search

gu-lab20 / MD-ALL

Molecular Diagnosis (MD) of Acute Lymphoblastic Leukemia (ALL): An integrative ALL classification system based on RNA-seq.
4 stars 3 forks source link
acute-lymphoblastic-leukemia classification machine-learning visualization

readme

MD-ALL

Molecular Diagnosis of Acute Lymphoblastic Leukemia

MD-ALL is designed to perform subtyping of B-cell acute lymphoblastic leukemia (B-ALL) using bulk RNA-seq or single-cell RNA-seq (scRNA-seq) data. The minimum input required for bulk RNA-seq analysis is the gene expression read count output from bioinformatics tools, such as HTSeq-count and FeatureCount. By uploading additional VCF files and raw outputs of fusion callers (FusionCatcher and/or Cicero), MD-ALL can detect B-ALL-related mutations and fussions in test sample and run RNAseqCNV to determine chromosome-level CNVs as well as iAMP21. MD-ALL can classify B-ALL cases into a total of 26 subtypes using bulk RNA-seq data. For scRNA-seq analysis, only a gene (per row) x cell (per column) expression matrix is needed. MD-ALL is a one-stop platform for sensitive, accurate, and comprehensive B-ALL subtyping based on RNA-seq data. To generate the input files for MD-ALL, users can refer to this RNA-seq analysis pipeline from raw fastq files. MD-ALL can be run by both shinyAPP or command lines.

The workflow of MD-ALL:



1. Installation

1.1 Install R and RStudio

MD-ALL is implemented in R. To ensure a better user experience, we recommend installing RStudio. Before using MD-ALL, please ensure that you have installed R and RStudio.

1.2 Install required packages

The following codes will check the packages already exist and install all the missing ones and MD-ALL. Please note that RTools is also required for the installation of the MD-ALL package.

#Get installed list of pacakges
installedPackages=  installed.packages()[,"Package"]

#Install CRAN R packages
list.of.packages=c("devtools", "BiocManager","dplyr","stringr","Seurat",
                   "ggplot2","ggrepel","cowplot","umap",
                   "shiny","shinyjs","shinydashboard")

new.packages = list.of.packages[!(list.of.packages %in% installedPackages)]
if(length(new.packages)){
  install.packages(new.packages)
}

#Install Bioconductor R packages
list.of.bioconductor.packages = c("DESeq2", "SingleR","SummarizedExperiment")
new.bioconductor.packages = list.of.bioconductor.packages[!(list.of.bioconductor.packages %in% installedPackages)]

if (length(new.bioconductor.packages)) {
  BiocManager::install(new.bioconductor.packages)
}

#Install PhenoGrapth
if(!"Rphenograph" %in% installedPackages){
  devtools::install_github("JinmiaoChenLab/Rphenograph")
}

#Install MD-ALL
devtools::install_github("gu-lab20/MD-ALL")

If you want to install using the script, you can download the installation script and run it in RStudio.

2. MD-ALL ShinyApp

Luanch the APP

library(MDALL)
run_shiny_MDALL()

After launching the Shiny app, users can select the analysis for either bulk RNA-seq or scRNA-seq data from the left sidebar. In the analysis for bulk RNA-seq, three modes are available: Single Sample, Multiple Samples, and Count Matrix Only.



2.1 Analysis for bulk RNA-seq data (Single Sample)

2.1.1 Upload data

To run the analysis for bulk RNA-seq data in the Single Sample mode, please upload at least the read count file. This mode is for the analysis of only one sample. For analysis of mulitple samples, please check section 4. MD-ALL also accepts VCF files and the raw outputs from fusion callers (MD-ALL supports FusionCatcher and Cicero) to perform more accurate B-ALL subtype classification. If the VCF and fusion calling files are missing, the output will be solely based on gene expression profile (GEP).

After uploading the input files, click the ‘Run’ button, and MD-ALL will start the analysis. The running time is around 3-5 minutes per sample using a standard desktop. Users can speed up the process by doing fewer rounds of GEP-based subtype prediction. The default parameter includes all gene number options. Please note that the parameters will not be displayed until the input files are successfully uploaded.

To test MD-ALL, users can download the testing files in the ‘tests.zip’ file from this GitHub repository.



The following are examples of code used in the analysis of GATK HaplotyperCaller to obtain the VCF file, which will serve as input in the subsequent RNAseqCNV and B-ALL mutations calling analysis. The codes start from .bam file. 

gatk SplitNCigarReads \
-R GRCh38.fa \
-I sample.bam \
-O sample.SplitNCigarReads.bam 

gatk HaplotypeCaller \
--dont-use-soft-clipped-bases \
-R GRCh38.fa \
-I sample.SplitNCigarReads.bam \
-O sample.HaplotypeCaller.vcf



2.1.2 Gene expression results

The ‘GEP Prediction’ tab contains results for GEP that are based on the read count file.

The top two panels show the head and tail of the raw count file and the results after normalization using VST from DESeq2.

The left plot panel contains the prediction heatmap on top and the UMAP at the bottom.
The color of the prediction heatmap indicates the predicted B-ALL subtypes for the testing sample.
The UMAP includes the 1821 reference B-ALL samples representing a total of 19 subtypes with distinct GEPs. The mapping of the testing sample is also highlighted on the UMAP.

The right panel contains a gene expression box plot, which allows users to check the expression levels of coding genes, particularly the feature genes for certain subtypes. For instance, NUTM1 is the feature gene for the NUTM1 subtype, so users can check its expression level when the prediction shows that the subtype is NUTM1 to confirm this subtype. The same condition also applies to CDX2 gene for CDX2/UBTF subtype and HLF gene for HLF subtype. Another scenario is to confirm the high expression of CRLF2 gene when a CRLF2 fusion is detected. It is also convenient to check the expression level of any genes of interest across B-ALL subtypes using this box plot.

Based on the GEP prediction results in this following figure, it can be inferred that the testing sample is likely to be a Ph/-like case with high CRLF2 expression.

2.1.3 RNAseqCNV results

The ‘RNAseqCNV’ tab contains results for the analysis of RNAseqCNV. Both the read count file and VCF file are required for this analysis. Please refer to the RNAseqCNV paper for more details. Briefly, chromosome gain leads to an overall higher expression of the genes on that chromosome, while chromosome loss leads to lower expression at the chromosome level. With chromosome gain or loss, the distribution of minor allele frequency (MAF) of common SNPs becomes skewed. The peak of the MAF distribution on normal diploid chromosomes is around 0.5. RNAseqCNV can also identify iAMP21 as we described in our manuscript.

In the following figure, we can see a higher overall expression for chr9 and chrX, as well as a skewed distribution (shown in red line) of MAF, indicating gain of chr9 and chrX.



2.1.4 Mutation results

The ‘Gene mutation’ tab will display B-ALL-related mutations detected in the VCF file.
The ‘B-ALL Mutations’ panel will show known B-ALL mutations (as we describe in our manuscript) detected in this sample.
The ‘B-ALL Subtype Defining Mutations’ panel will display the B-ALL subtype-defining mutations (PAX5:P80R, IKZF1:N159Y, and ZEB2:H1038R) tested for this purpose.

This figure indicates that no B-ALL mutations were detected in the VCF file of this possible Ph/-like case.



Another example of mutation results is as follows: a PAX5:P80R mutation was detected, which is the definitive mutation for the PAX5 P80R subtype.



2.1.5 Fusion results

If the output files from FusionCathcer and/or Cicero are uploaded, B-ALL-related fusions will be listed as in the following figure. Please be aware that fusions supported by a low number of reads may not be reliable.

2.1.6 Subtype summary

A summary of the analyses will be displayed under the “Summary” tab, along with the predicted final subtype, which will integrate the information from genetic alterations and GEP, shown in the last panel. Please note that the fusions listed in the ‘Genetic Alteration’ panel will only include fusions related to the final subtype. A final subtype will be generated automatically only if consistent predictions are obtained from both PhenoGraph and SVM. Otherwise, user judgment is required.

For this case, the GEP prediction indicated a Ph/-like subtype with high accuracy because both PhenoGraph and SVM achieved 100% confidence score. This case has no BCR::ABL1 fusion but multiple CRLF2 rearrangements, a signature event for Ph-like subtype, which lead to definitive classification of this test sample as Ph-like subtype.

2.2 Analysis for bulk RNA-seq data (Multiple Samples)

2.2.1 Upload data

For analysis of multiple samples, users can use the ‘Multiple Samples’ mode. Firstly, the users need to prepare the metadata table containing the filenames of the read count, VCF, FusionCatcher, and Cicero outputs, as shown below:



After uploading the correct metadata table, the parameters of MD-ALL will appear, just like in ‘Single Sample’ mode. Users can click the “Run” button to start the analysis.



2.2.2 Results for bulk RNA-seq analysis in ‘Multiple Samples’ mode

After all the analyses are done, users can check the results in the ‘Results’ tab. Select the sample ID in the top left panel, and the MD-ALL summary in the top right panel will update according to the selected sample ID. The bottom panel contains the results of GEP and RNAseqCNV; users can check these results using the ‘GEP’ and ‘RNAseqCNV’ tabs.

Results showing GEP in the ‘Multiple Samples’ mode:



Results showing RNAseqCNV in the ‘Multiple Samples’ mode:



2.3 Analysis for bulk RNA-seq data (Count Matrix Only)

2.3.1 Upload data

For users who only have the gene read count matrix, MD-ALL offers the ‘Count Matrix Only’ mode. The input data should be a read count matrix with rows representing genes and columns representing samples; the first column should be the ENSG gene IDs. After uploading the read count matrix, the parameters will appear and users can click the ‘Run’ button to start the analysis.



2.3.2 Results for bulk RNA-seq analysis in Count Matrix Only mode

The ‘Count Matrix Only’ mode only contains the results of GEP, since no other types of input are used. Users can check the results after the analysis is done. Users can still select the sample IDs in the top left panel, and the other parts will update accordingly:



2.4 Analysis for scRNA-seq data

2.4.1 Upload data for scRNA-seq analysis

The input file for scRNA-seq analysis is the count matrix of single cells with rows for genes and columns for cells. The ‘countMatrix_singlecell.tsv’ file in the same ‘tests.zip’ zip file can be used for testing.

2.4.2 Results of B-ALL subtyping using scRNA-seq data

Three UMAPs of single cells are shown on the left side. The top UMAP is colored by different cell types. The middle UMAP shows the cells belonging to pre-B and pro-B cells, which were used for the purpose of B-ALL subtyping. The bottom UMAP shows the cells colored by the annotation of different B-ALL subtypes.

The bar plot on the right side gives the percentages of each cell type with the pre-B and pro-B cells as one big group and the rest as the other group. The heatmap on the right side shows the scaled correlation score calculated by the SingleR package. The higher the correlation score, the more likely that the cell belongs to that subtype.

As shown in the bottom UMAP and the heatmap, most of the pre-B and pro-B cells are annotated as Hyperdiploid subtype with high correlation scores. The percentage of the cells annotated to Hyperdiploid among pre-B and pro-B cells were more than 97%, indicating that this sample is a Hyperdiploid case based on scRNA-seq GEP.



3. MD-ALL Command line

3.1 Analysis for bulk RNA-seq data

3.1.1 Read count data

library(MDALL)
df_count=read_input(file_count,delimiter = "\t",header = F)

3.1.2 Normalization with reference data

df_vst=get_vst_values(obj_in = obj_234_HTSeq,df_count = df_count)

3.1.3 Get normalized expression values for feature genes

df_feateure_exp=get_geneExpression(df_vst = df_vst,genes = c("CDX2","CRLF2","NUTM1"))

3.1.4 Get gene expression box plot

obj_boxplot=obj_merge(obj_in = obj_1821,df_in = df_vst,assay_name_in = "vst")
draw_BoxPlot(obj_in = obj_boxplot,group.by = "diag_raw1",features = "CRLF2",highlightLevel = "TestSample",plot_title = "Box Plot of Expression")

3.1.5 Imputation

df_vst_i=f_imputation(obj_ref = obj_234_HTSeq,df_in = df_vst)

3.1.6 Add testing sample to reference dataset for subtype prediction

obj_=obj_merge(obj_in = obj_1821,df_in = df_vst_i,assay_name_in = "vst")

3.1.7 Draw uMAP plot

obj_=run_umap(obj_in = obj_,out_label = "umap",n_neighbors = 10,variable_n = c(1058),feature_panel = "keyFeatures")
draw_DimPlot(obj_,group.by = "diag_raw",reduction = "umap",highlightLevel = "TestSample")

3.1.8 Run PhenoGraph clustering and SVM prediction

df_out_phenograph=get_PhenoGraphPreds(obj_in = obj_,feature_panel = "keyFeatures",SampleLevel = "TestSample",
                                      neighbor_k = 10,
                                      # variable_n_list = c(seq(100,1000,100),1058)
                                      variable_n_list = c(100,1058)
                                      )

df_out_svm=get_SVMPreds(models_svm,df_in = df_vst_i)

df_pred=bind_rows(df_out_phenograph,df_out_svm) %>% mutate(N=sprintf("%04d",featureN))

gg_tilePlot(df_in = df_pred,x = "N",y = "method",var_col = "pred",x_tick_label_var = "featureN",title = "Predition Heatmap")

3.1.9 Run RNAseqCNV

RNAseqCNV_out=run_RNAseqCNV(df_count = df_count,snv_file = file_vcf)

CNV_label=paste0(RNAseqCNV_out$df_cnv_out$gender,";\n",RNAseqCNV_out$df_cnv_out$chrom_n,",",RNAseqCNV_out$df_cnv_out$alterations)
chrom_n=RNAseqCNV_out$df_cnv_out$chrom_n

print(CNV_label)
print(chrom_n)

Draw RNAseqCNV Plot

get_RNAseqCNV_plot(RNAseqCNV_out=RNAseqCNV_out)

3.1.10 Get mutations

out_mutation=get_BALL_mutation(file_vcf)
out_mutation$out_text_BALLmutation

3.1.11 Get fusions

For fusionCaller:

fusion_fc=get_BALL_fusion(file_fusioncatcher,type = "fc")
fusion_fc

For cicero:

fusion_c=get_BALL_fusion(file_cicero,type = "fc")
fusion_c

3.1.12 Get GEP prediction summarise

df_sum=get_subtype_final(
  id="TestSample",
  df_feateure_exp = df_feateure_exp,
                  df_out_phenograph = df_out_phenograph,df_out_svm = df_out_svm,
                  out_mutation = out_mutation,
                  chrom_n = chrom_n,CNV_label = CNV_label,
                  fusion_fc = fusion_fc,fusion_c = fusion_c)
df_sum

3.2 Wrap of analysis for one sample

df_out_testOne=run_one_sample(sample_id = "TestId",file_count = file_count,
               file_vcf = file_vcf,
               file_fusioncatcher = file_fusioncatcher,
               file_cicero = file_cicero,
               featureN_PG = c(100,1058))
df_out_testOne

3.3 Wrap of analysis for multiple samples

Prepare the listing file of input files

df_listing=read.table("test/file_list.tsv",sep  = "\t",header = T)
df_listing

Analysis for multiple samples

out_testMul=run_multiple_samples(file_listing = "test/file_list.tsv",featureN_PG = c(100,1058))

check results

df_out_testMul=out_testMul$df_sums
df_out_testMul

3.4 BALL subtyping for single cell data


count_sc=read_sc_file("test/countMatrix_singlecell.tsv")
sc_report=get_SC_subtypes(count_matrix = count_sc,SE_celltype = SE_celltype,SE_BALL = SE_BALL)
sc_report

Contact:

Zunsong Hu: zuhu@coh.org

Zhaohui Gu: zgu@coh.org