The binary response we will be considering is called “pathological complete response” (pCR). After a patient has undergone cancer treatment, a pathologist will examine the patient’s tissue for any remnant cancer cells. If the pathologist finds no evidence of cancer cells remaining, the patient will be given a label of pCR = 1. If cancer cells were found after treatment, the patient will be instead be given pCR = 0. It is of great interest to patients and clinicians to try to predict which pCR response a patient may end up having prior to receiving any cancer treatments.
For the binary response, we have 18 studies. Loading in the file pCR_binary.RData, we will find three list objects, pCR, XX_pCR, and pam50_pCR. pCR is a list of 18 binary response (pCR) vectors, so calling pCR[[1]] retrieves the pCR patient responses for the first study. XX_pCR is a list of 18 matrices, with each matrix containing the biomarker expressions for the patients of a study. Therefore, calling XX_pCR[[1]] yields a matrix of dimensions (number of biomarkers) by (number of patients for study 1). The row names of each matrix correspond to the biomarker names. We have been given 10,115 biomarkers for consideration (corresponding to the # of rows of each matrix), and these are the same ones across all 18 studies.
pam50_pCR is a list with the pam50 predictor vectors, coded as a categorical variable corresponding to five breast cancer subtypes. This predictor is commercialized by a company called Prosigna. Using a logistic regression with pam50_pCR as the predictor and pCR as the response, we will obtain an AUC that serves as a benchmark for classification performance. Note that some vectors of pam50_pCR have missing values NA so will only be able to evaluate our benchmark on a subset of all studies. The challenge given to us by Prof. Parmigiani is to train a classifier for pCR response using the matrices in XX_pCR with gene expressions for a total of 1,309 patients in 18 studies.
We will begin our data analysis by visualizing our gene expressions, response variables and addressing any problems that may arise.
First, let’s pool all of the data and do some basic EDA.
load("data/pCR_binary.RData")
# Initialize df to be filled using first study
XX_full = t(XX_pCR[[1]])
YY_full = pCR[[1]]
ZZ_full = rep(1, length(YY_full))
df_full = cbind(XX_full,YY_full,ZZ_full)
N_genes = dim(XX_full)[2]
# fill df
for (ii in 2:18) {
XX_full = t(XX_pCR[[ii]])
YY_full = pCR[[ii]]
ZZ_full = rep(ii, length(YY_full))
df = cbind(XX_full,YY_full,ZZ_full)
df_full = rbind(df_full, df)
}
XXX = df_full[,1:(ncol(df_full) - 2)]
YYY = df_full[,ncol(df_full) - 1]
ZZZ = df_full[,ncol(df_full)]
gene_names = colnames(XXX)Now that we have pooled expression levels, let’s visualize mean and standard deviation expression levels for our biomarkers.
avg_expr = apply(XXX, 2, mean)
sd_expr = apply(XXX, 2, sd)
hist(avg_expr, main = "Histogram of Mean Expression", col = "cornflowerblue")
hist(sd_expr, main = "Histogram of SD of Expression", col = "cornflowerblue")
It appears that the majority of low expression and low variance biomarkers have already been removed from our dataset. This allows us to proceed without throwing out any biomarkers due to low expression/variance.
By printing the response variable for each study, we notice a few troubling things right off the bat.
Ns = c()
for (ii in 1:18) {
print(summary(pCR[[ii]]))
Ns = c(Ns, length(pCR[[ii]]))
}## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1404 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 1.0000 0.6981 1.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.2107 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1786 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1613 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
## 0.0000 0.0000 0.0000 0.2602 1.0000 1.0000 5
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.00 0.00 0.00 0.15 0.00 1.00
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1639 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0 0 0 0 0 0
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1946 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0 0 0 0 0 0
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.2941 1.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.000 0.000 0.000 0.169 0.000 1.000
## Min. 1st Qu. Median Mean 3rd Qu. Max. NA's
## 0.0000 0.0000 0.0000 0.1667 0.0000 1.0000 1
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.000 0.000 0.000 0.125 0.000 1.000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.3519 1.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.2261 0.0000 1.0000
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.0000 0.0000 0.0000 0.1818 0.0000 1.0000
From our summary output, we can tell that we have missingness in the response for study 6 and study 14. Since the missingness occurs in only a few data points, I will simply omit these patients from consideration in my analysis.
my_pCR = pCR
my_XX_pCR = XX_pCR
my_pam50_pCR = pam50_pCR
# Replace without na in response
na.idx.6 = which(is.na(pCR[[6]]))
my_pCR[[6]] = pCR[[6]][-na.idx.6]
my_XX_pCR[[6]] = XX_pCR[[6]][,-na.idx.6]
my_pam50_pCR[[6]] = pam50_pCR[[6]][-na.idx.6]
na.idx.14 = which(is.na(pCR[[14]]))
my_pCR[[14]] = pCR[[14]][-na.idx.14]
my_XX_pCR[[14]] = XX_pCR[[14]][,-na.idx.14]
my_pam50_pCR[[14]] = pam50_pCR[[14]][-na.idx.14]Let’s do some exploration of the number of patients in each study.
table(YYY,ZZZ)## ZZZ
## YYY 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
## 0 98 16 206 23 26 91 17 102 16 178 6 12 59 20 14 35 89 9
## 1 16 37 55 5 5 32 3 20 0 43 0 5 12 4 2 19 26 2
which(Ns < 20)## [1] 9 11 12 15 18
Troublingly, studies 9 and 11 don’t have any patients with optimal response. Studies 9, 11, 12, 15, and 18 also have fewer than 20 participants in total, and there are fewer than 10 patients in the optimal response categories in studies 4, 5, 7, 9, 11, 12, 14, 15 and 18. These small sample sizes are concerning as they make stratified sampling based on studies and response classes for a train/test split infeasible.
As mentioned above, small sample sizes in studies make it very hard to stratify sampling based on studies and response classes in a train-test split. As a result, it would be advantageous for us to group certain studies together so that we can train-test split on larger groups of studies. This is justified if we can show that certain studies cluster in groups of similar characteristics when it comes to gene expression measurements. To see if such clusters exist, let’s reduce the dimensionality of our full dataset using a PCA decomposition and plotting each patient along the first two principal components.
# library(gmodels)
dim(XXX) # Takes a while to run
XXX.pca = fast.prcomp(XXX, center = TRUE,scale. = TRUE)
saveRDS(XXX.pca, file = "data/pca.rds")Plot patients:
For clarity, let’s also visualize the cluster centroids:
As you can see, three clusters appear in this two dimensional latent space: Studies 4, 5, 8, and 17; studies 6, 7, and 9; and studies 1, 2, 3, 10, 11, 12, 13, 14, 15, 16, and 18. These clusters could be driven by differences in technologies (microarrays, RT-PCR, etc.) or labs may result in different baseline levels of measurement error (bias/variance) when measuring mRNA levels. This may result in clusters of studies we observe. They could also be attributed to differences (age, gender, sick/healthy, genetic differences) in the samples of patients. We can leverage these clusters to perform a train-test split based our response classes and our three groups.[1]
Using the groups
group1 = c(5,17,8,4)
group2 = c(6,9,7)
group3 = c(1,2,3,10,11,12,13,14,15,16,18)
# Initialize df to be filled using first study
XX_full = pam50_full = YY_full = ZZ_full = Group = c()
df_new = matrix(ncol=N_genes + 3)
# fill df
for (ii in 1:18) {
XX_full = t(my_XX_pCR[[ii]])
pam50_row = my_pam50_pCR[[ii]]
if (ii %in% group1) {Groupnum=1} else if(ii %in% group2){Groupnum=2} else{Groupnum=3}
YY_full = my_pCR[[ii]]
Group = rep(Groupnum, length(YY_full))
ZZ_full = rep(ii, length(YY_full))
df = cbind(XX_full, Group, YY_full, ZZ_full)
df_new = rbind(df_new, df)
pam50_full = c(pam50_full, pam50_row)
}
df_new = df_new[-1,]
XXX = df_new[,1:(ncol(df_new) - 3)]
Groups = df_new[,(ncol(df_new) - 2)]
YYY = df_new[,ncol(df_new) - 1]
ZZZ = df_new[,ncol(df_new)]Check for good qualities:
table(YYY, Groups)## Groups
## YYY 1 2 3
## 0 240 124 653
## 1 56 35 195
dim(df_new)## [1] 1303 10118
Train-test split manually so that I can stratify my samples by study and outcome:
# Train - Val - Test split
TEST_TRAIN_ = 0.8
test_train_split = function(TEST_TRAIN = TEST_TRAIN_, seed = 117) {
df_train = c()
df_test = c()
pam50_train = c()
pam50_test = c()
set.seed(seed)
for (group in 1:3) {
for (Y in c(0,1)) {
# Subset dataframe
df_subset = df_new[intersect(which(Groups == group),which(YYY == Y)),]
pam50_subset = pam50_full[intersect(which(Groups == group),which(YYY == Y))]
# Add training rows to Training with assigned fold for 5-fold CV
rows_train = sample(1:nrow(df_subset), floor(TEST_TRAIN * nrow(df_subset)))
folds <- cut(seq(1,length(rows_train)),breaks=5,labels=FALSE)
df_train = rbind(df_train, cbind(df_subset[rows_train, ], folds))
pam50_train = c(pam50_train, pam50_subset[rows_train])
# Add test rows
df_test = rbind(df_test, df_subset[-rows_train,])
pam50_test = c(pam50_test, pam50_subset[-rows_train])
}
}
colnames(df_train) = c(colnames(df_new), "fold")
colnames(df_test) = colnames(df_new)
return(list(df_train, df_test, pam50_train, pam50_test))
}
test_train = test_train_split()
df_train = test_train[[1]]
df_test = test_train[[2]]
pam50_train = test_train[[3]]
pam50_test = test_train[[4]]Check dimensions of training and test data and cast as dataframes instead of matrices:
dim(df_train); length(pam50_train)## [1] 1041 10119
## [1] 1041
dim(df_test); length(pam50_test)## [1] 262 10118
## [1] 262
df_training = as.data.frame(df_train)
df_testing = as.data.frame(df_test)We have successfully split our dataset into test sets and training sets.
Let’s take a peak at pam50_pCR, so we have an idea of which studies Prosignia used to construct training and test data. It appears that studies 1-3 and 17-18 don’t have predictor values for the Prosignia studies for whatever reason.
for (ii in 1:18) {print(head(pam50_pCR[[ii]]))}## [1] NA NA NA NA NA NA
## [1] NA NA NA NA NA NA
## [1] NA NA NA NA NA NA
## [1] "Claudin" "Her2" "Basal" "Luminal A" "Her2" "Her2"
## [1] "Luminal B" "Basal" "Basal" "Luminal B" "Luminal B" "Her2"
## [1] "Basal" "Basal" "Basal" "Luminal B" "Luminal A" "Basal"
## [1] "Luminal A" "Luminal A" "Basal" "Luminal A" "Basal" "Her2"
## [1] "Luminal A" "Luminal B" "Luminal A" "Basal" "Luminal A" "Luminal A"
## [1] "Basal" "Basal" "Normal" "Claudin" "Basal" "Claudin"
## [1] "Luminal A" "Luminal B" "Basal" "Basal" "Luminal A" "Basal"
## [1] "Her2" "Luminal B" "Basal" "Luminal A" "Luminal B" "Luminal A"
## [1] "Her2" "Luminal A" "Basal" "Her2" "Her2" "Basal"
## [1] "Luminal B" "Luminal A" "Basal" "Luminal A" "Luminal A" "Luminal B"
## [1] "Luminal B" "Luminal B" "Luminal B" "Luminal A" "Luminal B" "Her2"
## [1] "Basal" "Luminal B" "Basal" "Luminal A" "Normal" "Luminal A"
## [1] "Luminal A" "Basal" "Luminal B" "Luminal B" "Luminal B" "Luminal B"
## [1] NA NA NA NA NA NA
## [1] NA NA NA NA NA NA
Therefore, we can only get baselines based on the other studies. Let’s remove bad studies, train a logistic regression using pam50 as suggested, and get a benchmark test AUC/accuracy. Let’s do this on one instantiation of the test set.
# library(cvAUC)
bad_studies = c(1,2,3,17,18)
idx.rm_train = which(df_training$ZZ_full %in% bad_studies)
train_baseline_Y = df_training[-idx.rm_train,c("YY_full","Group")]
train_pam = pam50_train[-idx.rm_train]
df_train_baseline = data.frame(Y = train_baseline_Y$YY_full, X = train_pam)
idx.rm_test = which(df_testing$ZZ_full %in% bad_studies)
test_baseline_Y = df_testing[-idx.rm_test,c("YY_full","Group")]
test_pam = pam50_test[-idx.rm_test]
df_test_baseline = data.frame(X = test_pam)
model_baseline = glm(Y~X, data=df_train_baseline,family="binomial")
preds = c()
AUCs = c()
accuracies = c()
classifications = c()
for (group in 1:3) {
idx_train = which(test_baseline_Y$Group == group)
idx_test = which(test_baseline_Y$Group == group)
# Vary predictions by group? Uncomment
# model_baseline = glm(Y~X, data=df_train_baseline[idx_train,],family="binomial")
pred = predict(model_baseline, newdata=data.frame(X = df_test_baseline$X[idx_test]))
preds = c(preds, pred)
auc_ROCR <- .5 + abs(as.numeric(AUC(pred, test_baseline_Y$YY_full[idx_test])) - 0.5)
AUCs <- c(AUCs, auc_ROCR)
accuracy = mean(round(predict(model_baseline,
data.frame(X = df_test_baseline$X[idx_test]),
type="response")) ==
test_baseline_Y$YY_full[idx_test])
accuracies = c(accuracies, accuracy)
}
AUC_full=AUC(preds, test_baseline_Y$YY_full)
accuracy_full=mean(round(predict(model_baseline,data.frame(X = df_test_baseline$X),
type="response"))==test_baseline_Y$YY_full)
# By study Test AUCs and Accuracies
rbind(AUCs, accuracies)## [,1] [,2] [,3]
## AUCs 0.7419355 0.6914286 0.7753134
## accuracies 0.8611111 0.7812500 0.7820513
rbind(AUC_full, accuracy_full)## [,1]
## AUC_full 0.7537577
## accuracy_full 0.8013699
Our benchmark for AUC on this instantiation of the test set is 0.754, and our benchmark classification accuracy using a probability cutoff of 0.5 on the test set is 0.801. We can also see that the metrics as they vary across our group labeling. The main metric that I will be focusing on when training my models will be AUC, but I will always include accuracy as a sanity check. I will also be using on this particular test-train split in my analysis as well, but I can use Monte Carlo cross validation to get a better estimate of the true test AUC (see below). This will allow me to get a measure of uncertainty on my estimates for these metrics.
AUC_mc = c()
Accuracy_mc = c()
for (ii in 1:20) {
test_train = test_train_split(seed = ii)
df_train = test_train[[1]]
df_test = test_train[[2]]
pam50_train = test_train[[3]]
pam50_test = test_train[[4]]
df_training = as.data.frame(df_train)
df_testing = as.data.frame(df_test)
idx.rm_train = which(df_training$ZZ_full %in% bad_studies)
train_baseline_Y = df_training[-idx.rm_train,c("YY_full","Group")]
train_pam = pam50_train[-idx.rm_train]
df_train_baseline = data.frame(Y = train_baseline_Y$YY_full, X = train_pam)
idx.rm_test = which(df_testing$ZZ_full %in% bad_studies)
test_baseline_Y = df_testing[-idx.rm_test,c("YY_full","Group")]
test_pam = pam50_test[-idx.rm_test]
df_test_baseline = data.frame(X = test_pam)
model_baseline = glm(Y~X, data=df_train_baseline,family="binomial")
preds = predict(model_baseline, newdata=data.frame(X = df_test_baseline$X))
AUC_full=AUC(preds, test_baseline_Y$YY_full)
accuracy_full=mean(round(predict(model_baseline,data.frame(X = df_test_baseline$X),
type="response"))==test_baseline_Y$YY_full)
AUC_mc = c(AUC_mc, AUC_full)
Accuracy_mc = c(Accuracy_mc, accuracy_full)
}Get means and standard deviations of our metrics across 20 train-test splits.
mean(AUC_mc); sd(AUC_mc)## [1] 0.7260011
## [1] 0.03368138
mean(Accuracy_mc)## [1] 0.795831
Our benchmark average AUC of 20 different test sets is 0.726 with a standard deviation of 0.0336, and our average benchmark classification accuracy is 0.796. It appears that our particular choice of the train-test split resulted in a larger AUC than average.
Due to computational constraints, we need to choose one instantiation of the train/test set for modeling. I am going to use a seed of 117. The chunk below loads in our instantiation of the train-test split with seed 117, the function’s default.
test_train = test_train_split()
df_train = test_train[[1]]
df_test = test_train[[2]]
pam50_train = test_train[[3]]
pam50_test = test_train[[4]]Before we model, we need to do some dimension reductionality and feature engineering. Since the number of predictors that we have is much larger than the number of patients that we have (1,309 patients, 10,117 genes), most of our traditional statistical methods like logistic regression would likely perform poorly. Though I could fit some sort of Artificial Neural Network (ANN) on the full data set, I doubt that would perform very well either due to the size of our dataset. Furthermore, it would be impossible to tell which biomarkers are more important than others in that kind of an ANN, making our model largely uninterpretable.
As a result, I have decided to search for individual genes and gene pairs that may be good predictors of our response before modeling. Once I have identified these good genes and gene pairs as good predictors, I will fit a variety of models (ANN, Random Forest, XGBoost) on smaller training and testing datasets including only our good genes and gene pairs.
Now that we know more about the studies in our dataset, let’s start fishing for useful individual biomarkers in our training set using the CompScores function. Let’s use study 1 as our primer for now, getting AUC, negative log p-values from t-tests, and fold changes as our metrics:
# library(ROCR)
CompScores = function(XX,YY){
NGenes = nrow(XX)
SSS = data.frame(matrix(NA,NGenes,4))
colnames(SSS) = c("AUC","nlpvalueT","FoldChange","Var")
for (gg in 1:NGenes){
# AUC
result_AUC <- try(ROCR::prediction(as.vector(XX[gg,]),YY))
if (class(result_AUC) != "try-error") {
OCpred = ROCR::prediction(as.vector(XX[gg,]),YY)
result_AUC2 <- try(performance(OCpred,measure="auc"))
if (class(result_AUC2) != "try-error") {
OCauc = performance(OCpred,measure="auc")
SSS[gg,"AUC"] = .5 + abs(as.numeric(OCauc@y.values)-.5)
}
}
# Try to run t-test
result_t <- try(t.test(as.vector(XX[gg,YY==1]),as.vector(XX[gg,YY==0])))
if (class(result_t) != "try-error") {
OCt = t.test(as.vector(XX[gg,YY==1]),as.vector(XX[gg,YY==0]))
# T-test negative log p-value
SSS[gg,"nlpvalueT"] = -log(OCt$p.value)
# Difference in means (aka fold change as data are on log scale)
SSS[gg,"FoldChange"] = OCt$estimate[1] - OCt$estimate[2]
SSS[gg,"Var"] = var(XX[gg,])
}
}
return(SSS)
}
set.seed(117)
Scores = CompScores(my_XX_pCR[[1]],my_pCR[[1]])I am going to set an AUC cutoff of 0.7, negative log pvalue cutoff of 6, and fold change cutoff of 0.5. Genes above all three of these cutoffs will be returned in a list of potentially useful genes. The list of potentially useful genes as determined by study 1 is printed in the chunk below.
pcut1 = 6
acut1 = 0.7
fcut1 = 0.5
table(abs(Scores[,"FoldChange"]) > fcut1,
Scores[,"nlpvalueT"] > pcut1,
Scores[,"AUC"] > acut1,
dnn=c("FoldChange", "nlpvalueT", "AUC"))
fcut_good = which(abs(Scores[,"FoldChange"]) > fcut1)
acut_good = which(Scores[,"AUC"] > acut1)
pcut_good = which(Scores[,"nlpvalueT"] > pcut1)
promising_genes1 = gene_names[intersect(intersect(fcut_good, acut_good), pcut_good)]
promising_genes1The chunk below functionalizes this process and applies to all studies, getting lists of potentially useful genes in each study:
get_promising_genes = function(study_idx,
pcut=pcut1,
acut=acut1,
fcut = fcut1,
seed = 117){
set.seed(seed)
Scores = CompScores(my_XX_pCR[[study_idx]],my_pCR[[study_idx]])
fcut_good = which(abs(Scores[,"FoldChange"]) > fcut)
acut_good = which(Scores[,"AUC"] > acut)
pcut_good = which(Scores[,"nlpvalueT"] > pcut)
promising_genes = gene_names[intersect(intersect(fcut_good, acut_good), pcut_good)]
return(promising_genes)
}
# I tried to run this in a loop and it took up too much RAM :(
promising_genes2 = get_promising_genes(2)
promising_genes3 = get_promising_genes(3)
promising_genes4 = get_promising_genes(4)
promising_genes5 = get_promising_genes(5)
promising_genes6 = get_promising_genes(6)
promising_genes7 = get_promising_genes(7)
promising_genes8 = get_promising_genes(8)
promising_genes10 = get_promising_genes(10)
promising_genes12 = get_promising_genes(12)
promising_genes13 = get_promising_genes(13)
promising_genes14 = get_promising_genes(14)
promising_genes15 = get_promising_genes(15)
promising_genes16 = get_promising_genes(16)
promising_genes17 = get_promising_genes(17)
promising_genes18 = get_promising_genes(18)
save.image(file = "data/genes.RData") # save to diskNow that we have genes for each study, let’s look for genes that pop up as significant in at least 3 studies.
load("data/genes.RData")
promising_genes = c(promising_genes1, promising_genes2,
promising_genes3, promising_genes4,
promising_genes5, promising_genes6,
promising_genes7, promising_genes8,
promising_genes10, promising_genes12,
promising_genes13, promising_genes14,
promising_genes15, promising_genes16,
promising_genes17, promising_genes18)
table(promising_genes)[table(promising_genes) > 3]## promising_genes
## ABAT ACADSB ANP32E ASPM CDC20 CDCA8 CENPA DNAJC12 DNALI1 ERBB4
## 4 5 4 4 6 4 4 4 4 5
## ESR1 EZH2 GATA3 GLI3 MAPT MCM5 MCM6 MELK PTK7 RAB17
## 4 4 4 4 4 4 4 4 4 5
## RFC4 SCUBE2 SLC22A5 SOX11 TCEAL1 TTK TXNDC5
## 4 4 6 4 4 5 4
genes_mult_studies = names(table(promising_genes)[table(promising_genes) > 3])We have now found 27 genes were deemed to be above all of our criterion in at least 3 of the 16 studies. This is indicative of genes with relatively strong and consistent signals across heterogenous studies. The best of these genes were CDC20 and SLC22A5 (significant in 6 studies). These genes may be good candidates for classifiers.
Now that we have found individual genes that may be good predictors for the response variable, let’s look for pairs of genes to use in our analysis as well. We can use the switchBox package to find the 10 best top scoring pairs in our training dataset. The choice of 10 was made based on which choice of N top scoring pairs yielded the highest AUC using the TSP algorithm. We identify these pairs as useful predictors for the response.
# library(switchBox)
N_pairs = 10
idx_val <- which(df_training$fold == 1)
valData <- df_training[idx_val, ]
trainData <- df_training[-idx_val, ]
classifier1=SWAP.KTSP.Train(t(trainData[,1:(ncol(df_new) - 2)]),
as.factor(trainData[,(ncol(df_new) - 1)]),
krange=c(N_pairs))
print(classifier1$TSPs)## [,1] [,2]
## [1,] "NUP153" "CCND1"
## [2,] "C11orf75" "SLC7A8"
## [3,] "PSME4" "DCTN1"
## [4,] "CSRP2" "FAH"
## [5,] "RABGAP1L" "CA12"
## [6,] "STMN1" "POLD4"
## [7,] "FANCG" "GAMT"
## [8,] "YEATS2" "MGRN1"
## [9,] "NUP160" "TSC2"
## [10,] "MCM7" "UGCG"
tsp_result=SWAP.GetKTSP.Result(classifier1, t(valData[,1:(ncol(df_new) - 2)]),
as.factor(valData[,(ncol(df_new) - 1)]))
tsp_AUC=tsp_result$stats[5]
tsp_Accuracy=tsp_result$stats[1]
tsp_AUC; tsp_Accuracy## auc
## 0.6939543
## accuracy
## 0.6682464
In our modeling train and test datasets, we want to retain important individual biomarkers as well as the difference of our identified gene pairs. I’ve written a function to do just that below. In addition, I also have decided to standardize my datasets based on the mean and standard deviation in the training data.
get_feat_engineered_df = function(df, classify, N_pairs_ = N_pairs) {
df_out = df[,genes_mult_studies]
N_fished_genes = length(genes_mult_studies)
TSP_names = c()
for (ii in 1:N_pairs_) {
gene1 = classify$TSPs[ii,][1]
gene2 = classify$TSPs[ii,][2]
TSP = df[,gene1] - df[,gene2]
df_out = cbind(df_out, TSP)
TSP_names = c(TSP_names, paste("TSP",ii, sep =""))
}
colnames(df_out)[1:N_fished_genes] <- genes_mult_studies
colnames(df_out)[(N_fished_genes + 1):(N_fished_genes + N_pairs_)] <- TSP_names
return(df_out)
}
train_scalar = function(df, df_train){
df_out = c()
for (name in colnames(df)) {
col_new = (df[,name] - mean(df_train[,name]))/sd(df_train[,name])
df_out = cbind(df_out, col_new)
}
colnames(df_out) = colnames(df)
return(df_out)
}These chunks of code save our standardized dataframes as csv files.
df_train_new = get_feat_engineered_df(df_training, classifier1)
df_train_new_scaled = train_scalar(df_train_new, df_train_new)
df_train_for_nn = data.frame(df_train_new_scaled,
YY=df_training$YY_full)
df_test_new = get_feat_engineered_df(df_testing, classifier1)
df_test_new_scaled = train_scalar(df_test_new, df_train_new)
df_test_for_nn = data.frame(df_test_new_scaled,
YY=df_testing$YY_full)
dim(df_train_for_nn)## [1] 1041 38
dim(df_test_for_nn)## [1] 262 38
write.csv(df_train_for_nn,"data_created/df_train_full.csv", row.names = FALSE)
write.csv(df_test_for_nn,"data_created/df_test.csv", row.names = FALSE)I performed my modeling in python. I have attached an appendix with a jupyter notebook containing executable models and hyperparameter training. I will describe my methods at a high level below. As I mentioned before, I will be training my models specifically to improve validation/test AUC, but I will include accuracy metrics as a sanity check.
First, I attempted to do this classification task using an artificial neural network. In general, multi-level feed-forward neural networks tend to perform really well on classification tests with large datasets. Though we don’t have a very large dataset, I was interested to see how a FFNN would perform on our classification task. Using 5-fold cross validation on the training set, I found a mean validation AUC of 0.781 and a mean validation accuracy of 0.801.
After retraining my ANN on the entire training set and evaluating its performance on the test set, I found a test AUC of 0.781 and test accuracy of 0.782.
Decision tree based models like Random Forest classifiers and Gradient Boosting tend to out-perform ANNs on classification tasks with small to medium sized datasets. I decided to fit a random forest classifier and XGBoost classifier to our task and evaluate their performance.
First, I needed tuned the hyperparameters for my random forest model using cross validation. For computational ease, I used 3-fold cross validation and a random search algorithm over potential permutations of a variety of values for the number of trees, max features per tree, max depth of trees, in samples per split, min samples per leaf, and whether or not to bootstrap our data.
After tuning these parameters, I fit a random forest model to the whole training dataset and found a test AUC of 0.774 and a test classification accuracy of 0.786.
After fitting this tuned random forest model, I plotted the variable importance measures based on mean decreases in impurity for each feature (shown below).
# All defaults
img2_path <- "images/rf_full_var_imp.png"
include_graphics(img2_path)
Given these relative variable importances, I thought it would be interesting to refit the random forest model using only the top 10 most important predictors. Using the same hyperparameters, I found a test AUC of 0.769 and test accuracy of 0.779, respectively.
I also replotted the relative variable importances (shown below).
# All defaults
img3_path <- "images/rf_pithy_var_imp.png"
include_graphics(img3_path)
I began by iteratively tuning my hyperparameters beginning with max depth and min child weight, then tuning subsample and colsample, and ending with the learning rate eta.
After tuning these parameters, I fit a random forest model to the whole training dataset and found a test AUC of 0.773 and a test classification accuracy of 0.786. In addition, XGBoost allows us to plot the relative gains of each feature to the model (shown below).
# All defaults
img4_path <- "images/gain_xgboost.png"
include_graphics(img4_path)
Below, I will plot all of my test metrics side-by-side for comparison. Note that the first benchmark bar is the average metrics found through Monte Carlo Cross validation while the second benchmark bar is the particular AUC/Accuracy in our train/test set.
# All defaults
img5_path <- "images/comparison.png"
include_graphics(img5_path)
From the plot, it appears that our best model in terms of AUC on the test set is the ANN using a multi-layer feed forward neural network. As a result, I chose this to be my final model. I will elaborate on this in the next section.
Interestingly, the random forest model with all of the features and the XGBoost model perform remarkably similarly. Though the random forest model with only 10 features performs worse of our created models, it is worth noting that I did not retune hyperparameters given new features.[2] I think given the fact that it uses nearly a fourth of the parameters that our other models use, this pithy model performs really well.
As previously mentioned, multi-level feed-forward neural networks are the tool of choice when it comes to classification tasks with large datasets. I will describe my implementation below:
Using the keras library in the tensorflow package in python, I trained neural network to perform our classification task. I spent a while exploring different architectures for this neural network, finally settlling on an architecture beginning with an input layer of our original 37 features. Then, the 37 features are densed into two hidden layers of 20 and 12 neurons accounting for 760 and 252 estimated weights, respectively. Both hidden layers include 20% dropout during training to prevent overfitting and use nonlinear ReLU activation. Finally, the twelve outputs from the final hidden layer are combined into an output layer with dimensionality 1 and sigmoid activation to ensure the output is between 0 and 1.
A diagram of the architecture is given below:
# All defaults
img1_path <- "images/ann_diagram.png"
include_graphics(img1_path)
The optimizer I chose for this ANN was the Adam algorithm to leverage adaptive learning rates after exploring SGD but finding worse results. The objective function was the natural choice of binary cross entropy as it is uniquely suited for classification tasks. To prevent overfitting, I also added early stopping to the model, monitoring validation AUC and restoring the weights from the best epoch. I also chose to use a batch size of 10 which seems reasonable given the size of our dataset.
To tune the architecure and the aforementioned hyperparameters, I used 5-fold cross validation on the training set to get estimates of the mean validation AUCs and accuracies. I found a mean validation AUC of 0.781 and a mean validation accuracy of 0.801.
Finally, I retraining my ANN on the entire training set, tracking training and validation AUCs. I plotted the training and test to check if there is evidence of overfitting. The black line represents the epoch at which the algorithm triggered early stopping.
# All defaults
img6_path <- "images/history_NN.png"
include_graphics(img6_path)
The training and testing AUC plots are encouraging as it does take a number of epochs to trigger early stopping. Also, early stopping clearly helps us prevent overfitting as we have chosen a point to stop changing the weights when the validation/test AUC is highest.
Upon evaluating its performance on the test set, I found that this model has a test AUC of 0.781 test accuracy of 0.782. This was an in increase in test AUC of 0.028 on our instantiation of the train-test set which is pretty good all things considered.
There are still a few questions I have that remain unanswered after completing this project. I leave them here as points of discussion.
In our random forest and XGBoost models, we were able to create feature importance plots based on things like mean decrease impurity in trees and relative gains. As a reminder, these plots are reproduced below:
# All defaults
img2_path <- "images/rf_full_var_imp.png"
include_graphics(img2_path)
# All defaults
img4_path <- "images/gain_xgboost.png"
include_graphics(img4_path)
Our best model in terms of AUC performance when predicting on out-of-sample data is the ANN Model. However, ANNs have their shortcomings on the inferential side of things.With 1,025 total parameters to make connections from our inputs to our outputs, it’s harder for us to make statements about which features in our set of 37 are more important than others as directly. As a result, I turned to permutation importance (implemented with the eli5 library) to attempt to get variable importances for our ANNs. Permutation importance functions by destroying the information in a given predictor and calculating how much the metric (AUC in this case) decreases. The decreases for each predictor are plotted below. Notably, the most important features in our ANN are TSP1, MAPT, TSP10, and ESR1.
# All defaults
img6_path <- "images/NN_importances.png"
include_graphics(img6_path)
First of all, it is important to note that these variable importance metrics are not typically robust to multicollinearity in predictors. For example, if TSP2 and TSP6 have similar importances, it is not clear if that is because they are both important on their own or if they have a similar effect together and are being ranked similarly. Secondly, it is important to interpret these metrics only within a model as relative importances.
With that being said, many of the predictors deemed important are not the same across models. For example, both of our tree-based methods deem TSP2 as a very important prdictor while our ANN does not. There is also considerable disagreement over TSP5.
Some of these discrepencies stem from the fact that permutation importance for an ANN is more variable to different instantiations of the ANN (training seeds) and permutations in each feature than feature importance in tree based methods due to the nature of its design. However, the discrepencies are large enough to warrant more of exploration in a future analysis.
I standardized my gene expression data for computational ease when training our neural network and to prevent vanishing or exploding gradients during training. Though our decision tree models would likely be more robust to unstandardized data due to their design, I still trained them on the standardized expressions. It would be an interesting extension to this project to train our decision tree models with unstandardized data or train our neural network with unstandardized data to check for robustness.
For further information on modeling and hyperparameter tuning, see the .ipynb at https://github.com/sethbilliau/pCR_classification.
[1] A potential drawback to this PCA approach is that our PCA components capture variation in the dataset at a very high level - the variation captured in the PCA components may not accurately represent the variation that drives differences in responses. For this reason, I have decided to leverage this group labeling only to perform a splitting of the data and not to do any modeling.
[2] Hyperparameter tuning took around 15 minutes using a GPU for the full model - I don’t have the bandwidth to do it at this time.