15 #include "SingleRead.h"
16 #include "SingleReadQ.h"
17 #include "PairedEndRead.h"
18 #include "PairedEndReadQ.h"
20 #include "SingleHit.h"
21 #include "PairedEndHit.h"
24 #include "SingleModel.h"
25 #include "SingleQModel.h"
26 #include "PairedEndModel.h"
27 #include "PairedEndQModel.h"
29 #include "Transcript.h"
30 #include "Transcripts.h"
33 #include "GroupInfo.h"
34 #include "HitContainer.h"
35 #include "ReadIndex.h"
36 #include "ReadReader.h"
38 #include "ModelParams.h"
40 #include "HitWrapper.h"
41 #include "BamWriter.h"
45 const double STOP_CRITERIA = 0.001;
46 const int MAX_ROUND = 10000;
47 const int MIN_ROUND = 20;
51 void *reader, *hitv, *ncpv, *mhp, *countv;
55 int m, M; // m genes, M isoforms
56 int N0, N1, N2, N_tot;
60 bool genBamF; // If user wants to generate bam file, true; otherwise, false.
61 bool updateModel, calcExpectedWeights;
62 bool genGibbsOut; // generate file for Gibbs sampler
64 char refName[STRLEN], imdName[STRLEN], outName[STRLEN];
65 char refF[STRLEN], groupF[STRLEN], cntF[STRLEN], tiF[STRLEN];
66 char mparamsF[STRLEN], bmparamsF[STRLEN];
69 char *pt_fn_list, *pt_chr_list;
70 char inpSamF[STRLEN], outBamF[STRLEN], fn_list[STRLEN], chr_list[STRLEN];
72 char out_for_gibbs_F[STRLEN];
74 vector<double> theta, eel; // eel : expected effective length
76 double *probv, **countvs;
80 Transcripts transcripts;
84 template<class ReadType, class HitType, class ModelType>
85 void init(ReadReader<ReadType> **&readers, HitContainer<HitType> **&hitvs, double **&ncpvs, ModelType **&mhps) {
86 int nReads, nHits, rt;
87 int nrLeft, nhT, curnr; // nrLeft : number of reads left, nhT : hit threshold per thread, curnr: current number of reads
91 char readFs[2][STRLEN];
92 ReadIndex *indices[2];
95 readers = new ReadReader<ReadType>*[nThreads];
96 genReadFileNames(imdName, 1, read_type, s, readFs);
97 for (int i = 0; i < s; i++) {
98 indices[i] = new ReadIndex(readFs[i]);
100 for (int i = 0; i < nThreads; i++) {
101 readers[i] = new ReadReader<ReadType>(s, readFs);
102 readers[i]->setIndices(indices);
105 hitvs = new HitContainer<HitType>*[nThreads];
106 for (int i = 0; i < nThreads; i++) {
107 hitvs[i] = new HitContainer<HitType>();
110 sprintf(datF, "%s.dat", imdName);
112 if (!fin.is_open()) { fprintf(stderr, "Cannot open %s! It may not exist.\n", datF); exit(-1); }
113 fin>>nReads>>nHits>>rt;
114 if (nReads != N1) { fprintf(stderr, "Number of alignable reads does not match!\n"); exit(-1); }
115 //assert(nReads == N1);
116 if (rt != read_type) { fprintf(stderr, "Data file (.dat) does not have the right read type!\n"); exit(-1); }
117 //assert(rt == read_type);
119 //A just so so strategy for paralleling
120 nhT = nHits / nThreads;
124 ncpvs = new double*[nThreads];
125 for (int i = 0; i < nThreads; i++) {
126 int ntLeft = nThreads - i - 1; // # of threads left
127 if (!readers[i]->locate(curnr)) { fprintf(stderr, "Read indices files do not match!\n"); exit(-1); }
128 //assert(readers[i]->locate(curnr));
130 while (nrLeft > ntLeft && hitvs[i]->getNHits() < nhT) {
131 if (!hitvs[i]->read(fin)) { fprintf(stderr, "Cannot read alignments from .dat file!\n"); exit(-1); }
132 //assert(hitvs[i]->read(fin));
134 if (verbose && nrLeft % 1000000 == 0) { printf("DAT %d reads left!\n", nrLeft); }
136 ncpvs[i] = new double[hitvs[i]->getN()];
137 memset(ncpvs[i], 0, sizeof(double) * hitvs[i]->getN());
138 curnr += hitvs[i]->getN();
140 if (verbose) { printf("Thread %d : N = %d, NHit = %d\n", i, hitvs[i]->getN(), hitvs[i]->getNHits()); }
145 mhps = new ModelType*[nThreads];
146 for (int i = 0; i < nThreads; i++) {
147 mhps[i] = new ModelType(mparams, false); // just model helper
150 probv = new double[M + 1];
151 countvs = new double*[nThreads];
152 for (int i = 0; i < nThreads; i++) {
153 countvs[i] = new double[M + 1];
157 if (verbose) { printf("EM_init finished!\n"); }
160 template<class ReadType, class HitType, class ModelType>
161 void* E_STEP(void* arg) {
162 Params *params = (Params*)arg;
163 ModelType *model = (ModelType*)(params->model);
164 ReadReader<ReadType> *reader = (ReadReader<ReadType>*)(params->reader);
165 HitContainer<HitType> *hitv = (HitContainer<HitType>*)(params->hitv);
166 double *ncpv = (double*)(params->ncpv);
167 ModelType *mhp = (ModelType*)(params->mhp);
168 double *countv = (double*)(params->countv);
170 bool needCalcConPrb = model->getNeedCalcConPrb();
174 int N = hitv->getN();
176 vector<double> fracs; //to remove this, do calculation twice
179 if (needCalcConPrb || updateModel) { reader->reset(); }
180 if (updateModel) { mhp->init(); }
182 memset(countv, 0, sizeof(double) * (M + 1));
183 for (int i = 0; i < N; i++) {
184 if (needCalcConPrb || updateModel) {
185 if (!reader->next(read)) {
186 fprintf(stderr, "Can not load a read!\n");
189 //assert(reader->next(read));
191 fr = hitv->getSAt(i);
192 to = hitv->getSAt(i + 1);
193 fracs.resize(to - fr + 1);
197 if (needCalcConPrb) { ncpv[i] = model->getNoiseConPrb(read); }
198 fracs[0] = probv[0] * ncpv[i];
199 if (fracs[0] < EPSILON) fracs[0] = 0.0;
201 for (int j = fr; j < to; j++) {
202 HitType &hit = hitv->getHitAt(j);
203 if (needCalcConPrb) { hit.setConPrb(model->getConPrb(read, hit)); }
205 fracs[id] = probv[hit.getSid()] * hit.getConPrb();
206 if (fracs[id] < EPSILON) fracs[id] = 0.0;
210 if (sum >= EPSILON) {
212 countv[0] += fracs[0];
213 if (updateModel) { mhp->updateNoise(read, fracs[0]); }
214 if (calcExpectedWeights) { ncpv[i] = fracs[0]; }
215 for (int j = fr; j < to; j++) {
216 HitType &hit = hitv->getHitAt(j);
219 countv[hit.getSid()] += fracs[id];
220 if (updateModel) { mhp->update(read, hit, fracs[id]); }
221 if (calcExpectedWeights) { hit.setConPrb(fracs[id]); }
224 else if (calcExpectedWeights) {
226 for (int j = fr; j < to; j++) {
227 HitType &hit = hitv->getHitAt(j);
236 template<class ReadType, class HitType, class ModelType>
237 void* calcConProbs(void* arg) {
238 Params *params = (Params*)arg;
239 ModelType *model = (ModelType*)(params->model);
240 ReadReader<ReadType> *reader = (ReadReader<ReadType>*)(params->reader);
241 HitContainer<HitType> *hitv = (HitContainer<HitType>*)(params->hitv);
242 double *ncpv = (double*)(params->ncpv);
245 int N = hitv->getN();
248 assert(model->getNeedCalcConPrb());
251 for (int i = 0; i < N; i++) {
252 if (!reader->next(read)) {
253 fprintf(stderr, "Can not load a read!\n");
256 fr = hitv->getSAt(i);
257 to = hitv->getSAt(i + 1);
259 ncpv[i] = model->getNoiseConPrb(read);
260 for (int j = fr; j < to; j++) {
261 HitType &hit = hitv->getHitAt(j);
262 hit.setConPrb(model->getConPrb(read, hit));
269 template<class ModelType>
270 void calcExpectedEffectiveLengths(ModelType& model) {
272 double *pdf = NULL, *cdf = NULL, *clen = NULL; // clen[i] = sigma_{j=1}^{i}pdf[i]*(lb+i)
274 model.getGLD().copyTo(pdf, cdf, lb, ub, span);
275 clen = new double[span + 1];
277 for (int i = 1; i <= span; i++) {
278 clen[i] = clen[i - 1] + pdf[i] * (lb + i);
282 eel.resize(M + 1, 0.0);
283 for (int i = 1; i <= M; i++) {
284 int totLen = refs.getRef(i).getTotLen();
285 int fullLen = refs.getRef(i).getFullLen();
286 int pos1 = max(min(totLen - fullLen + 1, ub) - lb, 0);
287 int pos2 = max(min(totLen, ub) - lb, 0);
289 if (pos2 == 0) { eel[i] = 0.0; continue; }
291 eel[i] = fullLen * cdf[pos1] + ((cdf[pos2] - cdf[pos1]) * (totLen + 1) - (clen[pos2] - clen[pos1]));
293 if (eel[i] < MINEEL) { eel[i] = 0.0; }
301 template<class ModelType>
302 void writeResults(ModelType& model, double* counts) {
304 char modelF[STRLEN], thetaF[STRLEN];
308 sprintf(modelF, "%s.model", outName);
311 sprintf(thetaF, "%s.theta", outName);
312 fo = fopen(thetaF, "w");
313 fprintf(fo, "%d\n", M + 1);
314 for (int i = 0; i < M; i++) fprintf(fo, "%.15g ", theta[i]);
315 fprintf(fo, "%.15g\n", theta[M]);
318 //calculate tau values
319 double *tau = new double[M + 1];
320 memset(tau, 0, sizeof(double) * (M + 1));
323 for (int i = 1; i <= M; i++)
324 if (eel[i] >= EPSILON) {
325 tau[i] = theta[i] / eel[i];
328 if (denom <= 0) { fprintf(stderr, "No alignable reads?!\n"); exit(-1); }
330 for (int i = 1; i <= M; i++) {
334 //isoform level results
335 sprintf(outF, "%s.iso_res", imdName);
336 fo = fopen(outF, "w");
337 for (int i = 1; i <= M; i++) {
338 const Transcript& transcript = transcripts.getTranscriptAt(i);
339 fprintf(fo, "%s%c", transcript.getTranscriptID().c_str(), (i < M ? '\t' : '\n'));
341 for (int i = 1; i <= M; i++)
342 fprintf(fo, "%.2f%c", counts[i], (i < M ? '\t' : '\n'));
343 for (int i = 1; i <= M; i++)
344 fprintf(fo, "%.15g%c", tau[i], (i < M ? '\t' : '\n'));
345 for (int i = 1; i <= M; i++) {
346 const Transcript& transcript = transcripts.getTranscriptAt(i);
347 fprintf(fo, "%s%c", transcript.getLeft().c_str(), (i < M ? '\t' : '\n'));
352 sprintf(outF, "%s.gene_res", imdName);
353 fo = fopen(outF, "w");
354 for (int i = 0; i < m; i++) {
355 const string& gene_id = transcripts.getTranscriptAt(gi.spAt(i)).getGeneID();
356 fprintf(fo, "%s%c", gene_id.c_str(), (i < m - 1 ? '\t' : '\n'));
358 for (int i = 0; i < m; i++) {
359 double sumC = 0.0; // sum of counts
360 int b = gi.spAt(i), e = gi.spAt(i + 1);
361 for (int j = b; j < e; j++) sumC += counts[j];
362 fprintf(fo, "%.2f%c", sumC, (i < m - 1 ? '\t' : '\n'));
364 for (int i = 0; i < m; i++) {
365 double sumT = 0.0; // sum of tau values
366 int b = gi.spAt(i), e = gi.spAt(i + 1);
367 for (int j = b; j < e; j++) sumT += tau[j];
368 fprintf(fo, "%.15g%c", sumT, (i < m - 1 ? '\t' : '\n'));
370 for (int i = 0; i < m; i++) {
371 int b = gi.spAt(i), e = gi.spAt(i + 1);
372 for (int j = b; j < e; j++) {
373 fprintf(fo, "%s%c", transcripts.getTranscriptAt(j).getTranscriptID().c_str(), (j < e - 1 ? ',' : (i < m - 1 ? '\t' :'\n')));
380 if (verbose) { printf("Expression Results are written!\n"); }
383 template<class ReadType, class HitType, class ModelType>
384 void release(ReadReader<ReadType> **readers, HitContainer<HitType> **hitvs, double **ncpvs, ModelType **mhps) {
386 for (int i = 0; i < nThreads; i++) {
391 for (int i = 0; i < nThreads; i++) {
403 inline bool doesUpdateModel(int ROUND) {
404 //return false; // never update, for debugging only
405 return ROUND <= 20 || ROUND % 100 == 0;
408 //Including initialize, algorithm and results saving
409 template<class ReadType, class HitType, class ModelType>
414 double bChange = 0.0, change = 0.0; // bChange : biggest change
417 ModelType model(mparams); //master model
418 ReadReader<ReadType> **readers;
419 HitContainer<HitType> **hitvs;
421 ModelType **mhps; //model helpers
423 Params fparams[nThreads];
424 pthread_t threads[nThreads];
429 //initialize boolean variables
430 updateModel = calcExpectedWeights = false;
433 theta.resize(M + 1, 0.0);
434 init<ReadType, HitType, ModelType>(readers, hitvs, ncpvs, mhps);
436 //set initial parameters
438 theta[0] = max(N0 * 1.0 / (N_tot - N2), 1e-8);
439 double val = (1.0 - theta[0]) / M;
440 for (int i = 1; i <= M; i++) theta[i] = val;
442 model.estimateFromReads(imdName);
444 for (int i = 0; i < nThreads; i++) {
445 fparams[i].model = (void*)(&model);
447 fparams[i].reader = (void*)readers[i];
448 fparams[i].hitv = (void*)hitvs[i];
449 fparams[i].ncpv = (void*)ncpvs[i];
450 fparams[i].mhp = (void*)mhps[i];
451 fparams[i].countv = (void*)countvs[i];
454 /* set thread attribute to be joinable */
455 pthread_attr_init(&attr);
456 pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
462 updateModel = doesUpdateModel(ROUND);
464 for (int i = 0; i <= M; i++) probv[i] = theta[i];
467 for (int i = 0; i < nThreads; i++) {
468 rc = pthread_create(&threads[i], &attr, E_STEP<ReadType, HitType, ModelType>, (void*)(&fparams[i]));
469 if (rc != 0) { fprintf(stderr, "Cannot create thread %d at ROUND %d! (numbered from 0)\n", i, ROUND); exit(-1); }
473 for (int i = 0; i < nThreads; i++) {
474 rc = pthread_join(threads[i], &status);
475 if (rc != 0) { fprintf(stderr, "Cannot join thread %d at ROUND %d! (numbered from 0)\n", i, ROUND); exit(-1); }
479 model.setNeedCalcConPrb(false);
481 for (int i = 1; i < nThreads; i++) {
482 for (int j = 0; j <= M; j++) {
483 countvs[0][j] += countvs[i][j];
492 for (int i = 0; i <= M; i++) sum += countvs[0][i];
493 assert(sum >= EPSILON);
494 for (int i = 0; i <= M; i++) theta[i] = countvs[0][i] / sum;
498 for (int i = 0; i < nThreads; i++) { model.collect(*mhps[i]); }
503 bChange = 0.0; totNum = 0;
504 for (int i = 0; i <= M; i++)
505 if (probv[i] >= 1e-7) {
506 change = fabs(theta[i] - probv[i]) / probv[i];
507 if (change >= STOP_CRITERIA) ++totNum;
508 if (bChange < change) bChange = change;
511 if (verbose) printf("ROUND = %d, SUM = %.15g, bChange = %f, totNum = %d\n", ROUND, sum, bChange, totNum);
512 } while (ROUND < MIN_ROUND || totNum > 0 && ROUND < MAX_ROUND);
513 //while (ROUND < MAX_ROUND);
515 if (totNum > 0) fprintf(stderr, "Warning: RSEM reaches %d iterations before meeting the convergence criteria.\n", MAX_ROUND);
517 //calculate expected effective lengths for each isoform
518 calcExpectedEffectiveLengths<ModelType>(model);
520 //correct theta vector
522 for (int i = 1; i <= M; i++)
523 if (eel[i] < EPSILON) { theta[i] = 0.0; }
524 else sum += theta[i];
525 if (sum < EPSILON) { fprintf(stderr, "No Expected Effective Length is no less than %.6g?!\n", MINEEL); exit(-1); }
526 for (int i = 0; i <= M; i++) theta[i] /= sum;
528 //generate output file used by Gibbs sampler
530 if (model.getNeedCalcConPrb()) {
531 for (int i = 0; i < nThreads; i++) {
532 rc = pthread_create(&threads[i], &attr, calcConProbs<ReadType, HitType, ModelType>, (void*)(&fparams[i]));
533 if (rc != 0) { fprintf(stderr, "Cannot create thread %d when generate files for Gibbs sampler! (numbered from 0)\n", i); exit(-1); }
535 for (int i = 0; i < nThreads; i++) {
536 rc = pthread_join(threads[i], &status);
537 if (rc != 0) { fprintf(stderr, "Cannot join thread %d when generate files for Gibbs sampler! (numbered from 0)\n", i); exit(-1); }
540 model.setNeedCalcConPrb(false);
542 sprintf(out_for_gibbs_F, "%s.ofg", imdName);
543 FILE *fo = fopen(out_for_gibbs_F, "w");
544 fprintf(fo, "%d %d\n", M, N0);
545 for (int i = 0; i < nThreads; i++) {
546 int numN = hitvs[i]->getN();
547 for (int j = 0; j < numN; j++) {
548 int fr = hitvs[i]->getSAt(j);
549 int to = hitvs[i]->getSAt(j + 1);
552 if (ncpvs[i][j] > 0.0) { ++totNum; fprintf(fo, "%d %.15g ", 0, ncpvs[i][j]); }
553 for (int k = fr; k < to; k++) {
554 HitType &hit = hitvs[i]->getHitAt(k);
555 if (hit.getConPrb() >= EPSILON) {
557 fprintf(fo, "%d %.15g ", hit.getSid(), hit.getConPrb());
561 if (totNum > 0) { fprintf(fo, "\n"); }
567 //calculate expected weights and counts using learned parameters
568 updateModel = false; calcExpectedWeights = true;
569 for (int i = 0; i < nThreads; i++) {
570 rc = pthread_create(&threads[i], &attr, E_STEP<ReadType, HitType, ModelType>, (void*)(&fparams[i]));
571 if (rc != 0) { fprintf(stderr, "Cannot create thread %d when calculate expected weights! (numbered from 0)\n", i); exit(-1); }
574 for (int i = 0; i < nThreads; i++) {
575 rc = pthread_join(threads[i], &status);
576 if (rc != 0) { fprintf(stderr, "Cannot join thread %d! (numbered from 0) when calculate expected weights!\n", i); exit(-1); }
579 model.setNeedCalcConPrb(false);
580 for (int i = 1; i < nThreads; i++) {
581 for (int j = 0; j <= M; j++) {
582 countvs[0][j] += countvs[i][j];
587 /* destroy attribute */
588 pthread_attr_destroy(&attr);
592 double *mw = model.getMW();
594 for (int i = 0; i <= M; i++) {
595 theta[i] = (mw[i] < EPSILON ? 0.0 : theta[i] / mw[i]);
598 assert(sum >= EPSILON);
599 for (int i = 0; i <= M; i++) theta[i] /= sum;
601 writeResults<ModelType>(model, countvs[0]);
604 sprintf(outBamF, "%s.bam", outName);
605 if (transcripts.getType() == 0) {
606 sprintf(chr_list, "%s.chrlist", refName);
607 pt_chr_list = (char*)(&chr_list);
610 BamWriter writer(inpSamType, inpSamF, pt_fn_list, outBamF, pt_chr_list);
611 HitWrapper<HitType> wrapper(nThreads, hitvs);
612 writer.work(wrapper, transcripts);
615 release<ReadType, HitType, ModelType>(readers, hitvs, ncpvs, mhps);
618 int main(int argc, char* argv[]) {
623 printf("Usage : rsem-run-em refName read_type imdName outName [-p #Threads] [-b samInpType samInpF has_fn_list_? [fn_list]] [-q] [--gibbs-out]\n\n");
624 printf(" refName: reference name\n");
625 printf(" read_type: 0 single read without quality score; 1 single read with quality score; 2 paired-end read without quality score; 3 paired-end read with quality score.\n");
626 printf(" imdName: name for all upstream/downstream user-unseen files. (different files have different suffices)\n");
627 printf(" outName: name for all output files. (different files have different suffices)\n");
628 printf(" -p: number of threads which user wants to use. (default: 1)\n");
629 printf(" -b: produce bam format output file. (default: off)\n");
630 printf(" -q: set it quiet\n");
631 printf(" --gibbs-out: generate output file use by Gibbs sampler. (default: off)\n");
632 printf("// model parameters should be in imdName.mparams.\n");
636 time_t a = time(NULL);
638 strcpy(refName, argv[1]);
639 read_type = atoi(argv[2]);
640 strcpy(imdName, argv[3]);
641 strcpy(outName, argv[4]);
647 pt_fn_list = pt_chr_list = NULL;
649 for (int i = 5; i < argc; i++) {
650 if (!strcmp(argv[i], "-p")) { nThreads = atoi(argv[i + 1]); }
651 if (!strcmp(argv[i], "-b")) {
653 inpSamType = argv[i + 1][0];
654 strcpy(inpSamF, argv[i + 2]);
655 if (atoi(argv[i + 3]) == 1) {
656 strcpy(fn_list, argv[i + 4]);
657 pt_fn_list = (char*)(&fn_list);
660 if (!strcmp(argv[i], "-q")) { quiet = true; }
661 if (!strcmp(argv[i], "--gibbs-out")) { genGibbsOut = true; }
663 if (nThreads <= 0) { fprintf(stderr, "Number of threads should be bigger than 0!\n"); exit(-1); }
664 //assert(nThreads > 0);
669 sprintf(refF, "%s.seq", refName);
672 sprintf(groupF, "%s.grp", refName);
676 sprintf(tiF, "%s.ti", refName);
677 transcripts.readFrom(tiF);
679 sprintf(cntF, "%s.cnt", imdName);
681 if (!fin.is_open()) { fprintf(stderr, "Cannot open %s! It may not exist.\n", cntF); exit(-1); }
682 fin>>N0>>N1>>N2>>N_tot;
685 if (N1 <= 0) { fprintf(stderr, "There are no alignable reads!\n"); exit(-1); }
687 if (nThreads > N1) nThreads = N1;
689 //set model parameters
691 mparams.N[0] = N0; mparams.N[1] = N1; mparams.N[2] = N2;
692 mparams.refs = &refs;
694 sprintf(mparamsF, "%s.mparams", imdName);
696 if (!fin.is_open()) { fprintf(stderr, "Cannot open %s! It may not exist.\n", mparamsF); exit(-1); }
697 fin>> mparams.minL>> mparams.maxL>> mparams.probF;
698 int val; // 0 or 1 , for estRSPD
700 mparams.estRSPD = (val != 0);
701 fin>> mparams.B>> mparams.mate_minL>> mparams.mate_maxL>> mparams.mean>> mparams.sd;
702 fin>> mparams.seedLen;
707 case 0 : EM<SingleRead, SingleHit, SingleModel>(); break;
708 case 1 : EM<SingleReadQ, SingleHit, SingleQModel>(); break;
709 case 2 : EM<PairedEndRead, PairedEndHit, PairedEndModel>(); break;
710 case 3 : EM<PairedEndReadQ, PairedEndHit, PairedEndQModel>(); break;
711 default : fprintf(stderr, "Unknown Read Type!\n"); exit(-1);
714 time_t b = time(NULL);