def makeRelDiffPlot(truthCol, predCol, df, outBase, method, measure):
plt.cla()
plt.clf()
if measure == "tpm":
outFile = "{}_tpm_rel_diff.pdf".format(outBase)
elif measure == "num_reads":
outFile = "{}_num_reads_rel_diff.pdf".format(outBase)
# Get the relative difference numbers
rdV, nzV = AnalysisUtils.relDiff(truthCol, predCol, df)
try:
_ = plt.hist(rdV["relDiff"], bins=100, histtype="stepfilled")
mdText = "mean rel diff = {0:.2f}\nmean |rel diff| = {1:.2f}".format(
rdV["relDiff"].mean(), rdV["relDiff"].abs().mean())
plt.figtext(0.15, 0.85, mdText)
ax = plt.axes()
ax.set_ylabel("frequency")
ax.set_xlabel("relative difference")
# Get rid of axis spines
sns.despine()
plt.savefig(outFile)
except:
f = open(outFile, 'w')
f.close()
print("Error making relative difference plot {}".format(outFile))
def rsemRelDiffPlots(D, measure, methods, simIDs, outdir, plotname):
sns.set_style('white')
sns.set_context('poster')
plt.clf()
plt.cla()
md = {}
for m in methods:
md[m] = []
for i in simIDs:
rd, _ = AnalysisUtils.relDiff("{}_truth{}".format(measure, str(i)),
"{}_{}{}".format(measure, m, str(i)), D)
md[m].append(rd['relDiff'].abs().mean())
ax = None
for mn, mv in md.items():
weights = np.ones_like(mv)/len(mv)
ax = sns.distplot(mv, hist_kws={"histtype": "stepfilled"},kde=False, rug=True, label=mn, ax=ax, norm_hist=False)
sns.despine()
ax = plt.axes()
ax.set_xlabel('mean absolute relative difference')
ax.set_ylabel('frequency')
plt.legend()
plt.savefig('{}/{}.pdf'.format(outdir, plotname))
def makeTable(methodDict, outpath, outfile, measure, annotPath):
import pandas as pd
import seaborn as sns
import ParsingUtils
import AnalysisUtils
dframes = []
for k, v in methodDict.items():
if k.upper().startswith('SALMON'):
d = ParsingUtils.readSalmon(v, '_{}'.format(k))
elif k.upper().startswith('KALLISTO'):
d = ParsingUtils.readKallisto(v, '_{}'.format(k))
elif k.upper().startswith('EXPRESS'):
d = ParsingUtils.readExpress(v, '_{}'.format(k))
elif k.upper() == 'SAILFISH':
d = ParsingUtils.readSailfish(v, '_{}'.format(k))
elif k.upper() == 'SAILFISH (QUASI)':
d = ParsingUtils.readSalmon(v, '_{}'.format(k))
elif k.upper().startswith('TRUTH'):
suffix = '_{}'.format(k)
d = ParsingUtils.readProFile(v, suffix)
d["TPM{}".format(suffix)] = 1000000.0 * (d["ExpFrac{}".format(suffix)] / d["ExpFrac{}".format(suffix)].sum())
# Flux sim thinks paired-end = 2 reads . . . sinh
d["NumReads{}".format(suffix)] = d["SeqNum{}".format(suffix)] * 0.5
# Add this dataframe to the list
dframes.append(d)
M = dframes[0].join(dframes[1:])
# Filter eXpress results
minVal = np.inf
for mn in set(methodDict.keys()) - set(["Truth", "eXpress"]):
newMin = M.loc[M["{}_{}".format(measure, mn)]>0, "{}_{}".format(measure,mn)].min()
minVal = min(minVal, newMin)
print("filtering eXpress results < {} {}".format(minVal, measure))
AnalysisUtils.filterValues("{}_{}".format(measure, "eXpress"), M, minVal)
org = outfile.split('/')[-1].split('_')[0]
print("org = {}".format(org))
if org == 'human':
plotStratifiedDiffs(M, methodDict, annotPath, outpath, measure)
mrdName = 'abs. mean rel. diff.'
corrName = 'Spearman corr.'
propName = 'Proportionality corr.'
tpefName = 'TP error fraction'
tpMedErrorName = 'TP median per. error'
res = pd.DataFrame(data={ m : {tpMedErrorName : np.nan, tpefName : np.nan, mrdName : np.nan, corrName : np.nan, propName : np.nan} for m in (methodDict.keys() - set('Truth'))})
import scipy as sp
import scipy.stats
for k in methodDict:
if k.upper() != "TRUTH":
c = sp.stats.spearmanr(M["{}_Truth".format(measure)], M["{}_{}".format(measure, k)])[0]
res[k][corrName] = c
mrd, _ = AnalysisUtils.relDiff("{}_Truth".format(measure), "{}_{}".format(measure, k), M)
res[k][mrdName] = mrd["relDiff"].abs().mean()
pc = AnalysisUtils.proportionalityCorrelation("{}_Truth".format(measure), "{}_{}".format(measure, k), M)
res[k][propName] = pc
tpind = M[M["{}_Truth".format(measure)] >= 1]
y = tpind["{}_{}".format(measure, k)]
x = tpind["{}_Truth".format(measure)]
ef = 10.0
re = (y - x) / x
are = 100.0 * (y - x).abs() / x
tpef = len(are[are > ef]) / float(len(are))
res[k][tpefName] = tpef
res[k][tpMedErrorName] = re.median()
res.drop('Truth', axis=1, inplace=True)
print(res)
res.to_csv(outfile+".csv")
with open(outfile, 'w') as ofile:
ofile.write(res.to_latex(float_format=lambda x: "{0:.2f}".format(x)))
print("wrote {}".format(outpath))
def plotStratifiedDiffs(M, methodDict, annotPath, outpath, measure):
import pickle
import seaborn as sns
tgmap = {}
with open('{}/tgmap.txt'.format(annotPath),'r') as ifile:
for l in ifile:
toks = l.split()
tgmap[toks[0]] = toks[1]
genes = pd.DataFrame([(k,v) for k,v in tgmap.items()], columns=['Name', 'Gene'])
genes.set_index('Name', inplace=True)
M = M.join(genes)
relDiffDict = {}
for mn, mf in methodDict.items():
rdiffs, _ = AnalysisUtils.relDiff('{}_Truth'.format(measure), '{}_{}'.format(measure, mn), M)
M['RelDiff_{}'.format(mn)] = rdiffs
mByGenes = M.groupby('Gene')
groups = dict(list(mByGenes))
ntmap = {}
methods = ["Sailfish", "Salmon", "eXpress"]
def retainedMethod(mn):
return mn in methods
sns.set_palette(sns.color_palette([methodcolor[m] for m in methods]))
numExpressedGenes = 0
for i,(k,group) in enumerate(groups.items()):
if i % 10000 == 0:
print("processing group {} of {}".format(i, len(groups)))
## remove 0s?
if group['NumReads_Truth'].sum() == 0:
continue
numExpressedGenes += 1
numTran = len(group)
if numTran not in ntmap:
ntmap[numTran] = {}
for mn, mf in methodDict.items():
if retainedMethod(mn):
#if not mn.upper().endswith('TRUTH'):
ntmap[numTran][mn] = []
for mn, mf in methodDict.items():
if retainedMethod(mn):
#if not mn.upper().endswith('TRUTH'):
ntmap[numTran][mn].append(group['RelDiff_{}'.format(mn)].abs().mean())
print("There were {} expressed genes (genes with at least 1 expressed isoform)".format(numExpressedGenes))
nt = [set([])]
totInClass = 0
classSize = 1000
for k in sorted(ntmap.keys()):
if k > 1:
vd = ntmap[k]
#num = len(vd['Kallisto'])
num = len(vd['Sailfish'])
if len(nt[-1]) == 0 or totInClass <= classSize:
nt[-1].add(k)
totInClass += num
if totInClass > classSize:
totInClass = 0
nt.append(set([]))
print(nt)
ntClasses = {}
for c in nt:
# make a name for the class
minVal = min(list(c))
maxVal = max(list(c))
if minVal == maxVal:
ntClasses[minVal] = str(minVal)
else:
name = str(minVal)+"-"+str(maxVal)
for x in c:
ntClasses[x] = name
from pprint import pprint
pprint(ntClasses)
data = []
for k, v in ntmap.items():
if k > 1:
for mn, vals in v.items():
for x in vals:
cname = ntClasses[k]
data.append((cname, mn, x))
ntxp = '# txp. per gene'
D = pd.DataFrame(data, columns=[ntxp, 'Method', 'MARDs'])
sns.set_style('white')
#sns.set_palette('pastel', desat=0.7)
plt.clf()
plt.cla()
plt.figure(figsize=(12,8))
cnames = sorted(list(set([v for k,v in ntClasses.items()])))
print("Classes = ")
pprint(cnames)
x_order = sorted(cnames, key=lambda x : int(x.split('-')[0]))
import matplotlib
matplotlib.rc("lines", markersize=0, markeredgewidth=0)
g = sns.factorplot(ntxp, "MARDs", "Method", D, kind='box', markers='', x_order=x_order)
#g = sns.facetgrid(ntxp, "MARDs", "Method", D)
g.set_xticklabels(rotation=30)
#g.despine(offset=10, trim=True)
#plt.tight_layout()
plt.gcf().subplots_adjust(bottom=0.15)
plt.savefig('{}/GeneStratRD.pdf'.format(outpath))
