Data processing and quality control

What we produced: FASTQ files (if starting with a different file format), 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).

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).

Figure 8: Enrichment in GO terms, downregulated genes.

Figure 9: Enrichment in GO terms, upregulated genes.

Variant discovery

What we produced:

• - Bash scripts for variant calling, quality control (filtering steps), and variant association analysis
• - VCF files (before and after QC)
• - Table: identified variants (exonic, non-synonymous)

Figure 10: schematic overview of the pipeline for variant discovery and evaluation.

Variant calling

Genetic variants were called with samtools and bcftools, using sorted .bam files as input. This simple variant calling is followed by a stringent filtering step.

Quality control (filtering steps)

As recommended in the GATK Best Practices Guideline for variant discovery using RNA-Seq data, we applied hard filters to the raw variants obtained after variant calling, in an attempt to optimise both high sensitivity and specificity.

Furthermore, as we only have 4 samples, we decided to use quite stringent parameters / thresholds to filter the data, hoping to retain “true” and of as high a quality as possible variants. Filtering was performed using scripts from GATK and VCFtools.

Filters:

• (1) Diallelic variants only.
• (2) Hardy-Weinberg equilibrium (HWE) deviation test. It is a common practice to remove sites that deviate from HWE because the deviation can be caused by genotyping errors. Normally, for case-control data, only controls should be tested for deviation from HWE (because for cases, sites associated with disease status can deviate from HWE). In our case, as all tests were performed in a bidirectional manner, deviation from HWE was tested in all the samples and we excluded sites with a HWE p-value < 1.10^-7.
• (3) Call rate (percentage of samples with a non-missing genotype, CR). The proportion of missing genotypes is an useful indicator of poor genotype quality. We decided to keep variants with a CR > 98%, which allows to keep good quality variants only. As mean CR in raw data was of about 64%, we discarded over 60% of variants using this filter.
• (4) Filtering based on Fisher Strand values (FS > 30.0) and Quality by Depth (QD < 2.0), as well as filtering out clusters of at least 3 SNPs in a window of 35 bases between them.

In order to assess the quality gain at each QC step, we estimated the ratio of transitions (Ti, purine to purine or pyrimidine to pyrimidine mutation) to transversions (Tv, purine to pyrimidine or vice versa) in the identified single nucleotide variants (SNVs). Particularly in coding regions, a higher number of transitions is expected, as transversions are more likely to change the underlying amino acid and lead to a deleterious mutation. Ti/Tv ratios are an approximate measure of quality: higher Ti/Tv ratios are associated with lower false positives.

Figure 11: Number of variants retained and Ti/Tv ratio for every QC step.

Table 3: Number of variants retained and Ti/Tv ratio for every QC step.

Annotation

Annotation attributes such as genomic region, gene name, variant type and consequence are attached to the variants list according to the reference hg19 using ANNOVAR (AnnotateVariation perl script). The primary genomic effects that are annotated include splice sites, nonsense, nonsynonymous and synonymous variants.

Association testing between individual variants and phenotypic traits

Here, common variants were defined as being those that are present in more than one sample. Of the 294 034 variants retained after quality control, 233 294 were identified as common. We identified 24 347 exonic variants only, over 19 000 of these were common. Thus, we decided to work only on common variants.
We performed standard single variant test to assess association: logistic regression and fisher’s exact test.

We found 531 exonic non-synonymous variants having a Fisher’s p-value < 0.05 (p = 0.02, being the lowest value we could get with 4 samples). 315 of these variants were only present in the melanoma cell lines (all were homozygous variants).

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.