Experimental Design

Why experimental design matters before you sequence

No amount of bioinformatics analysis can rescue a poorly designed experiment. The decisions you make before sequencing determine what questions you can answer. They also determine how confident you can be in your results.

A common mistake is to focus on the analysis and treat the experimental design as an afterthought. In practice, most failed RNA-seq studies fail because of design problems, not analysis problems. Spending time on design before you submit samples will save weeks of troubleshooting later.

Biological vs technical replicates

Replicates are the foundation of any statistical analysis. Without them, you cannot distinguish true biological signal from noise.

Technical replicates are repeated measurements of the same biological sample. You split one RNA extract into two aliquots and sequence both. Technical replicates tell you about the variability introduced by the sequencing process itself. In modern RNA-seq, technical variability is very low. Sequencing the same library twice produces nearly identical results.

Biological replicates are independent biological samples from the same condition. If you are studying the effect of a drug on gene expression, each replicate is a different animal, a different patient, or a different cell culture flask treated with the same drug. Biological replicates capture the natural variation between individuals. This variation is what your statistical test needs to account for.

Use at least 3 biological replicates per condition. This is the absolute minimum. With fewer than 3, you cannot estimate variance reliably, and your statistical power will be very low. Many funding agencies and journals now expect at least 3 replicates. If your budget allows, 4 to 6 replicates per condition will substantially improve your ability to detect differentially expressed genes.

Technical replicates are generally not necessary for differential expression analysis. The money spent on technical replicates is almost always better spent on additional biological replicates.

Sample size considerations

The number of replicates you need depends on several factors.

Effect size. If the expression differences between your conditions are large, fewer replicates are sufficient. If you expect subtle changes, you need more replicates to detect them reliably.

Biological variability. Some systems are highly variable. Human patient samples vary far more than inbred mouse lines. Highly variable systems need more replicates.

Number of conditions. If you have many conditions, you may need to balance replicate number against the total cost. A study with 2 conditions and 6 replicates each is more powerful than a study with 6 conditions and 2 replicates each.

Power analysis tools like RNASeqPower in R can help you estimate the number of replicates needed for your specific scenario. If you have pilot data or published data from a similar experiment, use it to estimate the variance and expected effect sizes.

Batch effects

Batch effects are systematic technical differences between groups of samples that were processed at different times, by different people, or on different equipment. They are one of the biggest threats to RNA-seq data quality.

Common sources of batch effects include:

• RNA extraction date. Samples extracted on Monday may differ systematically from samples extracted on Friday, even if the biological conditions are the same.
• Library preparation batch. Reagent lots, incubation times, and technician handling all introduce variation.
• Sequencing lane or flowcell. Different lanes on the same flowcell can have slightly different error profiles. Different flowcells can have larger differences.
• Sample processing order. If all treatment samples are processed first and all control samples are processed second, any technical drift will look like a biological effect.

How to mitigate batch effects

Randomize. The most important step. Distribute your treatment and control samples randomly across extraction batches, library preparation batches, and sequencing lanes. Never process all samples from one condition together.

Record everything. Track which samples were processed together, by whom, on which date, and on which lane. This metadata is essential for identifying and correcting batch effects computationally.

Process samples together when possible. If your sample count is small enough, extract all RNA on the same day, prepare all libraries in one batch, and sequence them on the same flowcell. This eliminates batch effects entirely.

Use balanced designs. If you must process samples in batches, make sure each batch contains an equal number of samples from every condition.

Library preparation methods

Two main strategies exist for enriching mRNA from total RNA.

Poly-A selection uses oligo-dT beads to capture mRNA molecules by their poly-A tails. This is the most common method for standard gene expression studies. It selectively captures messenger RNA and excludes ribosomal RNA, transfer RNA, and most non-coding RNAs.

Ribosomal RNA depletion uses probes to remove ribosomal RNA from total RNA. Everything else remains, including mRNA, long non-coding RNAs, and other RNA species.

Choose poly-A selection when you are focused on protein-coding gene expression. It produces cleaner data with less ribosomal contamination. Choose rRNA depletion when you need to study non-coding RNAs, when your RNA is partially degraded, or when you are working with organisms that lack poly-A tails on their transcripts.

Sequencing depth recommendations

Sequencing depth is the total number of reads you generate per sample. More reads cost more money, but they give you better sensitivity to detect lowly expressed genes.

10 to 20 million reads per sample is sufficient for detecting moderate to large expression changes in well-expressed genes. This is a reasonable starting point for pilot experiments.

20 to 30 million reads per sample is the standard recommendation for differential expression analysis. This depth captures most expressed genes and provides good statistical power.

50 million reads or more per sample may be necessary for detecting rare transcripts, studying alternative splicing, or working with complex tissues that express many genes at low levels.

Beyond 30 million reads, you get diminishing returns for standard differential expression. The additional reads mostly add coverage to already well-covered genes. If you have a fixed sequencing budget, it is almost always better to add more biological replicates at moderate depth than to sequence fewer replicates at very high depth.

Multiplexing and barcoding

Modern sequencing runs produce far more reads than a single sample needs. Multiplexing allows you to sequence multiple samples on the same lane or flowcell.

Each sample receives a unique short DNA sequence called a barcode or index during library preparation. After sequencing, the reads are separated by barcode in a process called demultiplexing. This assigns each read back to its sample of origin.

Multiplexing reduces cost per sample. A single lane on an Illumina NovaSeq can produce 800 million to 1 billion reads. If you need 30 million reads per sample, you can multiplex roughly 25 to 30 samples per lane.

When multiplexing, make sure your barcodes are compatible and sufficiently different from each other. Most library preparation kits provide validated barcode sets. Your sequencing facility will have guidelines for barcode selection.

Common mistakes to avoid

Too few replicates. Two replicates per condition is not enough. You cannot estimate variance with only two observations. Budget for at least three, and aim for more.

Confounding batch with condition. If you process all treated samples in one batch and all controls in another, you cannot separate the treatment effect from the batch effect. Always randomize.

Ignoring RNA quality. Low-quality RNA produces biased results. Always check RNA integrity before proceeding with library preparation. Discard samples with low RIN scores.

Inconsistent protocols. Small differences in how samples are handled can introduce systematic bias. Use the same RNA extraction kit, the same library prep protocol, and the same sequencing platform for all samples in your study.

Sequencing too deeply instead of adding replicates. Statistical power for differential expression depends more on the number of replicates than on sequencing depth. Three replicates at 20 million reads each will outperform two replicates at 60 million reads each.

Not recording metadata. You will need to know which batch, lane, and date each sample was processed on. Record this information at the time of processing. Trying to reconstruct it later is painful and error-prone.

Next steps

With a solid experimental design in place, the next page covers nf-core Pipelines for processing your sequencing data.