Re: [ccp4bb] model bias

2017-10-11 Thread Edward A. Berry

And given that all these maps (Fo, Fc, and their differences)
are made without the F000 term and so have average value zero,
it is highly likely that Fc will be negative in the region
of the deleted subunit (solvent is the lowest density in the model
except for the chinks of vacuum between the atoms in the protein,
and averaging over the volume of the protein you expect protein
to be electron-denser than solvent).
If Fc is negative then Fo-Fc negative implies Fo is still
more negative, and 2fo-fc cannot be positive.

It would be instructive to make Fo and Fc maps and check
their values in the region of the deleted subunit to resolve
the contradiction.


On 10/11/2017 04:47 PM, Edward A. Berry wrote:

This is hard to understand/impossible, at least in terms of
simple 2fo-fc and fo-fc maps.
We can think of a difference map as the difference between two maps,
like fo-fc map = fo map - fc map.
There is no model in the region, since you deleted it,
so Fc=0, and Fo-Fc negative => Fo<0, so 2fo-fc cannot be positive.

No, wait, there will be solvent, so Fc = Rho(solvent)
Fo-Fc is negative, so Fo  2Fo>Rho(solv)
2Fo > Fc > Fo
Fo > 1/2 Fc
So this would be possible if the density of your protein is less than
that of the solvent but more than half that of the solvent.
Partial occupancy? but where the protein is missing there
would be solvent (in Fo if not in Fc),
so you still couldn't get below density of the solvent.
Maybe its a problem with the solvent model?

On 10/11/2017 09:48 AM, Karsten Dreifus wrote:

Dear all,
I have a 120 aa protein. Matthews coefficient indicates 3 mol in asu.
Molrep (template with 70 % seq identity)  finds three NCS molecules
(the template model has only 1 chain as it is in different space
group). Now, REFMAC refines it to 25/30 % R/RFREE. I was just worried
about model bias, so I just deleted chain C and re-ran the refinement.
I observe that in the place of chain C, the DELFWT map in Coot shows
negative density (red colour). Shouldnt the deleted regions show
POSITIVE density (green colour)? The 2fo-fc map in the deleted region
shows usual blue colour. I shook the 2chain model with Phenix simple
dynamics and then ran refinement with same results.

How do I verify the model is correct?
Karsten





Re: [ccp4bb] model bias

2017-10-11 Thread Edward A. Berry

This is hard to understand/impossible, at least in terms of
simple 2fo-fc and fo-fc maps.
We can think of a difference map as the difference between two maps,
like fo-fc map = fo map - fc map.
There is no model in the region, since you deleted it,
so Fc=0, and Fo-Fc negative => Fo<0, so 2fo-fc cannot be positive.

No, wait, there will be solvent, so Fc = Rho(solvent)
Fo-Fc is negative, so Fo  2Fo>Rho(solv)
2Fo > Fc > Fo
Fo > 1/2 Fc
So this would be possible if the density of your protein is less than
that of the solvent but more than half that of the solvent.
Partial occupancy? but where the protein is missing there
would be solvent (in Fo if not in Fc),
so you still couldn't get below density of the solvent.
Maybe its a problem with the solvent model?

On 10/11/2017 09:48 AM, Karsten Dreifus wrote:

Dear all,
I have a 120 aa protein. Matthews coefficient indicates 3 mol in asu.
Molrep (template with 70 % seq identity)  finds three NCS molecules
(the template model has only 1 chain as it is in different space
group). Now, REFMAC refines it to 25/30 % R/RFREE. I was just worried
about model bias, so I just deleted chain C and re-ran the refinement.
I observe that in the place of chain C, the DELFWT map in Coot shows
negative density (red colour). Shouldnt the deleted regions show
POSITIVE density (green colour)? The 2fo-fc map in the deleted region
shows usual blue colour. I shook the 2chain model with Phenix simple
dynamics and then ran refinement with same results.

How do I verify the model is correct?
Karsten



Re: [ccp4bb] model bias

2017-10-11 Thread Pavel Afonine
A round of refinement with simulated annealing followed by minimization
should address your concern.
Pavel

On Wed, Oct 11, 2017 at 4:48 PM, Karsten Dreifus 
wrote:

> Dear all,
> I have a 120 aa protein. Matthews coefficient indicates 3 mol in asu.
> Molrep (template with 70 % seq identity)  finds three NCS molecules
> (the template model has only 1 chain as it is in different space
> group). Now, REFMAC refines it to 25/30 % R/RFREE. I was just worried
> about model bias, so I just deleted chain C and re-ran the refinement.
> I observe that in the place of chain C, the DELFWT map in Coot shows
> negative density (red colour). Shouldnt the deleted regions show
> POSITIVE density (green colour)? The 2fo-fc map in the deleted region
> shows usual blue colour. I shook the 2chain model with Phenix simple
> dynamics and then ran refinement with same results.
>
> How do I verify the model is correct?
> Karsten
>


[ccp4bb] model bias

2017-10-11 Thread Karsten Dreifus
Dear all,
I have a 120 aa protein. Matthews coefficient indicates 3 mol in asu.
Molrep (template with 70 % seq identity)  finds three NCS molecules
(the template model has only 1 chain as it is in different space
group). Now, REFMAC refines it to 25/30 % R/RFREE. I was just worried
about model bias, so I just deleted chain C and re-ran the refinement.
I observe that in the place of chain C, the DELFWT map in Coot shows
negative density (red colour). Shouldnt the deleted regions show
POSITIVE density (green colour)? The 2fo-fc map in the deleted region
shows usual blue colour. I shook the 2chain model with Phenix simple
dynamics and then ran refinement with same results.

How do I verify the model is correct?
Karsten


[ccp4bb] model bias

2015-04-29 Thread Aleksandar Bijelic

Dear CCP4 users,

I am currently solving a structure (2.8-2.9 A resolution) of a protein 
complexed with a ligand using MR with the apo-form of this protein as 
model (resolution of the model is 2.4). After MR-phasing I performed a 
regular autobuild run giving me good outputs and thus I refined the best 
pdb leading to good values according to R-values and geometry, however, 
the denstiy doesn´t look well (but I think it´s due to the moderate 
resolution). Now I want to get sure if the side chains which are 
involved in the ligand binding are correctly positioned. However, the 
active site is suspicously similar to the active site of the model 
(apo-form) and so I am afraid that this could be due to model bias.  My 
question is how to check and to get rid of the bias (if present) at this 
stage (after several refinements). I read the publication of Terwilliger 
about iterative-build OMIT maps but since I am a bloody novice in this 
field I didn´t really understand it. I originally thought 
iterative-build OMIT maps are performed to compare the output map with 
one´s map in order to detect uncertainties, but what to do next? Or 
should I start from the beginning but how to proceed than, what should I 
do (I am using Phenix via GUI) ... Is it possible and reasonable to run 
autobuild with iterative omit map option? Or is it only reasonable if 
experimental phases are available? I didn´t run iterative-build OMIT 
maps yet because I am not sure how to run it correctly (what method is 
the best?) and at my institute the run will take more than 1 day and I 
don´t want to block one computer until I am not sure if it is 
reasonable. I hope you can give me some advice and help me. Thank you in 
advance.


Regards,

Aleks

--
---
Aleksandar Bijelic, MSc.

Institut für Biophysikalische Chemie
Universität Wien
Althanstrasse 14
A-1090 Wien

Tel: +43 1 4277 52536
e-Mail: aleksandar.bije...@univie.ac.at




Re: [ccp4bb] model bias

2015-04-29 Thread Eleanor Dodson
If you know what residues are likely to be involved, then set their
occupancies to 0.0 (In coot go to Measures - Residue info - edit occ)
then do a few cycles of refinement with that model, and then see if there
is difference density for those residues...

Quicker than an omit map procedure.
Eleanor


On 29 April 2015 at 12:16, Aleksandar Bijelic 
aleksandar.bije...@univie.ac.at wrote:

 Dear CCP4 users,

 I am currently solving a structure (2.8-2.9 A resolution) of a protein
 complexed with a ligand using MR with the apo-form of this protein as model
 (resolution of the model is 2.4). After MR-phasing I performed a regular
 autobuild run giving me good outputs and thus I refined the best pdb
 leading to good values according to R-values and geometry, however, the
 denstiy doesn´t look well (but I think it´s due to the moderate
 resolution). Now I want to get sure if the side chains which are involved
 in the ligand binding are correctly positioned. However, the active site is
 suspicously similar to the active site of the model (apo-form) and so I am
 afraid that this could be due to model bias.  My question is how to check
 and to get rid of the bias (if present) at this stage (after several
 refinements). I read the publication of Terwilliger about iterative-build
 OMIT maps but since I am a bloody novice in this field I didn´t really
 understand it. I originally thought iterative-build OMIT maps are performed
 to compare the output map with one´s map in order to detect uncertainties,
 but what to do next? Or should I start from the beginning but how to
 proceed than, what should I do (I am using Phenix via GUI) ... Is it
 possible and reasonable to run autobuild with iterative omit map option? Or
 is it only reasonable if experimental phases are available? I didn´t run
 iterative-build OMIT maps yet because I am not sure how to run it correctly
 (what method is the best?) and at my institute the run will take more than
 1 day and I don´t want to block one computer until I am not sure if it is
 reasonable. I hope you can give me some advice and help me. Thank you in
 advance.

 Regards,

 Aleks

 --
 ---
 Aleksandar Bijelic, MSc.

 Institut für Biophysikalische Chemie
 Universität Wien
 Althanstrasse 14
 A-1090 Wien

 Tel: +43 1 4277 52536
 e-Mail: aleksandar.bije...@univie.ac.at

 



Re: [ccp4bb] model bias

2015-04-29 Thread vellieux

Hello,

I would certainly try the usual approaches (map coefficients that are 
less sensitive to model-bias, i.e. Sigmaa; OMIT maps - there are several 
of these which can be calculated). In addition, you may have a look at 
the approach of Liu and Xiong (2014, J. Mol. Biol. 426, 980-993). It is 
a well known technique (applying a negative temperature factor to 
amplitudes, for map sharpening) but this paper describes a systematic 
study indicating improvement whatever the situation.


As usual in macromolecular crystallography, confidence is gained by the 
use of several approaches.


HTH,

Fred.

On 29/04/15 13:16, Aleksandar Bijelic wrote:

Dear CCP4 users,

I am currently solving a structure (2.8-2.9 A resolution) of a protein 
complexed with a ligand using MR with the apo-form of this protein as 
model (resolution of the model is 2.4). After MR-phasing I performed a 
regular autobuild run giving me good outputs and thus I refined the 
best pdb leading to good values according to R-values and geometry, 
however, the denstiy doesn´t look well (but I think it´s due to the 
moderate resolution). Now I want to get sure if the side chains which 
are involved in the ligand binding are correctly positioned. However, 
the active site is suspicously similar to the active site of the model 
(apo-form) and so I am afraid that this could be due to model bias.  
My question is how to check and to get rid of the bias (if present) at 
this stage (after several refinements). I read the publication of 
Terwilliger about iterative-build OMIT maps but since I am a bloody 
novice in this field I didn´t really understand it. I originally 
thought iterative-build OMIT maps are performed to compare the output 
map with one´s map in order to detect uncertainties, but what to do 
next? Or should I start from the beginning but how to proceed than, 
what should I do (I am using Phenix via GUI) ... Is it possible and 
reasonable to run autobuild with iterative omit map option? Or is it 
only reasonable if experimental phases are available? I didn´t run 
iterative-build OMIT maps yet because I am not sure how to run it 
correctly (what method is the best?) and at my institute the run will 
take more than 1 day and I don´t want to block one computer until I am 
not sure if it is reasonable. I hope you can give me some advice and 
help me. Thank you in advance.


Regards,

Aleks




--
Fred. Vellieux (B.Sc., Ph.D., hdr)

IBS / ELMA
Campus EPN
71 avenue des Martyrs
CS 10090
F-38044 Grenoble Cedex 9
Tel: +33 457428605
Fax: +33 476501890


Re: [ccp4bb] model bias

2015-04-29 Thread Shibom Basu
Dear Aleksandar,

I did use SA-OMIT map feature in Phenix, implemented by Tom Terwilliger.
I would strongly recommend you to use this feature.
Remove the suspicious residues from the pdb and run first SA-omit map in
phenix with default settings, and later with different starting
temperatures.
Make sure also to use harmonic restrains to prohibit neighboring residues
to creep into the omitted region.
Also, use MR solution and the raw mtz file with only FP and SIGFP that you
have for the omit map. I have experienced that if you use refined model
then it does not completely get rid of biases.
At max, you can use rigid body refined one.

On the other side, if your molecule is not too big, then you can try the
CNS. In that case, I would recommend remove the 'suspicious' resiudes and
1st use segmented rigid body refinement, i.e, consider each chain or
subunit as rigid entity.
Then, you can try group B factor with restrains along with DEN mode for
simulated annealing. I strongly recommend the DEN mode which can help you
to remove bias.
Please note that it is necessary to try out several different temperatures
and different temperature gradient as well for simulated annealing in order
to reproduce the features.
Otherwise, you never know if you are stuck at local minima or not.

thanks
shibom

On Wed, Apr 29, 2015 at 4:16 AM, Aleksandar Bijelic 
aleksandar.bije...@univie.ac.at wrote:

 Dear CCP4 users,

 I am currently solving a structure (2.8-2.9 A resolution) of a protein
 complexed with a ligand using MR with the apo-form of this protein as model
 (resolution of the model is 2.4). After MR-phasing I performed a regular
 autobuild run giving me good outputs and thus I refined the best pdb
 leading to good values according to R-values and geometry, however, the
 denstiy doesn´t look well (but I think it´s due to the moderate
 resolution). Now I want to get sure if the side chains which are involved
 in the ligand binding are correctly positioned. However, the active site is
 suspicously similar to the active site of the model (apo-form) and so I am
 afraid that this could be due to model bias.  My question is how to check
 and to get rid of the bias (if present) at this stage (after several
 refinements). I read the publication of Terwilliger about iterative-build
 OMIT maps but since I am a bloody novice in this field I didn´t really
 understand it. I originally thought iterative-build OMIT maps are performed
 to compare the output map with one´s map in order to detect uncertainties,
 but what to do next? Or should I start from the beginning but how to
 proceed than, what should I do (I am using Phenix via GUI) ... Is it
 possible and reasonable to run autobuild with iterative omit map option? Or
 is it only reasonable if experimental phases are available? I didn´t run
 iterative-build OMIT maps yet because I am not sure how to run it correctly
 (what method is the best?) and at my institute the run will take more than
 1 day and I don´t want to block one computer until I am not sure if it is
 reasonable. I hope you can give me some advice and help me. Thank you in
 advance.

 Regards,

 Aleks

 --
 ---
 Aleksandar Bijelic, MSc.

 Institut für Biophysikalische Chemie
 Universität Wien
 Althanstrasse 14
 A-1090 Wien

 Tel: +43 1 4277 52536
 e-Mail: aleksandar.bije...@univie.ac.at

 



Re: [ccp4bb] model bias

2015-04-29 Thread Xiaodi Yu
Hi Aleks:

Maybe you can try CNS ( Initial refinement by simulated annealing) also. It may 
help to get rid of the model bias and takes short time to run.

Xiaodi

 Date: Wed, 29 Apr 2015 13:52:53 +0200
 From: frederic.velli...@ibs.fr
 Subject: Re: [ccp4bb] model bias
 To: CCP4BB@JISCMAIL.AC.UK
 
 Hello,
 
 I would certainly try the usual approaches (map coefficients that are 
 less sensitive to model-bias, i.e. Sigmaa; OMIT maps - there are several 
 of these which can be calculated). In addition, you may have a look at 
 the approach of Liu and Xiong (2014, J. Mol. Biol. 426, 980-993). It is 
 a well known technique (applying a negative temperature factor to 
 amplitudes, for map sharpening) but this paper describes a systematic 
 study indicating improvement whatever the situation.
 
 As usual in macromolecular crystallography, confidence is gained by the 
 use of several approaches.
 
 HTH,
 
 Fred.
 
 On 29/04/15 13:16, Aleksandar Bijelic wrote:
  Dear CCP4 users,
 
  I am currently solving a structure (2.8-2.9 A resolution) of a protein 
  complexed with a ligand using MR with the apo-form of this protein as 
  model (resolution of the model is 2.4). After MR-phasing I performed a 
  regular autobuild run giving me good outputs and thus I refined the 
  best pdb leading to good values according to R-values and geometry, 
  however, the denstiy doesn´t look well (but I think it´s due to the 
  moderate resolution). Now I want to get sure if the side chains which 
  are involved in the ligand binding are correctly positioned. However, 
  the active site is suspicously similar to the active site of the model 
  (apo-form) and so I am afraid that this could be due to model bias.  
  My question is how to check and to get rid of the bias (if present) at 
  this stage (after several refinements). I read the publication of 
  Terwilliger about iterative-build OMIT maps but since I am a bloody 
  novice in this field I didn´t really understand it. I originally 
  thought iterative-build OMIT maps are performed to compare the output 
  map with one´s map in order to detect uncertainties, but what to do 
  next? Or should I start from the beginning but how to proceed than, 
  what should I do (I am using Phenix via GUI) ... Is it possible and 
  reasonable to run autobuild with iterative omit map option? Or is it 
  only reasonable if experimental phases are available? I didn´t run 
  iterative-build OMIT maps yet because I am not sure how to run it 
  correctly (what method is the best?) and at my institute the run will 
  take more than 1 day and I don´t want to block one computer until I am 
  not sure if it is reasonable. I hope you can give me some advice and 
  help me. Thank you in advance.
 
  Regards,
 
  Aleks
 
 
 
 -- 
 Fred. Vellieux (B.Sc., Ph.D., hdr)
 
 IBS / ELMA
 Campus EPN
 71 avenue des Martyrs
 CS 10090
 F-38044 Grenoble Cedex 9
 Tel: +33 457428605
 Fax: +33 476501890
  

Re: [ccp4bb] model bias

2015-04-29 Thread Aleksandar Bijelic

Dear All,

thank you all for your suggestions and support. The quick look (by 
removing interacting AA and subsequent refinement with and without 
little SA) revealed that some AAs are mislocated and indeed seem to be 
more directed to the density belonging to the ligand. So, I will try 
some of your sugestions to intervene bias directly after MR before 
further refinement. Thank you again!


Regards,

Aleks

--
---
Aleksandar Bijelic, MSc.

Institut für Biophysikalische Chemie
Universität Wien
Althanstrasse 14
A-1090 Wien

Tel: +43 1 4277 52536
e-Mail: aleksandar.bije...@univie.ac.at




Re: [ccp4bb] model bias

2009-08-12 Thread Eleanor Dodson

You dont give the resolution of your data.

There are always things to check
1) space group - could it be P222 or P21 2 2 or P 21 21 2 or P2 21 2 
etc etc - there are 8 possibilities.


You can use absences along the axial lines to help decide on the screw 
axes, but these can be misleading if you have more than 1 molecule in 
the asymmetric unit and the molecules are related by a 
non-crrystallographic translation. Phaser can be run in each space group 
in turn and usually that can decide on SG. Did you do that?



2)  Is there a problem with the data - severe anisotropy? twinning? etc etc

3) Maybe you just need to do more rebuilding - depending on 
theresolution some of the automated procedures can help - Arp/wARP, 
buccaneer and so on..


 Eleanor


ManojSaxena wrote:

Hi all,

I am working with a protein that have 28% similar to my MR template.
I have  processed data in HKL2000 for one of my crystal and I got unique sol in 
space group
P212121. with LLG 131 and TFZ score 13.5
 I have used  buccaneer and coot for model building and my  Rfee came to 45%.
I used the PDB file from this crystal ( Rfee of 45%) as my MR model for data  obtained from 
another crystal and got sol with TFZ score=40.5 and LLG=2305.


I used coot and did a round of refmac refinement now my Rfee is 41%.

My concern is 
a) wether my scheme is good or not because I am afraid that this will increase the model 
bias.
b)Now I am stuck at 41% Rfee  I have many chain breaks and loop regions have not good 
density. I tried TLS refinement it did not help. what else I shall try??


Thanks

Manoj 
PSU

Biochemistry
  


Re: [ccp4bb] model bias

2009-08-12 Thread David Briggs
Hi Manoj,

Following on from Poul's reply, and maybe whilst you are waiting to
get derivatives  ;)  you could try something like the following for
getting around the model-bias-after-borderline-MR-issue.

1) generate a prime--switch'd map from resolve.
2) use this map to prune your model (be quite brutal here).
3) Re-refine, using refmac and, in parallel, a simulated annealing
refinement protocol (CNS/Phenix).
4) Stick your pruned models and all these maps into coot and see what
you can build back.

I've found that schemes/strategies based around this can break the
back of refinement after borderline MR.

HTH

Dave


2009/8/11 ManojSaxena mks...@rediffmail.com:
 Hi all,

 I am working with a protein that have 28% similar to my MR template.
 I have processed data in HKL2000 for one of my crystal and I got unique sol
 in space group
 P212121. with LLG 131 and TFZ score 13.5
 I have used buccaneer and coot for model building and my Rfee came to 45%.
 I used the PDB file from this crystal ( Rfee of 45%) as my MR model for data
 obtained from
 another crystal and got sol with TFZ score=40.5 and LLG=2305.

 I used coot and did a round of refmac refinement now my Rfee is 41%.

 My concern is
 a) wether my scheme is good or not because I am afraid that this will
 increase the model
 bias.
 b)Now I am stuck at 41% Rfee I have many chain breaks and loop regions have
 not good
 density. I tried TLS refinement it did not help. what else I shall try??

 Thanks

 Manoj
 PSU
 Biochemistry




-- 

David C. Briggs PhD
Father  Crystallographer
http://drdavidcbriggs.googlepages.com/home
Skype: DocDCB



[ccp4bb] model bias

2009-08-11 Thread ManojSaxena
Hi all,



I am working with a protein that have 28% similar to my MR template.

I have  processed data in HKL2000 for one of my crystal and I got unique sol in 
space group

P212121. with LLG 131 and TFZ score 13.5

 I have used  buccaneer and coot for model building and my  Rfee came to 45%.

I used the PDB file from this crystal ( Rfee of 45%) as my MR model for data  
obtained from 

another crystal and got sol with TFZ score=40.5 and LLG=2305.



I used coot and did a round of refmac refinement now my Rfee is 41%.



My concern is 

a) wether my scheme is good or not because I am afraid that this will increase 
the model 

bias.

b)Now I am stuck at 41% Rfee  I have many chain breaks and loop regions have 
not good 

density. I tried TLS refinement it did not help. what else I shall try??



Thanks



Manoj 

PSU

Biochemistry

Re: [ccp4bb] model bias

2009-08-11 Thread Poul Nissen
Great that you have MR phases - it will help you identify heavy-atom  
sites for phasing and perhaps even the sulphur sites will be enough.


Poul
On 11/08/2009, at 20.33, ManojSaxena wrote:


Hi all,

I am working with a protein that have 28% similar to my MR template.
I have processed data in HKL2000 for one of my crystal and I got  
unique sol in space group

P212121. with LLG 131 and TFZ score 13.5
I have used buccaneer and coot for model building and my Rfee came  
to 45%.
I used the PDB file from this crystal ( Rfee of 45%) as my MR model  
for data obtained from

another crystal and got sol with TFZ score=40.5 and LLG=2305.

I used coot and did a round of refmac refinement now my Rfee is 41%.

My concern is
a) wether my scheme is good or not because I am afraid that this  
will increase the model

bias.
b)Now I am stuck at 41% Rfee I have many chain breaks and loop  
regions have not good
density. I tried TLS refinement it did not help. what else I shall  
try??


Thanks

Manoj
PSU
Biochemistry





[ccp4bb] model bias: phaser + prime-and-switch VS simulated annealing in phenix VS simulated annealing in cns + prime and switch

2008-05-02 Thread hari jayaram
First off sorry for the cross post across bbs

I have a molecular replacement solution for a single site mutant for data
that goes out to 2.8 A.
After molecular replacement in phaser I run the following and examine the
maps for bias from the model.

Option1: simluated annealing refinement in phenix using the phaser mtz

Option2: take the phaser mtz and then just run prime and switch in resolve
and look at the map

Option3: First part: take the phaser mtz and then run simulated annealing
and sigmaa in cns 1.2.2 . Second part: Run prime-and-switch using part1s
model phases and modified Fs

I am observing that Option 1 and Option 2 the maps are indistinguishable
from each other and show very little bias from the model.
In Option 3 just cns simulated annealing does not go as far at removing bias
as cns simulated annealing plus prime-and-switch.

Although my estimate of model bias is just a look and see feeling. Since I
lack a thorough understanding of the relative merits of these algorithms
I have the following questions

1)Does simulated annealing refinement as implemented in phenix use a
prime-and-switch style approach to modify phases at any point to guide the
refinement

2) Am I wrong in assuming that just simulated annealing and sigma-aa
weighted maps for cns (Option 3 part 1) are not as good as  cns simulated
annealing + sigmaa+ prime_and_switch ( option3 part 2)

Hoping to get a better idea on how well these approaches fair at removing
model bias

Hari Jayaram
Postdoc , Brandeis University