Before you order
How many sequences to order against one target
An antisense oligonucleotide binds one window of one transcript. Whether it silences the target depends on which window, and that cannot be settled in advance, so a working sequence is found by testing more than one.
This page has three parts and nothing else: arithmetic you can follow and set yourself, what the published studies on this site ordered, and what the literature says about why the arithmetic is a ceiling rather than a forecast.
The arithmetic
What ordering more than one does to the odds
Set two numbers. The first is how many sequences you order against the target. The second is the chance that any one of them reaches the knockdown you would count as a result, which depends on your target, your cells, and where you set that bar.
Each button is a count from one published screen. Move the control anywhere between them: the number is yours to choose, and it depends on your target, your cells and what you will count as working.
- At least one of the 3 works
- 58%
- 1 - 0.75^3 = 0.578
- At least two of the 3 work
- 16%
- 1 - 0.75^3 - 3 x 0.25 x 0.75^2 = 0.156
1 of 4 in that study: 80% knockdown of the microRNA at 1 μM in an acute brain slice. Reported by Baby et al., Aging Cell, 2020. It is that study's figure, for that target and that bar, and not a rate AUM claims for your target.
Both numbers assume the sequences succeed or fail independently of one another. They do not. Candidates against one transcript share its folding and the regions of it a protein can reach, so a poor target tends to defeat several of them together, and the true chance is lower than the arithmetic above. Read the number as a ceiling, not as a forecast.
| Sequences ordered | At least one works | At least two work |
|---|---|---|
| 1 | 25% | 0% |
| 2 | 44% | 6% |
| 3 | 58% | 16% |
| 4 | 68% | 26% |
| 5 | 76% | 37% |
| 6 | 82% | 47% |
The published record
What the published studies ordered
The site holds 71 publication records, each read from the paper itself and checked by a second reader. 4 are not counted here: 3 used no AUM product, and 1 used an oligonucleotide that copies part of the transcript and binds a protein, rather than binding the transcript as an antisense sequence does. Of the other 67: 34 state that more than one antisense sequence was made or used against one target, 25 name a single sequence, and 8 do not say.
| Sequences against one target | Studies |
|---|---|
| 2 | 11 |
| 3 | 4 |
| 4 | 10 |
| 5 | 4 |
| 6 | 1 |
| 8 | 1 |
| 19 | 1 |
One bar per number of sequences, its length the number of studies. 2 of the 34 give a range rather than one number and are deliberately not in the figure, because rounding a range into a bar would invent a number: Hunter et al., Frontiers in Agronomy, 2021 used about four per target, up to eight, and Marasca et al., Nature Genetics, 2022 used two to five, by target.
What came of the screen, where the study says so
8 of those studies say how many of their candidates worked. Each sentence below is this site's own summary of that study, and each links to the record, where the paper's own words and the figure it comes from are printed.
- Baby et al., Aging Cell, 20204 sequences
- One AUMantagomir sdASO design out of four reached 80% knockdown of miR-134-5p at 1 μM, bath applied to an acute brain slice with no transfection reagent of any kind.
- Pelisch et al., eNeuro, 20214 sequences
- Screening the four in lipopolysaccharide-stimulated primary macrophages found one that cut CCL3 messenger RNA by about 80% at 10 μM with no transfection reagent, and the other three did not move it, so that one carried the rest of the study.
- Margiotta et al., Frontiers in Physiology, 20203 per target, 6 in all
- Two of six AUMsilence sdASOs lowered chicken CRY1 and CRY2 transcripts by 87% and 91% in pectoral myotube cultures, with no effect on the related CRY4 transcript.
- Zhang et al., Science, 20235 sequences
- Five sequences gave five different depths of knockdown, and the loss of neuronal viability tracked the depth, which is a dose-response argument made with sequences rather than with concentration.
- Xu et al., Acta Neuropathologica, 20246 sequences
- Six independent self-delivering AUMsilence sdASOs against MSUT2 each cut MSUT2 protein in wild-type mouse primary neurons, so the effect was reproduced across sequences rather than resting on one.
- Page et al., Journal of Neuroimmunology, 20238 sequences
- Eight AUMsilence sdASOs were designed against MK2 and screened in stimulated mouse macrophages. Three of the eight cut the messenger RNA by 70-80% and five did nothing, which is the case for screening several sequences rather than trusting one.
- Akimova et al., Frontiers in Immunology, 202419 sequences
- Candidate oligonucleotides passing the three screening criteria: 11 of 19.
- Hunter et al., Frontiers in Agronomy, 2021about four per target, up to eight
- Sequence choice decided everything, and the paper reports the failures beside the successes: two of four sequences worked against one bacterial gene, one of four against another, and two of five against the insect gene, with the authors attributing the misses to target RNA folding.
Every study that used more than one
Read those counts as a floor rather than an estimate. A paper prints the sequence that worked, because that is what a paper is for, so a study that names one sequence may have screened several and published the one that silenced its target. The studies above that name the candidates which did nothing are the exception, and they are the most useful rows on the page.
The assumption
Candidates against one transcript are not independent
The arithmetic multiplies the chances as though each sequence were its own experiment. It is not. Candidates against one transcript share the same folding, the same regions a protein can reach, and the same expression in your cells, so a difficult target tends to defeat several of them together. That is why the number above is a ceiling.
The clearest published statement of it measured the same design method against four different messenger RNAs and found the rate of effective sequences differed by roughly ten-fold between them, at one stated threshold.
We observed that, even after all corrections, the hit rate varies significantly between different target mRNAs, indicating they are primarily driven by mRNA-specific features.
hit rate (SNCA: 38%; APP: 36%; MAPT: 10%; BACE1: 4%, Fig. 9)
Those percentages are counted at that study's own bar, which it states as siRNA hits (i.e. ≤ 35% target mRNA expression)
(Results, the threshold the hit rates above are counted at), in its own cells and chemistry. They are not a rate for your target and this page does not present them as one.
What drives it is where on the transcript a sequence binds. A study of 188,521 gapmer antisense sequences drawn from published patent tables found accessibility to be a central determinant of how well they work.
ASOs targeting unpaired regions showed 8.5% higher knockdown efficacy compared to those targeting base-paired regions, indicating that target accessibility significantly influences ASO performance
The four-gene siRNA study above reports something that follows from it, in that modality rather than this one: the sequences that worked sat in clusters along the transcript rather than scattered evenly.
clusters of effective siRNAs within 50 nts of one another
Clustering cuts both ways. It is why several candidates can fail together, and it is why a candidate near one that worked is worth trying.
Choosing the candidates
Where the candidates come from changes the odds
The published hit rates that sound alarming come from screens that tile a transcript without choosing between the windows. One such screen made 278 candidates against one gene and 234 against another, every 20-mer along the pre-messenger RNA, and counted how many reached a 40% fall in the target.
We identified 8 ASOs (out of 278; 2.8% hit rate) targeting APOE and 10 ASOs (out of 234; 4.2% hit rate) targeting TREM2 resulting in at least 40% reduction of the target RNA at the single concentration tested
How those candidates were made, and the conditions the screen was read at:
we designed 20-mer oligonucleotides tiling their respective pre-mRNA using a gliding window on their reference sequences without any restrictions.
Next, we treated THP-1 cells with a panel of APOE and TREM2 ASOs for 72 h at a single concentration of 5 µM to identify candidate lead ASOs
That is the scale clinical discovery works at, by design, and it is worth knowing before a research screen of three or five is judged against it.
For RNase H1-activating PS ASOs, a minimum of PS gapmer ASOs designed to bind to 400–500 sites in a target RNA are typically screened at a single dose level.
Choosing between the windows is what moves the odds. In one published comparison, 50 candidates against the same gene were tested side by side in the same cells: 18 taken from the top one percent of a scoring model, and 32 chosen at random from the whole library.
ASOs ranked in the top 1% by OligoAI achieved superior target knockdown efficacy (median = 81% [IQR: 64%–88%]) compared to randomly selected ASOs (median = 36% [IQR: 11%–62%])
Our analysis revealed that to match the median knockdown efficacy of our 18 top-scored ASOs, a standard random screening approach would require testing 103 ASOs (95% confidence interval, CI: 53–179).
That is a published model rather than ours, and the comparison is one gene in one cell line. It is on this page because it is the closest measurement anyone has published of the thing that matters here: scoring every window before you synthesize, rather than walking along the transcript. It is what the AUMsilence platform does with your target.
The four papers on this page were read through Europe PMC on 2026-09-15, and every quotation was matched against the full text that day.
What we recommend
More than one sequence against a target you have not silenced before
Three sequences
For AUMsilence
Five sequences
Five sequences are the choice for an experiment you cannot easily repeat: primary cells from one donor, an animal cohort, a transcript whose accessible regions are few. The order page calls them the first pass with the knockdown guarantee, on those same two products. Elsewhere the order page sells one to four against a target, which is the same argument one step short.
If none of the five reaches 70% knockdown of the target transcript under the protocol, AUM designs and supplies five more against the same target at no charge, once. The terms say how a claim is made. Terms and conditions
One sequence
One sequence is the right order when the sequence is already settled: one published against your target that you are reproducing, or a design you have used before. The order page calls it a design you already trust. It is not the place to start on a target nobody has silenced.
What to order alongside
A screen is only readable against its controls. Every product has its own, matched in chemistry and purification to the sequences you are ordering, and they are on the order page with their sizes and prices.
- A scramble control, every time: AUMscramble
sdASO for the self-delivering products, AUMscramble toASO for the transfection-optimized ones, AUMscramble siRNA for AUMsiRNA. It is the arm every knockdown is read against, and a screen without one measures the treatment rather than the target. - A positive control where the cells are new to you: AUMposctrl sdASO, or its toASO and siRNA equivalents. When nothing moves, it separates a target that resisted from an experiment that did not work.
- A fluorescently labeled scramble control, once per cell type. It shows uptake before the first knockdown experiment, which tells a delivery problem from a sequence problem.
For research use only. Not for use in diagnostic or therapeutic procedures.