def run_edgeR(gene_expression, bio_assignment, gene_names, batch_info=None, batch=True):
if batch_info is None:
batch = False
r_counts = conversion_pydataframe(gene_expression)
r_bio_group = conversion_pydataframe(bio_assignment)
r_dge = r.DGEList(counts=r.t(r_counts), genes=gene_names)
r.assign("dge", r_dge)
r.assign("bio_group", r.factor(r_bio_group))
r("dge$samples$bio_group <- bio_group")
if batch:
r_batch_group = conversion_pydataframe(batch_info)
r.assign("batch_group", r.factor(r_batch_group))
r("dge$samples$batch_group <- batch_group")
r("""dge <- suppressWarnings(edgeR::calcNormFactors(dge))""")
if not batch:
r("""design <- model.matrix(~bio_group, data = dge$samples)""")
r("""colnames(design) <- c("Intercept", "bio")""")
if batch:
r("""design <- model.matrix(~bio_group+batch_group, data = dge$samples)""")
r("""colnames(design) <- c("Intercept", "bio", "batch")""")
r("""dge <- estimateDisp(dge, design)""")
r("""fit <- glmFit(dge, design)""")
if not batch:
r("""lrt <- glmLRT(fit)""")
if batch:
r("""lrt <- glmLRT(fit, coef="bio")""")
return r("lrt$table$PValue")
def ancova(lm1, lm2, names=('lm1', 'lm2')):
"""
Compares the slopes and intercepts of two linear models. Currently this is
quite limited in that it only compares single-variable linear models that
have `x` and `y` attributes.
Returns (pval of slope difference, pval of intercept difference).
Recall that if the slope is significant, you can't really say anything
about the intercept.
"""
# R code, from the extremely useful blog:
# http://r-eco-evo.blogspot.com/2011/08/
# comparing-two-regression-slopes-by.html
#
# model1 = aov(y~x*factor, data=df)
# (interaction term on summary(model1)'s 3rd table line)
#
# model2 = aov(y~x+factor, data=df)
# (2nd table line for "factor" in summary(model2) is the sig of intercept
# diff)
#
# anova(model1, model2)
# does removing the interaction term affect the model fit?
# Construct variables suitable for ANOVA/ANCOVA
label1 = [names[0] for i in lm1.x]
label2 = [names[1] for i in lm2.x]
labels = r.factor(np.array(label1 + label2))
xi = np.concatenate((lm1.x, lm2.x))
yi = np.concatenate((lm1.y, lm2.y))
# The workflow is to populate the formula as a separate environment.
# This first formula includes the interaction term
fmla1 = robjects.Formula('yi~xi*labels')
fmla1.environment['xi'] = xi
fmla1.environment['yi'] = yi
fmla1.environment['labels'] = labels
result1 = r('aov(%s)' % fmla1.r_repr())
interaction_pval = r.summary(result1)[0].rx2('Pr(>F)')[2]
# No interaction term
fmla2 = robjects.Formula('yi~xi+labels')
fmla2.environment['xi'] = xi
fmla2.environment['yi'] = yi
fmla2.environment['labels'] = labels
result2 = r('aov(%s)' % fmla2.r_repr())
intercept_pval = r.summary(result2)[0].rx2('Pr(>F)')[1]
# TODO: anova(result1, result2)?
return interaction_pval, intercept_pval
rownames=['true'],
colnames=['predicted'])
# <headingcell level=4>
# Using non-base packages in Rpy2
# <codecell>
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
r = robjects.r
e1071 = importr('e1071')
Yr = np2r(iris['Type'])
Yr = r.factor(Yr)
svm = e1071.svm(Xr, Yr)
yhat = r.predict(svm, Xr)
print r.table(yhat, Yr)
# <headingcell level=4>
# ggplot2 in python with Rpy2
# <markdowncell>
# Thanks to [Fei Yu](http://www.thefeiyu.com/) for this vignette.
# <codecell>
import rpy2.robjects as robjects
print pd.crosstab(iris['Type'], yhat_hclust, rownames=['true'], colnames=['predicted'])
# <headingcell level=4>
# Using non-base packages in Rpy2
# <codecell>
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
r = robjects.r
e1071 = importr('e1071')
Yr = np2r(iris['Type'])
Yr = r.factor(Yr)
svm = e1071.svm(Xr, Yr)
yhat = r.predict(svm, Xr)
print r.table(yhat, Yr)
# <headingcell level=4>
# ggplot2 in python with Rpy2
# <markdowncell>
# Thanks to [Fei Yu](http://www.thefeiyu.com/) for this vignette.
# <codecell>
import rpy2.robjects as robjects
