void BranchAcceptor::load(istream &is)
{
double cutoff;
Strand strand;
SignalType signalType;
BOOM::String p;
int consensusOffset;
is >> signalType;
is >> p;
cutoff=p.asDouble();
is >> strand;
is >> consensusOffset;
setSignalType(signalType);
setStrand(strand);
setCutoff(cutoff);
BOOM::String dummy;
is>>dummy; // will always be "WWAM"
branchPoint=new WWAM(getGC(),is);
is>>dummy; // will always be "WAM"
acceptor=new WAM(getGC(),is);
int contextWindowLength=branchPoint->getContextWindowLength()+
acceptor->getContextWindowLength();
setSizes(2,consensusOffset,contextWindowLength);
}
/**
* Reimplemented from UMLWidget::loadFromXMI to load
* SignalWidget from XMI.
*/
bool SignalWidget::loadFromXMI( QDomElement & qElement )
{
if( !UMLWidget::loadFromXMI( qElement ) )
return false;
setName(qElement.attribute( "signalname", "" ));
setDocumentation(qElement.attribute( "documentation", "" ));
QString type = qElement.attribute( "signaltype", "" );
setSignalType((SignalType)type.toInt());
return true;
}
BranchAcceptor::BranchAcceptor(GarbageCollector &gc,
BOOM::Vector<TrainingSequence*> &seqs,
int branchPointOrder,
int acceptorOrder,
int branchContextLength,
int minSampleSize,
int consensusOffset,
SignalType signalType)
: SignalSensor(gc),
branchPoint(NULL),
acceptor(NULL)
{
// ctor
// Misc. initialization
setStrand(FORWARD_STRAND);
setSignalType(signalType);
int sensorLength=seqs[0]->getLength();
int acceptorContextLength=sensorLength-branchContextLength;
setSizes(2,consensusOffset,sensorLength);
// Split training sequences into branch point windows and acceptor
// windows
BOOM::Vector<TrainingSequence*> branchPoints, acceptors;
BOOM::Vector<TrainingSequence*>::iterator cur=seqs.begin(),
end=seqs.end();
for(; cur!=end ; ++cur)
{
TrainingSequence &S=**cur;
TrainingSequence *branchPoint=new TrainingSequence();
TrainingSequence *acceptor=new TrainingSequence();
S.getSubsequence(0,branchContextLength,*branchPoint);
S.getSubsequence(branchContextLength,acceptorContextLength,*acceptor);
branchPoints.push_back(branchPoint);
acceptors.push_back(acceptor);
}
// Train the branch point sensor & the acceptor sensor
acceptor=new WAM(gc,acceptors,acceptorOrder,minSampleSize,
signalType,consensusOffset-branchContextLength,2);
branchPoint=new WWAM(gc,branchPoints,branchPointOrder,minSampleSize,
signalType,0,0);
// Delete the training subsequences
cur=branchPoints.begin(); end=branchPoints.end();
for(; cur!=end ; ++cur) delete *cur;
cur=acceptors.begin(); end=acceptors.end();
for(; cur!=end ; ++cur) delete *cur;
}
BWM::BWM(GarbageCollector &gc,BOOM::Vector<TrainingSequence*> &sequences,
SignalType signalType,int consensusOffset,int consensusLength,
float gcContent,float alpha)
: WMM(gc)
{
/*
This constructor performs training of the BWM
*/
// Set some things in the base class
setStrand(FORWARD_STRAND);
setSignalType(signalType);
// Compute background single-nucleotide probabilities
float atContent=1-gcContent;
float AT=atContent/2, GC=gcContent/2;
BOOM::Array1D<float> background(alphabet.getNumElements());
background.setAllTo(0);
Symbol A=alphabet.lookup('A'), T=alphabet.lookup('T');
Symbol C=alphabet.lookup('C'), G=alphabet.lookup('G');
background[A]=AT;
background[T]=AT;
background[C]=GC;
background[G]=GC;
// Allocate the underlying WMM matrix
float pseudocount=0;
int n=sequences.size();
int len=sequences[0]->getLength();
setSizes(consensusLength,consensusOffset,len);
int nAlpha=alphabet.getNumElements();
matrix.resize(len,nAlpha);
matrix.setAllTo(pseudocount);
// Iterate through the training sequences & collect counts
BOOM::FloatArray1D effectiveSize(len);
effectiveSize.setAllTo(0);
for(int i=0 ; i<n ; ++i)
{
TrainingSequence &seq=*sequences[i];
int l=seq.getLength();
if(l!=len) throw BOOM::String("length mismatch in BWM: ")+len+" vs "+l;
for(int pos=0 ; pos<len ; ++pos)
{
Symbol s=seq[pos];
int count=seq.getBoostCount();
matrix[pos][s]+=count;
effectiveSize[pos]+=count;
}
}
// Perform a binomial test at each position to see if frequencies
// are significantly different from the background frequencies
numSignificantPositions=0;
for(int pos=0 ; pos<len ; ++pos)
{
// First, we have to identify the most extreme count
int mostExtremeDiff=0;
Symbol mostExtremeSymbol=0;
int sampleSize=int(effectiveSize[pos]+5/9.0);
for(Symbol s=0 ; s<nAlpha ; ++s)
{
int expected=int(background[s]*sampleSize+5/9.0);
int observed=(int)matrix[pos][s];
int diff=abs(expected-observed);
if(diff>mostExtremeDiff)
{
mostExtremeDiff=diff;
mostExtremeSymbol=s;
}
}
// Apply binomial distribution to get P-value for this count
float p=background[mostExtremeSymbol];
int expected=int(p*sampleSize+5/9.0);
int observed=expected+mostExtremeDiff; // +/- are symmetric, so use +
double P=
BinomialDistribution::rightTailedPValue(observed,sampleSize,p);
// If not significantly different from background, then the observed
// frequencies are probably just a statistical fluctation due to
// small sample size, so use the background probabilities instead
if(P>alpha)
{
effectiveSize[pos]=0;
for(Symbol s=0 ; s<nAlpha ; ++s)
{
matrix[pos][s]=background[s]*sampleSize;
effectiveSize[pos]+=matrix[pos][s];
}
}
else
{
cout<<pos<<" : using OBSERVED frequencies, p="<<P<<endl;
++numSignificantPositions;
}
}
// Normalize counts into probabilities
for(int pos=0 ; pos<len ; ++pos)
for(Symbol s=0 ; s<nAlpha ; ++s)
matrix[pos][s]/=effectiveSize[pos];
// Convert probabilities to log-probabilities
convertToLogs();
}
