Difference between revisions of "Team:Bielefeld-CeBiTec/Model"

Line 201: Line 201:
 
<li> <a href="https://www.rosettacommons.org/demos/latest/tutorials/prepare_ligand/prepare_ligand_tutorial"> </a>“.params”-file: </br>
 
<li> <a href="https://www.rosettacommons.org/demos/latest/tutorials/prepare_ligand/prepare_ligand_tutorial"> </a>“.params”-file: </br>
 
A conformer ensemble has to be generated using information about the ligand, as the non-canonical amino acids are not generally available in databases like <a href=""http://www.rcsb.org/pdb/home/home.do>PDB</a>, making it necessary to build them manually  
 
A conformer ensemble has to be generated using information about the ligand, as the non-canonical amino acids are not generally available in databases like <a href=""http://www.rcsb.org/pdb/home/home.do>PDB</a>, making it necessary to build them manually  
−
using tools like <a href="https://pymol.org/2/">pymol</a><a href="https://avogadro.cc/">Avogadro</a> or <a href="http://www.cambridgesoft.com/software/overview.aspx">Chemdraw</a>. Using these tools,
+
using tools like <a href="https://pymol.org/2/">pymol</a>, <a href="https://avogadro.cc/">Avogadro</a> or <a href="http://www.cambridgesoft.com/software/overview.aspx">Chemdraw</a>. Using these tools,
 
files can be saved in the desired format. The ligand needs to be specified in the “.sdf”, “.mol” or “.mol2” file format. Such a  
 
files can be saved in the desired format. The ligand needs to be specified in the “.sdf”, “.mol” or “.mol2” file format. Such a  
 
file can be obtained automatically by converting the relevant information from a “.pdb” file, if available. This conversion process usually also involves augmenting the data with hydrogen atoms  
 
file can be obtained automatically by converting the relevant information from a “.pdb” file, if available. This conversion process usually also involves augmenting the data with hydrogen atoms  
Line 208: Line 208:
 
  stores all bonds between the individual atoms, including the binding angles and binding distances. Rosetta cannot generate the conformer ensemble by itself, so an additional tool is needed.  
 
  stores all bonds between the individual atoms, including the binding angles and binding distances. Rosetta cannot generate the conformer ensemble by itself, so an additional tool is needed.  
 
  Different tools are capable of creating the conformer ensemble automatically, but it is best to manually define constraints for the chi1, chi2 and backbone psi torsion angles that define the  
 
  Different tools are capable of creating the conformer ensemble automatically, but it is best to manually define constraints for the chi1, chi2 and backbone psi torsion angles that define the  
−
  orientation of the ligand in the binding pocket. For this, we know of three tools: The first is OpenEye Omega, but the full license is very costly and the free version is hard to obtain.  
+
  orientation of the ligand in the binding pocket. For this, we know of three tools: The first is <a href="https://www.eyesopen.com/omega">OpenEye Omega</a>, but the full license is very costly and the academic free version is hard to obtain.  
−
  The second tool is Accelrys Discovery Studio, but Accerlys does not provide a free license. The third tool is TINKER, which is free, but poorly documented and depends on a specific keyfile,  
+
  The second tool is <a href="http://accelrys.com/">Accelrys Discovery Studio</a>, but Accerlys does not provide a free license. The third tool is <a href="https://dasher.wustl.edu/tinker/">TINKER</a>, which is free, but poorly documented and depends on a specific keyfile,  
−
  which requires a high amount of chemical expertise to generate. Conformers might also be generated without constrains, for which different tools are available, in our case, we used ConFlex.  
+
  which requires a high amount of chemical expertise to generate. Conformers might also be generated without constrains, for which different tools are available, in our case, we used <a href="http://www.conflex.net/">ConFlex</a>.  
 
  Conformers need to be stored in one file (“.sdf”, “.mol”, or “.mol2”).
 
  Conformers need to be stored in one file (“.sdf”, “.mol”, or “.mol2”).
 
<li> “.pdb”-file: </br>
 
<li> “.pdb”-file: </br>
 
The input-file for the scaffold, in our case the tRNA synthetase, can be downloaded in PDB format from Protein Data Bank (PDB). It is then necessary to delete the natural ligand from the PDB-file,  
 
The input-file for the scaffold, in our case the tRNA synthetase, can be downloaded in PDB format from Protein Data Bank (PDB). It is then necessary to delete the natural ligand from the PDB-file,  
−
as we need to incorporate our own aaRS. and,Additionally, it is advised to relax the  preferably, the structure should be relaxedin order to allow for flexibility with regards to the simulation outcomes.  
+
as we need to incorporate our own aaRS. Additionally, it is advised to relax the  preferably, the structure should be relaxed in order to allow for flexibility with regards to the simulation outcomes.  
−
For further details, see the (documentation: https://www.rosettacommons.org/docs/latest/application_documentation/structure_prediction/relax.)
+
For further details, see the <a href="https://www.rosettacommons.org/docs/latest/application_documentation/structure_prediction/relax.">ROSETTA Relaxing documentation.</a>
 
<li> “.cst”-file: </br>
 
<li> “.cst”-file: </br>
 
The .cst-file defines the potential hydrogen bonds between the ligand and the amino acid. For example, the code block  characterized by the tags “CST::BEGIN” and “CST::END”, specifies the orientation or  
 
The .cst-file defines the potential hydrogen bonds between the ligand and the amino acid. For example, the code block  characterized by the tags “CST::BEGIN” and “CST::END”, specifies the orientation or  

Revision as of 19:10, 1 November 2017

Modeling

Organization of our modeling projects

On this page, we describe our main modeling project, which was integral for our whole project. However, besides this complex modeling, we also conducted and applied several straight-forward stochastic and statistical models to support and guide our laboratory work. These modeling projects are briefly described here; for further information, please check the corresponding link leading to the part of our project the model has been an element of.
Discriminant function model for the ICG prediction:
We conducted a discriminant function analysis for the recognition of which base – natural or unnatural – is present at a specified position of a base sequence. This model is part of our ICG model software, found here.
Calculation of an effective library size for the selection system:
We used a combination of combinatorics and statistics to calculate the optimal library size for the selection process, such that it is expected to contain all possible sequence mutations, and therefore easily all possible resulting amino acids, at least once. This calculation is part of the translational system, found here.
Comparison of mRFP production for the positive selection system (BBa_K2201373):
We modeled and visually compared the mRFP production over time for the normal signaling and the enhanced signaling circuit of the positive selection system. The system and plot can be found here.

Short Summary

As our project explores possibilities of an expanded genetic code via unnatural bases and non-canonical amino acids, we set out to complement our lab work via modeling of novel amino acyl tRNA synthetases (aaRS) for a non-canonical amino acids we synthetized in the lab. In order to incorporate non-canonical amino acids into proteins via the translational process, the aaRS has to attach the amino acid to the respective tRNA. Thus, we designed aaRS sequences which were meant to link our own non-canonical amino acid to a fitting tRNA. As a result, we obtained a couple of sequences of possible aaRS candidates, which we evaluated, based on a ROSETTA score, and ordered via gene synthesis. n practice, our modeling consisted of the following steps:

Figure (NUMMER ANGEBEN!): ABBILDUNGSTITEL
BILDUNTERSCHRIFT

Step Software/Method Meaning
1. Ligand Preparation Manually via Avogadro Due to the novelty of our amino acid, no information on the ligand is available in databases. Therefore, all information has to be provided manually and then generate a conformer ensemble, containing for example all energetically useful arrangements of atoms within the molecule.
2. Scaffold categorization ROSETTA protocol The scaffold describes the rough layout of the synthetase. We downloaded the scaffold 1J1U , the aaRS of as a template, and then relaxed its structure to improve the outcome of the ROSETTA algorithm.
3. Set simulation constrains Manually via ROSETTA Constrains with regards to possible mutations of the synthetase ensure that the generated sequences fit to the amino acid. For example, we constrained the distance between certain atoms and their angle to a range optimal for hydrogen bonds.
4. Enzyme Matching ROSETTA protocol ROSETTA combines information about the ligand and constrains to find possible hydrogen bonding partners and propose the shape of the scaffold within the set constraints.
5. Enzyme Design ROSETTA protocol An algorithm uses the information from the previous step and information on the ligand to simulate the mutation process and generate sequences for optimized scaffolds with corresponding scores as measures of fit.
6. Evaluate results in silico Manually We evaluate the visual output and the score values and order the sequences with the most promising results via gene synthesis.
7. Evaluate results in vivo Manually The synthetases are validated in the lab with the corresponding ncAA via a positive-negative selection system..
As a result, we obtained a couple of sequences of possible aaRS candidates, which we evaluated, based on a ROSETTA score, and ordered via gene synthesis. Figure A describes our modeling project as a whole

Introduction

Overview

Figure (NUMMER ANGEBEN!): ABBILDUNGSTITEL
BILDUNTERSCHRIFT

As part of our iGEM project, we are faced with the challenge of adapting the tRNA synthetase to non-canonical amino acids. For this purpose, modelled possible candidates for synthetases as a preparation for carrying out a positive-negative selection according to (Liu et al., 2007) in the laboratory. Due to the rapid development in the field of protein and molecular structure analysis, there has been an increase in the availability of molecular 3D structure data. These data are organized in publicly available databases which provide a foundation for the modeling and simulation of chemical-biological processes in bioinformatics. As our non-canonical amino acid has been synthetized by ourselves, no such comprehensive information is available, yet. However, information of similarly structured amino acids can potentially serve as a basis for our modeling. As evaluating an expanded genetic code is a complex task, the practical laboratory work of our project is supplemented by a theoretical approach, involving modeling, simulation, and evaluation on the computerin silico. Specifically, we focused on simulation to designaimed at designing an aaRS tRNA synthetase for the new non-canonical amino acid CBT-ASP. Additionally to CBT, we also simulated the evolution process for the non-canonical amino acid NPA as a validation of our modeling procedure, altough as synthases for this ncAA are known and thus comparable to our in silico result, we can evaluate our modeling procedure. (Vielleicht hier ein wenig schöner) For this purposeOur core challenge was to evolve, the binding pocket must be evolved in a manner which effectively charges the tRNA with the amino acid, thus also recognizing this amino acid specifically.

Method

We used the open-source software "Rosetta" for the main part of our modeling project, which was introduced at the University of Washington by David Baker in 1997 (Simon et al.,1997), initially in the context of protein structure prediction. Since then, Rosetta has grown to include numerous modules and is currently widely used in research. In our application, we focus on the Rosetta module called the "Rosetta Enzyme Design Protocol"

ROSETTA Enzyme Design

Overview

Since the non-canonical amino acid synthesized in the laboratory is completely novel, there is no corresponding tRNA synthetase which can load the tRNA, yet. For this reason, we use the enzyme design protocol to design the binding pocket in a way that allows it to form an effective and specific enzyme. The protocol consists of two main steps: matching and designing (Richter et al., 2011) The enzyme design algorithm basically is summarized in Fig. B

Figure (2): Flowchart Enzym Design Protocol

Matching Step

Figure (NUMMER ANGEBEN!): ABBILDUNGSTITEL
BILDUNTERSCHRIFT

The meaning of the matching step is to match the amino acids which constrains to the ligand, following specific constrains which ensure that the result is sensible and feasible. For this, ROSETTA analyzes the structural formula of the non- canonical amino acid and offers the possible hydrogen binding partners.

Matching step inputs

For the matching step, the following input-files are needed:
  • a “.params”-file specifying information about the ligand
  • a “.pdb”-file providing a rough scaffold layout
  • a “.cst”-file to define the bindings between ligand and scaffold
  • a “.pos”-file to define the positions of the amino acids of the scaffold
  • a “.flags” file to control all inputs. These files are necessary, as they describe the ligand and backbone and specify the parameters of the algorithm
To read about each file in further detail, please click the technical detail button below:
SHOW TECHNICAL DETAILS
  • “.params”-file:
    A conformer ensemble has to be generated using information about the ligand, as the non-canonical amino acids are not generally available in databases like PDB, making it necessary to build them manually using tools like pymol, Avogadro or Chemdraw. Using these tools, files can be saved in the desired format. The ligand needs to be specified in the “.sdf”, “.mol” or “.mol2” file format. Such a file can be obtained automatically by converting the relevant information from a “.pdb” file, if available. This conversion process usually also involves augmenting the data with hydrogen atoms in case they are missing from the “.pdb” file. Alternatively, the ligand can be designed using "Simplified Molecular Input Line Entry Specification" (SMILES) or manually using tools such as Avogadro, as we did. In the next step, the ligand file is used to create a conformer ensemble that is in turn used to create a Rosetta parameter (“.params”) file. In addition to the specific names of all atoms present in the ligand, this parameter file also stores all bonds between the individual atoms, including the binding angles and binding distances. Rosetta cannot generate the conformer ensemble by itself, so an additional tool is needed. Different tools are capable of creating the conformer ensemble automatically, but it is best to manually define constraints for the chi1, chi2 and backbone psi torsion angles that define the orientation of the ligand in the binding pocket. For this, we know of three tools: The first is OpenEye Omega, but the full license is very costly and the academic free version is hard to obtain. The second tool is Accelrys Discovery Studio, but Accerlys does not provide a free license. The third tool is TINKER, which is free, but poorly documented and depends on a specific keyfile, which requires a high amount of chemical expertise to generate. Conformers might also be generated without constrains, for which different tools are available, in our case, we used ConFlex. Conformers need to be stored in one file (“.sdf”, “.mol”, or “.mol2”).
  • “.pdb”-file:
    The input-file for the scaffold, in our case the tRNA synthetase, can be downloaded in PDB format from Protein Data Bank (PDB). It is then necessary to delete the natural ligand from the PDB-file, as we need to incorporate our own aaRS. Additionally, it is advised to relax the preferably, the structure should be relaxed in order to allow for flexibility with regards to the simulation outcomes. For further details, see the ROSETTA Relaxing documentation.
  • “.cst”-file:
    The .cst-file defines the potential hydrogen bonds between the ligand and the amino acid. For example, the code block characterized by the tags “CST::BEGIN” and “CST::END”, specifies the orientation or catalytic function of the enzyme.
    More specifically, the first record of the block begins with “TEMPLATE::ATOM_MAP”, followed by either “atom_name” or “atom_type”, depending on whether a specific residue or a specific type of residue is provided. In the latter case, it is not important to choose specific atoms. Instead, a catalytic residue of the amino acid such as “OH” or “Nhis” is specified. The next lines of the TEMPLATE::ATOM_MAP record define the residues using one-letter or three-letter-codes that are prefixed by “residue1” or “residue3”, respectively. The second record, beginning with the tag “CONSTRAINT”, contains all relevant distance, angle and torsion constraints for the matching. Each constraint is described with five parameters. In the case of the distance constraint, the first parameter describes the optimal distance “x0” between the chosen residues, the second parameter describes the tolerance “xtol”, the third parameter defines the strength “k” and the fourth parameter specifies the type of bond (1 for a covalent bond, 0 otherwise). If the modulus of the difference between the actual distance “x” and the specified optimal distance is smaller than the tolerance, then the penality score is zero. Otherwise, the constraint consists of the term k* ( |x - x0| - xtol ) to the penality score. For the angle and torsion constraints, the description is similar. If necessary, additional hydrogen bonds to other atoms of the ligand are specified in terms of additional blocks, using the tag “VARIABLE::CST”. Finally, most of the blocks described above can be optionally followed by an “ALGORITHM_INFO” record that stores details of the matching algorithm by parameter values. We refer to the Rosetta documentation for further details.
  • ”.pos”-file:
    The “.pos” file contains the allowed locations in the scaffold for the chosen catalytic residues in each constraint block of the “.cst” file.

Matching step outputs

Figure (NUMMER ANGEBEN!): ABBILDUNGSTITEL
BILDUNTERSCHRIFT

The output generated in the matching step is the layout of the scaffold as well as one or more states of the amino acid which enable interaction with the ligand. This information is stored as a “.pdb” file and becomes part of the input for the design step.

Our results for this step

We used the “1j1u”-scaffold from PDB for our matching step. The “1j1u.pdb”-file contains the Tyrosyl-tRNA-synthetase, which is labeld under “Chain A”, the orthogonol tRNA under “Chain B” and the natural ligand Tyrosyl. For our project, we deleted the natural ligand and “Chain B”, because it was not neccerary to change their structure or sequence and it was a way to save computer time and power. We designed the ligands manually by using Avogadro, and for the .cst-file, we choose the default matching algorithm for simulations of both amino acids.

Design Step

The design step applies an algorithm such that the binding pocket and the near environment are mutated and the remaining scaffold is repacked. Additionally, a badness-of-fit score is generated which indicates how well the mutation fits the amino acid. For every file from the matching step, a model with a score and a “.pdb-file” will be generated, specifying where the sequence can be located, and the 3D-structure can be analyzed. Notably, the amino acid structure can be extracted separately. The following section describes the structure of the design step. For further details on each step, click the technical details button.
1. Optimizing the catalytic interactions
For the first alternative, the file can be generated either by the Rosetta standard or a manually created .”res”- file. For more details, we refer to the Rosetta documentation. (link:https://www.rosettacommons.org/manuals/archive/rosetta3.5_user_guide/d1/d97/resfiles.html).
For the latter alternative, residues are automatically categorized by their location of the Calpha.

TECHNICAL DETAILS
Residues are catagorized as follows:
  • residues that have their Calpha within a distance cut1 angstroms of any ligand heavyatom will be set to designable
  • res that have Calpha within a distance cut2 of any ligand heavyatom and the Cbeta closer to that ligand atom than the Calpha will be set to designable. cut2 has to be larger than cut1
  • res that have Calpha within a certain distance cut3 of any ligand heavyatom will be set to repackable. cut3 has to be larger than cut2
  • res that have Calpha within a distance cut4 of any ligand heavy atom and the Cbeta closer to that ligand atom will be set to repackable. cut4 has to be larger than cut3
  • all residues not in any of the above 4 groups are kept static.

2. Cycles of sequence design and minimazation within constrains
To optimize the structure we used applied an iterative optimization algorithm. This algorithm mutates all residues from the backbone, which are not part of the catalytic center, to alanine, and a small energy function refraction will place the ligand in an optimal position to the backbone.

TECHNICAL DETAILS
For this approach, bb_min and chi_min allow for backbone flexibility and the rotation of the torsions. An alternative for this minimization step is the Monte Carlo rigid body ligand sampling. For further information on this method, we refer to the ROSETTA documentation (https://www.rosettacommons.org/manuals/archive/rosetta3.5_user_guide/d6/dbc/enzyme_design.html).

Design step inputs

The following input files are relevant for the design procedure:
  • “.pdb”-file generated in the matching step
  • “.cst”-file for the ligand
  • “.params”-file for the ligand and the scaffold
  • “.flags” to coordinate the inputs
For further information on these files, please refer to step 2 above.

Figure (NUMMER ANGEBEN!): ABBILDUNGSTITEL
BILDUNTERSCHRIFT

Design step outputs

The output for the design step is a “.pdb”-file containing the mutated scaffold and a “.score”-file. For every PDB-file, a line in the score-file is generated, so it is easy to evaluate the given structure. The first score in the file is the total score of the model. After that, the number of hydrogen bonds in the protein as a whole and in the constraints is listed, followed by the number of dismissed polars in the catalytic residues as well in the whole protein and in the constraints. See the technical details below for a full overview of the output information
TECHNICAL DETAILS
  • total_score: energy (excluding the constraint energy)
  • fa_rep: full atom repulsive energy
  • hbond_sc: hbond sidechain energy
  • all_cst: all constraint energy
  • tot_pstat_pm: pack statistics, 0-1, 1 = fully packed
  • total_nlpstat_pm: pack statistics withouth the ligand present
  • tot_burunsat_pm: buried unsatisfied polar residues, higher = more buried unsat polars (just a count)
  • tot_hbond_pm: total number of hbonds
  • tot_NLconst_pm: total number of non-local contacts ( two residues form a nonlocal contact if they are farther than 8 residues apart in sequence but interact with a Rosetta score of lower than -1.0 )

Our results for this step

We choose our synthetases because of a good total score and a good ligand score. We checked the corresponding PDB-files, and rated the ligand and the binding pocket as satisfying, so that the ligand assumedly does not collide with residues in the near environment. The total scores for CBT are not as good as the scores for NPA. However, the ligand scores are acceptable in both cases. A visual evaluation confirms that the ligand fits into the binding pocket.
Our results for this step
We used this algorithm to simulate the evolution of the tyrosyl-tRNA with the amino acids Nitrophenylalanine and CBT-ASP.
NPA simulation:
We created one .cst-file-block for the nitrogroup of NPA. Since there are two oxygen-atoms in the nitrogroup, we defined two atom nametags. As several possibilities are useful, we defined two possible constraint partners for the hydrogen bonds. The first is asparagine (N) or glutamine (Q) and the second is glycine (G). We set the possible distance to 2.8 A, as it is the optimal distance for hydrogenbonds, and a tolerance level of 0.5 A. We set the angles to 120° with a tolerance of 40°, as recommended by Florian Richter during our talk in cologne. The torsion angles were set to 180° with a tolerance of 180° and a penalty of 0, such that the torsion angles can rotate completely freely.(Richter, unpublished data)
CBT-ASP simulation:
CBT-ASP can build hydrogen bonds in two ways. The first is a weak hydrogen bond on the sulphur atom and the other possibility is a normal hydrogen bond on the nitrogen (N2) after the C-gamma. We wrote three cst-files, one for a possible bond with sulpur, one for a possible bond with nitrogen, and one for both bonds. As possible corresponding amino acids, we chose serine, threonine, tyrosine, asparagine, glutamine, and glycine.
It is recommended to write a “.flags”-file, because there are several input- parameters to be defined, but it is also possible to define them via console user interface.
For the categorization of the scaffold, we chose the automatic determination and set the following cuts: cut1: 6 A, cut2: 8 A, cut3: 10 A and cut4: 12 A, like the baker-lab commonly used.

Results

Results in silico

We used this algrithm to simulate the evolution of the tyrosyl-tRNA with the amino acids Nitrophenylalanine and CBT-ASP We obtained 13 synthetase sequences for CBT-ASP, and 43 sequences for NPA, which fit well into the binding site according to the ROSETTA score.

Results in vivo

In order to test the functionality and specificity of our modeled aaRS, we translated a selection of the most promising amino acid sequences into DNA sequences optimized for E.coli and ordered them via gene synthesis. We then used a positive-negative selection system for characterization. The experiment proceeds as follows: Due to problems with regards to the protein- and salt-concentration, we retransformed the gensyntheses which had been cloned into pSB1C3. In a next step, these syntheses were cotransformed in E.coli(BL21) with our positive selection plasmid. With regards to CBT2, only the original colony could be transformed. From CBT4 and CBT5, were used each originally isolated clone and its retransformed counterpart. Due to the IPTC-induced promoter, we used variants without IPTG, and with 5 mM, 10 mM, and 15 mM added IPTS for all plasmids for the kanamycine resistance. We chose additional variants with regards to the antibiotics; one variant each of kanamycine, kanamycine and chloramphenicole, and kanamycine, chloramphenicole and tetracycline. The number of resulting colonies for each variant are summarized in figure X. Our in vivo results show that our in silico designed enzymes did not lead to a loss of functioning. We ordered seven synthetase sequences of NPA via IDT, and four synthetase sequences of CBT-Asparagine by courtesy of Genscript, where we had previously won a grant of 500€. Due to problems on the part of IDT, the sequences for NPA could not be synthetized. We still provide the sequences for further use below. All further descriptions therefore refer to the synthetase sequences for CBT-Asparagine.
Unfortunately, we were not able to amplify the four sequences or clone them into the detection system pSB3T5. A test digestion revealed that the length of the sequence in the plasmid puk57, which was the plasmid delivered by Genscript, did not correspond to the sequence length ordered. Therefore, we disregarded synthetase candidate number 1 for further tests.
For the remaining three sequences, we instead utilized the positive-negative selection plasmids for validation of our syntheses. After the first, positive selection cycle, colonies formed only if the non-canonical amino acid was present. Thus, through the positive selection, we could show that the synthetase did not lead to the loss of functioning of the enzyme. To show specificity, we conducted a negative selection as well. We managed to clone the synthetases into the negative selection plasmid, but were not able to verify this selection cycle. Therefore, further tests are needed to validate the specificity of the synthetases.

References

Liu, W., Brock, A., Chen, S., Chen, S., Schultz, P. G. ,(2007). Genetic incorporation of unnatural amino acids into proteins in mammalian cells. Nature methods, 4(3), 239-244.
Richter, F., Leaver-Fay, A., Khare, S. D., Bjelic, S., Baker, D. (2011). De novo enzyme design using Rosetta3. PloS one, 6(5): e19230.
Simons, K. T., Kooperberg, C., Huang, E., Baker, D. (1997). Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and Bayesian scoring functions. Journal of molecular biology, 268(1), 209-225.