RNA-Seq Experimental Design

What is RNA-seq?

RNA-seq is a method of measuring gene expression using shotgun sequencing. The process involves reverse transcribing RNA into cDNA, then sequencing fragments on a high-throughput platform such as Illumina to obtain a large number of short reads. For each sample, the reads are then aligned to a genome, and the number of reads aligned to each gene or feature is recorded.

A typical RNA-seq experiment aims to find differentially expressed genes between two conditions (e.g. up and down-regulated genes in knock-out mice compared to wild-type mice). RNA-seq can also be used to discover new transcripts, splice variants, and fusion genes.

Why is a good experimental design vital?

An RNA-seq experiment produces high dimensional data. This means we get a huge number of observations for a small number of samples. For example, the expression of ~20,000 genes could be measured for 6 samples (3 knock-out and 3 wild-type). A frequently used approach to analyse RNA-seq data is to fit each gene to a linear model where for each of the 20,000 genes, parameters need to be estimated using a small number of observations. To complicate matters, each measurement of gene expression is comprised of a mix of biological signal and unwanted noise. Thus, in order to perform a robust statistical analysis, the methodology must be carefully designed.

Before you begin any RNA-seq experiment, some questions you should ask yourself are:

• Why do you expect to find differentially expressed genes in the particular tissue?
• What types of genes do you expect to find differentially expressed?
• What are the sources of variability from your samples?
• Where do you expect most of your variation to come from?

A coherent experimental design is the groundwork of a successful experiment. You should invest time and thought in designing a robust experiment as failing to think this step through can lead to unusable data and wasted time, money, and effort.

It is also useful to think about the statistical methods you will use to analyse the data. If you’re planning to bring a data analyst or bioinformatician onboard for data analysis, you should include him or her in the experimental design stage.

Terminology

Before progressing, it may be useful to define some terms which are commonly used in RNA-seq.

Variability:	A measure of how much the data is spread around. Variance is mathematically defined as the average of the squared difference between observations and the expected value. Simply put, a larger variance means it is harder to identify differentially expressed genes.
Feature:	A defined genomic region. Usually a gene in RNA-seq, but can also refer to any region such as an exon or an isoform. In RNA-seq, an estimate of abundance is obtained for each feature.
Biological replicates:	Samples that have been obtained from biologically separate samples. This can mean different individual organisms (e.g. tissue samples from different mice), different samplings of the same tumour, or different population of cells grown separately from each other but originating from the same cell-line. For example, the samples obtained from three different knock-out mice could be considered biological replicates in a knock-out versus wild-type experiment. A biological replicate combines both technical and biological variability as it is also an independent case of all the technical steps.
Technical replicates:	Samples in which the starting biological sample is the same, but the replicates are processed separately. For example, if a biological sample is divided and two different library preps are processed and sequenced, those two samples would be considered technical replicates.
Covariate:	The term 'covariate' is often used interchangeably with 'factor' or 'variable' in RNA-seq. The term refers to a property of the sample which may have some influence on gene expression and should be represented in the RNA-seq model. Covariates in RNA-seq are often categorical (e.g. treatment condition, sex, batch), but continuous factors are also possible (e.g. time points, age). A linear model will contain terms to represent the relationships between covariates and each sample. Each possible value a factor can take is called a level (e.g. 'male' and 'female' are two levels in the factor 'sex'). Factors can either be directly of interest to the experiment (e.g. treatment condition) or not of interest (also known as nuisance variables) (e.g. sex, batch). The purpose of covariates is to explain the variance seen in samples.
Confounding variable:	A confounding variable is a nuisance variable that is associated with the factor of interest. Possible confounding factors should be controlled for so they don't interfere with analysis. For example, if all knock-out mice samples were harvested in the morning and all wild-type mice samples were harvested in the afternoon, the time of sample collection would be a confounding factor as the effects from sample collection time and from the knock-out cannot be separated.
Statistical power:	The ability to identify differentially expressed genes when there really is a difference. This is partly dependent on variance and therefore is affected by the number of replicates available and sequencing depth.

The importance of replicates to estimate variance

When performing a differential gene expression analysis, we look at the expression values of each gene and try to determine if the expression is significantly different in the different conditions (e.g. knock-out and wild-type). The ability to distinguish whether a gene is differentially expressed is partly determined by the estimates of variability obtained by using multiple observations in each condition.

Variability is present in two forms: technical variability and biological variability.

Combined biological and technical variability is measured using biological replicates. Biological variability is the main source of variability and is due to natural variation in the population and within cells. This includes different individuals having different levels of a particular gene and the stochastic nature of expression levels in different cells.

Technical variability is measured using technical replicates. Technical variability is often very small compared to biological variability. Usually the question is whether an observed difference is greater than the total variability (i.e. significant). As combined variability is measured by biological replicates technical replicates are only important if you need to know the degree of biological variability or technical variability. An example of wanting technical variability would be method development. The main source of technical variation comes from RNA processing and from library prep. Variability from sequencing in different flow cells or different lanes is usually minimal. Generally, creating technical replicates from multiple library preps is unnecessary for RNA-seq experiments.

The amount of variance between your biological replicates will affect the outcome of your analysis. Ideally, you aim to have minimal variability between samples so you only measure the effect of the condition of interest. Too much variability between samples can drown out the signal of truly differentially expressed genes. Controlling for possible confounding factors between conditions is also important to prevent falsely attributing differential expression to the condition of interest.

Strategies to minimise variation between samples and to control confounding variables include:

• choosing organisms from the same litter,
• choosing organisms of the same sex if possible,
• using a constant sample collection time,
• having the same laboratory technician perform each library prep,
• randomising samples to prevent a confounding batch effect if all samples can’t be processed at one time.

If variation between samples can not be removed it should be balanced between conditions of interest as much as possible, and carefully recorded to allow its effect to be measured and potentially removed during analysis.

How many replicates and how many reads do I need?

Two very common question asked are:

• how many biological replicates do I need, and
• what sequencing depth is needed for each sample

in order to have enough statistical power for my RNA-seq experiment?

These questions cannot be precisely answered without a pilot study. A small amount of data (minimum of two biological replicates for each condition with at least 10M reads) can estimate the amount of biological variation, which determines how many biological replicates are required. Performing a pilot study is highly recommended to estimate statistical power and identify possible problems before investing more time and money into the project.

Scotty is a web-based tool that uses data generated from a pilot study to optimize a design for statistical power. With a limited budget, one must balance sequence coverage and number of biological replicates. Scotty also has a cost estimate feature which returns the most powerful design within budget constraints.

As a general rule, the number of biological replicates should never be below 3. For a basic RNA-seq differential expression experiment, 10M to 20M reads per sample is usually enough. If similar data exists it can be helpful to check the read counts for key genes of interest to estimate the required depth.

Biological variability is usually the largest effect limiting the power of RNA-seq analysis. The most improvement in an experiment will usually be achieved by increasing the biological replication to improve estimation of the biological variation.

It is often possible to design experiments where the analysis is done incrementally such that a pilot study is added to with an additional block of samples or a pool of libraries is sequenced to additional depth. In these cases care must be taken to balance the design in a manner that each stage is a valid experiment in its own right. This can allow a focused question to be answered in the first stage, with an ability to either address issues or progress to a second stage with additional questions.

Sequencing options to consider

How much total RNA is needed:
Many sequencing centres such as AGRF recommend at least 250ng of total RNA for RNA sequencing. It is possible to go as low as 100ng of total RNA, but results are not guaranteed. The quality of RNA is also important when making libraries. A RNA Integrity Number (RIN) is a number from 1 (poor) to 10 (good) and can indicate how much degradation there is in the sample. A poor score can lead to over representation at the 3’ end of the transcript and low yield. Samples with low RIN scores (below 8) are not recommended for sequencing. Care should also be taken to ensure RIN is consistent between conditions to avoid confounding this technical effect with the biological question.

Choosing an enrichment method:
Ribosomal RNA makes up >95% of total cellular RNA, so a preparation for RNA-seq must either enrich for mRNA using poly-A enrichment, or deplete rRNA. Poly-A enrichment is recommended for most standard RNA-seq experiments, but will not provide information about microRNAs and other non-coding RNA species. In general, ribo-depleted RNA-seq data will contain more noise, however, the protocol is recommended if you have poor or variable quality of RNA as the 3’ bias of poly-A enrichment will be more pronounced with increased RNA degradation. The amount of RNA needed for each method differs. For Poly-A enrichment a minimum of 100ng is needed while for ribo-depletion, a minimum of 200ng is recommended.

Choosing read type:
For basic differential expression analysis RNA-seq experiments, single-end sequencing is recommended to obtain gene transcript counts. In more advanced experiments, paired-ends are useful for determining transcript structure and discovering splice variants.

Choosing strandedness:

With a non-directional (unstranded) protocol, there is no way to identify whether a read originated from the coding strand or its reverse complement. Non-directional protocols allow mapping of a read to a genomic location, but not the direction in which the RNA was transcribed. They are therefore used to count transcripts for known genes, and are recommended for basic RNA-seq experiments. Directional protocols (stranded) preserve strand information and are useful for novel transcript discovery.

Multiplexing:
Multiplexing is an approach to sequence multiple samples in the same sequencing lane. By sequencing all samples in the same lane, multiplexing can also minimise bias from lane effects.

Spike-in controls:
RNA-seq spike-in controls are a set of synthetic RNAs of known concentration which act as negative or positive controls. These controls have been used for normalisation and quality control, but recent work has shown that the amount of technical variability in their use dramatically reduces their utility.

Summary

• A good experimental design is vital for a successful experiment. If you’re planning to work with a data analyst or bioinformatician, include them in the design stage.
• Aim to minimise variability by identifying possible sources of variance in your samples.
• Biological replicates are important. The number of biological replicates you should have should never be below 3. Technical replicates are often unnecessary.
• Pilot studies are highly recommended for identifying how many replicates and how many reads you should have for enough statistical power in your experiment.
• For basic RNA-seq experiments, poly-A enriched, single-ended, unstranded sequencing at depths of 10M to 20M is probably what you want.