Final_Project.Rmd

```{r}
library(tidyverse)
library(randomForest)
library(data.table)
library(gbm)
library(knitr)
library(keras)
```

```{r}
profile <- read_csv("mutation_profiles.csv")
data <- read_csv("data/full_data.csv")
```

```{r}
mat_dat <- data %>% select(case, gene) %>% unique() 
case_vec <- mat_dat %>% select(case) %>% unique()
gene_vec <- mat_dat %>% select(gene) %>% count(gene) %>% filter(n > 7) %>% select(gene)

for (i in 1:14573) {
  profile[,i] <- lapply(profile[,i], factor)
}
```

```{r}
case_vec <- case_vec %>% mutate(row = row_number())

case_vec <- case_vec %>% mutate(cancer_type = ifelse(row < 101, "lung",
                                              ifelse(row >= 101 & row < 201, "brain",
                                              ifelse(row >= 201 & row < 301, "kidney",
                                              ifelse(row >= 301 & row < 401, "leukemia",
                                              ifelse(row >= 401 & row < 501, "lung",
                                              ifelse(row >= 501 & row < 601, "brain",
                                              ifelse(row >= 601 & row < 701, "kidney",
                                              ifelse(row >= 701 & row < 743, "leukemia",
                                              ifelse(row >= 743 & row < 843, "brain",
                                              ifelse(row >= 843 & row < 960, "kidney",
                                              ifelse(row >= 960, "lung", "error"))))))))))))


profile <- profile %>% mutate(cancer_type = factor(case_vec$cancer_type))

write_csv(profile, "data/mutation_profiles.csv")
```

## Take Performance Data Out    ----!! Only Once !!----

```{r}
# Only excecute this once!

indices <- sample(1:nrow(profile), 211)
performance <- profile[indices,]
train_data <- profile[-indices,]

write_csv(performance, "data/performance_data.csv")
write_csv(train_data, "data/train_data.csv")
```

## Read in data

```{r}
performance <- read_csv("data/performance_data.csv")
for (i in 1:14574) {
  performance[,i] <- lapply(performance[,i], factor)
}

train_data <- read_csv("data/train_data.csv")
for (i in 1:14574) {
  train_data[,i] <- lapply(train_data[,i], factor)
}

reference <- tibble(gene = colnames(train_data), name = newnames)
```

## Random Foresting

```{r}
indices <- sample(1:nrow(train_data), 200)
train <- train_data[-indices,]
train_val <- train_data[indices,]
train_val_targets <- train_val$V_14574

train_matrix <- model.matrix(~ ., data = train)

train_matrix <- train_matrix[,-1]

newnames <- c(str_c("V_", 1:14574))

setnames(train, old = c(colnames(train)), new = newnames)

# Make the random forest

mod01 <- randomForest(V_14574 ~ ., data = train, mtry = 1000, n.trees = 1000)

preds <- predict(mod01, train_val)

table <- table(preds, train_val_targets)

accuracy <- (table[1,1] + table[2,2] + table[3,3] + table[4,4]) / sum(table)
error <- 1-accuracy
```


## Boosting

```{r}
indices <- sample(1:nrow(train_data), 200)
train <- train_data[-indices,]
train_val <- train_data[indices,]
train_val_targets <- train_val$V_14574

mod02 <- gbm(V_14574 ~ ., data = train, distribution = "multinomial", n.trees = 10000, interaction.depth = 3, shrinkage = .001)

preds <- predict(mod02, train_val, n.trees = 10000)

predictions <- apply(preds, 1, which.max)

train_val_targets

predictions_2 <- factor(ifelse(predictions == 4, "lung",
                             ifelse(predictions == 3, "leukemia",
                                    ifelse(predictions == 2, "kidney",
                                           ifelse(predictions == 1, "brain", NA)))))


tib <- table(predictions_2, train_val_targets)

(tib[1,1] + tib[2,2] + tib[3,3] + tib[4,4]) / sum(tib)

tib %>% kable()

mod02$trees
```

```{r}
# Save models
saveRDS(mod02, "boost_01.rds")
saveRDS(mod01, "randomForest_01.rds")
indices <- as.data.frame(indices)
write_csv(indices, "indices_for_mod01_mod02.csv")

mod02_pt2 <- readRDS("boost_01.rds")
mod01 <- readRDS("randomForest_01.rds")

tib <- importance(mod01, type = 2)

tib <- tibble(name = rownames(tib),
              imp = tib[,1])

tib <- tib %>% arrange(desc(imp)) 

tib <- full_join(tib, reference, by = "name")

#rm(mod02_pt2)

indices <- read_csv("indices_for_mod01_mod02.csv")

vector <- as_vector(indices[,1])
```


## Deep Learning Model

```{r}
# Read in data

train_data <- read_csv("data/train_data.csv")
for (i in 1:14574) {
  train_data[,i] <- lapply(train_data[,i], factor)
}
```

```{r}
indices <- sample(1:nrow(train_data), 200)

train <- model.matrix(~ ., data = train_data[-indices,1:14573])

train <- train[,-1]

to_one_hot <- function(labels, dimension = 4) {
  results <- matrix(0, nrow = length(labels), ncol = dimension)
  for (i in 1:length(labels))
    results[i, labels[[i]]] <- 1
  results
}

train_val <- as_vector(train_data[-indices,14574])
train_val <- to_one_hot(train_val)


test <- model.matrix(~ ., data = train_data[indices,1:14573])
test <- test[,-1]
test_targets <- as_vector(train_data[indices,14574])

test_targets <- to_one_hot(test_targets)
```

```{r}
# verify model dimensions
ncol(test)
ncol(train)
train_val
test_targets

model <- keras_model_sequential() %>%
    layer_dense(units = 1024, activation = "relu", input_shape = dim(train)[[2]]) %>%
    layer_dense(units = 1024, activation = "relu") %>%
    layer_dense(units = 1024, activation = "relu") %>%
    layer_dense(units = 512, activation = "relu") %>%
    layer_dense(units = 256, activation = "relu") %>%
    layer_dense(units = 128, activation = "relu") %>%
    layer_dense(units = 4, activation = "softmax")

model %>% compile(
    optimizer = "rmsprop",
    loss = "categorical_crossentropy",
    metrics = c("accuracy")
  )

model %>% fit(train, train_val,
                epochs = 20, batch_size = 128, validation_data = list(test, test_targets))
```