Team:Evry/Software/Pipeline
Pipeline
All the information presented on this page (quality-control, differential expression analysis, data visualisation, variant discovery) is also available as a PDF file.
Data processing and quality control
What we produced: FASTQ files (if we don't have them), FASTQC reports, BAM and SAM files.
Figure 1: schematic overview of the pipeline for RNA-seq data analysis.
Differential expression analysis
What we produced: script for differential expression analysis, table with read counts (tab separated format, 7 columns, ENSG ids).
RNA-seq data can be difficult to interpret (especially in terms of differential expression quantitation). Thus, we decided to adopt a simple method for the analysis, based on counting, for each gene and for each sample, the number of available reads and then testing for significant differences between two experimental conditions or groups.
We wrote an R script that automatically creates a PDF file (in the current directory) with all the figures necessary for visual inspection and result interpretation. The input is a tab separated file with reads counts.
ensembl_id melanocyte_1 melanocyte_2 melanoma_1 melanoma_2 ENSG00000000003 1964 2409 2328 2451 ENSG00000000005 0 2 10 12 ENSG00000000419 15122 19592 38225 36654 ENSG00000000457 12129 14893 7483 7812 ENSG00000000460 21930 25575 13123 13840 ENSG00000000938 48 58 26 42 ENSG00000000971 125 229 124 236 ENSG00000001036 11611 14125 14067 13518 ENSG00000001084 11429 13795 3549 3279
Figure 2: Example input format for DE analysis.
We tested two designs, as illustrated in the tables below: normal cells vs cancerous cells (4 samples), cancerous cells vs cancerous drug treated (4 samples).
Sample name | Condition |
---|---|
melanocyte_1 | M |
melanocyte_2 | M |
melanoma_1 | C |
melanoma_2 | C |
Sample name | Condition |
---|---|
melanoma_1 | C |
melanoma_2 | C |
melanoma_drug_1 | D |
melanoma_drug_2 | D |
Table 1 and 2: tested designs.
Visual exploration of the samples
Prior to checking distances between our samples, we applied a regularized-logarithm transformation (rlog) to stabilise the variance across the mean. The effects of the transformation are shown in the figure below.
Figure 3: Effect of the regularized-logarithm transformation on 'melanocyte_1' and 'melanocyte_2' samples.
We noticed that this step was particularly important for genes with low read counts.
We then checked the distances between our samples by performing Principal Components Analysis of the count data.
Figure 4: Principal Components Analysis (PCA) plot, normal vs cancerous cells.
We observed that differences between groups (normal vs cancerous cells represented in the PCA plot above) were greater than intra-groups differences, which is expected in this kind of design. However, as the inter-group differences were so pronounced, we figured that a great amount of genes would appear as differentially expressed. This is why we decided to apply really stringent thresholds for the detection:
- - log2 fold change (logFC) > 5 for upregulated genes or log2 fold change (logFC) < -5 for downregulated genes.
- - AND adjusted-p-value < 0.01
Differential expression analysis
Firstly, we took a look at the raw data (prior to any kind of normalization). We calculated mean counts for each gene and by condition and then the log2 fold change.
Prior to normalization, we filtered the data set to remove rows with very little or no information (remove genes with no counts or with just a single count). This allows to eliminate 17,386 transcripts already.
Using the DESeq R package (from Bioconductor), we were able to perform normalization of our data after calculation of size factors and we then were able to calculate mean counts for each gene and by condition and finally the logFC.
Figure 5: Distribution of logFC(cancerous/normal) values - raw data.
Figure 6: Distribution of logFC(cancerous/normal) values - normalized data.
Finally, we applied the nbinomWaldTest() function from the DESeq package to test for significance of coefficients in a negative binomial GLM, the model we used to assess differences in expression. As previously stated, selection of significantly up- or downregulated genes was based on the establishment of two selection thresholds: logFC and adjusted p-value (Wald test M vs C).
Figure 7: Differential expression as a function of mean expression. Left panel: threshold set at logFC > 2 or < -2. Right panel: threshold set at logFC > 5 or < -5. The red dots indicate genes for which the logFC was significantly higher than 5 or lower than -5. The circled point indicates the gene with the lowest adj-p-value.
We obtained a list of 1,649 differentially expressed genes: 931 upregulated genes and 718 downregulated genes.
Enrichment analysis
We retrieved the list of the 931 unregulated genes and the list of the 718 downregulated genes and looked for significantly enriched GO (Gene Ontology) terms in these lists (independently).
After idenfication of genes that are both overexpressed and mutated in tumor samples, we want to know if good candidate antigens can be predicted. Read more about the prediction step.