Introduction to the iimi package

Haochen Ning

First created on 2023-04-29. Updated on 2024-07-23

1. Introduction

This vignette aims to give an introduction on how to use the iimi package for plant virus diagnostics and how to visualize the coverage profile for the sample mapping. We also included a tutorial on creating unreliable regions.

1.1. Installation

First, let’s install necessary packages. You may skip this step if you have installed the packages before.

# install iimi
install.packages(c("iimi", "dplyr"))

# install Biostrings
if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("Biostrings")

1.2. Loading packages

We will load necessary packages before we start any analysis.

library(iimi)

library(Biostrings)

library(dplyr)

1.3. Data Pre-processing

We provide three example coverage profiles to demonstrate how to use iimi. These files are sourced from the dataset used in the VirHunter paper (Sukhorukov et al. 2022), which we also utilized for external validation in our manuscript (Ning et al. 2024). You can download these files directly from the Recherche Data Gouv website (Candresse, Marais-Colombel, and Brault 2022). We recommend storing all BAM files in a single folder for ease of access.

To get started creating coverage profiles and feature-extracted data frames, the process involves four steps: (1) Downloading the raw data in FASTQ format from the link above (2) Use Bowtie2 to map the FASTQ files (paired-end or single-end) against the official Virtool virus data base (ver. 1.4.0) to get SAM outputs (3) Convert SAM to indexed and sorted BAM using Samtools (4) Generate coverage profiles and feature-extracted data frames using iimi functions (tutorials in the next section)

We recommend Bowtie2 or minimap2 since we have tried both and they yield similar result with minimap2 having a slight decrease. We let both software to report all alignments (-a mode for Bowtie2, --secondary=yes for minimap2). You can also use other mapping tools.

Next, we provide a short tutorial to guide you through using iimi functions to make predictions on the real data.

2. Converting BAM file(s) into coverage profiles and feature-extracted data frame

Let’s convert the indexed and sorted BAM file(s) into coverage profiles and feature-extracted data frame from the previous section.

We will use the coverage profiles to visualize the mapping information. The feature-extracted data frame will be used in the model training and testing process.

Note that both training and testing data need to go through the conversion step.

In our example, we stored the conversion for both the testing and training datasets in the same object. You can do the conversion separately for your data.

Important: the example code does not work unless the path to the folder that stores your BAM files is provided.

2.1. State the path to the folder of your BAM files

If you already have coverage profiles in run-length encoding (RLE) format, go to section 2.2.

path_to_bamfiles <- list.files(
  path = "path/to/your/BAM/files/folder",
  pattern = "bam$",
  full.names = TRUE,
  include.dirs = TRUE
)

2.2. Create a data frame that contains the coverage profiles.

2.2.1. Convert BAM files to a list of RLEs.

You may skip this step if you already have converted them to RLE format.

example_cov <- convert_bam_to_rle(bam_file = "path_to_bamfiles")

2.2.2. Convert the list of RLEs to a feature-extracted data frame.

This section explains how to convert the RLE format to feature-extracted dataframes, with options for handling unreliable regions.

By default, the package uses provided unreliable regions and enables mappability profiling and filtering to eliminate false peaks. This is the recommended approach for most users. * To use the default settings, no additional code is required. * To disable this feature, set unreliable_region_enabled = FALSE when calling convert_rle_to_df().

If you wish to use your own unreliable regions, refer to section 3.3 for instructions on creating custom unreliable regions and input your custom unreliable regions using the unreliable_region_df parameter in convert_rle_to_df().

Note: Customization of unreliable regions must be done before calling convert_rle_to_df(). Most users can skip this step and use the provided unreliable regions.

# Using default settings (recommended)
df <- convert_rle_to_df(example_cov)

# Disabling unreliable region processing
df <-
  convert_rle_to_df(example_cov, unreliable_region_enabled = FALSE)

# Using custom unreliable regions
# Refer to section 3.3 for details
custom_regions <- create_custom_unreliable_regions()
df <-
  convert_rle_to_df(example_cov, unreliable_region_df = custom_regions)

3. Predicting the plant sample(s)

To make predictions, use the converted mapping result of the sample(s) that you wish to detect as the input, newdata. Make sure you have converted the indexed and sorted BAM files into feature-extracted data frame from the section above.

After preparing your test sample, you can choose to test the data using our provided training model or the model you trained using train_iimi(). The tutorial of training your own model is provided in the next section.

Note: if you wish to customize unreliable regions, please go to 3.3.

3.1 Using pre-trained models and no customization

If you wish to use provided training model, only input your data to newdata and choose a method of your wish using predict_iimi().

There are three methods that you may choose from: xgb, en, and rf, which stand for pre-trained XGBoost, elastic net, and random forest models. The example below uses the pre-trained XGBoost model.

prediction_default <- predict_iimi(newdata = df, method = "xgb")

The detection of your plant sample(s) is finished. The prediction is TRUE if virus infected the sample, FALSE if virus did not infect the sample.

3.2. Customizing your own model

If you would like to train your own model and use this model to test your data, you can use the codes below to train a new model with your own data.

Ideally, the number of the samples used to train the model should be bigger than 100. However, we are only providing a tutorial on how to use the train_iimi() function, only two samples are used to train the model since example_cov() only contains three in-house data’s coverage information.

Now, we need to prepare our training data. We are using a 80/20 random split to split the three samples. This means that two samples are used as the training data, and one sample is used as the testing data. If you are training your own data, the training data is the data that you want to train the model on; the testing data is the data that you would like to have a prediction on.

Here are some definitions/explanation of the objects to input in train_iimi():

  1. train_x: the feature-extracted data frame of plant samples that you would like to train iimi model on. Make sure that you have mapped the samples to the virus database and converted the mapping result to sorted and indexes BAM files.

  2. train_y: the known truth or labels for your train_x data. Please label the data to make sure that it has a detection label for virus segments as well.

  3. test_x: the feature-extracted data frame of plant samples that you would like to predict using your trained iimi model. Make sure that you have mapped the samples to the virus database and converted the mapping result to sorted and indexes BAM files.

# preparing the train and test data

# spliting into 80-20 train and test data set with the three plant samples
set.seed(123)
train_names <- sample(levels(as.factor(df$sample_id)),
                      length(unique(df$sample_id)) * 0.8)

# trian data
train_x = df[df$sample_id %in% train_names,]
# test data
test_x = df[df$sample_id %in% train_names == F,]

# preparing labels
train_y = c()
for (ii in 1:nrow(train_x)) {
  train_y = append(train_y, example_diag[train_x$seg_id[ii],
                                         train_x$sample_id[ii]])
}

Then, we plug in the variables into the train_iimi function with the default XGBoost model to train your custom model.

fit <- train_iimi(train_x = train_x, train_y = train_y)

Now, we have a trained model using the toy data.

Then, the process to detect which viruses infect the plant sample(s) is the same as previously described, except we are using a customized trained model.

prediction_customized <-
  predict_iimi(newdata = test_x,
               trained_model = fit)

The detection of the plant sample(s) is finished. The interpretation is the same as above.

3.3. Customizing unreliable regions

Note: if you would like to create your own unreliable regions, please customize them first, then extract features to build a data frame from section 2.2.2. using customized unreliable regions.

If you would like to create your own unreliable regions besides from using your own training model, you may use create_mappability_profile() and high_nucleotide_regions(). Both functions’ output is a data frame with the start and end position of the unmappable region, the virus that the region is on, and the category that it belongs to.

# An example of the provided unreliable regions
unreliable_regions %>% group_by(Categories) %>% sample_n(2)
## # A tibble: 8 × 4
## # Groups:   Categories [4]
##   Start   End `Virus segment` Categories                
##   <dbl> <dbl> <fct>           <fct>                     
## 1  9128  9214 a4o62bj9        A% > 45%                  
## 2 13422 13518 csvw8bc9        A% > 45%                  
## 3  6629  6713 kc1k6klg        CG% > 60%                 
## 4   737   870 y3xt5m37        CG% > 60%                 
## 5  2290  2404 bqv0w6sv        Unmappable regions (host) 
## 6   719   800 hxo0d0c6        Unmappable regions (host) 
## 7   638   817 1u34mk0r        Unmappable regions (virus)
## 8  1658  1744 q6lvljsq        Unmappable regions (virus)

Unreliable regions contains (1) mappability profile (2) high nucleotide content regions.

Including these two profiles into iimi ensures that there are no false peaks like the ones described in the previous section.

3.3.1. Mappability profile

Here is a short tutorial to make mappability profile.

  1. Split each of the virus segment from the virus database into a sliding window series with window size of your choice and with step size 1. The default value for window size is 75. You may choose any window size you want.

  2. Map one virus segment with each other, until you finish mapping it to all virus segments in the virus database. Also map the virus segment with a host genome of your choice. We chose to use Arabidopsis Thaliana.

  3. Sort and index the resulted BAM files from the mapping step.

  4. Assemble the mappability profile:

# if you would like to keep unmappable regions that can be mapped to other
# viruses or the host genome separate into two data frames, you may use the
# following code:

# input your own path that you would want to store regions on a virus that can
# be mapped to another virus
# you can customize the name of this type of mappability profile
mappability_profile_virus <-
  create_mappability_profile("path/to/bam/files/folder/virus", category = "Unmappable region (virus)")

# input your own path that you would want to store regions on a virus that can
# be mapped to the host genome
# you can customize the name of this type of mappability profile
mappability_profile_host <-
  create_mappability_profile("path/to/bam/files/folder/host", category = "Unmappable region (host)")

# if you would like to keep everything in the same data frame, you may use the
# following code:
mappability_profile <-
  create_mappability_profile("path/to/bam/files/folder/of/both/types/", category = "Unmappable region")

3.3.2. High nucleotide content regions

Creating the high nucleotide content regions is much easier than the mappability profile. We only need to use create_high_nucleotide_content() function.

Here is an example:

high_nucleotide_regions <-
  create_high_nucleotide_content(gc = 0.6, a = 0.45)

The default threshold for GC content is 60% and is 45% for A%. The thresholds are customizable.

Now, we can combine these two regions into the final unreliable regions. You can use them to convert your training and testing data to feature-extracted data frames. Please refer to section 2.2.2. to see how to do so.

4. Visualizing the coverage profiles

Next, we can visualize the coverage profile by using the plot_cov() function.

oldpar <- par(mfrow = c(1, 2))

## if you wish to plot all segments of one sample, you can try:
# plot_cov(covs = example_cov["S1"])

## if you wish to plot all segments from all samples, you can try:
# plot_cov(covs = example_cov)

## if you wish to plot certain segments from one sample, you can try:
segs = c("42jtlrir", "m0kacxse")
covs_selected = list()
covs_selected$`305S` <-
  example_cov$`305S`[segs]

## if you have many segments that you would want to plot, you can try the following code with the numbers changed
## to find the index of your desired segments:

covs_selected$S1 <-
  example_cov$S1[names(example_cov$S1)[c(1,72)]]

par(mar = c(2, 4, 1, 1))
layout(matrix(c(1, 1, 2, 3, 3, 4), nrow = 3))
plot_cov(covs = covs_selected)

par(oldpar)

This gives us a general idea of what the potential viruses are.

5. References

Candresse, T, A Marais-Colombel, and V Brault. 2022. Sugar beet sequencing data (RNAseq).” Recherche Data Gouv. https://doi.org/10.15454/MK1JIW.
Ning, H., I. Boyes, I. Numanagić, M. Rott, L. Xing, and X. Zhang. 2024. “Diagnostics of Viral Infections Using High-Throughput Genome Sequencing Data.”
Sukhorukov, G, M Khalili, O Gascuel, T Candresse, A Marais-Colombel, and M Nikolski. 2022. “VirHunter: A Deep Learning-Based Method for Detection of Novel RNA Viruses in Plant Sequencing Data.” Front Bioinform 2 (May). https://doi.org/10.3389/fbinf.2022.867111.