How to train your model

Step 1: Obtain input data

First, a positive and negative dataset have to be obtained. In this example, we want to predict AMPs in bats and decide to train a model using protein sequences found in bats. The positive dataset are AMPs and the negative dataset are random sequences. Both datasets were obtained from UniProt:

• positive by using the search term keyword:antimicrobial taxonomy:“Chiroptera [9397]”
• negative by using the search term taxonomy:“Chiroptera [9397]”

For the positive dataset:

• read data
• add “positive” label
• remove non standard amino acids

library(ampir)

bat_pos <- read_faa(system.file("extdata/bat_positive.fasta.gz", package = "ampir"))
bat_pos$Label <- "Positive"
bat_pos <- remove_nonstandard_aa(bat_pos)

For the negative dataset:

• read data
• add “negative” lavel
• remove non standard amino acids
• remove sequences (if any) that are also present in the positive dataset
• randomly select the same number of sequences that are in the positive dataset

bat_neg <- read_faa(system.file("extdata/bat_negative.fasta.gz", package = "ampir"))
bat_neg$Label <- "Negative"
bat_neg <- remove_nonstandard_aa(bat_neg)
bat_neg <- bat_neg[!bat_neg$seq_aa %in% bat_pos$seq_aa,]
bat_neg <- bat_neg[sample(nrow(bat_neg),78),]

Combine the positive and negative dataset

bats <- rbind(bat_pos, bat_neg)

Calculate features on the combined positive and negative dataset and add the label column

bats_features <- calculate_features(bats)
bats_features$Label <- as.factor(bats$Label)
rownames(bats_features) <- NULL

library(caret)

Split feature set data and create train and test set with caret

trainIndex <-createDataPartition(y=bats_features$Label, p=.7, list = FALSE)
bats_featuresTrain <-bats_features[trainIndex,]
bats_featuresTest <-bats_features[-trainIndex,]

Resample method using repeated cross validation and adding in a probability calculation with caret

trctrl_prob <- trainControl(method = "repeatedcv", number = 10, repeats = 3,
                            classProbs = TRUE)

Train model using a support vector machine with radial kernel with caret. Note: Other classification models are supported too. For example, to use a random forest model in caret, method could be changed from “svmRadial” to “ranger”.

my_bat_svm_model <- train(Label~.,
                       data = bats_featuresTrain[,-1], # excluding seq_name column
                       method="svmRadial",
                       trControl = trctrl_prob,
                       preProcess = c("center", "scale"))

Test model to get an indication of how well the model performs on test data with caret

my_bat_pred <- predict(my_bat_svm_model, bats_featuresTest)
cm <- confusionMatrix(my_bat_pred, bats_featuresTest$Label, positive = "Positive")

Subset from cm$byClass

Convert the bat feature test data to the original FASTA type format containing just the sequence name and sequence as this is the required input data for ampir

bat_test_set <- bats[bats$seq_name %in% bats_featuresTest$seq_name,][,-3]
rownames(bat_test_set) <- NULL

Use the trained bat model in ampir’s predict_amps function on the bat test set

my_bat_AMPs <- predict_amps(bat_test_set, min_len = 5, model = my_bat_svm_model)

my_bat_AMPs sample