1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
|
#include "SoloFeature.h"
#include "serviceFuns.cpp"
#include "SimpleGoodTuring/sgt.h"
#include <math.h>
#include <unordered_set>
#include <map>
#include <random>
double logMultinomialPDFsparse(const vector<double> &ambProfileLogP, const vector<uint32> &countCellGeneUMI, const uint32 stride, const uint32 shift, const int64 start, const uint32 nGenes, const vector<double> &logFactorial);
void SoloFeature::emptyDrops_CR()
{
if (nCB<=pSolo.cellFilter.eDcr.indMin) {
P.inOut->logMain << "emptyDrops_CR filtering: no empty cells found: nCB=" << nCB <<" emptyCellMinIndex="<< pSolo.cellFilter.eDcr.indMin << "\n";
return;
};
time_t rawTime;
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) <<" ... starting emptyDrops_CR filtering" <<endl;
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//find genes that were detected in all cells
unordered_set<uint32> featDet;
for (uint32 icb=0; icb<nCB; icb++) {
for (uint32 ig=0; ig<nGenePerCB[icb]; ig++) {
uint32 irec=countCellGeneUMIindex[icb]+ig*countMatStride;
if (countCellGeneUMI[irec + pSolo.umiDedup.countInd.main] > 0)
featDet.insert(countCellGeneUMI[irec]); //gene is present if it's count > 0 for
};
};
uint32 featDetN=featDet.size(); //total number of detected genes - this should have been done already?
////////////////////////////////////////////////////////////////////////////////////////////////////////////////
//indCount - total UMI per cell sorted descending
typedef struct {uint32 index, count;} IndCount;
vector<IndCount> indCount(nCB);
for (uint32 ii=0; ii<nCB; ii++) {
indCount[ii].index=ii;
indCount[ii].count=nUMIperCB[ii];
};
std::sort(indCount.begin(), indCount.end(), [](const IndCount &ic1, const IndCount &ic2) {
return (ic1.count>ic2.count) || (ic1.count==ic2.count && ic1.index<ic2.index); //descending order by count, ascending by index
});
//////////////////////////////////////////////////////////////////////////////////////////
//ambient gene counts: sum gene expression over the collection of empty cells
vector<uint32> ambCount(featuresNumber,0);
for (auto icb=pSolo.cellFilter.eDcr.indMin; icb<min(nCB,pSolo.cellFilter.eDcr.indMax); icb++) {
auto icb1 = indCount[icb].index;
for (uint32 ig=0; ig<nGenePerCB[icb1]; ig++) {
auto irec = countCellGeneUMIindex[icb1]+ig*countMatStride;
ambCount[countCellGeneUMI[irec+0]] += countCellGeneUMI[irec + pSolo.umiDedup.countInd.main];
};
};
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) <<" ... finished ambient cells counting" <<endl;
////////////////////////////////////////////////////////////////////
//frequencies
map<uint32,uint32> ambCountFreq; //ordered map is not really needed
for (auto &ac: ambCount) {
ambCountFreq[ac]++;
};
if (ambCountFreq.size()<=1) {//only 0-frequency genes are in the empty cells. This is possible because nCB can contain ome cells with no genes - because of multigene
P.inOut->logMain << "emptyDrops_CR filtering: empty cells contain no genes\n";
return;
};
ambCountFreq[0] -= (featuresNumber-featDetN); //subtract genes that were not detected in *any* cells
uint32 maxFreq = ambCountFreq.rbegin()->first;
///////////////////////////////////////////////////////////////////////
//SGT
vector<double> ambCountFreqSGT(maxFreq+1);//up to max frequency
{//SGT estimate of ambient profile
SGT<uint32> sgt;
for (auto &cf: ambCountFreq) {
if (cf.first != 0)
sgt.add(cf.first, cf.second);
};
sgt.analyse();
for (uint32 freq=0; freq<=maxFreq; freq++) {
sgt.estimate(freq, ambCountFreqSGT[freq]);
};
ambCountFreqSGT[0] /= ambCountFreq[0]; //divide freq=0 probability equally among all undetected genes in ambient profile
};
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) <<" ... finished SGT"<<endl;
//ambient profile for all features
vector<double> ambProfileLogP(featuresNumber, 0.0);//logarithm
vector<double> ambProfilePnon0, ambProfileLogPnon0;//only non-0 genes
{
for (uint32 ig=0; ig<featuresNumber; ig++) {
if (featDet.count(ig)>0) {//this is only needed if normalization below is performed
ambProfileLogP[ig]=ambCountFreqSGT[ambCount[ig]];
};
};
double norm1 = accumulate(ambProfileLogP.begin(), ambProfileLogP.end(), 0.0);
ambProfileLogPnon0.reserve(ambProfileLogP.size());
ambProfilePnon0.reserve(ambProfileLogP.size());
for (auto &cf: ambProfileLogP) {
if (cf>0) {
cf /= norm1;//normalization is just in case
ambProfilePnon0.push_back(cf);
cf = std::log(cf);
ambProfileLogPnon0.push_back(cf);
};
};
};
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) <<" ... finished ambient profile"<<endl;
//select candidate cells
uint32 iCandFirst, iCandLast; //first/last candidate cell in the descending sorted indCount
{
iCandFirst=filteredCells.nCellsSimple;//candidates start right after the cutoff for the simple filtering
uint32 minUMI = int(pSolo.cellFilter.eDcr.umiMinFracMedian * nUMIperCBsorted[filteredCells.nCellsSimple/2]);//this is not exactly median
minUMI = max(pSolo.cellFilter.eDcr.umiMin, minUMI);
for (iCandLast=iCandFirst; iCandLast<iCandFirst+pSolo.cellFilter.eDcr.candMaxN; iCandLast++) {
if (indCount[iCandLast].count<minUMI)
break;
};
--iCandLast;
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) << " ... candidate cells: minUMI="<< minUMI << "; number of candidate cells=" << iCandLast-iCandFirst+1 <<endl;
if (iCandLast<iCandFirst)
return; //no candidate cells to consider
};
//calculate observed probability for each candidate
vector<double> obsLogProb(iCandLast-iCandFirst+1);
{
vector<double> logFactorial; //tabulate log-factorial
logFactorial.resize(indCount[iCandFirst].count+1);
logFactorial[1]=0;
for (uint32 cc=2; cc<logFactorial.size(); cc++)
logFactorial[cc]=logFactorial[cc-1]+std::log(cc);
for (uint32 icand=0; icand<obsLogProb.size(); icand++) {
auto icell=indCount[icand+iCandFirst].index;
obsLogProb[icand]=logMultinomialPDFsparse(ambProfileLogP, countCellGeneUMI, countMatStride, pSolo.umiDedup.countInd.main, countCellGeneUMIindex[icell], nGenePerCB[icell], logFactorial);
};
time(&rawTime);
}
P.inOut->logMain << timeMonthDayTime(rawTime) << " ... finished observed logProb" <<endl;
//simulate the probabilities for each cell count
vector<vector<double>> simLogProb(pSolo.cellFilter.eDcr.simN);
{
std::discrete_distribution<uint32> distrAmb ( ambProfilePnon0.begin(), ambProfilePnon0.end() );
auto maxCount=indCount[iCandFirst].count;
//#pragma omp parallel for num_threads(P.runThreadN) //does not increase speed significantly - might be useful for larger number of simulations
for (uint64 isim=0; isim<simLogProb.size(); isim++) {
simLogProb[isim].resize(maxCount+1);
simLogProb[isim][0]=0;
std::mt19937 rngGen(19760110LLU*(isim+1));
vector<uint32> currCounts(ambProfilePnon0.size(), 0);
for (uint32 ic=1; ic<=maxCount; ic++) {
uint32 ig1 = distrAmb(rngGen);
currCounts[ig1]++;
simLogProb[isim][ic] = simLogProb[isim][ic-1] + ambProfileLogPnon0[ig1] + std::log(ic) - std::log(currCounts[ig1]);
};
};
};
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) << " ... finished simulations" <<endl;
//p-values
typedef struct{uint32 index; double p; double padj;} IndPPadj;
vector<IndPPadj> pValues(obsLogProb.size());
{
for (uint32 icand=0; icand<obsLogProb.size(); icand++) {
pValues[icand].index=indCount[icand+iCandFirst].index;
auto count1=indCount[icand+iCandFirst].count;
//auto funSumLess = [&] (uint32 n, vector<double> sp) { return n + (sp[count1]<obsLogProb[icand]); };
//uint32 nLowerP = std::accumulate<uint32>(simLogProb.begin(), simLogProb.end(), 0, funSumLess);
uint32 nLowerP=0;
//for (uint64 isim=0; isim<simLogProb.size(); isim++) {
for (auto &sp: simLogProb) {
nLowerP += ( sp[count1]<obsLogProb[icand] );
};
pValues[icand].p=double(1+nLowerP)/(1+simLogProb.size());
};
//BH
std::sort(pValues.begin(), pValues.end(), [](const IndPPadj &ip1, const IndPPadj &ip2) {return (ip1.p < ip2.p);} );
uint32 rank=0;
for (auto &ip: pValues) {
rank++;
ip.padj=ip.p*pValues.size()/rank;
};
for (auto ip=pValues.rbegin()+1; ip!=pValues.rend(); ++ip)
ip->padj = min(ip->padj, (ip-1)->padj); //make it non-decreasing
};
uint32 extraCells=0;
for (auto &ip: pValues) {
if (ip.padj<=pSolo.cellFilter.eDcr.FDR) {
++extraCells;
filteredCells.filtVecBool[ip.index]=true;
};
};
time(&rawTime);
P.inOut->logMain << timeMonthDayTime(rawTime) << " ... finished emptyDrops_CR filtering: number of additional non-ambient cells=" << extraCells <<endl;
return;
};
double logMultinomialPDFsparse(const vector<double> &ambProfileLogP, const vector<uint32> &countCellGeneUMI, const uint32 stride, const uint32 shift, const int64 start, const uint32 nGenes, const vector<double> &logFactorial)
{
uint32 sumCount=0;
double sumLogFac=0.0, sumCountLogP=0.0;
for (uint32 ig=0; ig<nGenes; ig++) {
auto count1 = countCellGeneUMI[start+ig*stride+shift];
sumCount += count1;
sumLogFac += logFactorial[count1];
sumCountLogP += ambProfileLogP[countCellGeneUMI[start+ig*stride]] * count1;
};
return logFactorial[sumCount] - sumLogFac + sumCountLogP;
};
|
