Low-input RNA and amplification bias

A biopsy core, a sorted population, or a laser-captured patch can yield less RNA than a standard library method was built to accept. You can still make a library. The extra copying that makes the library visible is also what warps the proportions. This is a selection guide for that choice: when a low-input class is honest, and when it would confound the comparison you want. The wider path is from cells to a gene expression result.

Small RNA amounts force a class choice

Standard bulk RNA-seq methods publish an input window. Inside that window, the library is mostly a sampling of the RNA you put in, plus the ordinary biases of fragmentation, priming and a modest number of PCR cycles. Below the window, there is not enough material for those steps to behave as the method was qualified. A low-input kit, a preamplification step, or a deliberate increase in PCR cycles is a different method. It is chosen because the sample cannot meet the standard window, not because it is a more sensitive version of the same measurement.

The selection question is therefore practical. If you can obtain more tissue or more cells without changing the biology, a standard-input method is usually the cleaner comparison. If the biology is the small sample, a low-input class is the right tool and its bias has to be shared by every group in the contrast. Follow the input range on the method you actually use. Loading below the printed minimum is not a braver use of the same kit. It is an uncharacterised reaction.

Single-cell protocols are a further class, with empty droplets and ambient RNA as their own problems. This page stays with low-input bulk RNA, where many molecules from a small sample are still amplified together.

Extra copying is not neutral

Two amplification habits show up in low-input work. One is preamplification of cDNA before the library is finished. The other is extra PCR cycles after adapters are on, because the yield would otherwise be too low to sequence. Both increase the number of copies of whatever molecule happened to be primed early.

PCR is not a perfect photocopier. Sequences with awkward GC content, strong structure, or unlucky lengths amplify less well. Early in the reaction, when few molecules are present, chance decides which molecules get a head start. Extra cycles lock that head start in. The sequenced file then contains many reads from a smaller number of original molecules. Duplicate rate rises. Effective depth, meaning distinct molecules, falls even if the FASTQ file looks large. Proportions shift toward whatever amplified readily. A gene can look induced because it was copied efficiently, or look absent because it dropped out of a thin sample.

That distortion is the bias profile. It is reproducible enough to be confounded with biology when one group is amplified more than another, and noisy enough to create dropout differences between replicates of the same group. Public method write-ups on protocols.io show how differently laboratories structure these steps. The structure you used is the one that has to be written next to the counts.

UMIs count molecules, kits carry a bias profile

A unique molecular identifier is a short random sequence attached to a molecule before amplification. Copies of that molecule inherit the same tag. In analysis, reads that share a mapping position and a tag collapse to one count. Duplicates stop pretending to be extra depth. This is why low-input and single-cell methods so often include tags.

The tag does not make the kit neutral. Molecules that were never captured never receive a tag. Bias during reverse transcription, before the tag is added, remains. A tag that is too short will be shared by two independent molecules of a highly expressed gene, and the collapse will under-count that gene. Tags added after PCR label copies, not molecules, and collapsing them is theatre.

A low-input kit is a bias profile, not a free lunch. It buys access to a sample you could not otherwise sequence. In exchange you accept more dropout, a higher duplicate burden, and a composition that will not match a high-input library made from the same RNA. Unique molecular identifiers are part of how you count that library honestly. They are not a claim that the proportions equal a standard prep.

Reverse transcription and PCR enzymes are reagent classes in the molecular biology catalogue. The cycle range, the tag length and the input window belong to the method insert, not to a number copied from a paper that used a different kit.

Matching the method to the input you have

Measure the RNA, or count the cells, before you choose. A spectrophotometric number from a dirty prep is a weak guide. Integrity matters as much as mass. Degraded RNA plus heavy amplification is a particularly distorted profile, because the fragments that survive are then copied unevenly. How to judge whether the preparation is still worth amplifying is part of protecting RNA during extraction.

Then choose.

If the measured amount sits inside the standard window, use the standard method. Extra cycles "to be safe" only add duplicates.

If the amount sits inside a low-input window, use that method for every sample in the contrast, including controls. Prefer a tag chemistry if the question depends on molecule counts rather than on raw read depth. Record the cycle number actually used. A sample that needed four more cycles than its pair is not the same library class in practice, even if the kit name matches.

If the amount sits below every printed window you are willing to use, stop. Pooling biological units to scrape a minimum destroys the replicate structure. Switching only the small samples to a different kit destroys the contrast. The honest branch is a different experiment: more material, or a targeted assay of the few transcripts you already care about.

Targeted amplification, including a focused RT-qPCR panel, is often the better selection when the gene list is short and the RNA is scarce but the question is not a transcriptome-wide discovery. The cycle logic of PCR itself is in how polymerase chain reaction works. A targeted assay has its own bias. It does not pretend to represent every gene.

When the duplicate rate or the profile argues back

After sequencing, look at duplication and at the fraction of reads that collapse to a tag, if tags were used. A file with a very high duplicate rate and a small number of unique molecules is a low-information library, however large the read count. Do not send it into a differential test as if depth were equal to a standard sample.

If one group shows systematically higher duplication because those samples had less RNA, the design is confounded. The branch is to rebuild with matched inputs, or to stop claiming a treatment effect. Analysis cannot equalise a bias that was applied to only one side.

If the same RNA, split and prepared once with a standard method and once with a low-input method, produces different rankings, believe that the methods differ. That split is a useful bridge. It is not a nuisance to be averaged away. Studies deposited in the Sequence Read Archive are hard to reuse when this bridge, and the cycle count, were never recorded.

What each class preserves

Distortions that imitate biology

Dropout looks like biological absence. A gene at zero in two low-input replicates and present in a third may be a sampling miss, not a switch. Treating those zeros as confident absences overstates the data. High expression of a fragment that amplifies well can also pull apparent share away from other genes, a composition effect that extra cycles exaggerate.

Batch the amplification. If all treated samples saw two extra cycles because they were amplified on a different day, the calendar is the contrast. Interleave groups on the cycler. A short-read overview from Illumina explains the sequencing end of the workflow. It does not set your cycle number.

Limits of a research library

Low-input RNA work is a research method. It is not a diagnostic measurement of a scarce clinical specimen, and it does not set containment. Small samples from human or infected sources still carry the biosafety decision your institution makes. Amplification enzymes and cleanup reagents need the chemical handling those products specify. A molecule count is not a concentration in the cell unless a calibration was built to earn that unit.

Heat, small volumes and a cycler that restarted

Low-input reactions are often small volumes. In a warm room, a brief delay with the lid loose changes the concentration by evaporation more than the same delay would change a large prep. Keep the vessels closed, and follow the storage conditions on the enzyme you actually have rather than a room-temperature assumption printed for a cooler building.

A power cut mid-amplification is a failed cycle, not a pause you can ignore. Restarting the block does not tell you how many effective cycles the enzyme completed. Those samples need a new identifier if you repeat the amplification, so the extra copying is visible in the metadata. Do not pool a restarted tube with a completed one and call them technical replicates.

Asking for a low-input discussion

The useful brief states the largest and smallest RNA amounts you actually measured, the organism, whether tags are required, and whether every group can be prepared with one method. Library class can be discussed from the mRNA sequencing enquiry reference. Whether the resulting counts can support a contrast can be discussed from the differential expression analysis enquiry reference. Use those pages as the scientific prompt. Say what the sample is, and ask which input window a method would require.

The surrounding workflow sits on the nucleic acid analysis pathway. Put the measured inputs and the comparison in the quote request.