def main(
data_df,model_df,
start=0,end=None,err=False,coarse_graining_level=0):
dicttype, modeltype = qc.get_model_type(model_df)
seq_cols = qc.get_cols_from_df(data_df,'seqs')
if not len(seq_cols)==1:
raise SortSeqError('Dataframe has multiple seq cols: %s'%str(seq_cols))
seq_dict,inv_dict = utils.choose_dict(dicttype,modeltype=modeltype)
#set name of sequences column based on type of sequence
type_name_dict = {'dna':'seq','rna':'seq_rna','protein':'seq_pro'}
seq_col_name = type_name_dict[dicttype]
#Cut the sequences based on start and end, and then check if it makes sense
if (start != 0 or end):
data_df.loc[:,seq_col_name] = data_df.loc[:,seq_col_name].str.slice(start,end)
if modeltype=='MAT':
if len(data_df.loc[0,seq_col_name]) != len(model_df.loc[:,'pos']):
raise SortSeqError('model length does not match dataset length')
elif modeltype=='NBR':
if len(data_df.loc[0,seq_col_name]) != len(model_df.loc[:,'pos'])+1:
raise SortSeqError('model length does not match dataset length')
col_headers = utils.get_column_headers(data_df)
if 'ct' not in data_df.columns:
data_df['ct'] = data_df[col_headers].sum(axis=1)
data_df = data_df[data_df.ct != 0]
if not end:
seqL = len(data_df[seq_col_name][0]) - start
else:
seqL = end-start
data_df = data_df[data_df[seq_col_name].apply(len) == (seqL)]
#make a numpy array out of the model data frame
model_df_headers = ['val_' + str(inv_dict[i]) for i in range(len(seq_dict))]
value = np.transpose(np.array(model_df[model_df_headers]))
#now we evaluate the expression of each sequence according to the model.
seq_mat,wtrow = numerics.dataset2mutarray(data_df.copy(),modeltype)
temp_df = data_df.copy()
temp_df['val'] = numerics.eval_modelmatrix_on_mutarray(np.array(model_df[model_df_headers]),seq_mat,wtrow)
temp_sorted = temp_df.sort_values(by='val')
temp_sorted.reset_index(inplace=True,drop=True)
#we must divide by the total number of counts in each bin for the MI calculator
#temp_sorted[col_headers] = temp_sorted[col_headers].div(temp_sorted['ct'],axis=0)
MI = EstimateMutualInfoforMImax.alt4(temp_sorted,coarse_graining_level=coarse_graining_level)
if not err:
Std = np.NaN
else:
data_df_for_sub = data_df.copy()
sub_MI = np.zeros(15)
for i in range(15):
sub_df = data_df_for_sub.sample(int(len(data_df_for_sub.index)/2))
sub_df.reset_index(inplace=True,drop=True)
sub_MI[i],sub_std = main(
sub_df,model_df,err=False)
Std = np.std(sub_MI)/np.sqrt(2)
return MI,Std
def main(df,lm='IM',modeltype='MAT',LS_means_std=None,\
db=None,iteration=30000,burnin=1000,thin=10,\
runnum=0,initialize='LS',start=0,end=None,foreground=1,\
background=0,alpha=0,pseudocounts=1,test=False,drop_library=False,\
verbose=False):
# Determine dictionary
seq_cols = qc.get_cols_from_df(df,'seqs')
if not len(seq_cols)==1:
raise SortSeqError('Dataframe has multiple seq cols: %s'%str(seq_cols))
dicttype = qc.colname_to_seqtype_dict[seq_cols[0]]
seq_dict,inv_dict = utils.choose_dict(dicttype,modeltype=modeltype)
'''Check to make sure the chosen dictionary type correctly describes
the sequences. An issue with this test is that if you have DNA sequence
but choose a protein dictionary, you will still pass this test bc A,C,
G,T are also valid amino acids'''
#set name of sequences column based on type of sequence
type_name_dict = {'dna':'seq','rna':'seq_rna','protein':'seq_pro'}
seq_col_name = type_name_dict[dicttype]
lin_seq_dict,lin_inv_dict = utils.choose_dict(dicttype,modeltype='MAT')
#wtseq = utils.profile_counts(df.copy(),dicttype,return_wtseq=True,start=start,end=end)
#wt_seq_dict_list = [{inv_dict[np.mod(i+1+seq_dict[w],len(seq_dict))]:i for i in range(len(seq_dict)-1)} for w in wtseq]
par_seq_dict = {v:k for v,k in seq_dict.items() if k != (len(seq_dict)-1)}
#drop any rows with ct = 0
df = df[df.loc[:,'ct'] != 0]
df.reset_index(drop=True,inplace=True)
#If there are sequences of different lengths, then print error but continue
if len(set(df[seq_col_name].apply(len))) > 1:
sys.stderr.write('Lengths of all sequences are not the same!')
#select target sequence region
df.loc[:,seq_col_name] = df.loc[:,seq_col_name].str.slice(start,end)
df = utils.collapse_further(df)
col_headers = utils.get_column_headers(df)
#make sure all counts are ints
df[col_headers] = df[col_headers].astype(int)
#create vector of column names
val_cols = ['val_' + inv_dict[i] for i in range(len(seq_dict))]
df.reset_index(inplace=True,drop=True)
#Drop any sequences with incorrect length
if not end:
'''is no value for end of sequence was supplied, assume first seq is
correct length'''
seqL = len(df[seq_col_name][0]) - start
else:
seqL = end-start
df = df[df[seq_col_name].apply(len) == (seqL)]
df.reset_index(inplace=True,drop=True)
#Do something different for each type of learning method (lm)
if lm == 'ER':
emat = Berg_von_Hippel(
df,dicttype,foreground=foreground,background=background,
pseudocounts=pseudocounts)
if lm == 'LS':
'''First check that is we don't have a penalty for ridge regression,
that we at least have all possible base values so that the analysis
will not fail'''
if LS_means_std: #If user supplied preset means and std for each bin
means_std_df = io.load_meanstd(LS_means_std)
#change bin number to 'ct_number' and then use as index
labels = list(means_std_df['bin'].apply(add_label))
std = means_std_df['std']
std.index = labels
#Change Weighting of each sequence by dividing counts by bin std
df[labels] = df[labels].div(std)
means = means_std_df['mean']
means.index = labels
else:
means = None
#drop all rows without counts
df['ct'] = df[col_headers].sum(axis=1)
df = df[df.ct != 0]
df.reset_index(inplace=True,drop=True)
''' For sort-seq experiments, bin_0 is library only and isn't the lowest
expression even though it is will be calculated as such if we proceed.
Therefore is drop_library is passed, drop this column from analysis.'''
if drop_library:
try:
df.drop('ct_0',inplace=True)
col_headers = utils.get_column_headers(df)
if len(col_headers) < 2:
raise SortSeqError(
'''After dropping library there are no longer enough
columns to run the analysis''')
except:
raise SortSeqError('''drop_library option was passed, but no ct_0
column exists''')
#parameterize sequences into 3xL vectors
raveledmat,batch,sw = utils.genweightandmat(
df,par_seq_dict,dicttype,means=means,modeltype=modeltype)
#Use ridge regression to find matrix.
emat = Compute_Least_Squares(raveledmat,batch,sw,alpha=alpha)
if lm == 'IM':
seq_mat,wtrow = numerics.dataset2mutarray(df.copy(),modeltype)
#this is also an MCMC routine, do the same as above.
if initialize == 'rand':
if modeltype == 'MAT':
emat_0 = utils.RandEmat(len(df[seq_col_name][0]),len(seq_dict))
elif modeltype == 'NBR':
emat_0 = utils.RandEmat(len(df['seq'][0])-1,len(seq_dict))
elif initialize == 'LS':
emat_cols = ['val_' + inv_dict[i] for i in range(len(seq_dict))]
emat_0_df = main(df.copy(),lm='LS',modeltype=modeltype,alpha=alpha,start=0,end=None,verbose=verbose)
emat_0 = np.transpose(np.array(emat_0_df[emat_cols]))
#pymc doesn't take sparse mat
emat = MaximizeMI_memsaver(
seq_mat,df.copy(),emat_0,wtrow,db=db,iteration=iteration,burnin=burnin,
thin=thin,runnum=runnum,verbose=verbose)
#now format the energy matrices to get them ready to output
if (lm == 'IM' or lm == 'memsaver'):
if modeltype == 'NBR':
emat_typical = gauge.fix_neighbor(np.transpose(emat))
elif modeltype == 'MAT':
emat_typical = gauge.fix_matrix(np.transpose(emat))
elif lm == 'ER':
'''the emat for this format is currently transposed compared to other formats
it is also already a data frame with columns [pos,val_...]'''
emat_cols = ['val_' + inv_dict[i] for i in range(len(seq_dict))]
emat_typical = emat[emat_cols]
emat_typical = (gauge.fix_matrix((np.array(emat_typical))))
else: #must be Least squares
emat_typical = utils.emat_typical_parameterization(emat,len(seq_dict))
if modeltype == 'NBR':
emat_typical = gauge.fix_neighbor(np.transpose(emat_typical))
elif modeltype == 'MAT':
emat_typical = gauge.fix_matrix(np.transpose(emat_typical))
em = pd.DataFrame(emat_typical)
em.columns = val_cols
#add position column
if modeltype == 'NBR':
pos = pd.Series(range(start,start - 1 + len(df[seq_col_name][0])),name='pos')
else:
pos = pd.Series(range(start,start + len(df[seq_col_name][0])),name='pos')
output_df = pd.concat([pos,em],axis=1)
# Validate model and return
output_df = qc.validate_model(output_df,fix=True)
return output_df
M = pymc.MCMC([pymcdf,emat],db='sqlite',dbname=dbname)
else:
M = pymc.MCMC([pymcdf,emat])
M.use_step_method(GaugePreservingStepper,emat)
if not verbose:
M.sample = shutthefuckup(M.sample)
#M.sample(iteration,thin=thin,tune_interval=20000)
M.sample(iteration,thin=thin)
emat_mean = np.mean(M.trace('emat')[burnin:],axis=0)
return emat_mean
df = pd.io.parsers.read_csv(sys.argv[1],delim_whitespace=True)
temp_seq_mat,wtrow = numerics.dataset2mutarray(df.copy(),'MAT')
temp_seq_mat2 = temp_seq_mat.toarray()
#we need a parameter for the effect of mutation at each base pair.
#we remove 20 base pairs to remove the barcode sequence on the end of the sequence.
len_seq = len(df.loc[0,'seq'])
len_outputseq = len_seq - 20
len_barcode = 20
#We add 4 parameters for the barcode, because
total_params = len_outputseq + len_barcode*4
seq_mat = np.zeros((temp_seq_mat2.shape[0],total_params))
for i in range(len_outputseq):
seq_mat[:,i] = np.sum(temp_seq_mat2[:,i*4:(i*4+4)],axis=1)
seq_mat[:,len_outputseq:] = temp_seq_mat2[:,-len_barcode*4:]
seq_mat = scipy.sparse.csr_matrix(seq_mat)
emat_0 = np.zeros((4,len_seq))
def main(data_df,
model_df,
start=0,
end=None,
err=False,
coarse_graining_level=0,
rsquared=False,
return_freg=False):
#determine whether you are working with RNA, DNA, or protein.
#this also should determine modeltype (MAT, NBR, PAIR).
dicttype, modeltype = qc.get_model_type(model_df)
#get column header for the sequence column.
seq_cols = qc.get_cols_from_df(data_df, 'seqs')
if not len(seq_cols) == 1:
raise SortSeqError('Dataframe has multiple seq cols: %s' %
str(seq_cols))
#create dictionary that goes from, for example, nucleotide to number and
#visa versa.
seq_dict, inv_dict = utils.choose_dict(dicttype, modeltype=modeltype)
#set name of sequences column based on type of sequence
type_name_dict = {'dna': 'seq', 'rna': 'seq_rna', 'protein': 'seq_pro'}
seq_col_name = type_name_dict[dicttype]
if not end:
seqL = len(data_df[seq_col_name][0]) - start
else:
seqL = end - start
#throw out wrong length sequences.
#Cut the sequences based on start and end, and then check if it makes sense
if (start != 0 or end):
data_df.loc[:,seq_col_name] = \
data_df.loc[:,seq_col_name].str.slice(start,end)
right_length = data_df.loc[:, seq_col_name].apply(len) == (seqL)
if not right_length.all():
sys.stderr.write('''Not all sequences are the same length!
Throwing out incorrect sequences!''')
data_df = data_df.loc[right_length, :]
data_df = data_df.reset_index(drop=True)
if modeltype == 'MAT':
if seqL != len(model_df.loc[:, 'pos']):
raise SortSeqError(
'model length does not match dataset length')
elif modeltype == 'NBR':
if seqL != len(model_df.loc[:, 'pos']) + 1:
raise SortSeqError(
'model length does not match dataset length')
elif modeltype == 'PAIR':
if int(scipy.misc.comb(seqL, 2)) != len(model_df.loc[:, 'pos']):
raise SortSeqError(
'model length does not match dataset length')
#get column names of the counts columns (excluding total counts 'ct')
col_headers = utils.get_column_headers(data_df)
if 'ct' not in data_df.columns:
data_df['ct'] = data_df[col_headers].sum(axis=1)
#remove empty rows.
data_df = data_df[data_df.ct != 0]
#determine sequence length.
#make a numpy array out of the model data frame
model_df_headers = [
'val_' + str(inv_dict[i]) for i in range(len(seq_dict))
]
value = np.array(model_df[model_df_headers])
#now we evaluate the expression of each sequence according to the model.
#first convert to matrix representation of sequences
seq_mat, wtrow = numerics.dataset2mutarray(data_df.copy(), modeltype)
temp_df = data_df.copy()
#evaluate energy of each sequence
temp_df['val'] = numerics.eval_modelmatrix_on_mutarray(
value, seq_mat, wtrow)
#sort based on value
temp_sorted = temp_df.sort_values(by='val')
temp_sorted.reset_index(inplace=True, drop=True)
#freg is a regularized plot which show how sequences are distributed
#in energy space.
if return_freg:
fig, ax = plt.subplots()
MI, freg = EstimateMutualInfoforMImax.alt4(
temp_sorted,
coarse_graining_level=coarse_graining_level,
return_freg=return_freg)
plt.imshow(freg, interpolation='nearest', aspect='auto')
plt.savefig(return_freg)
else:
MI = EstimateMutualInfoforMImax.alt4(
temp_sorted,
coarse_graining_level=coarse_graining_level,
return_freg=return_freg)
#if we want to calculate error then use bootstrapping.
if not err:
Std = np.NaN
else:
data_df_for_sub = data_df.copy()
sub_MI = np.zeros(15)
for i in range(15):
sub_df = data_df_for_sub.sample(int(
len(data_df_for_sub.index) / 2))
sub_df.reset_index(inplace=True, drop=True)
sub_MI[i], sub_std = main(sub_df, model_df, err=False)
Std = np.std(sub_MI) / np.sqrt(2)
#we can return linfoot corrolation (rsquared) or return MI.
if rsquared:
return (1 - 2**(-2 * MI)), (1 - 2**(-2 * Std))
else:
return MI, Std
# Create sequences to test this on
wtseq = 'AAAAAAAGTGAGATGGCAATCTAATTCGGCACCCCAGGTTTTACACTTTATGCTTCCGGCTCGTATGTTGTGTGG'
L = len(wtseq)
modeltypes = ['MAT','NBR']
seqtypes = ['dna','protein']
numseqs_dict = {'dna':10000,'protein':1000}
for seqtype in seqtypes:
for modeltype in modeltypes:
for mutrate in [0.01,0.1,1]:
numseqs = numseqs_dict[seqtype]
dataset_df = simulate_library(wtseq,numseq=numseqs,mutrate=mutrate,tags=True,\
dicttype=seqtype)
seqarray = numerics.dataset2seqarray(dataset_df,\
modeltype=modeltype)
mutarray, wtrow = numerics.dataset2mutarray(dataset_df,\
modeltype=modeltype)
# Print compression results
seqarray_size = numerics.nbytes(seqarray)
mutarray_size = numerics.nbytes(mutarray)
# Create matrix for random model
alphabet = qc.seqtype_to_alphabet_dict[seqtype]
C = len(alphabet)
num_rows = (L-1) if modeltype=='NBR' else L
num_cols = C**2 if modeltype=='NBR' else C
modelmatrix = randn(num_rows,num_cols)
# Create model dataframe
val_cols = qc.model_parameters_dict[(modeltype,seqtype)]
model_df = pd.DataFrame(modelmatrix,columns=val_cols)
# Create sequences to test this on
wtseq = 'AAAAAAAGTGAGATGGCAATCTAATTCGGCACCCCAGGTTTTACACTTTATGCTTCCGGCTCGTATGTTGTGTGG'
L = len(wtseq)
modeltypes = ['MAT', 'NBR']
seqtypes = ['dna', 'protein']
numseqs_dict = {'dna': 10000, 'protein': 1000}
for seqtype in seqtypes:
for modeltype in modeltypes:
for mutrate in [0.01, 0.1, 1]:
numseqs = numseqs_dict[seqtype]
dataset_df = simulate_library(wtseq,numseq=numseqs,mutrate=mutrate,tags=True,\
dicttype=seqtype)
seqarray = numerics.dataset2seqarray(dataset_df,\
modeltype=modeltype)
mutarray, wtrow = numerics.dataset2mutarray(dataset_df,\
modeltype=modeltype)
# Print compression results
seqarray_size = numerics.nbytes(seqarray)
mutarray_size = numerics.nbytes(mutarray)
# Create matrix for random model
alphabet = qc.seqtype_to_alphabet_dict[seqtype]
C = len(alphabet)
num_rows = (L - 1) if modeltype == 'NBR' else L
num_cols = C**2 if modeltype == 'NBR' else C
modelmatrix = randn(num_rows, num_cols)
# Create model dataframe
val_cols = qc.model_parameters_dict[(modeltype, seqtype)]
model_df = pd.DataFrame(modelmatrix, columns=val_cols)