-
Notifications
You must be signed in to change notification settings - Fork 81
/
feature_training.R
218 lines (194 loc) · 8.04 KB
/
feature_training.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
###ctDNA_training uses multipe ML algorithms aimming to
###predict the esiitance of ctDNA for cancer early diagnosis
###and to selecte the most promissing features for prediction
library(randomForest)
library(MASS)
source('R/get_features.R')##this function also loads find_cutoff function
source('R/log_ratio_normal.R')
###Calculate features for patient data
###don't be confused by the varialbe name, this is also part of training data
# test_features=list()
# for (filename in list.files('test/', pattern = 'pileups.freq$', full.names = TRUE)){
# pileup=read.table(filename, header=TRUE, as.is=TRUE)
# if(mean(pileup$read.depth)>800)
# test_features[[filename]]=get_features(pileup)
# }
# saveRDS(test_features,'inputdata/test_feature_list.rds')
test_features=readRDS('inputdata/test_feature_list.rds')
# tr_features=list()
# for (filename in list.files('training/', pattern = 'pileup.txt$', full.names = TRUE)){
# pileup=read.table(filename, header=TRUE, as.is=TRUE)
# if(mean(pileup$read.depth)>800)
# tr_features[[filename]]=get_features(pileup)
# }
# saveRDS(tr_features,'inputdata/tr_feature_list.rds')
tr_features=readRDS('inputdata/tr_feature_list.rds')
##training random forest
tr_features=do.call(rbind, tr_features)
test_features=do.call(rbind, test_features)
all_features=rbind(tr_features, test_features)
all_features$y=c(rep('0', nrow(tr_features)), rep('1', nrow(test_features)))
all_features$y= as.factor(all_features$y)
# model1=randomForest(y ~ . , data =all_features, ntree =10000, mtry=2)
# importance(model1, type = 2, scale = TRUE)
##some visualizations
# for (i in colnames(all_features)){
# plot(all_features$y, all_features[,i], main=paste(i))
# }
###linear model forward selection
all_features$y=as.numeric(as.character(all_features$y))
pval=c()
for (i in 1:(length(colnames(all_features))-1)){
f <- as.formula(paste('y ~', colnames(all_features)[i]))
modelAllHexSubscales <- lm(f, all_features)
pvalue=summary(modelAllHexSubscales)$coefficients[2,4]
pval=c(pval, pvalue)
}
names(pval)=colnames(all_features)[-dim(all_features)[2]]
pval=sort(pval)
linear_features=names(pval)[1:15]
# ###pca (doesn't do the job at all)
# pr_test=t(t(all_features[,1:7])/colMeans(all_features[,1:7]))
# pr_test=data.frame(pr_test)
# pr = prcomp(pr_test)
# summary(pr,loadings=TRUE)
# pr$sdev
# features=data.frame(cbind(pr$x, y=all_features$y))
# features$y=factor(features$y)
#
# features=data.frame(cbind(pr_test, y=all_features$y))
# features$y=factor(features$y)
# model2=randomForest(y ~ ., data =features, ntree = 20000, mtry=4)
# model3=randomForest(y ~ ., data =all_features, ntree = 20000, mtry=4)
# ####logistic regression feature forward selection
# colnames(all_features)
#logistic regression
# logstic_pval=c()
# for (i in 1:(length(colnames(all_features))-1)){
# f <- as.formula(paste('y ~', colnames(all_features)[i]))
# modelAllHexSubscales <- glm(f, all_features, family = "binomial")
# pvalue=summary(modelAllHexSubscales)$coefficients[2,4]
# logstic_pval=c(logstic_pval, pvalue)
# }
# names(logstic_pval)=colnames(all_features)[-length(colnames(all_features))]
# sort(logstic_pval)
#
# plot(all_features[names(sort(logstic_pval))[1:15]])
#
###forward selection with glm
#summary(glm(y ~ mad_mean_depth+mut_count_2+panel_allbase_mr, data =all_features, family = "binomial"))
#####################
###Naive bayesian####
#####################
##fitting a Gaussian distribution for each feature in 2 conditions
### and save the parameters
###name feature_fit matrix for training and test group
feature_fit_tr=sapply(1:(ncol(all_features)-1), function(x) {
fitdistr(all_features[all_features$y==0,x],densfun = 'normal')$estimate
})
colnames(feature_fit_tr)=colnames(all_features)[1:ncol(feature_fit_tr)]
feature_fit_test=sapply(1:(ncol(all_features)-1), function(x) {
fitdistr(all_features[all_features$y==1,x],densfun = 'normal')$estimate
})
colnames(feature_fit_test)=colnames(all_features)[1:ncol(feature_fit_test)]
###defind clusters of features using hierarchical clustering
###for each feature get the log ratio
# performance=list()
# performance_feature=list()
# for (try in 1:10000){
# patient=c(); control=c();
# if (try %% 100 ==0) print(try)
# try_features=sample(1:48, 10)
# test_features_now=test_features[,try_features]
#
# ###for each person get features(for),
# ###for feature get the log ratio(sapply)
# for (i in 1:nrow(test_features_now)){
# dummy=test_features_now[i,]
# patient=c(patient, mean(sapply(1:length(dummy), function(x)
# log_ratio(dummy[x], feature_fit_test[,try_features][,x], feature_fit_tr[,try_features][,x]))))
# }
#
# ##the same try_features
# tr_features_now=tr_features[,try_features]
# for (i in 1:nrow(tr_features_now)){
# dummy=tr_features_now[i,]
# control=c(control, mean(sapply(1:length(dummy), function(x)
# log_ratio(dummy[x], feature_fit_test[,try_features][,x], feature_fit_tr[,try_features][,x]))))
# }
#
# performance[[try]]=c(patient, control)
# performance_feature[[try]]=try_features
# }
# saveRDS(performance,'inputdata/performance.rds')
# saveRDS(performance_feature,'inputdata/performance_feature.rds')
###performance contains 10000 simulations of random selection of 10 features
###and their estimated performance, performance_features contains the
###corresponding features randomly selected
performance=readRDS('inputdata/performance.rds')
performance_features=readRDS('inputdata/performance_feature.rds')
##the log values can be added up to select the best features
##the value is confined between -4 and 4 to avoid outliers
##for determin ctDNA abundance
aa=sapply(performance, function(x){
len1=nrow(test_features)
len2=nrow(tr_features)
sum(ifelse(x[1:len1]>4, 4, x[1:len1]))/len1-
sum(ifelse(x[(len1+1):(len1+len2)]<-4, -4, x[(len1+1):(len1+len2)]))/2
})
I=aa>3.6
best_f_index=performance_features[I]
bb=unlist(best_f_index)
best_f=as.numeric(names(sort(table(bb), decreasing = TRUE))[1:15])
best_features=colnames(all_features[,best_f])
#####test the best features
###remove correlated features in linear model
feature_matrix=all_features[,linear_features]
rm_1cor=function(feature_matrix){
cor_matrix=abs(cor(feature_matrix))
limit=rowSums(as.matrix(cor_matrix)>0.8)
if(all(limit<2)) return(feature_matrix)
del_feature=which(limit>1)[1]
feature_matrix=feature_matrix[,-del_feature]
rm_1cor(feature_matrix)
}
linear_features=colnames(rm_1cor(feature_matrix))
selected_features=unique(c(rev(linear_features[!linear_features %in% best_features]), rev(best_features)))
selected_features=colnames(rm_1cor(all_features[,selected_features]))
###for each person get features(for),
###for feature get the log ratio(sapply)
test_features_now=test_features[,selected_features]
patient=list()
for (i in 1:nrow(test_features_now)){
dummy=test_features_now[i,]
patient[[i]]=sapply(1:length(dummy), function(x)
log_ratio(dummy[x], feature_fit_test[,selected_features][,x], feature_fit_tr[,selected_features][,x]))
}
##the same try_features
tr_features_now=tr_features[,selected_features]
control=list()
for (i in 1:nrow(tr_features_now)){
dummy=tr_features_now[i,]
control[[i]]=sapply(1:length(dummy), function(x)
log_ratio(dummy[x], feature_fit_test[,selected_features][,x], feature_fit_tr[,selected_features][,x]))
}
##some visulalization
require(ggplot2)
aa=data.frame(value=c(sapply(patient,function(x) sum(x)),
sapply(control,function(x) sum(x))),
condition=c(rep('patient',length(patient)), rep('control', length(control))))
ggplot(aa, aes(x = condition, y = value))+geom_boxplot(aes(colour =condition))+
geom_jitter(aes(color=condition), position=position_jitterdodge(dodge.width=0.001))+
theme_bw() +
theme(axis.line = element_line(colour = "black"),
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
panel.border = element_blank(),
panel.background = element_blank())+
geom_hline(yintercept = 0)
##post training selection
Reduce('+', lapply(patient, function(x) x>0))
Reduce('+', lapply(control, function(x) x<0))
Reduce('+',patient)
Reduce('+',control)
sum(sapply(patient, function(x) sum(x)>1))