ML_AI_training/UQ_ML_data.py

537 lines
21 KiB
Python
Executable file

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Mar 6 13:41:54 2022
@author: tanu
"""
def setvars(gene,drug):
#https://stackoverflow.com/questions/51695322/compare-multiple-algorithms-with-sklearn-pipeline
import os, sys
import pandas as pd
import numpy as np
print(np.__version__)
print(pd.__version__)
import pprint as pp
from copy import deepcopy
from collections import Counter
from sklearn.impute import KNNImputer as KNN
from imblearn.over_sampling import RandomOverSampler
from imblearn.under_sampling import RandomUnderSampler
from imblearn.over_sampling import SMOTE
from sklearn.datasets import make_classification
from imblearn.combine import SMOTEENN
from imblearn.combine import SMOTETomek
from imblearn.over_sampling import SMOTENC
from imblearn.under_sampling import EditedNearestNeighbours
from imblearn.under_sampling import RepeatedEditedNearestNeighbours
from sklearn.metrics import make_scorer, confusion_matrix, accuracy_score, balanced_accuracy_score, precision_score, average_precision_score, recall_score
from sklearn.metrics import roc_auc_score, roc_curve, f1_score, matthews_corrcoef, jaccard_score, classification_report
from sklearn.model_selection import train_test_split, cross_validate, cross_val_score
from sklearn.model_selection import StratifiedKFold,RepeatedStratifiedKFold, RepeatedKFold
from sklearn.pipeline import Pipeline, make_pipeline
#%% GLOBALS
rs = {'random_state': 42}
njobs = {'n_jobs': 10}
scoring_fn = ({ 'mcc' : make_scorer(matthews_corrcoef)
, 'accuracy' : make_scorer(accuracy_score)
, 'fscore' : make_scorer(f1_score)
, 'precision' : make_scorer(precision_score)
, 'recall' : make_scorer(recall_score)
, 'roc_auc' : make_scorer(roc_auc_score)
, 'jcc' : make_scorer(jaccard_score)
})
skf_cv = StratifiedKFold(n_splits = 10
#, shuffle = False, random_state= None)
, shuffle = True,**rs)
rskf_cv = RepeatedStratifiedKFold(n_splits = 10
, n_repeats = 3
, **rs)
mcc_score_fn = {'mcc': make_scorer(matthews_corrcoef)}
jacc_score_fn = {'jcc': make_scorer(jaccard_score)}
#%% FOR LATER: Combine ED logo data
#%% FOR LATER: active aa site annotations **DONE on 15/05/2022 as part of generating merged_dfs
###########################################################################
rs = {'random_state': 42}
njobs = {'n_jobs': 10}
homedir = os.path.expanduser("~")
#==============
# directories
#==============
datadir = homedir + '/git/Data/'
indir = datadir + drug + '/input/'
outdir = datadir + drug + '/output/'
#=======
# input
#=======
infile_ml1 = outdir + gene.lower() + '_merged_df3.csv'
#infile_ml2 = outdir + gene.lower() + '_merged_df2.csv'
my_df = pd.read_csv(infile_ml1, index_col = 0)
my_df.dtypes
my_df_cols = my_df.columns
geneL_basic = ['pnca']
geneL_na = ['gid']
geneL_na_ppi2 = ['rpob']
geneL_ppi2 = ['alr', 'embb', 'katg']
#%% get cols
mycols = my_df.columns
# # change from numberic to
# num_type = ['int64', 'float64']
# cat_type = ['object', 'bool']
# if my_df['active_aa_pos'].dtype in num_type:
# my_df['active_aa_pos'] = my_df['active_aa_pos'].astype(object)
# my_df['active_aa_pos'].dtype
# FIXME: if this is not structural, remove from source..
# Drop NA where numerical cols have them
if gene.lower() in geneL_na_ppi2:
#D1148 get rid of
na_index = my_df['mutationinformation'].index[my_df['mcsm_na_affinity'].apply(np.isnan)]
my_df = my_df.drop(index=na_index)
# FIXME: either impute or remove!
# for embb (L114M, F115L, V123L, V125I, V131M) delete for now
if gene.lower() in ['embb']:
na_index = my_df['mutationinformation'].index[my_df['ligand_distance'].apply(np.isnan)]
my_df = my_df.drop(index=na_index)# RERUN embb with the 5 values now present
###########################################################################
#%% Add lineage calculation columns
#FIXME: Check if this can be imported from config?
total_mtblineage_uc = 8
lineage_colnames = ['lineage_list_all', 'lineage_count_all', 'lineage_count_unique', 'lineage_list_unique', 'lineage_multimode']
#bar = my_df[lineage_colnames]
my_df['lineage_proportion'] = my_df['lineage_count_unique']/my_df['lineage_count_all']
my_df['dist_lineage_proportion'] = my_df['lineage_count_unique']/total_mtblineage_uc
###########################################################################
#%% AA property change
#--------------------
# Water prop change
#--------------------
my_df['water_change'] = my_df['wt_prop_water'] + str('_to_') + my_df['mut_prop_water']
my_df['water_change'].value_counts()
water_prop_changeD = {
'hydrophobic_to_neutral' : 'change'
, 'hydrophobic_to_hydrophobic' : 'no_change'
, 'neutral_to_neutral' : 'no_change'
, 'neutral_to_hydrophobic' : 'change'
, 'hydrophobic_to_hydrophilic' : 'change'
, 'neutral_to_hydrophilic' : 'change'
, 'hydrophilic_to_neutral' : 'change'
, 'hydrophilic_to_hydrophobic' : 'change'
, 'hydrophilic_to_hydrophilic' : 'no_change'
}
my_df['water_change'] = my_df['water_change'].map(water_prop_changeD)
my_df['water_change'].value_counts()
#--------------------
# Polarity change
#--------------------
my_df['polarity_change'] = my_df['wt_prop_polarity'] + str('_to_') + my_df['mut_prop_polarity']
my_df['polarity_change'].value_counts()
polarity_prop_changeD = {
'non-polar_to_non-polar' : 'no_change'
, 'non-polar_to_neutral' : 'change'
, 'neutral_to_non-polar' : 'change'
, 'neutral_to_neutral' : 'no_change'
, 'non-polar_to_basic' : 'change'
, 'acidic_to_neutral' : 'change'
, 'basic_to_neutral' : 'change'
, 'non-polar_to_acidic' : 'change'
, 'neutral_to_basic' : 'change'
, 'acidic_to_non-polar' : 'change'
, 'basic_to_non-polar' : 'change'
, 'neutral_to_acidic' : 'change'
, 'acidic_to_acidic' : 'no_change'
, 'basic_to_acidic' : 'change'
, 'basic_to_basic' : 'no_change'
, 'acidic_to_basic' : 'change'}
my_df['polarity_change'] = my_df['polarity_change'].map(polarity_prop_changeD)
my_df['polarity_change'].value_counts()
#--------------------
# Electrostatics change
#--------------------
my_df['electrostatics_change'] = my_df['wt_calcprop'] + str('_to_') + my_df['mut_calcprop']
my_df['electrostatics_change'].value_counts()
calc_prop_changeD = {
'non-polar_to_non-polar' : 'no_change'
, 'non-polar_to_polar' : 'change'
, 'polar_to_non-polar' : 'change'
, 'non-polar_to_pos' : 'change'
, 'neg_to_non-polar' : 'change'
, 'non-polar_to_neg' : 'change'
, 'pos_to_polar' : 'change'
, 'pos_to_non-polar' : 'change'
, 'polar_to_polar' : 'no_change'
, 'neg_to_neg' : 'no_change'
, 'polar_to_neg' : 'change'
, 'pos_to_neg' : 'change'
, 'pos_to_pos' : 'no_change'
, 'polar_to_pos' : 'change'
, 'neg_to_polar' : 'change'
, 'neg_to_pos' : 'change'
}
my_df['electrostatics_change'] = my_df['electrostatics_change'].map(calc_prop_changeD)
my_df['electrostatics_change'].value_counts()
#--------------------
# Summary change: Create a combined column summarising these three cols
#--------------------
detect_change = 'change'
check_prop_cols = ['water_change', 'polarity_change', 'electrostatics_change']
#my_df['aa_prop_change'] = (my_df.values == detect_change).any(1).astype(int)
my_df['aa_prop_change'] = (my_df[check_prop_cols].values == detect_change).any(1).astype(int)
my_df['aa_prop_change'].value_counts()
my_df['aa_prop_change'].dtype
my_df['aa_prop_change'] = my_df['aa_prop_change'].map({1:'change'
, 0: 'no_change'})
my_df['aa_prop_change'].value_counts()
my_df['aa_prop_change'].dtype
#%% IMPUTE values for OR [check script for exploration: UQ_or_imputer]
#or_cols = ['or_mychisq', 'log10_or_mychisq', 'or_fisher']
sel_cols = ['mutationinformation', 'or_mychisq', 'log10_or_mychisq']
or_cols = ['or_mychisq', 'log10_or_mychisq']
print("count of NULL values before imputation\n")
print(my_df[or_cols].isnull().sum())
my_dfI = pd.DataFrame(index = my_df['mutationinformation'] )
my_dfI = pd.DataFrame(KNN(n_neighbors= 5, weights="uniform").fit_transform(my_df[or_cols])
, index = my_df['mutationinformation']
, columns = or_cols )
my_dfI.columns = ['or_rawI', 'logorI']
my_dfI.columns
my_dfI = my_dfI.reset_index(drop = False) # prevents old index from being added as a column
my_dfI.head()
print("count of NULL values AFTER imputation\n")
print(my_dfI.isnull().sum())
#-------------------------------------------
# OR df Merge: with original based on index
#-------------------------------------------
my_df['index_bm'] = my_df.index
mydf_imputed = pd.merge(my_df
, my_dfI
, on = 'mutationinformation')
mydf_imputed = mydf_imputed.set_index(['index_bm'])
my_df['log10_or_mychisq'].isna().sum()
mydf_imputed['log10_or_mychisq'].isna().sum()
mydf_imputed['logorI'].isna().sum()
len(my_df.columns)
len(mydf_imputed.columns)
#-----------------------------------------
# REASSIGN my_df after imputing OR values
#-----------------------------------------
my_df = mydf_imputed.copy()
#%%########################################################################
#==========================
# Data for ML
#==========================
my_df_ml = my_df.copy()
#==========================
# BLIND test set
#==========================
# Separate blind test set
my_df_ml[drug].isna().sum()
blind_test_df = my_df_ml[my_df_ml[drug].isna()]
blind_test_df.shape
training_df = my_df_ml[my_df_ml[drug].notna()]
training_df.shape
# Target1: dst
training_df[drug].value_counts()
training_df['dst_mode'].value_counts()
#%% Build X: input for ML
common_cols_stabiltyN = ['ligand_distance'
, 'ligand_affinity_change'
, 'duet_stability_change'
, 'ddg_foldx'
, 'deepddg'
, 'ddg_dynamut2'
, 'mmcsm_lig'
, 'contacts']
# Build stability columns ~ gene
if gene.lower() in geneL_basic:
X_stabilityN = common_cols_stabiltyN
cols_to_mask = ['ligand_affinity_change']
if gene.lower() in geneL_ppi2:
# X_stabilityN = common_cols_stabiltyN + ['mcsm_ppi2_affinity' , 'interface_dist']
geneL_ppi2_st_cols = ['mcsm_ppi2_affinity', 'interface_dist']
X_stabilityN = common_cols_stabiltyN + geneL_ppi2_st_cols
cols_to_mask = ['ligand_affinity_change', 'mcsm_ppi2_affinity']
if gene.lower() in geneL_na:
# X_stabilityN = common_cols_stabiltyN + ['mcsm_na_affinity']
geneL_na_st_cols = ['mcsm_na_affinity']
X_stabilityN = common_cols_stabiltyN + geneL_na_st_cols
cols_to_mask = ['ligand_affinity_change', 'mcsm_na_affinity']
if gene.lower() in geneL_na_ppi2:
# X_stabilityN = common_cols_stabiltyN + ['mcsm_na_affinity'] + ['mcsm_ppi2_affinity', 'interface_dist']
geneL_na_ppi2_st_cols = ['mcsm_na_affinity'] + ['mcsm_ppi2_affinity', 'interface_dist']
X_stabilityN = common_cols_stabiltyN + geneL_na_ppi2_st_cols
cols_to_mask = ['ligand_affinity_change', 'mcsm_na_affinity', 'mcsm_ppi2_affinity']
X_foldX_cols = [ 'electro_rr', 'electro_mm', 'electro_sm', 'electro_ss'
, 'disulfide_rr', 'disulfide_mm', 'disulfide_sm', 'disulfide_ss'
, 'hbonds_rr', 'hbonds_mm', 'hbonds_sm', 'hbonds_ss'
, 'partcov_rr', 'partcov_mm', 'partcov_sm', 'partcov_ss'
, 'vdwclashes_rr', 'vdwclashes_mm', 'vdwclashes_sm', 'vdwclashes_ss'
, 'volumetric_rr', 'volumetric_mm', 'volumetric_ss'
]
X_str = ['rsa'
#, 'asa'
, 'kd_values'
, 'rd_values']
X_ssFN = X_stabilityN + X_str + X_foldX_cols
X_evolFN = ['consurf_score'
, 'snap2_score'
, 'provean_score']
X_genomic_mafor = ['maf'
, 'logorI'
# , 'or_rawI'
# , 'or_mychisq'
# , 'or_logistic'
# , 'or_fisher'
# , 'pval_fisher'
]
X_genomic_linegae = ['lineage_proportion'
, 'dist_lineage_proportion'
#, 'lineage' # could be included as a category but it has L2;L4 formatting
, 'lineage_count_all'
, 'lineage_count_unique'
]
X_genomicFN = X_genomic_mafor + X_genomic_linegae
#%% Construct numerical and categorical column names
# numerical feature names
# numerical_FN = common_cols_stabiltyN + foldX_cols + X_strFN + X_evolFN + X_genomicFN
numerical_FN = X_ssFN + X_evolFN + X_genomicFN
#categorical feature names
categorical_FN = ['ss_class'
# , 'wt_prop_water'
# , 'mut_prop_water'
# , 'wt_prop_polarity'
# , 'mut_prop_polarity'
# , 'wt_calcprop'
# , 'mut_calcprop'
, 'aa_prop_change'
, 'electrostatics_change'
, 'polarity_change'
, 'water_change'
, 'drtype_mode_labels' # beware then you can use it to predict
#, 'active_aa_pos' # TODO?
]
###########################################################################
#=======================
# Masking columns:
# (mCSM-lig, mCSM-NA, mCSM-ppi2) values for lig_dist >10
#=======================
#%% Masking columns
# my_df_ml['mutationinformation'][my_df['ligand_distance']>10].value_counts()
# my_df_ml.groupby('mutationinformation')['ligand_distance'].apply(lambda x: (x>10)).value_counts()
# my_df_ml.loc[(my_df_ml['ligand_distance'] > 10), 'ligand_affinity_change'] = 0
# (my_df_ml['ligand_affinity_change'] == 0).sum()
my_df_ml['mutationinformation'][my_df_ml['ligand_distance']>10].value_counts()
my_df_ml.groupby('mutationinformation')['ligand_distance'].apply(lambda x: (x>10)).value_counts()
my_df_ml.loc[(my_df_ml['ligand_distance'] > 10), cols_to_mask].value_counts()
# mask the column ligand distance > 10
my_df_ml.loc[(my_df_ml['ligand_distance'] > 10), cols_to_mask] = 0
(my_df_ml['ligand_affinity_change'] == 0).sum()
mask_check = my_df_ml[['mutationinformation', 'ligand_distance'] + cols_to_mask]
# write file for check
mask_check.sort_values(by = ['ligand_distance'], ascending = True, inplace = True)
mask_check.to_csv(outdir + 'ml/' + gene.lower() + '_mask_check.csv')
#%% extracting dfs based on numerical, categorical column names
#----------------------------------
# WITHOUT the target var included
#----------------------------------
num_df = training_df[numerical_FN]
num_df.shape
cat_df = training_df[categorical_FN]
cat_df.shape
all_df = training_df[numerical_FN + categorical_FN]
all_df.shape
#------------------------------
# WITH the target var included:
#'wtgt': with target
#------------------------------
# drug and dst_mode should be the same thing
num_df_wtgt = training_df[numerical_FN + ['dst_mode']]
num_df_wtgt.shape
cat_df_wtgt = training_df[categorical_FN + ['dst_mode']]
cat_df_wtgt.shape
all_df_wtgt = training_df[numerical_FN + categorical_FN + ['dst_mode']]
all_df_wtgt.shape
#%%########################################################################
#============
# ML data
#============
#------
# X: Training and Blind test (BTS)
#------
X = all_df_wtgt[numerical_FN + categorical_FN] # training data ALL
X_bts = blind_test_df[numerical_FN + categorical_FN] # blind test data ALL
#X = all_df_wtgt[numerical_FN] # training numerical only
#X_bts = blind_test_df[numerical_FN] # blind test data numerical
#------
# y
#------
y = all_df_wtgt['dst_mode'] # training data y
y_bts = blind_test_df['dst_mode'] # blind data test y
#X_bts_wt = blind_test_df[numerical_FN + ['dst_mode']]
# Quick check
#(X['ligand_affinity_change']==0).sum() == (X['ligand_distance']>10).sum()
for i in range(len(cols_to_mask)):
ind = i+1
print('\nindex:', i, '\nind:', ind)
print('\nMask count check:'
, (my_df_ml[cols_to_mask[i]]==0).sum() == (my_df_ml['ligand_distance']>10).sum()
)
print('Original Data\n', Counter(y)
, 'Data dim:', X.shape)
###############################################################################
#%%
############################################################################
# RESAMPLING
###############################################################################
#------------------------------
# Simple Random oversampling
# [Numerical + catgeorical]
#------------------------------
oversample = RandomOverSampler(sampling_strategy='minority')
X_ros, y_ros = oversample.fit_resample(X, y)
print('Simple Random OverSampling\n', Counter(y_ros))
print(X_ros.shape)
#------------------------------
# Simple Random Undersampling
# [Numerical + catgeorical]
#------------------------------
undersample = RandomUnderSampler(sampling_strategy='majority')
X_rus, y_rus = undersample.fit_resample(X, y)
print('Simple Random UnderSampling\n', Counter(y_rus))
print(X_rus.shape)
#------------------------------
# Simple combine ROS and RUS
# [Numerical + catgeorical]
#------------------------------
oversample = RandomOverSampler(sampling_strategy='minority')
X_ros, y_ros = oversample.fit_resample(X, y)
undersample = RandomUnderSampler(sampling_strategy='majority')
X_rouC, y_rouC = undersample.fit_resample(X_ros, y_ros)
print('Simple Combined Over and UnderSampling\n', Counter(y_rouC))
print(X_rouC.shape)
#------------------------------
# SMOTE_NC: oversampling
# [numerical + categorical]
#https://stackoverflow.com/questions/47655813/oversampling-smote-for-binary-and-categorical-data-in-python
#------------------------------
# Determine categorical and numerical features
numerical_ix = X.select_dtypes(include=['int64', 'float64']).columns
numerical_ix
num_featuresL = list(numerical_ix)
numerical_colind = X.columns.get_indexer(list(numerical_ix) )
numerical_colind
categorical_ix = X.select_dtypes(include=['object', 'bool']).columns
categorical_ix
categorical_colind = X.columns.get_indexer(list(categorical_ix))
categorical_colind
k_sm = 5 # 5 is deafult
sm_nc = SMOTENC(categorical_features=categorical_colind, k_neighbors = k_sm, **rs, **njobs)
X_smnc, y_smnc = sm_nc.fit_resample(X, y)
print('SMOTE_NC OverSampling\n', Counter(y_smnc))
print(X_smnc.shape)
globals().update(locals()) # TROLOLOLOLOLOLS
#print("i did a horrible hack :-)")
###############################################################################
#%% SMOTE RESAMPLING for NUMERICAL ONLY*
# #------------------------------
# # SMOTE: Oversampling
# # [Numerical ONLY]
# #------------------------------
# k_sm = 1
# sm = SMOTE(sampling_strategy = 'auto', k_neighbors = k_sm, **rs)
# X_sm, y_sm = sm.fit_resample(X, y)
# print(X_sm.shape)
# print('SMOTE OverSampling\n', Counter(y_sm))
# y_sm_df = y_sm.to_frame()
# y_sm_df.value_counts().plot(kind = 'bar')
# #------------------------------
# # SMOTE: Over + Undersampling COMBINED
# # [Numerical ONLY]
# #-----------------------------
# sm_enn = SMOTEENN(enn=EditedNearestNeighbours(sampling_strategy='all', **rs, **njobs ))
# X_enn, y_enn = sm_enn.fit_resample(X, y)
# print(X_enn.shape)
# print('SMOTE Over+Under Sampling combined\n', Counter(y_enn))
###############################################################################
# TODO: Find over and undersampling JUST for categorical data