siRNA, shRNA and ASO Design Rules That Actually Matter
10 min read · Updated August 5, 2026
You need to knock down a transcript. There are three established ways to do it, they fail differently, and the design rules for the most common one are twenty years old and still the right starting point.
This guide covers how to choose between an siRNA, a vector-expressed shRNA and an antisense gapmer; which sequence rules have held up; why the seed region is where off-target effects come from; and what none of these rules can tell you, which is the part that determines how you plan the experiment.
Three mechanisms, three failure modes
Picking the wrong modality wastes more time than a suboptimal sequence within the right one, so start here.
An siRNA is a short synthetic double-stranded RNA, conventionally 21 nucleotides per strand with two-nucleotide 3' overhangs. You transfect it, one strand is loaded into RISC as the guide, and the target mRNA is cleaved. It is fast, cheap and transient — knockdown peaks in a couple of days and fades. Best for a quick test in an easily transfected cell line.
An shRNA is the same silencing mechanism delivered genetically. A hairpin is expressed from a Pol III promoter such as U6 or H1, processed intracellularly, and fed into the same pathway. Because it is encoded in DNA it can be delivered by virus, selected for, and made stable or inducible. That is what you want for hard-to-transfect cells, long time courses, or in vivo work — at the cost of cloning and viral work up front.
An antisense oligonucleotide gapmer works by an entirely different route. It is a single-stranded, chemically modified DNA-like oligo that hybridises to the target RNA and recruits RNase H1, which cleaves the RNA strand of the resulting DNA:RNA duplex. Crucially, this does not need the RNAi machinery, which is largely cytoplasmic — so a gapmer can act on nuclear-retained transcripts, non-polyadenylated RNAs and intronic sequence that RNAi struggles to reach. It is the right choice for those targets and for splice-modulation strategies.
The siRNA rules that have held up
Two rule sets from 2004 remain the most-cited basis for rational siRNA design, and they encode complementary ideas.
The Reynolds criteria are an efficacy score built from eight sequence features, checked and summed. The main components are a GC content in roughly the 30 to 52% range, an A/U-rich 3' end on the sense strand, particular base preferences at specific positions, and the absence of long GC runs. A score of 6 or above is the conventional threshold for a candidate worth ordering.
The Ui-Tei rules are four criteria aimed at a different problem — making sure the correct strand becomes the guide. An A or U at the guide strand's 5' end, a G or C at its 3' end, an A/U-rich seed region, and no long GC stretch.
The reason those Ui-Tei rules look arbitrary until you know the mechanism: RISC preferentially loads whichever strand has the less thermodynamically stable 5' end. A duplex with an A/U-rich 5' end on the strand you want as the guide, and a G/C-rich 5' end on the other, biases loading in your favour. Get this backwards and RISC loads the passenger strand, which silences whatever that strand happens to match — you get knockdown of something, just not your gene. This is called the asymmetry rule and it is the single most mechanistically important thing in siRNA design.
The seed region, and where off-targets come from
The dominant off-target effect in RNAi is not partial matching across the whole guide. It is microRNA-like: positions 2 to 8 of the guide strand, the seed region, behave like a microRNA seed and can silence any transcript with a complementary site in its 3' UTR, with no requirement for similarity anywhere else in the guide.
That has an uncomfortable implication. A BLAST search for near-full-length matches to your siRNA will come back clean while your reagent silences dozens of unintended transcripts through seed matches alone. The specificity problem is concentrated in seven nucleotides, and full-length homology searching is close to blind to it.
The established proxy is thermodynamic. The more stable the seed duplex, the more effectively it drives this off-target silencing, so the melting temperature of the seed region predicts risk. The siDirect approach sets a threshold around 21.5 °C: below it, seed-driven off-targeting is substantially reduced. A seed Tm at or above that is a flag worth acting on, usually by picking a different candidate.
Be clear about what this is: a sequence-thermodynamic proxy, not a search. It says the seed is stable enough to cause trouble; it does not tell you how many transcripts actually carry a complementary site. Both checks are worth doing, and they catch different things.
Going from siRNA to an shRNA cassette
If your candidate works as an siRNA and you want it stably expressed, you rebuild the same sequence as a DNA cassette. The architecture is conventional and each part has a job.
Read in order: the sense strand, then a loop, then the antisense strand, then a Pol III terminator. The loop is what makes it a hairpin — TTCAAGAGA is the most widely used, and the specific choice is not usually critical. The terminator is a run of thymines, conventionally TTTTT, because Pol III terminates at a T-rich tract rather than at a polyadenylation signal. The whole thing is cloned downstream of a U6 or H1 promoter.
One consequence of Pol III worth knowing: it makes no distinction between your intended transcript and any premature termination, so an internal run of four or more Ts in your sense or antisense sequence can truncate the transcript. It is worth screening for.
SeqBench's siRNA / shRNA / ASO Designer scores candidates against both rule sets, flags seed Tm at the siDirect threshold, and assembles the full cassette — sense, loop, antisense, terminator — as DNA ready to clone, so the conversion is not a manual string-building exercise.
Gapmer design: the 5-10-5 architecture
An RNase H gapmer has a specific layout dictated by its mechanism, and the standard is described as 5-10-5. The central ten nucleotides are unmodified DNA, and they have to be: RNase H1 only cleaves the RNA strand of a DNA:RNA duplex, so it needs a genuine DNA gap to recognise. The five-nucleotide wings on either side carry sugar modifications — 2'-MOE or LNA — that raise binding affinity and resist nucleases. The whole oligo typically sits on a phosphorothioate backbone for stability and uptake.
The wing chemistry is a synthesis decision rather than a sequence one. What design has to get right is the sequence and the liabilities it carries, and there are four classic screens:
- G-quadruplex motifs. Runs of guanines can form four-stranded structures that make the oligo behave unpredictably and cause non-specific protein binding. Avoid the motif rather than trying to work around it.
- CpG dinucleotides. Unmethylated CpG in a phosphorothioate oligo is recognised by innate immune sensing and can trigger an inflammatory response, confounding any phenotype you were trying to measure.
- Self-complementarity. An oligo that hybridises to itself is not available to hybridise to your target.
- GC extremes. Too low and binding is too weak; too high and you get non-specific binding and aggregation.
What none of these rules tell you
Everything above is sequence and thermodynamics. Two things that matter a great deal are outside that scope, and knowing which is which is what stops you over-trusting a score.
The first is target accessibility. An mRNA is a folded structure, and a site buried in a stable stem or occluded by a bound protein can be a perfect match on paper and inaccessible in a cell. Rule-based scores do not model folding. This is a real part of why a well-scored candidate sometimes does nothing, and it is not something a higher Reynolds score can rescue.
The second is transcriptome-wide specificity. A seed Tm flag is a proxy for off-target risk, not a count of affected transcripts. Actually knowing what else your reagent hits requires searching the transcriptome, which is a separate step.
The practical conclusion is about experimental design rather than sequence design. Rational design shifts the odds; it does not identify the winner. Order three or four candidates against different regions of the transcript rather than one highly-scored one, because the failure modes above are not predictable from sequence and testing is cheaper than iterating. Then confirm knockdown at the protein level, not just the transcript level — RNase H cleavage and RISC cleavage both degrade mRNA, but residual protein with a long half-life can leave you with a phenotype-free 90% knockdown. And include the controls that make the result interpretable: a non-targeting reagent matched for chemistry, and ideally rescue by re-expressing the target, which is the cleanest available evidence that the phenotype came from your intended gene rather than a seed-driven off-target.
Frequently asked questions
What is the difference between siRNA, shRNA and an ASO?
An siRNA is a synthetic double-stranded RNA you transfect for fast, transient knockdown through RISC. An shRNA is the same silencing pathway delivered genetically from a Pol III promoter, so it can be virally delivered, selected and made stable or inducible. An ASO gapmer is a single-stranded modified oligo that recruits RNase H1 to cleave the target — a different mechanism that does not need the RNAi machinery, so it reaches nuclear and non-polyadenylated transcripts that RNAi struggles with.
What are the Reynolds rules for siRNA design?
An eight-criterion efficacy score from Reynolds et al. 2004, summing sequence features: GC content in roughly the 30–52% range, an A/U-rich 3' end on the sense strand, specific base preferences at certain positions, and no long GC runs. A score of 6 or above is the conventional threshold for a candidate worth ordering.
Why does the guide strand need an A or U at its 5' end?
Because RISC preferentially loads whichever strand has the less thermodynamically stable 5' end. Making the intended guide's 5' end A/U-rich and the passenger's G/C-rich biases loading correctly. Get it backwards and RISC loads the passenger strand, which silences whatever that strand matches — you see knockdown of something other than your gene. This asymmetry rule is the mechanistic core of the Ui-Tei criteria.
What is the siRNA seed region and why does it cause off-target effects?
Positions 2 to 8 of the guide strand. They act like a microRNA seed and can silence any transcript with a complementary site in its 3' UTR, with no similarity needed anywhere else in the guide. That is why a clean BLAST search for full-length matches does not mean a specific reagent — the risk is concentrated in seven nucleotides that full-length homology searching largely misses.
What does a seed Tm of 21.5 °C mean?
It is the siDirect threshold for seed-driven off-target risk: the melting temperature of the duplex formed by guide positions 2–8, where a more stable seed drives microRNA-like silencing more effectively. At or above roughly 21.5 °C a candidate is flagged as higher risk. It is a thermodynamic proxy for how much trouble the seed can cause, not a count of how many transcripts actually carry a matching site.
What goes into an shRNA cassette?
In order: the sense strand, a loop (TTCAAGAGA is the most widely used), the antisense strand, and a Pol III terminator, conventionally TTTTT — Pol III terminates at a T-rich tract rather than a polyadenylation signal. The assembled DNA is cloned downstream of a U6 or H1 promoter. Watch for internal runs of four or more Ts in your own sequence, since Pol III can terminate there prematurely.
Why is an ASO gapmer built as 5-10-5?
Because the mechanism requires it. RNase H1 only cleaves the RNA strand of a DNA:RNA duplex, so the oligo needs a genuine central gap of about ten unmodified DNA nucleotides for the enzyme to recognise. The five-nucleotide wings carry sugar modifications (2'-MOE or LNA) that raise affinity and resist nucleases, usually on a phosphorothioate backbone.
Why did my well-scored siRNA fail to knock anything down?
The most common reason is target accessibility, which sequence rules do not model: an mRNA is folded, and a site buried in a stable stem or occluded by a bound protein can be a perfect match on paper and unreachable in a cell. This is why rational design is for shifting the odds rather than picking a winner — order three or four candidates against different regions of the transcript, and confirm knockdown at the protein level, since a long-half-life protein can survive a 90% mRNA reduction with no phenotype.
Related references
Related tools
Design knockdown reagents against an mRNA — Reynolds/Ui-Tei-scored siRNAs with ready shRNA cassettes, or 5-10-5 ASO gapmers screened for known liabilities.
Fold an RNA to its minimum-free-energy secondary structure — dot-bracket, MFE, base-pair list and an exportable arc diagram, without installing ViennaRNA.
Nearest-neighbor Tm, ΔG, hairpins and primer dimers for any oligo or primer pair.
Estimate primer Tm, GC% and molecular weight from a sequence.