Mercurial > repos > lsong10 > psiclass
view PsiCLASS-1.0.2/TranscriptDecider.cpp @ 0:903fc43d6227 draft default tip
Uploaded
author | lsong10 |
---|---|
date | Fri, 26 Mar 2021 16:52:45 +0000 |
parents | |
children |
line wrap: on
line source
#include "TranscriptDecider.hpp" void TranscriptDecider::OutputTranscript( int sampleId, struct _subexon *subexons, struct _transcript &transcript ) { int i, j ; // determine the strand std::vector<int> subexonInd ; transcript.seVector.GetOnesIndices( subexonInd ) ; // Determine the strand char strand[2] = "." ; int size = subexonInd.size() ; if ( size > 1 ) { // locate the intron showed up in this transcript. for ( i = 0 ; i < size - 1 ; ++i ) { /*int nextCnt = subexons[ subexonInd[i] ].nextCnt ; if ( nextCnt == 0 ) continue ; for ( j = 0 ; j < nextCnt ; ++j ) { int a = subexons[ subexonInd[i] ].next[j] ; if ( subexonInd[i + 1] == a && subexons[ subexonInd[i] ].end + 1 < subexons[a].start ) // avoid the case like ..(...[... { break ; } } if ( j < nextCnt )*/ if ( subexons[ subexonInd[i] ].end + 1 < subexons[ subexonInd[i + 1] ].start ) { if ( subexons[ subexonInd[i] ].rightStrand == 1 ) strand[0] = '+' ; else if ( subexons[ subexonInd[i] ].rightStrand == -1 ) strand[0] = '-' ; break ; } } } // TODO: transcript_id char *chrom = alignments.GetChromName( subexons[0].chrId ) ; char prefix[10] = "" ; struct _subexon *catSubexons = new struct _subexon[ size + 1 ] ; // Concatenate adjacent subexons catSubexons[0] = subexons[ subexonInd[0] ] ; j = 1 ; for ( i = 1 ; i < size ; ++i ) { if ( subexons[ subexonInd[i] ].start == catSubexons[j - 1].end + 1 ) { catSubexons[j - 1].end = subexons[ subexonInd[i] ].end ; } else { catSubexons[j] = subexons[ subexonInd[i] ] ; ++j ; } } size = j ; int gid = GetTranscriptGeneId( subexonInd, subexons ) ; if ( 0 ) //numThreads <= 1 ) { fprintf( outputFPs[sampleId], "%s\tCLASSES\ttranscript\t%d\t%d\t1000\t%s\t.\tgene_id \"%s%s.%d\"; transcript_id \"%s%s.%d.%d\"; Abundance \"%.6lf\";\n", chrom, catSubexons[0].start + 1, catSubexons[size - 1].end + 1, strand, prefix, chrom, gid, prefix, chrom, gid, transcriptId[ gid - baseGeneId ], transcript.FPKM ) ; for ( i = 0 ; i < size ; ++i ) { fprintf( outputFPs[ sampleId ], "%s\tCLASSES\texon\t%d\t%d\t1000\t%s\t.\tgene_id \"%s%s.%d\"; " "transcript_id \"%s%s.%d.%d\"; exon_number \"%d\"; Abundance \"%.6lf\"\n", chrom, catSubexons[i].start + 1, catSubexons[i].end + 1, strand, prefix, chrom, gid, prefix, chrom, gid, transcriptId[ gid - baseGeneId ], i + 1, transcript.FPKM ) ; } } else { struct _outputTranscript t ; int len = 0 ; t.chrId = subexons[0].chrId ; t.geneId = gid ; t.transcriptId = transcriptId[ gid - baseGeneId ] ; t.FPKM = transcript.FPKM ; t.sampleId = sampleId ; t.exons = new struct _pair32[size] ; for ( i = 0 ; i < size ; ++i ) { t.exons[i].a = catSubexons[i].start + 1 ; t.exons[i].b = catSubexons[i].end + 1 ; len += t.exons[i].b - t.exons[i].a + 1 ; } t.cov = transcript.abundance * alignments.readLen / len ; t.ecnt = size ; t.strand = strand[0] ; //printf( "%lf\n", transcript.correlationScore ) ; if ( numThreads > 1 ) outputHandler->Add( t ) ; else outputHandler->Add_SingleThread( t ) ; } ++transcriptId[ gid - baseGeneId ] ; delete[] catSubexons ; } int TranscriptDecider::GetFather( int f, int *father ) { if ( father[f] != f ) return father[f] = GetFather( father[f], father ) ; return f ; } int TranscriptDecider::GetTranscriptGeneId( std::vector<int> &subexonInd, struct _subexon *subexons ) { int i ; int size = subexonInd.size() ; for ( i = 0 ; i < size ; ++i ) if ( subexons[ subexonInd[i] ].geneId != -2 ) return subexons[ subexonInd[i] ].geneId ; // Some extreme case, where all the regions are mixture regions. for ( i = 0 ; i < size - 1 ; ++i ) if ( subexons[ subexonInd[i] ].end + 1 < subexons[ subexonInd[i + 1] ].start ) { return defaultGeneId[ ( subexons[ subexonInd[i] ].rightStrand + 1 ) / 2 ] ; } return defaultGeneId[0] ; } int TranscriptDecider::GetTranscriptGeneId( struct _transcript &t, struct _subexon *subexons ) { if ( subexons[ t.first ].geneId != -2 ) return subexons[ t.first ].geneId ; if ( subexons[ t.last ].geneId != -2 ) return subexons[ t.last ].geneId ; std::vector<int> subexonInd ; t.seVector.GetOnesIndices( subexonInd ) ; return GetTranscriptGeneId( subexonInd, subexons ) ; } void TranscriptDecider::InitTranscriptId() { int i ; for ( i = 0 ; i < usedGeneId - baseGeneId ; ++i ) transcriptId[i] = 0 ; } bool TranscriptDecider::IsStartOfMixtureStrandRegion( int tag, struct _subexon *subexons, int seCnt ) { int j, k ; int leftStrandCnt[2] = {0, 0}, rightStrandCnt[2] = {0, 0}; for ( j = tag + 1 ; j < seCnt ; ++j ) if ( subexons[j].start > subexons[j - 1].end + 1 ) break ; for ( k = tag ; k < j ; ++k ) { if ( subexons[k].leftStrand != 0 ) ++leftStrandCnt[ ( subexons[k].leftStrand + 1 ) / 2 ] ; if ( subexons[k].rightStrand != 0 ) ++rightStrandCnt[ ( subexons[k].rightStrand + 1 ) / 2 ] ; } if ( rightStrandCnt[0] > 0 && leftStrandCnt[0] == 0 && leftStrandCnt[1] > 0 ) return true ; if ( rightStrandCnt[1] > 0 && leftStrandCnt[1] == 0 && leftStrandCnt[0] > 0 ) return true ; return false ; } // Return 0 - uncompatible or does not overlap at all. 1 - fully compatible. 2 - Head of the constraints compatible with the tail of the transcript // the partial compatible case (return 2) mostly likely happen in DP where we have partial transcript. int TranscriptDecider::IsConstraintInTranscript( struct _transcript transcript, struct _constraint &c ) { //printf( "%d %d, %d %d\n", c.first, c.last, transcript.first, transcript.last ) ; if ( c.first < transcript.first || c.first > transcript.last || !transcript.seVector.Test( c.first ) || ( !transcript.partial && !transcript.seVector.Test( c.last ) ) ) // no overlap or starts too early or some chosen subexons does not compatible return 0 ; // Extract the subexons we should focus on. int s, e ; s = c.first ; e = c.last ; bool returnPartial = false ; if ( e > transcript.last ) // constraints ends after the transcript. { if ( transcript.partial ) { e = transcript.last ; returnPartial = true ; } else return 0 ; } /*printf( "%s: %d %d: (%d %d) (%d %d)\n", __func__, s, e, transcript.seVector.Test(0), transcript.seVector.Test(1), c.vector.Test(0), c.vector.Test(1) ) ;*/ compatibleTestVectorT.Assign( transcript.seVector ) ; //compatibleTestVectorT.MaskRegionOutsideInRange( s, e, transcript.first, transcript.last ) ; compatibleTestVectorT.MaskRegionOutside( s, e ) ; compatibleTestVectorC.Assign( c.vector ) ; if ( c.last > transcript.last ) { //compatibleTestVectorC.MaskRegionOutsideInRange( s, e, c.first, c.last ) ; //compatibleTestVectorC.MaskRegionOutside( s, e ) ; compatibleTestVectorC.MaskRegionOutside( 0, e ) ; // Because the bits before s are already all 0s in C. } /*printf( "after masking %d %d. %d %d %d %d:\n", s, e, transcript.first, transcript.last, c.first, c.last ) ; compatibleTestVectorT.Print() ; compatibleTestVectorC.Print() ; */ // Test compatible. int ret = 0 ; if ( compatibleTestVectorT.IsEqual( compatibleTestVectorC ) ) { if ( returnPartial ) ret = 2 ; else ret = 1 ; } return ret ; } int TranscriptDecider::IsConstraintInTranscriptDebug( struct _transcript transcript, struct _constraint &c ) { //printf( "%d %d, %d %d\n", c.first, c.last, transcript.first, transcript.last ) ; if ( c.first < transcript.first || c.first > transcript.last ) // no overlap or starts too early. return 0 ; printf( "hi\n" ) ; // Extract the subexons we should focus on. int s, e ; s = c.first ; e = c.last ; bool returnPartial = false ; if ( e > transcript.last ) // constraints ends after the transcript. { if ( transcript.partial ) { e = transcript.last ; returnPartial = true ; } else return 0 ; } /*printf( "%s: %d %d: (%d %d) (%d %d)\n", __func__, s, e, transcript.seVector.Test(0), transcript.seVector.Test(1), c.vector.Test(0), c.vector.Test(1) ) ;*/ compatibleTestVectorT.Assign( transcript.seVector ) ; compatibleTestVectorT.MaskRegionOutside( s, e ) ; compatibleTestVectorC.Assign( c.vector ) ; if ( e > transcript.last ) compatibleTestVectorC.MaskRegionOutside( s, e ) ; /*printf( "after masking: (%d %d) (%d %d)\n", compatibleTestVectorT.Test(0), compatibleTestVectorT.Test(1), compatibleTestVectorC.Test(0), compatibleTestVectorC.Test(1) ) ;*/ // Test compatible. int ret = 0 ; if ( compatibleTestVectorT.IsEqual( compatibleTestVectorC ) ) { if ( returnPartial ) ret = 2 ; else ret = 1 ; } compatibleTestVectorT.Print() ; compatibleTestVectorC.Print() ; printf( "ret=%d\n", ret ) ; return ret ; } int TranscriptDecider::SubTranscriptCount( int tag, struct _subexon *subexons, int *f ) { if ( f[tag] != -1 ) return f[tag] ; int ret = 0 ; int i ; if ( subexons[tag].canBeEnd ) ret = 1 ; for ( i = 0 ; i < subexons[tag].nextCnt ; ++i ) { ret += SubTranscriptCount( subexons[tag].next[i], subexons, f ) ; } if ( ret == 0 ) ret = 1 ; return f[tag] = ret ; } void TranscriptDecider::CoalesceSameTranscripts( std::vector<struct _transcript> &t ) { int i, k ; if ( t.size() == 0 ) return ; std::sort( t.begin(), t.end(), CompSortTranscripts ) ; int size = t.size() ; k = 0 ; for ( i = 1 ; i < size ; ++i ) { if ( t[k].seVector.IsEqual( t[i].seVector ) ) { t[k].abundance += t[i].abundance ; t[i].seVector.Release() ; } else { ++k ; if ( i != k ) t[k] = t[i] ; } } t.resize( k + 1 ) ; } void TranscriptDecider::EnumerateTranscript( int tag, int strand, int visit[], int vcnt, struct _subexon *subexons, SubexonCorrelation &correlation, double correlationScore, std::vector<struct _transcript> &alltranscripts, int &atcnt ) { int i ; visit[ vcnt ] = tag ; //printf( "%s: %d %d %d %d. %d %d\n", __func__, vcnt, tag, subexons[tag].nextCnt, strand, subexons[tag].start, subexons[tag].end ) ; // Compute the correlation score double minCor = correlationScore ; for ( i = 0 ; i < vcnt ; ++i ) { double tmp = correlation.Query( visit[i], visit[vcnt] ) ; if ( tmp < minCor ) minCor = tmp ; } if ( subexons[tag].canBeEnd ) { struct _transcript &txpt = alltranscripts[atcnt] ; for ( i = 0 ; i <= vcnt ; ++i ) txpt.seVector.Set( visit[i] ) ; txpt.first = visit[0] ; txpt.last = visit[vcnt] ; txpt.partial = false ; txpt.correlationScore = minCor ; //printf( "%lf %d %d ", txpt.correlationScore, vcnt, visit[0] ) ; //txpt.seVector.Print() ; ++atcnt ; } for ( i = 0 ; i < subexons[tag].nextCnt ; ++i ) { int a = subexons[tag].next[i] ; if ( !SubexonGraph::IsSameStrand( subexons[tag].rightStrand, strand ) && subexons[a].start > subexons[tag].end + 1 ) continue ; int backupStrand = strand ; if ( subexons[a].start > subexons[tag].end + 1 && strand == 0 ) strand = subexons[tag].rightStrand ; EnumerateTranscript( subexons[tag].next[i], strand, visit, vcnt + 1, subexons, correlation, minCor, alltranscripts, atcnt ) ; strand = backupStrand ; } } void TranscriptDecider::SearchSubTranscript( int tag, int strand, int parents[], int pcnt, struct _dp &pdp, int visit[], int vcnt, int extends[], int extendCnt, std::vector<struct _constraint> &tc, int tcStartInd, struct _dpAttribute &attr ) { int i ; int size ; double cover ; bool keepSearch = true ; bool belowMin = false ; struct _subexon *subexons = attr.subexons ; visit[vcnt] = tag ; ++vcnt ; struct _dp visitdp ; visitdp.cover = -1 ; struct _transcript &subTxpt = attr.bufferTxpt ; subTxpt.seVector.Reset() ; for ( i = 0 ; i < pcnt ; ++i ) subTxpt.seVector.Set( parents[i] ) ; subTxpt.first = parents[0] ; subTxpt.last = parents[ pcnt - 1] ; for ( i = 0 ; i < vcnt ; ++i ) subTxpt.seVector.Set( visit[i] ) ; subTxpt.last = visit[ vcnt - 1 ] ; subTxpt.partial = true ; // Adjust the extendsCnt /*printf( "%s: %d %d %d\n", __func__, vcnt , extendCnt, extends[ extendCnt - 1] ) ; subTxpt.seVector.Print() ; tc[extends[extendCnt - 1]].vector.Print() ; printf( "Adjust extend:\n") ;*/ for ( i = extendCnt - 1 ; i >= 0 ; --i ) { if ( tc[ extends[i] ].last <= tag || ( tc[ extends[i] ].vector.Test( tag ) && IsConstraintInTranscript( subTxpt, tc[ extends[i] ] ) != 0 ) ) break ; } extendCnt = i + 1 ; // If the extension ends. subTxpt.partial = false ; if ( subexons[tag].nextCnt > 0 && ( extendCnt == 0 || tag >= tc[ extends[ extendCnt - 1 ] ].last ) ) { // Solve the subtranscript beginning with visit. // Now we got the optimal transcript for visit. visitdp = SolveSubTranscript( visit, vcnt, strand, tc, tcStartInd, attr ) ; keepSearch = false ; } //printf( "%s %d %d: visitdp.cover=%lf\n", __func__, parents[0], tag, visitdp.cover ) ; // the constraints across the parents and visit. size = tc.size() ; if ( visitdp.cover >= 0 ) { cover = visitdp.cover ; // Reset the subTxpt, since its content is modofitied in SolveSubTxpt called above. subTxpt.seVector.Reset() ; for ( i = 0 ; i < pcnt ; ++i ) subTxpt.seVector.Set( parents[i] ) ; subTxpt.seVector.Or( visitdp.seVector ) ; subTxpt.first = parents[0] ; subTxpt.last = visitdp.last ; subTxpt.partial = false ; if ( !attr.forAbundance && attr.minAbundance > 0 ) { for ( i = 0 ; i < pcnt - 1 ; ++i ) { if ( attr.uncoveredPair.find( parents[i] * attr.seCnt + parents[i + 1] ) != attr.uncoveredPair.end() ) belowMin = true ; } for ( i = -1 ; i < vcnt - 1 ; ++i ) { if ( i == -1 && pcnt >= 1 ) { if ( attr.uncoveredPair.find( parents[pcnt - 1] * attr.seCnt + visit[0] ) != attr.uncoveredPair.end() ) belowMin = true ; } else { if ( attr.uncoveredPair.find( visit[i] * attr.seCnt + visit[i + 1] ) != attr.uncoveredPair.end() ) belowMin = true ; } } if ( attr.forAbundance && belowMin ) cover = 1e-6 ; } for ( i = tcStartInd ; i < size ; ++i ) { if ( tc[i].first > parents[ pcnt - 1] ) break ; if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 1 ) { if ( tc[i].normAbund <= attr.minAbundance ) { belowMin = true ; cover = -2 ; break ; } if ( tc[i].abundance <= 0 ) continue ; if ( attr.forAbundance ) { if ( tc[i].normAbund < cover || cover == 0 ) cover = tc[i].normAbund ; } else { ++cover ; } } } if ( belowMin && pdp.cover == -1 ) { pdp.cover = -2 ; pdp.seVector.Assign( subTxpt.seVector ) ; pdp.first = subTxpt.first ; pdp.last = subTxpt.last ; pdp.strand = strand ; } else if ( cover > pdp.cover ) { pdp.cover = cover ; pdp.seVector.Assign( subTxpt.seVector ) ; pdp.first = subTxpt.first ; pdp.last = subTxpt.last ; pdp.strand = strand ; } } else if ( visitdp.cover == -2 && pdp.cover == -1 ) // no valid extension from visit { subTxpt.seVector.Reset() ; for ( i = 0 ; i < pcnt ; ++i ) subTxpt.seVector.Set( parents[i] ) ; subTxpt.seVector.Or( visitdp.seVector ) ; subTxpt.first = parents[0] ; subTxpt.last = visitdp.last ; pdp.cover = -2 ; pdp.seVector.Assign( subTxpt.seVector ) ; pdp.first = subTxpt.first ; pdp.last = subTxpt.last ; pdp.strand = strand ; } if ( subexons[tag].canBeEnd && ( visitdp.cover < 0 || attr.forAbundance ) ) // This works is because that the extension always covers more constraints. So we only go this branch if the extension does not work // and it goes this branch if it violates minAbundance // But we need to go here when we want to compute the maxAbundance transcript. // This part also works as the exit point of the recurive function. { bool belowMin = false ; subTxpt.seVector.Reset() ; for ( i = 0 ; i < pcnt ; ++i ) subTxpt.seVector.Set( parents[i] ) ; for ( i = 0 ; i < vcnt ; ++i ) subTxpt.seVector.Set( visit[i] ) ; subTxpt.first = parents[0] ; subTxpt.last = visit[ vcnt - 1] ; subTxpt.partial = false ; cover = 0 ; if ( attr.forAbundance || attr.minAbundance > 0 ) { for ( i = 0 ; i < pcnt - 1 ; ++i ) { if ( attr.uncoveredPair.find( parents[i] * attr.seCnt + parents[i + 1] ) != attr.uncoveredPair.end() ) belowMin = true ; } for ( i = -1 ; i < vcnt - 1 ; ++i ) { if ( i == -1 && pcnt >= 1 ) { if ( attr.uncoveredPair.find( parents[pcnt - 1] * attr.seCnt + visit[0] ) != attr.uncoveredPair.end() ) belowMin = true ; } else { if ( attr.uncoveredPair.find( visit[i] * attr.seCnt + visit[i + 1] ) != attr.uncoveredPair.end() ) belowMin = true ; } } //if ( belowMin == true ) // printf( "turned belowMin. %d. %d %d: %d %d %d\n", attr.uncoveredPair.size(), pcnt, vcnt, parents[0], visit[0], visit[ vcnt - 1] ) ; if ( attr.forAbundance && belowMin ) cover = 1e-6 ; } for ( i = tcStartInd ; i < size ; ++i ) { // note that the value is parents[ pcnt - 1], because // in above the part of "visit" is computed in SolveSubTranscript( visit ). if ( tc[i].first > visit[ vcnt - 1] ) break ; if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 1 ) { if ( tc[i].normAbund <= attr.minAbundance ) { belowMin = true ; cover = -2 ; break ; } if ( tc[i].abundance <= 0 ) continue ; if ( attr.forAbundance ) { if ( tc[i].normAbund < cover || cover == 0 ) cover = tc[i].normAbund ; } else { ++cover ; } } } if ( belowMin && pdp.cover == -1 ) { pdp.cover = -2 ; pdp.seVector.Assign( subTxpt.seVector ) ; pdp.first = subTxpt.first ; pdp.last = subTxpt.last ; pdp.strand = strand ; } else if ( cover > pdp.cover ) { pdp.cover = cover ; pdp.seVector.Assign( subTxpt.seVector ) ; pdp.first = subTxpt.first ; pdp.last = subTxpt.last ; pdp.strand = strand ; } } //printf( "%s %d: pdp.cover=%lf\n", __func__, tag, pdp.cover ) ; // keep searching. if ( keepSearch ) { for ( i = 0 ; i < subexons[tag].nextCnt ; ++i ) { int b = subexons[tag].next[i] ; if ( ( SubexonGraph::IsSameStrand( subexons[tag].rightStrand, strand ) && SubexonGraph::IsSameStrand( subexons[b].leftStrand, strand ) ) || subexons[b].start == subexons[tag].end + 1 ) { int backupStrand = strand ; if ( subexons[b].start > subexons[tag].end + 1 ) strand = subexons[tag].rightStrand ; SearchSubTranscript( subexons[tag].next[i], strand, parents, pcnt, pdp, visit, vcnt, extends, extendCnt, tc, tcStartInd, attr ) ; strand = backupStrand ; } } } return ; } struct _dp TranscriptDecider::SolveSubTranscript( int visit[], int vcnt, int strand, std::vector<struct _constraint> &tc, int tcStartInd, struct _dpAttribute &attr ) { int i ; int size ; /*printf( "%s: ", __func__ ) ; for ( i = 0 ; i < vcnt ; ++i ) printf( "%d ", visit[i] ) ; printf( ": %lf %d %d", attr.f1[ visit[0] ].cover, attr.f1[ visit[0] ].timeStamp, attr.timeStamp ) ; printf( "\n" ) ;*/ // Test whether it is stored in dp if ( vcnt == 1 ) { if ( attr.f1[ visit[0] ].cover != -1 && attr.f1[ visit[0] ].strand == strand && ( attr.f1[ visit[0] ].timeStamp == attr.timeStamp || ( attr.f1[ visit[0] ].minAbundance < attr.minAbundance && attr.f1[visit[0]].cover == -2 ) ) ) //even given lower minAbundance threshold, it fails { return attr.f1[ visit[0] ] ; } } else if ( vcnt == 2 && attr.f2 ) { int a = visit[0] ; int b = visit[1] ; if ( attr.f2[a][b].cover != -1 && attr.f2[a][b].strand == strand && ( attr.f2[a][b].timeStamp == attr.timeStamp || ( attr.f2[a][b].minAbundance < attr.minAbundance && attr.f2[a][b].cover == -2 ) ) ) { return attr.f2[a][b] ; } } else { int key = 0 ; for ( i = 0 ; i < vcnt ; ++i ) key = ( key * attr.seCnt + visit[i] ) % hashMax ; if ( key < 0 ) key += hashMax ; if ( attr.hash[key].cover != -1 && attr.hash[key].cnt == vcnt && attr.hash[key].strand == strand && ( attr.hash[key].first == visit[0] ) && ( attr.hash[key].timeStamp == attr.timeStamp || ( attr.hash[key].minAbundance < attr.minAbundance && attr.hash[key].cover == -2 ) ) ) { struct _transcript subTxpt = attr.bufferTxpt ; subTxpt.seVector.Reset() ; for ( i = 0 ; i < vcnt ; ++i ) subTxpt.seVector.Set( visit[i] ) ; //subTxpt.seVector.Print() ; //attr.hash[key].seVector.Print() ; subTxpt.seVector.Xor( attr.hash[key].seVector ) ; subTxpt.seVector.MaskRegionOutside( visit[0], visit[ vcnt - 1] ) ; //printf( "hash test: %d %d\n", key, subTxpt.seVector.IsAllZero() ) ; if ( subTxpt.seVector.IsAllZero() ) { return attr.hash[key] ; } // Can't use the code below, because vcnt is the header of subexons. /*for ( i = 0 ; i < vcnt ; ++i ) if ( !attr.hash[key].seVector.Test( visit[i] ) ) break ; if ( i >= vcnt ) return attr.hash[key] ;*/ } } // adjust tcStartInd size = tc.size() ; for ( i = tcStartInd ; i < size ; ++i ) if ( tc[i].first >= visit[0] ) break ; tcStartInd = i ; struct _subexon *subexons = attr.subexons ; struct _dp visitdp ; visitdp.seVector.Init( attr.seCnt ) ; visitdp.cover = -1 ; struct _transcript &subTxpt = attr.bufferTxpt ; // This happens when it is called from PickTranscriptsByDP, the first subexon might be the end. subTxpt.seVector.Reset() ; for ( i = 0 ; i < vcnt ; ++i ) subTxpt.seVector.Set( visit[i] ) ; subTxpt.first = visit[0] ; subTxpt.last = visit[vcnt - 1] ; if ( subexons[ visit[vcnt - 1] ].canBeEnd ) { subTxpt.partial = false ; double cover = 0 ; for ( i = tcStartInd ; i < size ; ++i ) { if ( tc[i].first > subTxpt.last ) break ; if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 1 ) { if ( tc[i].normAbund <= attr.minAbundance ) { cover = -2 ; break ; } if ( tc[i].abundance <= 0 ) continue ; if ( attr.forAbundance ) { if ( tc[i].normAbund < cover || cover == 0 ) cover = tc[i].normAbund ; } else ++cover ; } } visitdp.seVector.Assign( subTxpt.seVector ) ; visitdp.cover = cover ; visitdp.first = subTxpt.first ; visitdp.last = subTxpt.last ; visitdp.strand = strand ; } // Now we extend. size = tc.size() ; int *extends = new int[tc.size() - tcStartInd + 1] ; int extendCnt = 0 ; subTxpt.partial = true ; for ( i = tcStartInd ; i < size ; ++i ) { if ( tc[i].first > subTxpt.last ) break ; if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 2 ) { extends[extendCnt] = i ; ++extendCnt ; } } // Sort the extend by the index of the last subexon. if ( extendCnt > 0 ) { struct _pair32 *extendsPairs = new struct _pair32[extendCnt] ; for ( i = 0 ; i < extendCnt ; ++i ) { extendsPairs[i].a = extends[i] ; extendsPairs[i].b = tc[ extends[i] ].last ; } qsort( extendsPairs, extendCnt, sizeof( struct _pair32 ), CompPairsByB ) ; for ( i = 0 ; i < extendCnt ; ++i ) extends[i] = extendsPairs[i].a ; delete[] extendsPairs ; } size = subexons[ visit[vcnt - 1] ].nextCnt ; int nextvCnt = 1 ; if ( extendCnt > 0 && tc[ extends[ extendCnt - 1 ] ].last - visit[ vcnt - 1 ] > 1 ) nextvCnt = tc[ extends[ extendCnt - 1 ] ].last - visit[ vcnt - 1 ] ; int *nextv = new int[ nextvCnt ] ; for ( i = 0 ; i < size ; ++i ) { int a = visit[vcnt - 1] ; int b = subexons[a].next[i] ; if ( ( SubexonGraph::IsSameStrand( subexons[a].rightStrand, strand ) && SubexonGraph::IsSameStrand( subexons[b].leftStrand, strand ) ) || subexons[b].start == subexons[a].end + 1 ) { int backupStrand = strand ; if ( subexons[b].start > subexons[a].end + 1 ) strand = subexons[a].rightStrand ; SearchSubTranscript( subexons[ visit[vcnt - 1] ].next[i], strand, visit, vcnt, visitdp, nextv, 0, extends, extendCnt, tc, tcStartInd, attr ) ; strand = backupStrand ; } } //printf( "%s %d(%d) %d %d %d: %lf\n", __func__, visit[0], subexons[ visit[vcnt - 1] ].canBeEnd, size, extendCnt, strand, visitdp.cover ) ; delete[] nextv ; delete[] extends ; // store the result in the dp structure. // We return the structure stored in dp to simplify the memory access pattern. // In other words, we assume the structure returned from this function always uses the memory from attr.dp if ( vcnt == 1 ) { SetDpContent( attr.f1[ visit[0] ], visitdp, attr ) ; visitdp.seVector.Release() ; return attr.f1[ visit[0] ] ; } else if ( vcnt == 2 && attr.f2 ) { SetDpContent( attr.f2[ visit[0] ][ visit[1] ], visitdp, attr ) ; visitdp.seVector.Release() ; return attr.f2[ visit[0] ][ visit[1] ] ; } else { int key = 0 ; for ( i = 0 ; i < vcnt ; ++i ) key = ( key * attr.seCnt + visit[i] ) % hashMax ; if ( key < 0 ) key += hashMax ; //static int hashUsed = 0 ; //if ( attr.hash[key].cover == -1 ) // ++hashUsed ; //printf( "%d/%d\n", hashUsed, HASH_MAX) ; //printf( "hash write: %d\n", key ) ; SetDpContent( attr.hash[key], visitdp, attr ) ; attr.hash[key].cnt = vcnt ; visitdp.seVector.Release() ; return attr.hash[key] ; } } void TranscriptDecider::PickTranscriptsByDP( struct _subexon *subexons, int seCnt, int iterBound, Constraints &constraints, SubexonCorrelation &correlation, struct _dpAttribute &attr, std::vector<struct _transcript> &alltranscripts ) { int i, j, k ; std::vector<struct _transcript> transcripts ; std::vector<struct _constraint> &tc = constraints.constraints ; int tcCnt = tc.size() ; int coalesceThreshold = 1024 ; //printf( "tcCnt=%d\n", tcCnt ) ; attr.timeStamp = 1 ; attr.bufferTxpt.seVector.Init( seCnt ) ; attr.subexons = subexons ; attr.seCnt = seCnt ; double maxAbundance = -1 ; // Initialize the dp data structure /*memset( attr.f1, -1, sizeof( struct _dp ) * seCnt ) ; for ( i = 0 ; i < seCnt ; ++i ) memset( attr.f2[i], -1, sizeof( struct _dp ) * seCnt ) ; memset( attr.hash, -1, sizeof( struct _dp ) * HASH_MAX ) ;*/ for ( i = 0 ; i < seCnt ; ++i ) ResetDpContent( attr.f1[i] ) ; for ( i = 0 ; i < seCnt && attr.f2 ; ++i ) for ( j = i ; j < seCnt ; ++j ) ResetDpContent( attr.f2[i][j] ) ; for ( i = 0 ; i < hashMax ; ++i ) ResetDpContent( attr.hash[i] ) ; // Set the uncovered pair attr.uncoveredPair.clear() ; BitTable bufferTable( seCnt ) ; k = 0 ; for ( i = 0 ; i < seCnt ; ++i ) { for ( ; k < tcCnt ; ++k ) { if ( tc[k].last >= i ) break ; } if ( k >= tcCnt || tc[k].first > i ) { for ( j = 0 ; j < subexons[i].nextCnt ; ++j ) { attr.uncoveredPair[i * seCnt + subexons[i].next[j] ] = 1 ; } continue ; } for ( j = 0 ; j < subexons[i].nextCnt ; ++j ) { bool covered = false ; int l, n ; n = subexons[i].next[j] ; for ( l = k ; l < tcCnt ; ++l ) { if ( tc[l].first > i ) break ; if ( tc[l].vector.Test( i ) && tc[l].vector.Test( n ) ) { if ( n == i + 1 ) { covered = true ; break ; } else { bufferTable.Assign( tc[l].vector ) ; bufferTable.MaskRegionOutside( i + 1, n - 1 ) ; if ( bufferTable.IsAllZero() ) { covered = true ; break ; } } } } if ( !covered ) { //printf( "set!: (%d: %d %d) (%d: %d %d)\n", i, subexons[i].start, subexons[i].end, n, subexons[n].start, subexons[n].end ) ; attr.uncoveredPair[ i * seCnt + n ] = 1 ; } } } bufferTable.Release() ; // Find the max abundance attr.forAbundance = true ; attr.minAbundance = 0 ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].canBeStart ) { int visit[1] = {i} ; struct _dp tmp ; tmp = SolveSubTranscript( visit, 1, 0, tc, 0, attr ) ; if ( tmp.cover > maxAbundance ) maxAbundance = tmp.cover ; } } //PrintLog( "maxAbundance=%lf", maxAbundance ) ; //exit( 1 ) ; // Pick the transcripts. Quantative Set-Cover // Notice that by the logic in SearchSubTxpt and SolveSubTxpt, we don't need to reinitialize the data structure. attr.forAbundance = false ; int *coveredTc = new int[tcCnt] ; int coveredTcCnt ; struct _dp maxCoverDp ; struct _dp bestDp ; std::map<double, struct _dp> cachedCoverResult ; maxCoverDp.seVector.Init( seCnt ) ; bestDp.seVector.Init( seCnt ) ; int iterCnt = 0 ; while ( 1 ) { double bestScore ; // iterately assign constraints attr.minAbundance = 0 ; // Find the best candidate transcript. bestDp.cover = -1 ; bestScore = -1 ; while ( 1 ) { // iterate the change of minAbundance if ( cachedCoverResult.find( attr.minAbundance ) != cachedCoverResult.end() ) { struct _dp tmp = cachedCoverResult[ attr.minAbundance ] ; SetDpContent( maxCoverDp, tmp, attr ) ; } else { maxCoverDp.cover = -1 ; ++attr.timeStamp ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].canBeStart == false ) continue ; int visit[1] = {i} ; struct _dp tmp ; tmp = SolveSubTranscript( visit, 1, 0, tc, 0, attr ) ; if ( tmp.cover > maxCoverDp.cover && tmp.cover > 0 ) { SetDpContent( maxCoverDp, tmp, attr ) ; } //if ( subexons[i].start == 6870264 || subexons[i].start == 6872237 ) // printf( "%d: %lf\n", i, tmp.cover ) ; } if ( maxCoverDp.cover == -1 ) break ; struct _dp ccr ; ccr.seVector.Init( seCnt ) ; SetDpContent( ccr, maxCoverDp, attr ) ; cachedCoverResult[ attr.minAbundance ] = ccr ; } // the abundance for the max cover txpt. double min = -1 ; struct _transcript &subTxpt = attr.bufferTxpt ; subTxpt.seVector.Assign( maxCoverDp.seVector ) ; subTxpt.first = maxCoverDp.first ; subTxpt.last = maxCoverDp.last ; for ( i = 0 ; i < tcCnt ; ++i ) { if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 1 ) { if ( tc[i].normAbund < min || min == -1 ) min = tc[i].normAbund ; } } if ( attr.minAbundance == 0 ) { std::vector<int> subexonIdx ; maxCoverDp.seVector.GetOnesIndices( subexonIdx ) ; int size = subexonIdx.size() ; for ( i = 0 ; i < size - 1 ; ++i ) if ( attr.uncoveredPair.find( subexonIdx[i] * seCnt + subexonIdx[i + 1] ) != attr.uncoveredPair.end() ) { min = 1e-6 ; break ; } } double score = ComputeScore( maxCoverDp.cover, 1.0, min, maxAbundance, 0 ) ; if ( bestScore == -1 || score > bestScore ) { bestScore = score ; SetDpContent( bestDp, maxCoverDp, attr ) ; } else if ( score < bestScore ) { if ( ComputeScore( maxCoverDp.cover, 1.0, maxAbundance, maxAbundance, 0 ) < bestScore ) break ; } //PrintLog( "normAbund=%lf maxCoverDp.cover=%lf score=%lf timeStamp=%d", min, maxCoverDp.cover, score, attr.timeStamp ) ; attr.minAbundance = min ; } // end of iteration for minAbundance. if ( bestDp.cover == -1 ) break ; // Assign the constraints. coveredTcCnt = 0 ; double update = -1 ; struct _transcript &subTxpt = attr.bufferTxpt ; subTxpt.seVector.Assign( bestDp.seVector ) ; subTxpt.first = bestDp.first ; subTxpt.last = bestDp.last ; subTxpt.partial = false ; for ( i = 0 ; i < tcCnt ; ++i ) { if ( IsConstraintInTranscript( subTxpt, tc[i] ) == 1 ) { if ( tc[i].abundance > 0 && ( tc[i].abundance < update || update == -1 ) ) { update = tc[i].abundance ; } coveredTc[ coveredTcCnt ] = i ; ++coveredTcCnt ; } /*else { printf( "%d: ", i ) ; tc[i].vector.Print() ; if ( i == 127 ) { printf( "begin debug:\n" ) ; IsConstraintInTranscriptDebug( subTxpt, tc[i] ) ; } }*/ } update *= ( 1 + iterCnt / 50 ) ;//* ( 1 + iterCnt / 50 ) ; //PrintLog( "%d: update=%lf %d %d. %d %d %d", iterCnt, update, coveredTcCnt, tcCnt, // bestDp.first, bestDp.last, subexons[ bestDp.first ].start ) ; //bestDp.seVector.Print() ; struct _transcript nt ; nt.seVector.Duplicate( bestDp.seVector ) ; nt.first = bestDp.first ; nt.last = bestDp.last ; nt.partial = false ; nt.abundance = 0 ; for ( i = 0 ; i < coveredTcCnt ; ++i ) { j = coveredTc[i] ; if ( tc[j].abundance > 0 ) { double tmp = ( tc[j].abundance > update ? update : tc[j].abundance ) ; tc[j].abundance -= tmp ; double factor = 1 ; nt.abundance += ( tc[j].support * update / tc[j].normAbund * factor ) ; if ( tc[j].abundance <= 0 ) { std::vector<double> removeKey ; for ( std::map<double, struct _dp>::iterator it = cachedCoverResult.begin() ; it != cachedCoverResult.end() ; ++it ) { subTxpt.seVector.Assign( it->second.seVector ) ; subTxpt.first = it->second.first ; subTxpt.last = it->second.last ; subTxpt.partial = false ; if ( IsConstraintInTranscript( subTxpt, tc[j] ) == 1 ) { it->second.seVector.Release() ; removeKey.push_back( it->first ) ; } } int size = removeKey.size() ; int l ; for ( l = 0 ; l < size ; ++l ) cachedCoverResult.erase( removeKey[l] ) ; } } if ( tc[j].abundance < 0 ) tc[j].abundance = 0 ; } transcripts.push_back( nt ) ; if ( transcripts.size() >= transcripts.capacity() && (int)transcripts.size() >= coalesceThreshold ) { CoalesceSameTranscripts( transcripts ) ; if ( transcripts.size() >= transcripts.capacity() / 2 ) coalesceThreshold *= 2 ; } ++iterCnt ; if ( iterCnt >= iterBound ) break ; } CoalesceSameTranscripts( transcripts ) ; int size = transcripts.size() ; // Compute the correlation score for ( i = 0 ; i < size ; ++i ) { std::vector<int> subexonInd ; transcripts[i].seVector.GetOnesIndices( subexonInd ) ; double cor = 2.0 ; int s = subexonInd.size() ; for ( j = 0 ; j < s ; ++j ) for ( k = j + 1 ; k < s ; ++k ) { double tmp = correlation.Query( subexonInd[j], subexonInd[k] ) ; if ( tmp < cor ) cor = tmp ; } if ( cor > 1 ) cor = 0 ; transcripts[i].correlationScore = cor ; } // store the result for ( i = 0 ; i < size ; ++i ) alltranscripts.push_back( transcripts[i] ) ; // Release the memory for ( std::map<double, struct _dp>::iterator it = cachedCoverResult.begin() ; it != cachedCoverResult.end() ; ++it ) { it->second.seVector.Release() ; } attr.bufferTxpt.seVector.Release() ; delete[] coveredTc ; maxCoverDp.seVector.Release() ; bestDp.seVector.Release() ; } // Add the preifx/suffix of transcripts to the list void TranscriptDecider::AugmentTranscripts( struct _subexon *subexons, std::vector<struct _transcript> &alltranscripts, int limit, bool extend ) { int i, j, k ; int size = alltranscripts.size() ; if ( size >= limit ) return ; // Augment suffix, prefix transcripts for ( i = 0 ; i < size ; ++i ) { std::vector<int> subexonIdx ; alltranscripts[i].seVector.GetOnesIndices( subexonIdx ) ; int seIdxCnt = subexonIdx.size() ; // suffix for ( j = 1 ; j < seIdxCnt ; ++j ) { if ( subexons[ subexonIdx[j] ].canBeStart ) { struct _transcript nt ; nt.first = subexonIdx[j] ; nt.last = alltranscripts[i].last ; nt.seVector.Duplicate( alltranscripts[i].seVector ) ; nt.seVector.MaskRegionOutside( nt.first, nt.last ) ; nt.partial = false ; nt.correlationScore = 0 ; nt.abundance = 0 ; nt.constraintsSupport = NULL ; alltranscripts.push_back( nt ) ; if ( alltranscripts.size() >= limit ) return ; } } // prefix for ( j = 0 ; j < seIdxCnt - 1 ; ++j ) { if ( subexons[ subexonIdx[j] ].canBeEnd ) { struct _transcript nt ; nt.first = alltranscripts[i].first ; nt.last = subexonIdx[j] ; nt.seVector.Duplicate( alltranscripts[i].seVector ) ; nt.seVector.MaskRegionOutside( nt.first, nt.last ) ; nt.partial = false ; nt.correlationScore = 0 ; nt.abundance = 0 ; nt.constraintsSupport = NULL ; alltranscripts.push_back( nt ) ; if ( alltranscripts.size() >= limit ) return ; } } if ( extend ) { //Extentions right. for ( j = 0 ; j < seIdxCnt ; ++j ) { if ( subexons[ subexonIdx[j] ].nextCnt > 1 ) { for ( k = 0 ; k < subexons[ subexonIdx[j] ].nextCnt ; ++k ) { int idx = subexons[ subexonIdx[j] ].next[k] ; if ( alltranscripts[i].seVector.Test( idx ) ) continue ; int l ; std::vector<int> visited ; while ( 1 ) { if ( subexons[idx].nextCnt > 1 || subexons[idx].prevCnt > 1 ) { break ; } visited.push_back( idx ) ; if ( subexons[idx].canBeEnd && subexons[idx].nextCnt == 0 ) { struct _transcript nt ; nt.first = alltranscripts[i].first ; nt.last = idx ; nt.seVector.Duplicate( alltranscripts[i].seVector ) ; nt.seVector.MaskRegionOutside( nt.first, subexonIdx[j] ) ; int visitedSize = visited.size() ; for ( l = 0 ; l < visitedSize ; ++l ) nt.seVector.Set( visited[l] ) ; nt.partial = false ; nt.correlationScore = 0 ; nt.abundance = 0 ; nt.constraintsSupport = NULL ; alltranscripts.push_back( nt ) ; if ( alltranscripts.size() >= limit ) return ; } if ( subexons[idx].nextCnt == 1 ) idx = subexons[idx].next[0] ; else break ; } } } } // Extension towards left for ( j = 0 ; j < seIdxCnt ; ++j ) { if ( subexons[ subexonIdx[j] ].prevCnt > 1 ) { for ( k = 0 ; k < subexons[ subexonIdx[j] ].prevCnt ; ++k ) { int idx = subexons[ subexonIdx[j] ].prev[k] ; if ( alltranscripts[i].seVector.Test( idx ) ) continue ; int l ; std::vector<int> visited ; while ( 1 ) { if ( subexons[idx].nextCnt > 1 || subexons[idx].prevCnt > 1 ) { break ; } visited.push_back( idx ) ; if ( subexons[idx].canBeStart && subexons[idx].prevCnt == 0 ) { struct _transcript nt ; nt.first = idx ; nt.last = alltranscripts[i].last ; nt.seVector.Duplicate( alltranscripts[i].seVector ) ; nt.seVector.MaskRegionOutside( subexonIdx[j], nt.last ) ; int visitedSize = visited.size() ; for ( l = 0 ; l < visitedSize ; ++l ) nt.seVector.Set( visited[l] ) ; nt.partial = false ; nt.correlationScore = 0 ; nt.abundance = 0 ; nt.constraintsSupport = NULL ; alltranscripts.push_back( nt ) ; if ( alltranscripts.size() >= limit ) return ; } if ( subexons[idx].prevCnt == 1 ) idx = subexons[idx].prev[0] ; else break ; } } } } } // for if-extend } CoalesceSameTranscripts( alltranscripts ) ; } // Pick the transcripts from given transcripts. void TranscriptDecider::PickTranscripts( struct _subexon *subexons, std::vector<struct _transcript> &alltranscripts, Constraints &constraints, SubexonCorrelation &seCorrelation, std::vector<struct _transcript> &transcripts ) { int i, j, k ; std::vector<int> chosen ; std::vector<struct _matePairConstraint> &tc = constraints.matePairs ; int atcnt = alltranscripts.size() ; int tcCnt = tc.size() ; // transcript constraints int seCnt = 0 ; if ( tcCnt == 0 ) return ; if ( atcnt > 0 ) seCnt = alltranscripts[0].seVector.GetSize() ; else return ; double inf = -1 ; // infinity int coalesceThreshold = 1024 ; int *transcriptSeCnt = new int[ atcnt ] ; int *transcriptLength = new int[atcnt] ; double *transcriptAbundance = new double[atcnt] ; // the roughly estimated abundance based on constraints. double *avgTranscriptAbundance = new double[atcnt] ; // the average normAbund from the compatible constraints. BitTable *btable = new BitTable[ atcnt ] ; //BitTable lowCovSubexon ; // force the abundance to 0 for the transcript contains the subexon. double *coveredPortion = new double[atcnt] ; memset( avgTranscriptAbundance, 0 ,sizeof( double ) * atcnt ) ; for ( i = 0 ; i < atcnt ; ++i ) btable[i].Init( tcCnt ) ; for ( j = 0 ; j < tcCnt ; ++j ) { int a = constraints.matePairs[j].i ; int b = constraints.matePairs[j].j ; if ( constraints.constraints[a].support > inf ) inf = constraints.constraints[a].support ; if ( constraints.constraints[b].support > inf ) inf = constraints.constraints[b].support ; if ( tc[j].normAbund > inf ) inf = tc[j].normAbund ; tc[j].abundance = tc[j].normAbund ; } ++inf ; bool btableSet = false ; for ( i = 0 ; i < atcnt ; ++i ) { //printf( "correlation %d: %lf\n", i, alltranscripts[i].correlationScore ) ; /*for ( int l = 0 ; l < subexonInd.size() ; ++l ) { for ( int m = l ; m < subexonInd.size() ; ++m ) printf( "%lf ", seCorrelation.Query( l, m ) ) ; printf( "\n" ) ; }*/ for ( j = 0 ; j < tcCnt ; ++j ) { int a = tc[j].i ; int b = tc[j].j ; //printf( "try set btble[ %d ].Set( %d ): %d %d\n", i, j, a, b ) ; //alltranscripts[i].seVector.Print() ; //constraints.constraints[a].vector.Print() ; //constraints.constraints[b].vector.Print() ; if ( IsConstraintInTranscript( alltranscripts[i], constraints.constraints[a] ) == 1 && IsConstraintInTranscript( alltranscripts[i], constraints.constraints[b] ) == 1 ) { //printf( "set btble[ %d ].Set( %d ): %d %d\n", i, j, a, b ) ; btable[i].Set( j ) ; btableSet = true ; } } transcriptSeCnt[i] = alltranscripts[i].seVector.Count() ; } if ( btableSet == false ) { for ( i = 0 ; i < atcnt ; ++i ) btable[i].Release() ; delete[] btable ; return ; } double maxAbundance = -1 ; // The abundance of the most-abundant transcript double *adjustScore = new double[atcnt] ; memset( adjustScore, 0, sizeof( double ) * atcnt ) ; if ( atcnt > 0 /*&& alltranscripts[0].abundance == -1*/ ) { struct _pair32 *chain = new struct _pair32[seCnt] ; bool *covered = new bool[seCnt] ; bool *usedConstraints = new bool[constraints.constraints.size() ] ; std::vector<BitTable> togetherChain ; // those subexons is more likely to show up in the same transcript, like an IR with overhang, should be together to represent a 3'/5'-end /*lowCovSubexon.Init( seCnt ) ; double *avgDepth = new double[seCnt ] ; memset( avgDepth, 0, sizeof( double ) * seCnt ) ; int size = constraints.constraints.size() ; for ( i = 0 ; i < size ; ++i ) { std::vector<int> subexonIdx ; constraints.constraints[i].GetOnesIndices( subexonIdx ) ; int seIdxCnt = subexonidx.size() ; for ( j = 0 ; j < seIdxCnt ; ++j ) avgDepth[ subexonidx[j] ] += constraints.constraints[i].support ; } for ( i = 0 ; i < seCnt ; ++i ) { if ( avgDepth[i] * alignments.readLen / (double)( subexons[i].end - subexons[i].start + 1 ) < 1 ) }*/ struct _pair32 firstRegion, lastRegion ; for ( i = 0 ; i < seCnt ; ) { for ( j = i + 1 ; j < seCnt ; ++j ) { if ( subexons[j].start > subexons[j - 1].end + 1 ) break ; } int cnt = 0 ; for ( k = i ; k < j ; ++k ) { if ( ( subexons[k].leftType == 2 && subexons[k].rightType == 1 ) || ( subexons[k].leftType == 0 && subexons[k].rightType == 1 ) || ( subexons[k].leftType == 2 && subexons[k].rightType == 0 ) ) ++cnt ; } if ( cnt <= 1 ) { i = j ; continue ; } BitTable tmpTable( seCnt ) ; for ( k = i ; k < j ; ++k ) { if ( ( subexons[k].leftType == 2 && subexons[k].rightType == 1 ) || ( subexons[k].leftType == 0 && subexons[k].rightType == 1 ) || ( subexons[k].leftType == 2 && subexons[k].rightType == 0 ) ) tmpTable.Set( k ) ; } togetherChain.push_back( tmpTable ) ; i = j ; } for ( i = 0 ; i < atcnt ; ++i ) { double value = inf ; int tag = -1 ; alltranscripts[i].abundance = 0 ; alltranscripts[i].constraintsSupport = new double[tcCnt] ; std::vector<int> subexonIdx ; alltranscripts[i].seVector.GetOnesIndices( subexonIdx ) ; int seIdxCnt = subexonIdx.size() ; transcriptLength[i] = 0 ; firstRegion.a = subexonIdx[0] ; for ( j = 1 ; j < seIdxCnt ; ++j ) { if ( subexons[ subexonIdx[j] ].start > subexons[ subexonIdx[j - 1] ].end + 1 ) break ; } firstRegion.b = subexonIdx[j - 1] ; lastRegion.b = subexonIdx[ seIdxCnt - 1 ] ; for ( j = seIdxCnt - 2 ; j >= 0 ; --j ) { if ( subexons[ subexonIdx[j] ].end < subexons[ subexonIdx[j + 1] ].start - 1 ) break ; } lastRegion.a = subexonIdx[j + 1] ; for ( j = 0 ; j < seIdxCnt ; ++j ) transcriptLength[i] += subexons[ subexonIdx[j] ].end - subexons[ subexonIdx[j] ].start + 1 ; //for ( j = firstRegion.b ; j < lastRegion.a ; ++j ) for ( j = 0 ; j < seIdxCnt - 1 ; ++j ) { chain[j].a = subexonIdx[j] ; chain[j].b = subexonIdx[j + 1] ; covered[j] = false ; } memset( usedConstraints, false, sizeof( bool ) * constraints.constraints.size() ) ; int compatibleCnt = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { alltranscripts[i].constraintsSupport[j] = 0 ; if ( btable[i].Test(j) && tc[j].abundance > 0 ) { ++compatibleCnt ; double adjustAbundance = tc[j].abundance ; if ( seIdxCnt > 1 ) { if ( tc[j].i == tc[j].j && ( constraints.constraints[ tc[j].i ].first + constraints.constraints[ tc[j].i ].last == 2 * alltranscripts[i].first || constraints.constraints[ tc[j].i ].first + constraints.constraints[ tc[j].i ].last == 2 * alltranscripts[i].last ) ) { adjustAbundance = inf ; } else if ( tc[j].i != tc[j].j && ( constraints.constraints[ tc[j].i ].first + constraints.constraints[ tc[j].i ].last == 2 * alltranscripts[i].first || constraints.constraints[ tc[j].i ].first + constraints.constraints[ tc[j].i ].last == 2 * alltranscripts[i].last ) ) { adjustAbundance = constraints.constraints[ tc[j].j ].normAbund ; } else if ( tc[j].i != tc[j].j && ( constraints.constraints[ tc[j].j ].first + constraints.constraints[ tc[j].j ].last == 2 * alltranscripts[i].first || constraints.constraints[ tc[j].j ].first + constraints.constraints[ tc[j].j ].last == 2 * alltranscripts[i].last ) ) { adjustAbundance = constraints.constraints[ tc[j].i ].normAbund ; } } if ( adjustAbundance < value ) /*!( seIdxCnt > 1 && ( ( ( constraints.constraints[ tc[j].i ].first >= firstRegion.a && constraints.constraints[ tc[j].i ].last <= firstRegion.b ) && ( constraints.constraints[ tc[j].j ].first >= firstRegion.a && constraints.constraints[ tc[j].j ].last <= firstRegion.b ) ) || ( ( constraints.constraints[ tc[j].i ].first >= lastRegion.a && constraints.constraints[ tc[j].i ].last <= lastRegion.b ) && ( constraints.constraints[ tc[j].j ].first >= lastRegion.a && constraints.constraints[ tc[j].j ].last <= lastRegion.b ) ) ) ) )*/ { // Not use the constraints totally within the 3'/5'-end in the transcript value = adjustAbundance ; tag = j ; } avgTranscriptAbundance[i] += tc[j].abundance ; if ( !usedConstraints[ tc[j].i ] ) { struct _constraint &c = constraints.constraints[ tc[j].i ] ; for ( k = 0 ; k < seIdxCnt - 1 ; ++k ) { // Note that since the constraint is already compatible with the txpt, // chain[k].a/b must be also adjacent in this constraint. if ( c.vector.Test( chain[k].a ) && c.vector.Test( chain[k].b ) ) covered[k] = true ; } usedConstraints[ tc[j].i ] = true ; } if ( !usedConstraints[ tc[j].j ] ) { struct _constraint &c = constraints.constraints[ tc[j].j ] ; for ( k = 0 ; k < seIdxCnt - 1 ; ++k ) { if ( c.vector.Test( chain[k].a ) && c.vector.Test( chain[k].b ) ) covered[k] = true ; } usedConstraints[ tc[j].j ] = true ; } } } // Get some penalty if something should together did not show up together int size = togetherChain.size() ; if ( size > 0 ) { BitTable bufferTable( seCnt ) ; for ( j = 0 ; j < size ; ++j ) { bufferTable.Assign( togetherChain[j] ) ; bufferTable.And( alltranscripts[i].seVector ) ; //if ( !bufferTable.IsAllZero() && !bufferTable.IsEqual( togetherChain[j] ) ) // value /= 2 ; if ( !bufferTable.IsAllZero() ) { if ( bufferTable.IsEqual( togetherChain[j] ) ) //printf( "nice together!\n" ) ; ; else value /= 2 ; //printf( "bad together!\n" ) ; } } bufferTable.Release() ; } // Every two-subexon chain should be covered by some reads if a transcript is expressed highly enough int cnt = 0 ; for ( j = 0 ; j < seIdxCnt - 1 ; ++j ) if ( covered[j] == false ) // && j >= firstRegion.b && j <= lastRegion.a - 1 ) { value = 0 ; } else ++cnt ; if ( seIdxCnt > 1 ) coveredPortion[i] = (double)cnt / (double)( seIdxCnt - 1 ) ; else coveredPortion[i] = 1 ; if ( coveredPortion[i] == 0 ) coveredPortion[i] = (double)0.5 / ( seIdxCnt ) ; // For short subexon (readLength-subexon_length-1>30), we further require a constraint cover three conseuctive subexon /*memset( usedConstraints, false, sizeof( bool ) * constraints.constraints.size() ) ; for ( j = 1 ; j < seIdxCnt - 1 ; ++j ) { int k = subexonIdx[j] ; if ( alignments.readLen - ( subexons[k].end - subexons[k].start + 1 ) - 1 <= 30 ) continue ; // We need at least one of the side subexons are adjacent to the center one. if ( subexons[ subexonIdx[j - 1] ].end + 1 < subexons[k].start && subexons[k].end + 1 < subexons[ subexonIdx[j + 1] ].start ) continue ; int l = 0 ; for ( l = 0 ; l < tcCnt ; ++l ) { if ( btable[i].Test(l) && tc[l].abundance > 0 ) { if ( !usedConstraints[ tc[l].i ] ) { struct _constraint &c = constraints.constraints[ tc[l].i ] ; if ( c.vector.Test( subexonIdx[j - 1] ) && c.vector.Test( subexonIdx[j] ) && c.vector.Test( subexonIdx[j + 1] ) ) break ; usedConstraints[ tc[l].i ] = true ; } if ( !usedConstraints[ tc[l].j ] ) { struct _constraint &c = constraints.constraints[ tc[l].j ] ; if ( c.vector.Test( subexonIdx[j - 1] ) && c.vector.Test( subexonIdx[j] ) && c.vector.Test( subexonIdx[j + 1] ) ) break ; usedConstraints[ tc[l].j ] = true ; } } } // It is not covered if ( l >= tcCnt ) { int residual = alignments.readLen - ( subexons[k].end - subexons[k].start + 1 ) - 1 ; //printf( "residual: %d %d %lf\n", k, residual, value ) ; if ( value * residual > 2 ) { value = 1 / (double)residual ; } } }*/ if ( tag == -1 ) value = 0 ; if ( value > maxAbundance ) maxAbundance = value ; transcriptAbundance[i] = value ; if ( tag != -1 ) avgTranscriptAbundance[i] /= compatibleCnt ; //printf( "abundance %d: %lf %lf ", i, value, avgTranscriptAbundance[i] ) ; //alltranscripts[i].seVector.Print() ; } if ( maxAbundance == 0 ) { for ( i = 0 ; i < atcnt ; ++i ) { transcriptAbundance[i] = coveredPortion[i] ; } maxAbundance = 1 ; } //printf( "%s: %lf\n", __func__, maxAbundance ) ; int size = togetherChain.size() ; for ( j = 0 ; j < size ; ++j ) togetherChain[j].Release() ; delete[] usedConstraints ; delete[] covered ; delete[] chain ; } else { for ( i = 0 ; i < atcnt ; ++i ) { transcriptAbundance[i] = alltranscripts[i].abundance ; if ( transcriptAbundance[i] > maxAbundance ) maxAbundance = transcriptAbundance[i] ; coveredPortion[i] = 1 ; } if ( maxAbundance == 0 ) maxAbundance = 1 ; } // Obtain the prefix, suffix information of the transcripts. int *nextSuffix, *nextPrefix ; struct _pair32 *txptRank ; nextSuffix = new int[atcnt] ; nextPrefix = new int[atcnt] ; txptRank = new struct _pair32[atcnt] ; memset( nextSuffix, -1, sizeof( int ) * atcnt ) ; memset( nextPrefix, -1, sizeof( int ) * atcnt ) ; /*for ( i = 0 ; i < atcnt ; ++i ) { std::vector<int> subexonIdx ; txptRank[i].a = i ; alltranscripts[i].seVector.GetOnesIndices( subexonIdx ) ; txptRank[i].b = subexonIdx.size() ; } qsort( txptRank, atcnt, sizeof( struct _pair32 ), CompPairsByB) ; BitTable bufferTable( seCnt ) ; for ( i = atcnt - 1 ; i >= 0 ; --i ) { int a = txptRank[i].a ; for ( j = i - 1 ; j >= 0 ; --j ) { if ( txptRank[i].b == txptRank[j].b ) continue ; int b = txptRank[j].a ; if ( alltranscripts[b].last != alltranscripts[a].last ) continue ; bufferTable.Assign( alltranscripts[a].seVector ) ; bufferTable.MaskRegionOutside( alltranscripts[b].first, alltranscripts[b].last ) ; if ( bufferTable.IsEqual( alltranscripts[b].seVector ) ) { nextSuffix[a] = b ; break ; } } } for ( i = atcnt - 1 ; i >= 0 ; --i ) { int a = txptRank[i].a ; for ( j = i - 1 ; j >= 0 ; --j ) { if ( txptRank[i].b == txptRank[j].b ) continue ; int b = txptRank[j].a ; if ( alltranscripts[b].first != alltranscripts[a].first ) continue ; bufferTable.Assign( alltranscripts[a].seVector ) ; bufferTable.MaskRegionOutside( alltranscripts[b].first, alltranscripts[b].last ) ; if ( bufferTable.IsEqual( alltranscripts[b].seVector ) ) { nextPrefix[a] = b ; break ; } } } bufferTable.Release() ;*/ delete[] txptRank ; // Quantative Set-Cover int iterCnt = -1 ; double *coverCnt = new double[atcnt] ; for ( i = 0 ; i < atcnt ; ++i ) coverCnt[i] = -1 ; int *list = new int[atcnt] ; int listCnt ; while ( 1 ) { double max = -1 ; int maxtag = -1 ; double maxcnt = -1 ; ++iterCnt ; // Find the optimal candidate. for ( i = 0 ; i < atcnt ; ++i ) { double value = inf ; double cnt = 0 ; if ( coverCnt[i] == -1 ) { for ( j = 0 ; j < tcCnt ; ++j ) { if ( tc[j].abundance > 0 && btable[i].Test( j ) ) { cnt += tc[j].effectiveCount ; } } /*else { std::vector<int> tcIdx ; btable[i].GetOnesIndices( tcIdx ) ; int size = tcIdx.size() ; for ( j = 0 ; j < size ; ++j ) { if ( tc[ tcIdx[j] ].abundance > 0 ) { cnt += tc[ tcIdx[j] ].effectiveCount ; } } }*/ coverCnt[i] = cnt ; } else { cnt = coverCnt[i] ; } value = transcriptAbundance[i] ; if ( cnt < 1 ) // This transcript does not satisfy any undepleted constraints. continue ; cnt *= coveredPortion[i] ; double weight = 1 ; //* seCnt / transcriptSeCnt[i] ; //if ( maxAbundance >= 1 && value / maxAbundance >= 0.2 ) // seCntAdjust = sqrt( (double)( transcriptSeCnt[i] ) / seCnt ) ;//< 0.5 ? 0.5 : (double)( transcriptSeCnt[i] ) / seCnt ; if ( alltranscripts[i].FPKM > 0 && sampleCnt > 1 ) weight = ( 1 + alltranscripts[i].FPKM / sampleCnt ) ; double score = ComputeScore( cnt, weight, value, maxAbundance, alltranscripts[i].correlationScore ) ; if ( cnt > maxcnt ) maxcnt = cnt ; score += adjustScore[i] ; if ( score > max ) { max = score ; maxtag = i ; } else if ( score == max ) { if ( avgTranscriptAbundance[maxtag] < avgTranscriptAbundance[i] ) { max = score ; maxtag = i ; } } //printf( "score: %d %lf -> %lf\n", i, cnt, score ) ; } if ( maxcnt == 0 || maxtag == -1 ) break ; // Find the constraint that should be depleted. double update = inf ; int updateTag = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( btable[ maxtag ].Test( j ) && tc[j].abundance > 0 && tc[j].abundance <= update ) { update = tc[j].abundance ; updateTag = j ; } } // Search suffix and prefix to see whether these fit better. int p = nextSuffix[ maxtag] ; while ( p != -1 ) { if ( transcriptAbundance[p] >= 10.0 * transcriptAbundance[maxtag] && btable[p].Test( updateTag ) ) { //printf( "%d\n", p ) ; maxtag = p ; break ; } p = nextSuffix[p] ; } p = nextPrefix[maxtag] ; while ( p != -1 ) { if ( transcriptAbundance[p] >= 10.0 * transcriptAbundance[maxtag] && btable[p].Test( updateTag ) ) { maxtag = p ; break ; } p = nextPrefix[p] ; } // Update the abundance. int supportCnt = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( btable[maxtag].Test( j ) ) { if ( tc[j].abundance > 0 ) { tc[j].abundance -= 1 * update ; double factor = tc[j].effectiveCount ; double tmp = ( tc[j].support * update / tc[j].normAbund * factor ) ; alltranscripts[maxtag].constraintsSupport[j] += tmp ; alltranscripts[maxtag].abundance += tmp ; if ( tc[j].abundance <= 0 ) { int l ; for ( l = 0 ; l < atcnt ; ++l ) { if ( btable[l].Test(j) ) coverCnt[l] -= tc[j].effectiveCount ; } } ++supportCnt ; } else if ( alltranscripts[maxtag].constraintsSupport[j] == 0 ) { double sum = 0 ; double takeOut = 0 ; double factor = tc[j].effectiveCount ; listCnt = 0 ; for ( i = 0 ; i < atcnt ; ++i ) { if ( i == maxtag ) continue ; if ( alltranscripts[i].abundance > 0 && btable[i].Test(j) ) { sum += alltranscripts[i].constraintsSupport[j] ; double tmp = ( alltranscripts[i].constraintsSupport[j] + alltranscripts[maxtag].constraintsSupport[j] ) * transcriptAbundance[maxtag] / ( transcriptAbundance[maxtag] + transcriptAbundance[i] ) - alltranscripts[maxtag].constraintsSupport[j] ; if ( tmp > 0 ) { list[ listCnt ] = i ; ++listCnt ; takeOut += tmp ; //alltranscripts[i].constraintsSupport[j] * transcriptAbundance[maxtag] / ( transcriptAbundance[maxtag] + transcriptAbundance[i] ) ; } } } double ratio = 1 ; double takeOutFactor = 0.5 ; if ( update < tc[j].normAbund ) { if ( takeOut > ( tc[j].support * update / tc[j].normAbund * factor ) * takeOutFactor ) ratio = ( tc[j].support * update / tc[j].normAbund * factor ) * takeOutFactor / takeOut ; } else { if ( takeOut > ( tc[j].support * factor ) * takeOutFactor ) ratio = ( tc[j].support * factor ) * takeOutFactor / takeOut ; } if ( 1 ) //update < tc[j].normAbund ) { for ( i = 0 ; i < listCnt ; ++i ) { //double tmp = ( tc[j].support * update / tc[j].normAbund * factor ) * // ( alltranscripts[ list[i] ].constraintsSupport[j] / sum ) ; //if ( alltranscripts[ list[i] ].constraintsSupport[j] < tmp ) // printf( "WARNING! %lf %lf, %lf\n", alltranscripts[ list[i] ].constraintsSupport[j], sum, tmp ) ; //double tmp = alltranscripts[ list[i] ].constraintsSupport[j] * transcriptAbundance[maxtag] / ( transcriptAbundance[maxtag] + transcriptAbundance[ list[i] ] ) * ratio ; double tmp = ( ( alltranscripts[ list[i] ].constraintsSupport[j] + alltranscripts[maxtag].constraintsSupport[j] ) * transcriptAbundance[maxtag] / ( transcriptAbundance[maxtag] + transcriptAbundance[ list[i] ] ) - alltranscripts[maxtag].constraintsSupport[j] ) * ratio ; alltranscripts[ list[i] ].constraintsSupport[j] -= tmp ; alltranscripts[ list[i] ].abundance -= tmp ; } //double tmp = ( tc[j].support * update / tc[j].normAbund * factor ) ; //printf( "%lf %lf. %lf %lf\n", takeOut, ratio, update, tc[j].normAbund ) ; double tmp = takeOut * ratio ; alltranscripts[maxtag].constraintsSupport[j] += tmp ; alltranscripts[maxtag].abundance += tmp ; } /*else { double tmp = ( tc[j].support / (double)( listCnt + 1 ) ) * factor ; for ( i = 0 ; i < listCnt ; ++i ) { alltranscripts[ list[i] ].abundance -= alltranscripts[ list[i] ].constraintsSupport[j] ; alltranscripts[ list[i] ].constraintsSupport[j] = tmp ; alltranscripts[ list[i] ].abundance += tmp ; } alltranscripts[maxtag].constraintsSupport[j] += tmp ; alltranscripts[maxtag].abundance += tmp ; }*/ } } if ( tc[j].abundance < 0 ) { tc[j].abundance = 0 ; } } tc[ updateTag ].abundance = 0 ; if ( supportCnt == 0 ) break ; //adjustScore[maxtag] += 1 / (double)tcCnt ; //printf( "maxtag=%d %lf %d\n", maxtag, update, updateTag ) ; } for ( i = 0 ; i < atcnt ; ++i ) { if ( alltranscripts[i].abundance > 0 ) { struct _transcript nt = alltranscripts[i] ; nt.seVector.Nullify() ; nt.seVector.Duplicate( alltranscripts[i].seVector ) ; nt.constraintsSupport = NULL ; if ( transcriptAbundance[i] == 0 ) nt.correlationScore = -1 ; else nt.correlationScore = 0 ; nt.id = i ; transcripts.push_back( nt ) ; } } // Release the memory of btable. for ( i = 0 ; i < atcnt ; ++i ) { delete[] alltranscripts[i].constraintsSupport ; btable[i].Release() ; } delete[] btable ; delete[] list ; delete[] transcriptSeCnt ; delete[] transcriptLength ; delete[] transcriptAbundance ; delete[] avgTranscriptAbundance ; delete[] coveredPortion ; delete[] adjustScore ; delete[] coverCnt ; delete[] nextPrefix ; delete[] nextSuffix ; // Redistribute weight if there is some constraints that are unbalanced. /*tcnt = transcripts.size() ; for ( i = 0 ; i < tcnt ; ++i ) { int maxRatio = -1 ; for ( j = 0 ; j < tcCnt ; ++j ) if ( transcripts[i].constraintsSupport[j] > 0 ) { double factor = tc[j].effectiveCount ; if ( transcripts[]) } }*/ } void TranscriptDecider::AbundanceEstimation( struct _subexon *subexons, int seCnt, Constraints &constraints, std::vector<struct _transcript> &transcripts ) { int tcnt = transcripts.size() ; int size ; int i, j ; if ( tcnt <= 0 ) return ; std::vector<struct _matePairConstraint> &tc = constraints.matePairs ; int tcCnt = tc.size() ; // transcript constraints BitTable *btable = new BitTable[ tcnt ] ; int *transcriptLength = new int[tcnt] ; int *compatibleList = new int[tcnt] ; double *rho = new double[tcnt] ; // the abundance. int iterCnt = 0 ; for ( i = 0 ; i < tcnt ; ++i ) transcripts[i].constraintsSupport = new double[ tcCnt ] ; for ( i = 0 ; i < tcnt ; ++i ) { btable[i].Init( tcCnt ) ; double min = -1 ; for ( j = 0 ; j < tcCnt ; ++j ) { int a = tc[j].i ; int b = tc[j].j ; if ( IsConstraintInTranscript( transcripts[i], constraints.constraints[a] ) == 1 && IsConstraintInTranscript( transcripts[i], constraints.constraints[b] ) == 1 ) { //printf( "set btble[ %d ].Set( %d ): %d %d\n", i, j, a, b ) ; btable[i].Set( j ) ; if ( min == -1 || tc[j].normAbund < min ) min = tc[j].normAbund ; } } std::vector<int> subexonIdx ; transcripts[i].seVector.GetOnesIndices( subexonIdx ) ; int subexonIdxCnt = subexonIdx.size() ; int len = 0 ; for ( j = 0 ; j < subexonIdxCnt ; ++j ) len += subexons[ subexonIdx[j] ].end - subexons[ subexonIdx[j] ].start + 1 ; transcriptLength[i] = len - alignments.fragLen + 2 * alignments.fragStdev ; if ( transcriptLength[i] < 1 ) transcriptLength[i] = 1 ; rho[i] = transcripts[i].abundance / transcriptLength[i] ; // use the rough estimation generated before. if ( transcripts[i].correlationScore == -1 && rho[i] > 0.1 / (double)alignments.readLen ) rho[i] = 0.1 / (double)alignments.readLen ; } while ( 1 ) { for ( i = 0 ; i < tcnt ; ++i ) for ( j = 0 ; j < tcCnt ; ++j ) { transcripts[i].constraintsSupport[j] = 0 ; } for ( j = 0 ; j < tcCnt ; ++j ) { int clCnt = 0 ; double sum = 0 ; for ( i = 0 ; i < tcnt ; ++i ) { if ( btable[i].Test(j) ) { compatibleList[ clCnt ] = i ; ++clCnt ; sum += rho[i] ; } } for ( i = 0 ; i < clCnt ; ++i ) { double factor = tc[j].effectiveCount ; transcripts[ compatibleList[i] ].constraintsSupport[j] = ( rho[ compatibleList[i] ] / sum ) * tc[j].support * factor ; } } double diff = 0 ; for ( i = 0 ; i < tcnt ; ++i ) { double newAbund = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) newAbund += transcripts[i].constraintsSupport[j] ; double old = rho[i] ; rho[i] = newAbund / transcriptLength[i] ; //printf( "rho[%d]=%lf\n", i, rho[i] ) ; if ( transcripts[i].correlationScore == -1 && rho[i] > 0.1 / (double)alignments.readLen ) rho[i] = 0.1 / (double)alignments.readLen ; double tmp = ( old - rho[i] ) ; diff += tmp < 0 ? -tmp : tmp ; } //printf( "%lf\n", diff ) ; if ( diff < 1e-3) break ; ++iterCnt ; if ( iterCnt >= 1000 ) break ; } for ( i = 0 ; i < tcnt ; ++i ) { //printf( "%lf=>", transcripts[i].abundance ) ; transcripts[i].abundance = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { transcripts[i].abundance += transcripts[i].constraintsSupport[j] ; } //printf( "%lf. (%lf)\n", transcripts[i].abundance, transcripts[i].correlationScore ) ; //transcripts[i].seVector.Print() ; } for ( i = 0 ; i < tcnt ; ++i ) delete[] transcripts[i].constraintsSupport ; // Release the memory of btable. for ( i = 0 ; i < tcnt ; ++i ) { btable[i].Release() ; } delete[] compatibleList ; delete[] btable ; delete[] transcriptLength ; delete[] rho ; } int TranscriptDecider::RefineTranscripts( struct _subexon *subexons, int seCnt, bool aggressive, std::map<int, int> *subexonChainSupport, int *txptSampleSupport, std::vector<struct _transcript> &transcripts, Constraints &constraints ) { int i, j, k ; int tcnt = transcripts.size() ; if ( tcnt == 0 ) return 0 ; int tcCnt = constraints.matePairs.size() ; std::vector<struct _matePairConstraint> &tc = constraints.matePairs ; std::vector<struct _constraint> &scc = constraints.constraints ; //single-end constraints.constraints // Remove transcripts whose FPKM are too small. //printf( "%d %d\n", usedGeneId, baseGeneId ) ; double *geneMaxFPKM = new double[usedGeneId - baseGeneId ] ; int *geneMaxFPKMTag = new int[usedGeneId - baseGeneId ] ; double *nonOverlapMaxFPKM = new double[ usedGeneId - baseGeneId ] ; // the max FPKM among all the transcripts not overlapping with maxFPKMTag transcripts. memset( geneMaxFPKM, 0, sizeof( double ) * ( usedGeneId - baseGeneId ) ) ; memset( geneMaxFPKMTag, 0, sizeof( int ) * ( usedGeneId - baseGeneId ) ) ; memset( nonOverlapMaxFPKM, 0, sizeof( double ) * ( usedGeneId - baseGeneId ) ) ; double *geneMaxCov = new double[ usedGeneId - baseGeneId ] ; memset( geneMaxCov, 0, sizeof( double ) * ( usedGeneId - baseGeneId ) ) ; int *txptGid = new int[tcnt] ; /*for ( i = 0 ; i < tcnt ; ++i ) { printf( "%d: %lf ", i, transcripts[i].FPKM ) ; transcripts[i].seVector.Print() ; }*/ /*================================================================== Remove transcripts that has too few relative FPKM. (-f) ====================================================================*/ for ( i = 0 ; i < tcnt ; ++i ) { int gid = GetTranscriptGeneId( transcripts[i], subexons ) ; int len = GetTranscriptLengthFromAbundanceAndFPKM( transcripts[i].abundance, transcripts[i].FPKM ) ; //printf( "gid=%d\n", gid ) ; //printf( "%lf %lf %d\n", transcripts[i].abundance, transcripts[i].FPKM, len ) ; if ( transcripts[i].FPKM > geneMaxFPKM[gid - baseGeneId ] ) { geneMaxFPKM[ gid - baseGeneId ] = transcripts[i].FPKM ; geneMaxFPKMTag[ gid - baseGeneId ] = i ; } if ( transcripts[i].abundance * alignments.readLen / len > geneMaxCov[gid - baseGeneId ] ) geneMaxCov[gid - baseGeneId] = ( transcripts[i].abundance * alignments.readLen ) / len ; txptGid[i] = gid ; } for ( i = 0 ; i < tcnt ; ++i ) { int tag = txptGid[i] - baseGeneId ; if ( ( transcripts[i].last < transcripts[ geneMaxFPKMTag[ tag ] ].first || transcripts[i].first > transcripts[ geneMaxFPKMTag[tag] ].last ) && transcripts[i].FPKM > nonOverlapMaxFPKM[tag] ) nonOverlapMaxFPKM[tag] = transcripts[i].FPKM ; } BitTable bufferTable ; bufferTable.Duplicate( transcripts[0].seVector ) ; if ( !aggressive ) { // Rescue the transcripts covering unique constraints. int cnt = 0 ; int tag = 0 ; int *uniqCount = new int[tcnt] ; memset( uniqCount, 0, sizeof( int ) * tcnt ) ; for ( j = 0 ; j < tcCnt ; ++j ) { cnt = 0 ; if ( tc[j].uniqSupport <= 5 ) continue ; for ( i = 0 ; i < tcnt ; ++i ) { if ( IsConstraintInTranscript( transcripts[i], scc[ tc[j].i ] ) && IsConstraintInTranscript( transcripts[i], scc[ tc[j].j] ) ) { tag = i ; ++cnt ; } if ( cnt >= 2 ) break ; } if ( cnt == 1 ) { ++uniqCount[tag] ; } } for ( i = 0 ; i < tcnt ; ++i ) { if ( uniqCount[i] >= 2 ) { transcripts[i].abundance *= 4 ; transcripts[i].FPKM *= 4 ; } } delete[] uniqCount ; } int sccCnt = scc.size() ; double filterFactor = 1.0 ; for ( i = 0 ; i < tcnt ; ++i ) { //printf( "%d: %lf %lf\n", txptGid[i], transcripts[i].abundance, geneMaxFPKM[ txptGid[i] - baseGeneId ] ) ; if ( transcripts[i].FPKM < filterFactor * FPKMFraction * geneMaxFPKM[ txptGid[i] - baseGeneId ] ) { /*int cnt = 0 ; int coverCnt = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( transcripts[i].constraintsSupport[j] > 0 ) ++coverCnt ; double factor = tc[j].effectiveCount ; if ( transcripts[i].constraintsSupport[j] >= factor * tc[j].support - 1e-3 && tc[j].support >= 10 && tc[j].uniqSupport >= 0.95 * tc[j].support ) { ++cnt ; } } //cnt = 0 ; if ( cnt >= 2 ) { ; } else*/ transcripts[i].abundance = -transcripts[i].abundance ; } //if ( transcripts[i].FPKM >= 0.8 * geneMaxFPKM[ txptGid[i] - baseGeneId ] && geneMaxCov[ txptGid[i] - baseGeneId ] >= txptMinReadDepth ) // continue ; } if ( nonOverlapMaxFPKM != 0 ) { // Go two iterations to rescue, the first iteration should be just for marking. std::vector<int> rescueList ; for ( i = 0 ; i < tcnt ; ++i ) { if ( transcripts[i].abundance >= 0 ) continue ; for ( j = 0 ; j < tcnt ; ++j ) { if ( transcripts[j].abundance < 0 || txptGid[i] != txptGid[j] ) continue ; if ( transcripts[i].first <= transcripts[j].last && transcripts[i].last >= transcripts[j].first ) /*bufferTable.Assign( transcripts[i].seVector ) ; bufferTable.And( transcripts[j].seVector ) ; if ( !bufferTable.IsAllZero() )*/ break ; } if ( j >= tcnt && transcripts[i].FPKM >= FPKMFraction * nonOverlapMaxFPKM[ txptGid[i] - baseGeneId ] ) { //transcripts[i].abundance = -transcripts[i].abundance ; rescueList.push_back( i ) ; } } int size = rescueList.size() ; for ( i = 0 ; i < size ; ++i ) transcripts[ rescueList[i] ].abundance *= -1 ; } /*================================================================== Remove transcripts that has too few read coverage (-d) ====================================================================*/ for ( i = 0 ; i < tcnt ; ++i ) { if ( transcripts[i].abundance >= 0 ) { int len = GetTranscriptLengthFromAbundanceAndFPKM( transcripts[i].abundance, transcripts[i].FPKM ) ; double cov = ( transcripts[i].abundance * alignments.readLen ) / len ; //printf( "%d: %d %d %lf %lf\n", i, len, transcripts[i].seVector.Count(), cov, geneMaxCov[ txptGid[i] - baseGeneId ] ) ; if ( ( tcnt > 1 || len <= 1000 || transcripts[i].seVector.Count() <= 3 ) && cov < txptMinReadDepth ) { //if ( usedGeneId == baseGeneId + 1 && /*transcripts[i].seVector.Count() > 3 // && len > 1000 &&*/ geneMaxCov[ txptGid[i] - baseGeneId ] == cov ) if ( geneMaxCov[ txptGid[i] - baseGeneId ] == cov ) continue ; // Test whether it has some very abundant constraints. /*int cnt = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( transcripts[i].constraintsSupport[j] >= tc[j].support / 2.0 && tc[j].support >= 10 && tc[j].uniqSupport >= 0.95 * tc[j].support && tc[j].normAbund >= 1 ) { ++cnt ; } } if ( cnt >= 1 ) { continue ; }*/ // Test whether this transcript is fully covered. If so ,we can filter it. if ( geneMaxCov[ txptGid[i] - baseGeneId ] <= 5 ) { bufferTable.Reset() ; for ( j = 0 ; j < sccCnt ; ++j ) { if ( !IsConstraintInTranscript( transcripts[i], scc[j] ) ) continue ; bufferTable.Or( scc[j].vector ) ; } if ( bufferTable.IsEqual( transcripts[i].seVector ) ) transcripts[i].abundance = -transcripts[i].abundance ; } else transcripts[i].abundance = -transcripts[i].abundance ; /*else { transcripts[i].seVector.Print() ; bufferTable.Print() ; OutputTranscript( stderr, subexons, transcripts[i] ) ; }*/ } } } /*================================================================== Remove transcripts that is too short ====================================================================*/ for ( i = 0 ; i < tcnt ; ++i ) { if ( transcripts[i].abundance <= 0 ) continue ; int len = GetTranscriptLengthFromAbundanceAndFPKM( transcripts[i].abundance, transcripts[i].FPKM ) ; if ( len < 200 ) { transcripts[i].abundance = -transcripts[i].abundance ; } } // Rescue transcripts that showed up in many samples. /*for ( i = 0 ; i < tcnt ; ++i ) { if ( transcripts[i].abundance > 0 ) continue ; if ( txptSampleSupport[ transcripts[i].id ] >= 3 && txptSampleSupport[transcripts[i].id ] >= (int)( sampleCnt / 2 ) ) transcripts[i].abundance = -transcripts[i].abundance ; }*/ // Rescue some transcripts covering subexon chains showed up in many samples, but missing after filtration. struct _constraint tmpC ; tmpC.vector.Init( seCnt ) ; std::vector< struct _pair32 > missingChain ; std::vector<int> recoverCandidate ; bool *used = new bool[tcnt] ; memset( used, false, sizeof( bool ) * tcnt ) ; // Obtain the list of transcripts that should be recovered. for ( i = 0 ; i < seCnt && sampleCnt > 1 ; ++i ) { double maxFPKM = -1 ; for ( std::map<int, int>::iterator it = subexonChainSupport[i].begin() ; it != subexonChainSupport[i].end() ; ++it ) { if ( sampleCnt >= 0 && ( it->second < 3 || it->second < (int)( 0.5 * sampleCnt ) ) && it->second <= sampleCnt / 2 ) continue ; bool recover = true ; tmpC.vector.Reset() ; tmpC.vector.Set( i ) ; tmpC.vector.Set( it->first ) ; tmpC.first = i ; tmpC.last = it->first ; for ( j = 0 ; j < tcnt ; ++j ) { if ( transcripts[j].abundance < 0 ) continue ; if ( IsConstraintInTranscript( transcripts[j], tmpC ) ) { recover = false ; break ; } if ( recover ) { for ( j = 0 ; j < tcnt ; ++j ) { if ( transcripts[j].abundance > 0 ) continue ; //printf( "%d %lf\n", IsConstraintInTranscript( transcripts[j], tmpC ), transcripts[j].FPKM ) ; if ( IsConstraintInTranscript( transcripts[j], tmpC ) ) { /*if ( maxTag == -1 ) maxTag = j ; else { if ( txptSampleSupport[ transcripts[j].id ] > txptSampleSupport[ transcripts[maxTag ].id ] ) maxTag = j ; else if ( txptSampleSupport[ transcripts[j].id ] == txptSampleSupport[ transcripts[maxTag ].id ]) { if ( transcripts[j].FPKM > transcripts[maxTag].FPKM ) maxTag = j ; } }*/ struct _pair32 np ; np.a = i ; np.b = it->first ; missingChain.push_back( np ) ; if ( !used[j] ) { recoverCandidate.push_back( j ) ; used[j] = true ; } } } /*if ( maxTag != -1 && txptSampleSupport[ transcripts[maxTag].id ] > 1 ) { //printf( "recover %d %d\n", maxTag, txptSampleSupport[ transcripts[maxTag].id ] ) ; transcripts[maxTag].abundance *= -1 ; }*/ } } } } int size = recoverCandidate.size() ; memset( used, false, sizeof( bool ) * tcnt ) ; // Recover the candidates in the order of reliability int *geneRecoverCnt = new int[ usedGeneId - baseGeneId ] ; memset( geneRecoverCnt, 0, sizeof( int ) * ( usedGeneId - baseGeneId ) ) ; int round = 1 ; if ( aggressive && size > 1 ) round = 1 ; for ( i = 0 ; i < size ; ++i ) { int maxTag = -1 ; int maxCover = -1 ; for ( j = 0 ; j < size ; ++j ) { if ( !used[ recoverCandidate[j] ] ) { /*int cover = 0 ; k = missingChain.size() ; int l ; for ( l = 0 ; l < k ; ++l ) { if ( missingChain[l].a == -1 ) continue ; tmpC.vector.Reset() ; tmpC.vector.Set( missingChain[l].a ) ; tmpC.vector.Set( missingChain[l].b ) ; tmpC.first = missingChain[l].a ; tmpC.last = missingChain[l].b ; if ( IsConstraintInTranscript( transcripts[ recoverCandidate[j] ], tmpC ) ) { ++cover ; } }*/ if ( maxTag == -1 ) { maxTag = recoverCandidate[j] ; //maxCover = cover ; continue ; } /*if ( cover > maxCover ) { maxTag = recoverCandidate[j] ; maxCover = cover ; } else if ( cover == maxCover ) {*/ if ( txptSampleSupport[ transcripts[ recoverCandidate[j] ].id ] > txptSampleSupport[ transcripts[ maxTag ].id ] ) maxTag = recoverCandidate[j] ; else if ( txptSampleSupport[ transcripts[ recoverCandidate[j] ].id ] == txptSampleSupport[ transcripts[ maxTag ].id ] ) { if ( transcripts[ recoverCandidate[j] ].FPKM > transcripts[ maxTag ].FPKM ) maxTag = recoverCandidate[j] ; } /*else if ( transcripts[ recoverCandidate[j] ].FPKM > transcripts[ maxTag ].FPKM ) maxTag = recoverCandidate[j] ; else if ( transcripts[ recoverCandidate[j] ].FPKM == transcripts[ maxTag ].FPKM ) { if ( txptSampleSupport[ transcripts[ recoverCandidate[j] ].id ] > txptSampleSupport[ transcripts[ maxTag ].id ] ) maxTag = recoverCandidate[j] ; }*/ //} } } if ( maxTag == -1 || txptSampleSupport[ transcripts[ maxTag ].id ] <= 2 || txptSampleSupport[ transcripts[maxTag].id ] < 0.5 * sampleCnt ) break ; used[maxTag] = true ; if ( geneRecoverCnt[ txptGid[maxTag] - baseGeneId ] >= round ) continue ; ++geneRecoverCnt[ txptGid[maxTag] - baseGeneId ] ; k = missingChain.size() ; int cnt = 0 ; for ( j = 0 ; j < k ; ++j ) { if ( missingChain[j].a == -1 ) continue ; tmpC.vector.Reset() ; tmpC.vector.Set( missingChain[j].a ) ; tmpC.vector.Set( missingChain[j].b ) ; tmpC.first = missingChain[j].a ; tmpC.last = missingChain[j].b ; if ( IsConstraintInTranscript( transcripts[maxTag], tmpC ) ) { missingChain[j].a = -1 ; ++cnt ; } } int len = GetTranscriptLengthFromAbundanceAndFPKM( transcripts[maxTag].abundance, transcripts[maxTag].FPKM ) ; double cov = ( transcripts[maxTag].abundance * alignments.readLen ) / len ; if ( cnt >= 1 && cov > 1.0 ) { transcripts[maxTag].abundance *= -1 ; } } delete[] used ; delete[] geneRecoverCnt ; tmpC.vector.Release() ; tcnt = RemoveNegativeAbundTranscripts( transcripts ) ; delete []geneMaxCov ; bufferTable.Release() ; delete []geneMaxFPKM ; delete []geneMaxFPKMTag ; delete []nonOverlapMaxFPKM ; delete []txptGid ; /*================================================================== Remove transcripts that seems duplicated ====================================================================*/ for ( i = 0 ; i < tcnt ; ++i ) { int support = 0 ; int uniqSupport = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( !IsConstraintInTranscript( transcripts[i], scc[ tc[j].i ] ) || !IsConstraintInTranscript( transcripts[i], scc[ tc[j].j ] ) ) continue ; //support += scc[ tc[j].i ].support + scc[ tc[j].j ].support ; //uniqSupport += scc[ tc[j].i ].uniqSupport + scc[ tc[j].j ].uniqSupport ; support += tc[j].support ; uniqSupport += tc[j].uniqSupport ; //printf( "constraint uniqness: %d: %d %d\n", i, tc[j].uniqSupport, tc[j].support ) ; } //printf( "%d: %d %d\n", i, uniqSupport, support ) ; if ( (double)uniqSupport < 0.03 * support ) transcripts[i].abundance = -1 ; } tcnt = RemoveNegativeAbundTranscripts( transcripts ) ; /*================================================================== Remove shadow transcripts, the abnormal 2-exon txpt whose intron is very close to the true one or one of the anchor exon is shorter than 25bp.... ====================================================================*/ int minusCnt = 0, plusCnt = 0 ; int mainStrand ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].rightStrand == 1 ) ++plusCnt ; else if ( subexons[i].rightStrand == -1 ) ++minusCnt ; } if ( plusCnt > minusCnt ) mainStrand = 1 ; else mainStrand = -1 ; for ( i = 0 ; i < tcnt ; ++i ) { std::vector<int> subexonIdx ; transcripts[i].seVector.GetOnesIndices( subexonIdx ) ; int size = subexonIdx.size() ; int intronCnt = 0 ; int anchorIdx = 0 ; // the subexon adjacent to the only intron. for ( j = 0 ; j < size - 1 ; ++j ) { if ( subexons[ subexonIdx[j] ].end + 1 < subexons[ subexonIdx[j + 1] ].start ) { ++intronCnt ; anchorIdx = j ; } } if ( intronCnt != 1 ) continue ; int anchorExonLength[2] = {0, 0}; int tag = 0 ; for ( j = 0 ; j < size ; ++j ) { anchorExonLength[tag] += subexons[ subexonIdx[j] ].end - subexons[ subexonIdx[j] ].start + 1 ; if ( tag == 0 && subexons[ subexonIdx[j] ].end + 1 < subexons[ subexonIdx[j + 1] ].start ) ++tag ; } int flag = 0 ; if ( subexons[ subexonIdx[anchorIdx] ].rightStrand == mainStrand ) { j = subexonIdx[ anchorIdx ] ; if ( subexons[j].end - subexons[j].start + 1 <= 20 || ( subexons[j+ 1].start == subexons[j].end + 1 && subexons[j + 1].end - subexons[j + 1].start + 1 <= 20 && subexons[j + 1].rightStrand == mainStrand ) ) ++flag ; j = subexonIdx[ anchorIdx + 1 ] ; if ( subexons[j].end - subexons[j].start + 1 <= 20 || ( subexons[j].start == subexons[j - 1].end + 1 && subexons[j - 1].end - subexons[j - 1].start + 1 <= 20 && subexons[j - 1].leftStrand == mainStrand ) ) ++flag ; } if ( anchorExonLength[0] <= 25 || anchorExonLength[1] <= 25 ) flag = 2 ; // the alignment support the intron must be unique and has enough support. int support = 0 ; int uniqSupport = 0 ; for ( j = 0 ; j < tcCnt ; ++j ) { if ( !IsConstraintInTranscript( transcripts[i], scc[ tc[j].i ] ) || !IsConstraintInTranscript( transcripts[i], scc[ tc[j].j ] ) ) continue ; if ( ( scc[ tc[j].i ].vector.Test( subexonIdx[ anchorIdx ] ) && scc[ tc[j].i ].vector.Test( subexonIdx[ anchorIdx + 1 ] ) ) || ( scc[ tc[j].j ].vector.Test( subexonIdx[ anchorIdx ] ) && scc[ tc[j].j ].vector.Test( subexonIdx[ anchorIdx + 1 ] ) ) ) { support += tc[j].support ; uniqSupport += tc[j].uniqSupport ; } } if ( (double)uniqSupport < 0.3 * support || support < txptMinReadDepth ) { flag = 2 ; } if ( flag == 2 ) transcripts[i].abundance = -1 ; } tcnt = RemoveNegativeAbundTranscripts( transcripts ) ; return transcripts.size() ; } void TranscriptDecider::ComputeTranscriptsScore( struct _subexon *subexons, int seCnt, std::map<int, int> *subexonChainSupport, std::vector<struct _transcript> &transcripts ) { int i, j ; int tcnt = transcripts.size() ; struct _constraint tmpC ; tmpC.vector.Init( seCnt ) ; for ( i = 0 ; i < tcnt ; ++i ) transcripts[i].correlationScore = 0 ; for ( i = 0 ; i < seCnt ; ++i ) { for ( std::map<int, int>::iterator it = subexonChainSupport[i].begin() ; it != subexonChainSupport[i].end() ; ++it ) { if ( sampleCnt >= 0 && ( it->second < 3 || it->second < (int)( 0.1 * sampleCnt ) ) && it->second <= sampleCnt / 2 ) continue ; tmpC.vector.Reset() ; tmpC.vector.Set( i ) ; tmpC.vector.Set( it->first ) ; tmpC.first = i ; tmpC.last = it->first ; for ( j = 0 ; j < tcnt ; ++j ) { if ( IsConstraintInTranscript( transcripts[j], tmpC ) ) ++transcripts[j].correlationScore ; } } } tmpC.vector.Release() ; } int TranscriptDecider::Solve( struct _subexon *subexons, int seCnt, std::vector<Constraints> &constraints, SubexonCorrelation &subexonCorrelation ) { int i, j, k ; int cnt = 0 ; int *f = new int[seCnt] ; // this is a general buffer for a type of usage. bool useDP = false ; compatibleTestVectorT.Init( seCnt ) ; // this is the bittable used in compatible test function. compatibleTestVectorC.Init( seCnt ) ; for ( i = 0 ; i < seCnt ; ++i ) { subexons[i].canBeStart = subexons[i].canBeEnd = false ; if ( subexons[i].prevCnt == 0 ) subexons[i].canBeStart = true ; else if ( subexons[i].leftClassifier < canBeSoftBoundaryThreshold && subexons[i].leftClassifier != -1 && subexons[i].leftStrand != 0 ) // The case of overhang. { // We then look into whether there is a left-side end already showed up before this subexon in this region of subexons. bool flag = true ; for ( j = i - 1 ; j >= 0 ; --j ) { if ( subexons[j].end + 1 != subexons[j + 1].start ) break ; if ( subexons[i].canBeStart == true ) { flag = false ; break ; } } subexons[i].canBeStart = flag ; } if ( subexons[i].nextCnt == 0 ) subexons[i].canBeEnd = true ; else if ( subexons[i].rightClassifier < canBeSoftBoundaryThreshold && subexons[i].rightClassifier != -1 && subexons[i].rightStrand != 0 ) { subexons[i].canBeEnd = true ; } // Remove other soft end already showed up in this region of subexons. if ( subexons[i].canBeEnd == true ) { for ( j = i - 1 ; j >= 0 ; --j ) { if ( subexons[j].end + 1 != subexons[j + 1].start ) break ; if ( subexons[j].canBeEnd == true ) { subexons[j].canBeEnd = false ; break ; } } } //printf( "%d: %d %lf\n", subexons[i].canBeStart, subexons[i].prevCnt, subexons[i].leftClassifier ) ; } // Go through the cases of mixture region to set canBeStart/End. // e.g: +[...]+_____+[....]-...]+____+[..)_____-[...]- // ^ then we need to force a start point here. // Do we need to associate a strand information with canBeStart, canBeEnd? for ( i = 0 ; i < seCnt ; ) { // [i, j) is a region. for ( j = i + 1 ; j < seCnt ; ++j ) if ( subexons[j].start > subexons[j - 1].end + 1 ) break ; if ( subexons[i].canBeStart == false ) // then subexons[i] must has a hard left boundary. { int leftStrandCnt[2] = {0, 0} ; for ( k = i ; k < j ; ++k ) { if ( !SubexonGraph::IsSameStrand( subexons[k].rightStrand, subexons[i].leftStrand ) ) break ; if ( subexons[k].leftStrand != 0 ) ++leftStrandCnt[ ( subexons[k].leftStrand + 1 ) / 2 ] ; } if ( k < j && leftStrandCnt[ ( subexons[k].rightStrand + 1 ) / 2 ] == 0 ) subexons[i].canBeStart = true ; } if ( subexons[j - 1].canBeEnd == false ) { int rightStrandCnt[2] = {0, 0} ; for ( k = j - 1 ; k >= i ; --k ) { if ( !SubexonGraph::IsSameStrand( subexons[k].leftStrand, subexons[j - 1].rightStrand ) ) break ; if ( subexons[k].rightStrand != 0 ) ++rightStrandCnt[ ( subexons[k].rightStrand + 1 ) / 2 ] ; } if ( k >= i && rightStrandCnt[ ( subexons[k].leftStrand + 1 ) / 2 ] == 0 ) subexons[j - 1].canBeEnd = true ; } //if ( subexons[i].start == 6870264) // printf( "hi %d %d\n",i , subexons[i].canBeStart ) ; i = j ; } /*for ( i = 0 ; i < seCnt ; ++i ) { printf( "%d %d: %d %d\n", subexons[i].start, subexons[i].end, subexons[i].canBeStart, subexons[i].canBeEnd ) ; }*/ // Find the gene ids. baseGeneId = subexons[0].lcCnt ; usedGeneId = subexons[0].rcCnt ; defaultGeneId[0] = defaultGeneId[1] = -1 ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].geneId < 0 ) continue ; //if ( baseGeneId == -1 || subexons[i].geneId < baseGeneId ) // baseGeneId = subexons[i].geneId ; //if ( subexons[i].geneId > usedGeneId ) // usedGeneId = subexons[i].geneId ; if ( ( subexons[i].rightStrand == -1 || subexons[i].leftStrand == -1 ) && ( defaultGeneId[0] == -1 || subexons[i].geneId < defaultGeneId[0] ) ) defaultGeneId[0] = subexons[i].geneId ; if ( ( subexons[i].rightStrand == 1 || subexons[i].leftStrand == 1 ) && ( defaultGeneId[1] == -1 || subexons[i].geneId < defaultGeneId[1] ) ) defaultGeneId[1] = subexons[i].geneId ; } if ( defaultGeneId[0] == -1 ) defaultGeneId[0] = baseGeneId ; if ( defaultGeneId[1] == -1 ) defaultGeneId[1] = usedGeneId - 1 ; // Go through the constraints to find the chain of subexons that should be kept. std::map<int, int> *subexonChainSupport = new std::map<int, int>[ seCnt ] ; for ( i = 0 ; i < sampleCnt ; ++i ) { std::vector<int> subexonIdx ; std::vector<struct _pair32> chain ; int tcCnt = constraints[i].constraints.size() ; int size ; for ( j = 0 ; j < tcCnt ; ++j ) { struct _constraint c = constraints[i].constraints[j] ; if ( c.uniqSupport < 0.95 * c.support || c.support < 3 ) continue ; subexonIdx.clear() ; c.vector.GetOnesIndices( subexonIdx ) ; size = subexonIdx.size() ; for ( k = 0 ; k < size - 1 ; ++k ) { struct _pair32 p ; p.a = subexonIdx[k] ; p.b = subexonIdx[k + 1] ; //if ( subexons[p.a].end + 1 == 113235898 && subexons[ p.b ].start + 1 == 113236121 ) // printf( "bad bad %d %d %d\n", i, c.uniqSupport, c.support ) ; if ( subexons[ p.a ].end + 1 < subexons[ p.b ].start ) chain.push_back( p ) ; } } // Remove redundancy. sort( chain.begin(), chain.end(), CompSortPairs ) ; size = chain.size() ; k = 0 ; for ( j = 1 ; j < size ; ++j ) { if ( chain[j].a == chain[k].a && chain[j].b == chain[k].b ) continue ; else { ++k ; chain[k] = chain[j] ; } } chain.resize( k + 1 ) ; // Add those to sample count size = k + 1 ; for ( j = 0 ; j < size ; ++j ) { if ( subexonChainSupport[ chain[j].a ].count( chain[j].b ) ) { ++subexonChainSupport[ chain[j].a ][ chain[j].b ] ; } else subexonChainSupport[ chain[j].a ][ chain[j].b ] = 1 ; } } /*for ( i = 0 ; i < seCnt ; ++i ) { printf( "%d:", i ) ; for ( std::map<int, int>::iterator it = subexonChainSupport[i].begin() ; it != subexonChainSupport[i].end() ; ++it ) printf( " (%d %d) ", it->first, it->second ) ; printf( "\n" ) ; }*/ //printf( "%d %d %d\n", defaultGeneId[0], baseGeneId, usedGeneId ) ; cnt = 0 ; memset( f, -1, sizeof( int ) * seCnt ) ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].canBeStart ) { cnt += SubTranscriptCount( i, subexons, f ) ; } } if ( cnt <= USE_DP ) { for ( i = 0 ; i < seCnt ; ++i ) if ( f[i] > USE_DP ) { useDP = true ; break ; } } else useDP = true ; if ( !useDP ) { for ( i = 0 ; i < sampleCnt ; ++i ) { double msize = constraints[i].matePairs.size() ; double csize = constraints[i].constraints.size() ; if ( cnt > ( csize / msize ) * ( csize / msize ) * seCnt && cnt > USE_DP / ( msize * msize ) && cnt > 50 ) { useDP = true ; break ; } } } int atCnt = cnt ; printf( "%d: atCnt=%d seCnt=%d %d %d %d\n", subexons[0].start + 1, atCnt, seCnt, useDP, (int)constraints[0].constraints.size(), (int)constraints[0].matePairs.size() ) ; fflush( stdout ) ; std::vector<struct _transcript> alltranscripts ; if ( !useDP ) { int origSize = atCnt ; alltranscripts.resize( atCnt ) ; for ( i = 0 ; i < atCnt ; ++i ) { alltranscripts[i].seVector.Init( seCnt ) ; alltranscripts[i].correlationScore = 1 ; } atCnt = 0 ; for ( i = 0 ; i < seCnt ; ++i ) { if ( subexons[i].canBeStart ) EnumerateTranscript( i, 0, f, 0, subexons, subexonCorrelation, 1, alltranscripts, atCnt ) ; } for ( i = atCnt ; i < origSize ; ++i ) alltranscripts[i].seVector.Release() ; alltranscripts.resize( atCnt ) ; //printf( "transcript cnt: %d\n", atCnt ) ; //printf( "%d %d\n", alltranscripts[0].seVector.Test( 1 ), constraints[0].matePairs.size() ) ; } else // Use dynamic programming to pick a set of candidate transcript. { std::vector<struct _transcript> sampleTranscripts ; // pre allocate the memory. struct _dpAttribute attr ; attr.f1 = new struct _dp[seCnt] ; if ( seCnt <= 10000 ) { attr.f2 = new struct _dp*[seCnt] ; for ( i = 0 ; i < seCnt ; ++i ) attr.f2[i] = new struct _dp[seCnt] ; } else attr.f2 = NULL ; hashMax = HASH_MAX ; if (seCnt > 500) hashMax = 1000003 ; else if (seCnt > 1000) hashMax = 10000019 ; else if (seCnt > 1500) hashMax = 20000003 ; attr.hash = dpHash ; if ( hashMax != HASH_MAX ) attr.hash = new struct _dp[hashMax] ; for ( i = 0 ; i < seCnt ; ++i ) { attr.f1[i].seVector.Nullify() ; attr.f1[i].seVector.Init( seCnt ) ; for ( j = i ; j < seCnt ; ++j ) { attr.f2[i][j].seVector.Nullify() ; attr.f2[i][j].seVector.Init( seCnt ) ; } } for ( i = 0 ; i < hashMax ; ++i ) { attr.hash[i].seVector.Nullify() ; attr.hash[i].seVector.Init( seCnt ) ; } // select candidate transcripts from each sample. struct _pair32 *sampleComplexity = new struct _pair32[ sampleCnt ] ; for ( i = 0 ; i < sampleCnt ; ++i ) { sampleComplexity[i].a = i ; sampleComplexity[i].b = constraints[i].constraints.size() ; } qsort( sampleComplexity, sampleCnt, sizeof( sampleComplexity[0] ), CompPairsByB ) ; int downsampleCnt = -1 ; for ( i = sampleCnt - 1 ; i >= 0 ; --i ) { sampleTranscripts.clear() ; int iterBound = constraints[ sampleComplexity[i].a ].constraints.size() ; if ( i < sampleCnt - 1 ) iterBound = 100 ; if ( i < sampleCnt - 10 && alltranscripts.size() > 1000 ) iterBound = 10 ; //printf( "%d %d: %d %d %d %d\n", subexons[0].start + 1, sampleComplexity[i].a, constraints[ sampleComplexity[i].a ].constraints.size(), constraints[ sampleComplexity[i].a ].matePairs.size(), // alltranscripts.size(), iterBound ) ; fflush( stdout ) ; if ( maxDpConstraintSize > 0 ) { Constraints truncatedConstraints ; truncatedConstraints.TruncateConstraintsCoverFrom( constraints[ sampleComplexity[i].a ], seCnt, maxDpConstraintSize ) ; PickTranscriptsByDP( subexons, seCnt, iterBound, truncatedConstraints, subexonCorrelation, attr, sampleTranscripts ) ; } else if ( ( constraints[ sampleComplexity[i].a ].constraints.size() > 1000 && constraints[ sampleComplexity[i].a ].constraints.size() * 10 < constraints[ sampleComplexity[i].a ].matePairs.size() ) || ( downsampleCnt > 0 && (int)constraints[ sampleComplexity[i].a ].constraints.size() >= downsampleCnt ) || seCnt >= 1500 ) { Constraints downsampledConstraints ; int stride = (int)constraints[ sampleComplexity[i].a ].matePairs.size() / (int)constraints[ sampleComplexity[i].a ].constraints.size() ; if ( downsampleCnt > 0 ) stride = (int)constraints[ sampleComplexity[i].a ].constraints.size() / downsampleCnt ; if ( stride < 1 ) stride = 1 ; downsampledConstraints.DownsampleConstraintsFrom( constraints[ sampleComplexity[i].a ], stride ) ; if ( downsampleCnt <= 0 ) downsampleCnt = downsampledConstraints.constraints.size() ; if ( iterBound <= 10 ) continue ; PickTranscriptsByDP( subexons, seCnt, iterBound, downsampledConstraints, subexonCorrelation, attr, sampleTranscripts ) ; } else { PickTranscriptsByDP( subexons, seCnt, iterBound, constraints[ sampleComplexity[i].a ], subexonCorrelation, attr, sampleTranscripts ) ; } int size = sampleTranscripts.size() ; for ( j = 0 ; j < size ; ++j ) alltranscripts.push_back( sampleTranscripts[j] ) ; // we can further pick a smaller subsets of transcripts here if the number is still to big. CoalesceSameTranscripts( alltranscripts ) ; AugmentTranscripts( subexons, alltranscripts, 1000, false ) ; } // release the memory. delete[] sampleComplexity ; for ( i = 0 ; i < seCnt ; ++i ) { attr.f1[i].seVector.Release() ; for ( j = i ; j < seCnt && attr.f2 ; ++j ) attr.f2[i][j].seVector.Release() ; } for ( i = 0 ; i < hashMax ; ++i ) attr.hash[i].seVector.Release() ; delete[] attr.f1 ; for ( i = 0 ; i < seCnt && attr.f2 ; ++i ) delete[] attr.f2[i] ; delete[] attr.f2 ; if (hashMax != HASH_MAX) delete[] attr.hash ; } transcriptId = new int[usedGeneId - baseGeneId] ; std::vector<struct _transcript> *predTranscripts = new std::vector<struct _transcript>[sampleCnt] ; atCnt = alltranscripts.size() ; for ( i = 0 ; i < atCnt ; ++i ) alltranscripts[i].FPKM = 0 ; for ( i = 0 ; i < sampleCnt ; ++i ) { int size = alltranscripts.size() ; for ( j = 0 ; j < size ; ++j ) alltranscripts[j].abundance = -1 ; //printf( "pick: %d: %d %d\n", i, constraints[i].matePairs.size(), alltranscripts.size() ) ; PickTranscripts( subexons, alltranscripts, constraints[i], subexonCorrelation, predTranscripts[i] ) ; /*double tmp = FPKMFraction ; FPKMFraction = 0 ; size = predTranscripts.size() ; for ( j = 0 ; j < size ; ++j ) { ConvertTranscriptAbundanceToFPKM( subexons, predTranscripts[j] ) ; } RefineTranscripts( subexons, seCnt, predTranscripts, constraints[i] ) ; FPKMFraction = tmp ;*/ } atCnt = alltranscripts.size() ; int *txptSampleSupport = new int[atCnt] ; memset( txptSampleSupport, 0, sizeof( int ) * atCnt ) ; for ( i = 0 ; i < sampleCnt ; ++i ) { int size = predTranscripts[i].size() ; for ( j = 0 ; j < size ; ++j ) { ++txptSampleSupport[ predTranscripts[i][j].id ] ; ++alltranscripts[ predTranscripts[i][j].id ].FPKM ; } } for ( i = 0 ; i < sampleCnt ; ++i ) { int size = alltranscripts.size() ; for ( j = 0 ; j < size ; ++j ) alltranscripts[j].abundance = -1 ; //printf( "pick: %d: %d %d\n", i, constraints[i].matePairs.size(), alltranscripts.size() ) ; size = predTranscripts[i].size() ; for ( j = 0 ; j < size ; ++j ) { predTranscripts[i][j].seVector.Release() ; } predTranscripts[i].clear() ; PickTranscripts( subexons, alltranscripts, constraints[i], subexonCorrelation, predTranscripts[i] ) ; } std::vector<int> *rawPredTranscriptIds = new std::vector<int>[sampleCnt] ; std::vector<double> *rawPredTranscriptAbundance = new std::vector<double>[sampleCnt] ; for ( i = 0 ; i < sampleCnt ; ++i ) { int size = predTranscripts[i].size() ; for ( j = 0 ; j < size ; ++j ) { rawPredTranscriptIds[i].push_back( predTranscripts[i][j].id ) ; rawPredTranscriptAbundance[i].push_back( predTranscripts[i][j].abundance ) ; } } // Do the filtration. for ( i = 0 ; i < sampleCnt ; ++i ) { int size = predTranscripts[i].size() ; for ( j = 0 ; j < size ; ++j ) { ConvertTranscriptAbundanceToFPKM( subexons, predTranscripts[i][j] ) ; } size = RefineTranscripts( subexons, seCnt, false, subexonChainSupport, txptSampleSupport, predTranscripts[i], constraints[i] ) ; // Recompute the abundance. AbundanceEstimation( subexons, seCnt, constraints[i], predTranscripts[i] ) ; for ( j = 0 ; j < size ; ++j ) ConvertTranscriptAbundanceToFPKM( subexons, predTranscripts[i][j] ) ; size = RefineTranscripts( subexons, seCnt, true, subexonChainSupport, txptSampleSupport, predTranscripts[i], constraints[i] ) ; //ComputeTranscriptsScore( subexons, seCnt, subexonChainSupport, predTranscripts[i] ) ; } // Rescue some filtered transcripts memset( txptSampleSupport, 0, sizeof( int ) * atCnt ) ; for ( i = 0 ; i < sampleCnt ; ++i ) { int size = predTranscripts[i].size() ; for ( j = 0 ; j < size ; ++j ) { ++txptSampleSupport[ predTranscripts[i][j].id ] ; } } bool *predicted = new bool[atCnt] ; for ( i = 0 ; i < sampleCnt ; ++i ) { memset( predicted, false, sizeof( bool ) * atCnt ) ; if ( predTranscripts[i].size() != rawPredTranscriptIds[i].size() ) { int psize = predTranscripts[i].size() ; int rsize = rawPredTranscriptIds[i].size() ; int tcCnt = constraints[i].matePairs.size() ; for ( j = 0 ; j < psize ; ++j ) predicted[ predTranscripts[i][j].id ] = true ; for ( j = 0 ; j < rsize ; ++j ) { int id = rawPredTranscriptIds[i][j] ; if ( predicted[ id ] == false && ( txptSampleSupport[ id ] >= 3 && txptSampleSupport[id] >= 0.25 * sampleCnt ) ) { struct _transcript nt = alltranscripts[id] ; nt.seVector.Nullify() ; nt.seVector.Duplicate( alltranscripts[id].seVector ) ; nt.constraintsSupport = NULL ; nt.correlationScore = -1 ; nt.abundance = rawPredTranscriptAbundance[i][j] ; nt.id = id ; predTranscripts[i].push_back( nt ) ; } } if ( psize != predTranscripts[i].size() ) AbundanceEstimation( subexons, seCnt, constraints[i], predTranscripts[i] ) ; } int size = predTranscripts[i].size() ; if ( 0 ) //size == 1 ) { //AugmentTranscripts( subexons, predTranscripts[i], false ) ; int l = predTranscripts[i].size() ; int tcCnt = constraints[i].matePairs.size() ; for ( j = 0 ; j < l ; ++j ) { predTranscripts[i][j].abundance = 1.0 / alignments.readLen ; } AbundanceEstimation( subexons, seCnt, constraints[i], predTranscripts[i] ) ; std::vector<int> subexonIdx ; for ( j = 0 ; j < l ; ++j ) { subexonIdx.clear() ; predTranscripts[i][j].seVector.GetOnesIndices( subexonIdx ) ; int subexonIdxCnt = subexonIdx.size() ; int len = 0 ; for ( k = 0 ; k < subexonIdxCnt ; ++k ) len += subexons[ subexonIdx[k] ].end - subexons[ subexonIdx[k] ].start + 1 ; if ( predTranscripts[i][j].abundance * alignments.readLen / len < 2.0 ) predTranscripts[i][j].abundance = -1 ; else ConvertTranscriptAbundanceToFPKM( subexons, predTranscripts[i][j] ) ; } RemoveNegativeAbundTranscripts( predTranscripts[i] ) ; } // Output size = predTranscripts[i].size() ; InitTranscriptId() ; for ( j = 0 ; j < size ; ++j ) { OutputTranscript( i, subexons, predTranscripts[i][j] ) ; } for ( j = 0 ; j < size ; ++j ) { predTranscripts[i][j].seVector.Release() ; } } delete []predicted ; delete []transcriptId ; delete []predTranscripts ; delete []rawPredTranscriptIds ; delete []rawPredTranscriptAbundance ; delete []txptSampleSupport ; atCnt = alltranscripts.size() ; for ( i = 0 ; i < atCnt ; ++i ) alltranscripts[i].seVector.Release() ; compatibleTestVectorT.Release() ; compatibleTestVectorC.Release() ; delete[] f ; delete[] subexonChainSupport ; return 0 ; } void *TranscriptDeciderSolve_Wrapper( void *a ) { int i ; struct _transcriptDeciderThreadArg &arg = *( (struct _transcriptDeciderThreadArg *)a ) ; TranscriptDecider transcriptDecider( arg.FPKMFraction, arg.classifierThreshold, arg.txptMinReadDepth, arg.sampleCnt, *( arg.alignments ) ) ; transcriptDecider.SetNumThreads( arg.numThreads + 1 ) ; transcriptDecider.SetMultiThreadOutputHandler( arg.outputHandler ) ; transcriptDecider.SetMaxDpConstraintSize( arg.maxDpConstraintSize ) ; transcriptDecider.Solve( arg.subexons, arg.seCnt, arg.constraints, arg.subexonCorrelation ) ; int start = arg.subexons[0].start ; int end = arg.subexons[ arg.seCnt - 1 ].end ; int chrId = arg.subexons[0].chrId ; // Release memory for ( i = 0 ; i < arg.seCnt ; ++i ) { delete[] arg.subexons[i].prev ; delete[] arg.subexons[i].next ; } delete[] arg.subexons ; // Put the work id back to the free threads queue. pthread_mutex_lock( arg.ftLock ) ; arg.freeThreads[ *( arg.ftCnt ) ] = arg.tid ; ++*( arg.ftCnt ) ; if ( *( arg.ftCnt ) == 1 ) pthread_cond_signal( arg.fullWorkCond ) ; pthread_mutex_unlock( arg.ftLock) ; printf( "Thread %d: %s %d %d finished.\n", arg.tid, arg.alignments->GetChromName(chrId), start + 1, end + 1 ) ; fflush( stdout ) ; pthread_exit( NULL ) ; }