Select the best model and generate statistical potentials

SOAP Model Selection module

class modelSelection.parlist(iterable=(), /)

class modelSelection.sps(modellist=[], evalPotFunc=None)

Evaluate a list of models on sge cluster

class modelSelection.spss(model=[], logfoldername='', evalPotFunc=None)

Select the best model.

Parameters:

• model: the model space to search the best model within

Common usage::

spl=spss(model=model1) spl.find_best_par() spl.log()

Note

The model and model space are defined using python lists and dictionaries.

Define the recovery function features, distance only, from 0 to 20A with bin size 0.05, see the features.

rfeature=[['d',0,20,0.05]] # 'd' - distance, 0 - start position, 20 - end position, 0.05 - bin size

Features can be defined using either list or string, list defintion will be converted into string defintion on the fly.

Define the spliens used for generating recoery functions. see recoveryFunction:

par={'uniform':4,'featurebins':rfeature[0],'distribute':'lownumberpriority','firstpoint':2.25}
slo={'key':'sacnfflex',
'valueset':{
                'sacnfflex':{'sftype':'sacnf','par':par,'parvalue':[1,1,1,1]},
                'sacnf52':{'sftype':'sacnf','par':[2.75,3.75],'parvalue':[1,1,1,1]},
}}

The most important parameter to vary is ‘uniform’, which defines the number of anchor points to use.

Define the recovery function, see recoveryFunction:

ref1={'type':'sf','features':rfeature,'sftype':slo,'par':slo,'parvalue':slo,'ratio':[1.0]}

Define the features for the statistics calculation from a set of structures, a158 represent residue dependent atom type:

sfeature=[rfeature[0],'a158','as158']

Define the processing method of the probabilistic table, ‘npend’ means normalized by the last bin value, see statsTable.scaledsp:

pm=['','npend']

Processing method can be defined using either list or string, list defintion will be converted into string definition on the fly.

Define the probabilistic table using for scoring:

scaledsp1={'type':'scaledsp','pdbset':'X_2.2A_0.25rfree','features':sfeature,'genmethod':'cs1','pm':pm,'refs':[ref1],'ratio':[1.0]}

The benchmark criteria, defines how the statistical potentials are benchmarked, see benchmark:

bmtype={'type':'dsscore','dslist':['fpdb','fpdu'],'criteria':'3xtop1_rmsdallif_mean_+2xtop3_rmsdallif_mean_','combine':'scoresum','bm':'cs1'}

• ‘type’ can be ‘dsscore’ or ‘refinescore’ (use SOAP for refine).

• ‘combine’ can only be ‘scoresum’ at the moment.

• ‘bm’ has the same meaning as genmethod.

Parameters to optimize:

search1={'object':ref1,'key':'parvalue','pos':[0,1,2,3],'InitialGenerator':{'type':'dfire','values':initvalues}}
search2={'object':ref1,'key':'par','pos':[0,1,2,3],'InitialGenerator':{'type':'dfire','values':initvalues}}

• ‘Object’: the object to optimize, needs to be a dict

• ‘key’: the key of the object to optimize, must be a list

• ‘pos’: the positions of the values in the list to optimize.

• ‘InitialGenerator’: ways to generate initial value

Discrete search dictionaries, defining the model space - the model variables/options to vary:

dsearch4={'object':par,'key':'uniform','valueset':[7,6,5,4,3]}

dsearch9={'object':[scaledsp1,scaledsp1],'key':['genmethod','pdbset'],'valueset':[['cs1','X2_2.2A_0.25rfree']]}#,['cs1','X2_2.2A_0.25rfree_30'],['cs1','X2_2.2A_0.25rfree_60'],['cs1','X2_2.2A_0.25rfree_95'],['bs20dsp','X_2.2A_0.25rfree'],['bs20dsp','X_2.2A_0.25rfree_30'],['bs20dsp','X_2.2A_0.25rfree_60'],['bs15dsp','X_2.2A_0.25rfree'],['bs10dsp','X_2.2A_0.25rfree']]}#,,'cs1'

• ‘Object’: the object to vary, needs to be a dict

• ‘key’: the key of the object to vary, often a string

• ‘valueset’: the set of values to search within.

Parameters controling sampling and optimization, please check the code in the sampling module for detailed meanings:

ni=40

initvalues=list(np.arange(0,ni)/10.0+1)

initvalues=np.array(initvalues)

inner={'improved':2,'maxloop':100,'minloop':2}
outer={'improved':4,'maxloop':5023}

td=list(np.sqrt(np.linspace(1,float(10)**2,ni)))
td=np.array(td)
tune_interval=200

sampleschedule={'inner':inner,'outer':outer}
ssm={'sm':'mcp','reset_betweenruns':2,'blockupdate':True, 'exchange_method':1,
      'sample_schedule':sampleschedule,
      'stepmethod':'mxmp2.0','tune_interval':200,'add_bias':False,'temperature_distribution':td}

ssm0={'sm':'mcs','reset_betweenruns':2,'blockupdate':False,'using_globalbest':True,
      'sample_schedule':sampleschedule,
      'stepmethod':'mxmp2.0','tune_interval':201,'add_bias':False,'temperature_distribution':td}


ssm2={'sm':'mcp','reset_betweenruns':2,'blockupdate':False, 'exchange_method':1,
      'sample_schedule':sampleschedule,
      'stepmethod':'mxmp2.0','tune_interval':200,'add_bias':False,'temperature_distribution':td}

ssm20={'sm':'mcs','reset_betweenruns':2,'blockupdate':True,'using_globalbest':True,
      'sample_schedule':sampleschedule,
      'stepmethod':'mxmp2.0','tune_interval':201,'add_bias':False,'temperature_distribution':td}


ssm1={'sm':'mca','reset_betweenruns':2,'blockupdate':True, 'using_globalbest':True,
      'sample_schedule':sampleschedule,
      'stepmethod':'mxmp2.0','tune_interval':201,'add_bias':False,'temperature_distribution':np.zeros(ni)+2}

ssm3={'sm':'powell','blockupdate':False}

sml=[ssm20,ssm2,ssm0,ssm,ssm0,ssm,ssm0,ssm, ssm1]

Define the final model:

model1={'scorers':[scaledsp1,ref1],'bmtype':bmtype,'searches':[search1,search2], 'runsperscorer':ni,
    'dsearches':[dsearch2,dsearch5,dsearch4,dsearch7,dsearch8,dsearch9],'sml':sml,'cvk':2,'repeat':1,'fold':3}

• ‘scorers’: scoring terms

• ‘bmtype’ : benchmark criteria for juding a statistical potential

• ‘searches’ : parameters to optimzie

• ‘dsearches’ : model options/values for model seleciton

• ‘sml’ : optimization method

• ‘runsperscorer’ : how many runs per replica exchange

• ‘cvk’ : k times cross validation

• ‘repeat’ : how many replica exchanges to carry out for each optimization run

• ‘fold’ : n fold cross validation

• ‘testperc’: if included, this percentage of decoys will be left out for final validation.