrete molecules, creating a solvated protein. This step is crucial to validate the structures, as the interaction with water is one of the primary mechamism for protein folding. Since neither candidate CASP12 nor S_14771 have been modeled with explicit water an according MD simulation is imperative, to verify the correctness of the candidates conformation. This of course is much more expensive in terms of computational ressources. As the protein has to be placed in a simulation box and said box is filled with water molecules. This is called solvation and is visualized for candidate S_14771 in figure eeeeee.

Figure x : Sortase A7M in a force field surrounded by discrete water molecules. Image was made with gmxSolvate.

We used GROMACS (GROningen MAchine for Chemical Simulations) as the tool for our molecular dynamic simulations. GROMACS solves Newtons equations of motion for individual atoms [1] . While this classical simulation is much more accurate than predictions made by the other methods, approximations are used nonetheless: Forces are cut after a certain radius and the system size is quite small. [1] Additionally, atoms are assumed to be classical particles, which is not the case, as quantum mechanics plays a role in particle-particle interactions. Still, this simulation is very computationally expensive. Therefore, only time periods less than one second could be simulated.

Methods

To perform the molecular dynamics simulations we mostly followed the GROMACS Lysosome tutorial as it serves our purpose perfectly. We created our simulation box to be of dodecahedral shape and a 0.7 nm distance of the solute to the box borders. We used periodic boundry conditions and a Na+ Cl- concentration of 0.012 mol/L. The main difference of our approach was that we used the CHARMM36 force field instead of the OPLS-AA/L force field and have adjusted our molecular dynamics parameters accordingly. The simulation was performed on a NVIDIA GTX 760 graphics card allowing us to simulate approximately 1 ns per hour.

To analyse the MD simulation we used the Python programming language and the Biotite package as well as GROMACS analysis tools as covar and anaeig. The first analyses are a root-mean-square deviation (RMSD), a root-mean-square fluctuation (RMSF) and a gyration radius analysis. RMSD calculations have been described in the structure prediction section. To compute the RMSF the movement distance of each residue is computed as a root-mean-square over time as:

Figure x : caption

where v(t)i is the position of atom i at time t. The radius of gyration is The final analysis performed on the MD simulation is called Principle Component Analysis (PCA). By applying PCA to a protein it is possible to gain insights into the relevant vibrational motions and thereby the physical mechanism of the protein .

Results

The first possible indicators of a stable protein structure are converging root-mean-square deviation (RMSD), small root-mean-square fluctuation (RMSF) values as well as converging radii of gyration. Using the Python software package and the module Biotite we calculated these quantities and plotted the results for both candidate S_14771 and candidate CASP12.

Figure x : The RMSD is one of three main indicators of a stable protein structure of the MD simulation of S_14771 over the period of 200,000 ps. As time progressed the RMSD increased with a smaller slope. The value stabilizes at a time of 110,000 ps and fluctuated around the value of 6 Å.

Figure x : At t = 40,000 ps already the RMSD has arived at a stable value, while at the same time the gyration (fig x) radius decreases over time continuously. This information suggests the protein might be folding and potentially develpoing secondary structures not present previously.

Figure x : The prominent fluctuations of the residues from ranges 105 to 115 might indicate a binding site or another form of functional structure. The radius of gyration, just as the RMSD fig xyz, stabilizes around a simulation time of of 110,000 ps and converges towards a value of 16.7 Å.

Figure x : As from t = 40,000 ps the radius of gyration decreases constantly. At the end of the simulation the gyration radius reaches a value of 17 Å. This behavior indicates folding of the protein structure.

Figure x : The fluctuations (RMSF) of most residues appear insignificant compared to the first, the last residues and the residues close to residue 110 . Typically the N- and C-terminus tend to fluctuate more intensively due to the lack of stabilizing structures. The prominent fluctuations in the range of residue 105 to 115 can indicate a binding site or another form of functional structure.

Figure x : The prominent fluctuations of the residues from ranges 105 to 115 might indicate a binding site or another form of functional structure. The radius of gyration, just as the RMSD fig xyz, stabilizes around a simulation time of of 110,000 ps and converges towards a value of 16.7 Å.

Typical RMSDs and radii of gyration converge towards a value dependent on the size of the protein. Convergence of those quantities can be interpreted as a stable state of the protein structure. As it can be seen in Figures x and y both the RMSD and the radius of gyration stabilize at the same time as the simulation reaches 110,000 ps (110 ns), suggesting a now stabilized structure of candidate S_14771 solvated in water. Another indicator of a functional protein is the RMSF. Instead of being averaged over all atoms, the RMSF is averaged over time with respect to each amino acid. It provides insights in both protein stability and functionality. Fig xzf reveals the RMSF of residues 105 to 115 to be significantly higher than that of other residues. This hints at the presence of a functional unit along these residues. As commented on in the section describing our structure prediction approaches, the N- and C-terminal regions tend to fluctuate more strongly as a result of the absence of stabilizing structures.

RMSD and gyration of radius calculations of candidate CASP12 (figures x and y) provide evidence of folding. However, the RMSF values show values significantly higher, an effect possibly caused by instability or refolding. Nevertheless, the strongest fluctuations, disregarding the terminal regions, can be seen in the region of residue 105 to 115. This insight consolidates the theory that residues 105 to 115 might be a part of a functional unit.

We were unsure whether candidate CASP12 can be considered a plausible structure and how to interpret the findings concerning the prominent fluctuations. Therefore, we decided to perform a Principle Component Analysis.

Principle component analysis

To analyze our system further Principle Component Analysis (PCA) was performed using GROMACS.

Animation 4: A Principle Component Analysis of a fast (blue) and a slow (red) mode showing the most prominent movements of the Cα-chain of candidate S_14771. Both modes show movement of the β6/β7 loop consisting of residues 105 to 115 towards the active site . Thus we can assume that the closing β6/β7 loop is involved in the reaction mechanism.

The results from the Principle Component Analysis of candidate S_14771 (animation xy) show a movement of the residues 105 to 115 towards the active site, supporting our theory that residues 105 to 115 are important for the reaction mechanism. Since the slow mode (red), which shows the most relevant movement of the sortase, moves further towards the active site, it is possible that the β6/β7 loop either closes the binding site of the ligand peptides or even transports one peptide towards the other.

Animation xyz shows the results of the Principle Component Analysis of candidate CASP12. As the RMSF calculations suggested (fig xyz), the whole protein seems to be moving randomly with no directed movement. In addition the active site amino acids are spread across the protein confirming our assumption that the protein is not in a stable or plausible conformation.

Conclusion

We gained evidence that at least on of our Sortase A7M models is a valid and stable candidate by performing various methods to analyse the structural stability and validity of our two Sortase A7M candidates. The candidate S_14771 that was generated using RosettaCM appears to be a fitting candidate not only due to successful analyses, but also since the residues of the active site are close enough to each other to catalyze a ligation reaction. Our model created through deep learning excelled only in terms of RMSD and gyration radius calculations. Not only the RMSF and Principle Component Analysis but also the conformation of the active site have proven candidate CASP12 to be of no use for further calculations as it does not portray a valid conformation of Sortase A7M.

Docking

Now that the binding site of the Sortase had been found, the peptide ligand needed to be inserted into the binding site to create a peptide-protein complex. The procedure of choice for the introduction of a ligand into the binding site of a protein is called docking. In the following sections, we will present the protocol and methods we used as well as the results they yielded.

Background

Enzymes are one of the most relevant macromolecules in biology. Their functionality is determined through the way they interact with their ligands. Although enzymes are highly specific concerning the ligands they interact with, similar compounds can often bind to the same enzyme albeit with different affinity. To determine the best possible binding conformation of the protein-ligand complex, we use FlexPepDock, an algorithm provided by the the RosettaCommons software package.

Procedure

The ab-initio FlexPepDock protocol consists of multiple steps and is documented on the RosettaCommons online documentation. We modified the protocol as the one provided did not work with our approach. The modified protocol has the following form:

  1. secondary structure determination
  2. complex creation
  3. FlexPepDock refinement

To determine the secondary structure of the peptide, fragment files (3- and 5-mers) had to be generated and a PSIPRED secondary structure prediction had to be performed. As the peptides had a sequence length less than 20 amino acids, we were not able to use the online services such as Robetta and the PSIPRED online service. Instead we used the Rosetta FragmentPicker application and the PSIPRED command line tool. The generated structures serve as the input for the refinement protocol.
The generation of the peptide-protein complex can be divided into three steps:

  • peptide creation
  • peptide relaxation
  • coarse complex creation

The peptide structure was created through ab-initio modeling. Initial creation of the peptide was followed by insertion of the peptide into the sortase binding site. This lead to a coarse model of the peptide sortase complex. Here we used insight gained from the molecular dynamics simulation to place the peptide close to the binding site.
In the final step the FlexPepDock refinement protocol is executed and 50,000 complex structures are generated. We used the inputs as described in {{fuhrman paper}}, written by the authors of the FlexPepDock documentation.
To get a better overview over our data we performed a clustering in python, using the scikit-learn package. We clustered the structures with respect to:

  • total score: the total score of the docking provided by the Rosetta scoring function
  • interface score: the sum of the energy of the residues in the interfacing region
  • reweighted score: a score calculated by double weighting the contribution of the residues in the interfacing region
  • root-mean-square deviation: the root-mean-square deviation of the peptides in relation to the structure with the highest score
  • peptide direction: the direction the peptide is facing

Here clustering is used to group the docking results and thereby descrease the samlple size. From the 50,000 results we picked the results with the 500 best total scores, the 500 best interface scores and the 500 best reweighted scores. As we aimed to create an unbiased set for clustering, the abscence of duplicates in the set was ensured. We decreased the sample size to 100 groups representing the best scoring structures from the three categories.

Results

For sequences MGGGGPPPPPP(M-polyG), GGGGPPPPPP(polyG) and PPPPPPLPETGG(LPETGG) 50,000 structures have been created and clustered. After the clustering the sample consisted of 100 structures of docked complexes.

Figure x : The three best scoring structures (total score, interface score, reweighted score) of the LPETGG-tag are shown. Only two results are visible as the best reweighted score candidate is identical to the best interface score candidate. The reacting section of the LPETGG-tag namely glycine is colored yellow as is the active site. The glycin of both ligand peptides is facing the active site.

Analysis of the scores has shown a similar score for all the three dockings. The best scoring results of the LPETGG docking show a tendency of the glycines to face the active site while also being in close proximity to the active site.

Figure x : The three best scoring structures (total score, interface score, reweighted score) of the poly-g peptide are shown. Only two results are visible as the best reweighted score candidate is identical to the best interface score candidate. Instead of facing the active site (yellow) the reacting glycines (yellow) appear to interact with the β6/β7 loop of the sortase.

Figure x : The three best scoring structures (total score, interface score, reweighted score) of the poly-g peptide are shown. Only two results are visible as the best reweighted score candidate is identical to the best interface score candidate. Concerning the M-poly-G peptide no uniform directional orientation can be observed. The structure with the best interface score (light blue) is oriendted towards the loop while the structure with the best total/reweighted (dark blue) is oriented towards the β-sheets.

Figure lpetgg shows the docking result of the LPETGG peptide to the sortase. The results shown are the best scoring structures of the clustering with respect to the total score, interface score and reweighted score. As the best scoring structure is the same for the total score and the reweighted score only two peptides are shown. This also applies to figures x and y. For both results the reacting glycin residues (yellow) are facing the active site. Additionally, the same residues are in close proximity to the active site.

The figures x ad y show the docking of the both polyG and M-polyG. While polyG results align well and seem to be interacting with the β6/β7 loop rather than with the active site, this does not seem to be the case for M-polyG. Instead of both structures interacting with the β6/β7 loop or active site one (best interaction score; dark blue) interacts with the β6/β7 loop and the other (best reweighted/total score; light blue-gray) appears to interact with the active site.

Figure x : The close up of the M-polyG peptide (best total/reweighted score) indicates an interaction of methionine with arginine139 and cysteine126.

Figure x : Methionine of the result with the best interface score interacted with the β6/β7 loop rather than the active site. Still the reactive glycine residues appear to be bound to the β6/β7 loop.

As can be seen in figure 16 visualizing the result of the the docking simulation total/reweighted score) suggests an interaction of methionine and two of the active sites namely arginine139 and cysteine126. Visualizing the result of the according docking simulation, as can be seen in figure 16, suggests an interaction between methionine and two active site residues, namely arginine139 and cysteine126. Figure 17 shows the interaction of M-polyG with the β6/β7 loop. The glycines still interact with the β6/β7 loop. Instead of binding above the β6/β7 loop, which is the case for polyG as illustrated in fig z, the interaction seems to be influenced by methionine. By interacting with the residues in the β-helix methionine could potentially hinder binding of glycine to the β6/β7 loop by partial immobilization of the peptide. Overall peptide binding and orientation is less uniform compared polyG without the leading methionine, which could be an indicator of lesser binding affinity of M-PolyG towards the β6/β7 loop.

Conclusion

To computationally investigate binding affinities of the polyG and M-polyG as well as the LPETGG tags we performed docking simulations using the Rosetta FlexPepDock application. We used a modified version of the recommended protocol as the modified version was easier to automate and served our purpose better than the standard protocol. From the calculated scores only, we could not see a difference in binding affinities. Thus, we inspected the best scoring structures regarding the total score, the interface score and the reweighted score using PyMOL. Since the best structures with respect to total score and reweighted score were the same for all simulations, only two structures have been inspected per run. A polyproline tag was appended to all the peptides to simulate the modification of the VLPs with a small peptide.

As expected, the results showed that for LPETGG, the glycines of both peptides oriented towards the active site. This is unsurprising as peptides with the sequence LPXTGG are known to be substrate of the Sortase. It was more surprising to see the polyG tag oriented away from the active site since polyG also is a known substrate of the sortase. Both polyG peptides were facing the β6/β7 loop (residues 105 to 115) uniformly and appeared to be interacting with it. The M-polyG peptides did not show a uniform orientation or interaction scheme. On one hand the visualization of the best result concerning the total and reweighted score has shown interaction of methionine with the cysteine126 and arginine139, two residues of the active site. On the other hand, the visualization of the best result with respect to the interface score shows the M-polyG facing the mobile β6/β7 loop. In contrast to the polyG peptide the lacking the methionine, the M-polyG peptide is pulled down below the β6/β7 loop by the methionine interacting with one of the β-sheets leading to the active site. This is not the case with the polgG results, which lie aligned in one plane with the β6/β7 loop.

Modeling Conclusion

generic filler text

[edit]
Categories
//awards/basic_part/winner
Parameters
None