My knockdown and knockout phenotypes disagree

Knockdown and knockout phenotypes disagree for five recurring reasons: genetic compensation, off-target silencing, acute versus chronic loss, residual protein in either reagent, and selection in a clonal line. The siRNA gives a strong phenotype. The knockout line gives nothing, or something different. This is common enough to have a literature, and the explanation matters because each possibility licenses a different claim.

The unhelpful instinct is to trust whichever result supports the hypothesis. The useful move is to work out which mechanism produced the discrepancy, because that is usually answerable with experiments you can do quickly — paralogue expression, a protein-level blot, a second reagent — and because a reviewer who reads a paper reporting only one of the two results will ask for the other.

How often knockdown and knockout disagree

Knockdown and knockout phenotypes disagree often enough that a mismatch is the expected outcome, not an anomaly. The clearest systematic evidence comes from zebrafish reverse genetics: Kok and colleagues compared mutant lines against published morpholino phenotypes for the same genes and reported poor correlation between the two, with most morphant defects not reproduced in mutant embryos (Developmental Cell 2015;32:97–108). Their conclusion was procedural: mutant phenotypes should be the standard metric of gene function, and morpholinos used for ancillary analysis once they reproduce it.

That is a zebrafish figure and does not transfer wholesale to mammalian cell lines, where the systematic comparisons that exist are screen-level rather than phenotype-level: Morgens and colleagues compared shRNA and CRISPR/Cas9 libraries in K562 cells and found that the two detected essential genes with similar precision but that the results showed little correlation with each other (Nature Biotechnology 2016;34:634–636). What does transfer is the base-rate intuition: a single knockdown phenotype that no genetic null reproduces is a common state of the literature rather than a rare accident, and a manuscript that reports one reagent and treats agreement as assumed is making a claim the field has already tested and found unreliable.

The practical consequence is that the discrepancy is not a reason to abandon the experiment. It is a phenotype in its own right with a small number of testable causes, and the sections below run through them in the order they are worth testing. See also reviewer wants orthogonal validation, which is the same requirement stated from the other side of the desk.

Genetic compensation

Genetic compensation is the best-characterised explanation, and the one most often overlooked. Mutations that trigger nonsense-mediated decay of the transcript can upregulate related genes, producing a knockout animal that is functionally buffered while the knockdown — which lowers protein without generating a degraded transcript — is not.

The signature is a knockout with a milder phenotype than the knockdown, plus upregulation of sequence-related paralogues. Look for it directly: measure paralogue expression in the knockout. If they are elevated, you have your answer, and it is publishable in itself.

Note the direction. Compensation makes the knockout look weaker. If your knockout phenotype is stronger, this is not it.

The founding observation is Rossi and colleagues’ zebrafish egfl7 work (Nature 2015;524:230–233): egfl7 mutants showed no obvious phenotype while egfl7 morphants had severe vascular defects, and comparing mutant and morphant proteomes identified extracellular matrix genes upregulated only in mutants — genes that could themselves rescue the morphants. The same paper contains the cleanest control in the literature on this point: egfl7 CRISPR interference, which obstructs transcript elongation rather than producing a degraded mutant transcript, caused severe vascular defects without upregulating those genes.

The mechanism was named two ways in 2019. El-Brolosy and colleagues showed the response requires mutant mRNA degradation: alleles that fail to transcribe the mutated gene show no transcriptional adaptation and give rise to more severe phenotypes than alleles displaying mutant mRNA decay, and the genes upregulated are disproportionately those with sequence similarity to the mutant transcript (Nature 2019;568:193–197). Ma and colleagues showed in capn3a and nid1a zebrafish that the response depends on a premature termination codon and on transgene sequence homology, and requires Upf3a and COMPASS components including Wdr5, with increased H3K4me3 at the transcription start sites of the compensating genes (Nature 2019;568:259–263).

Two consequences follow for your design. First, the decisive genetic test is an allele that removes the locus or its promoter rather than one that installs a premature termination codon: no mutant transcript, no transcriptional adaptation, and the phenotype you see is closer to the true null. Second, the upregulated paralogue is a mechanistic result, and reporting it converts a failed replication into a finding — which answers the standard mechanism not shown objection in advance.

One edge case follows from the mechanism itself, and it is worth working out before you assume the response is exclusive to mutants. If the trigger is degradation of the mutant transcript rather than the lesion in the DNA — which is what the allele comparison in El-Brolosy and colleagues establishes — then a knockdown method that also destroys the transcript is not automatically exempt. A gapmer antisense oligonucleotide or an RNA-targeting nuclease degrades its target; a translation-blocking morpholino and a classical RISC-loaded siRNA differ in how much degraded transcript they leave behind. Where your knockdown chemistry destroys the message, measure paralogue expression in the knockdown as well as the knockout rather than assuming the knockdown is the uncompensated arm of the comparison.

Off-target effects of the knockdown

Off-target silencing is the other leading candidate, and the one reviewers assume first. siRNAs silence unintended transcripts through seed-region matches, and a phenotype absent in a clean genetic null is exactly what an off-target effect looks like.

Distinguishing it is straightforward: use two or three independent siRNAs targeting different regions. A phenotype reproduced by all of them is unlikely to be off-target. A phenotype from one reagent is a hypothesis about that reagent.

The decisive experiment is rescue — restore the target with a construct resistant to the siRNA and ask whether the phenotype reverses.

The molecular basis is specific enough to design against. Jackson and colleagues showed by expression profiling that siRNAs produce reagent-specific rather than target-specific signatures, silencing non-targeted transcripts sharing as few as eleven contiguous nucleotides of identity (Nature Biotechnology 2003;21:635–637). Birmingham and colleagues then located the determinant: off-targeting tracks with perfect 3′ untranslated region matches to the hexamer or heptamer seed of the antisense strand, positions 2–7 or 2–8, and not with overall identity (Nature Methods 2006;3:199–204). Two siRNAs that happen to share a seed are therefore not independent reagents, however different their full sequences look, and checking the seeds takes one alignment.

Two controls are stronger than the scrambled sequence most manuscripts show, which cannot separate a seed effect from a target effect because it abolishes both at once. The C911 control replaces bases 9 through 11 of the siRNA with their complement, preserving the seed while destroying on-target activity, so a phenotype that survives the swap is seed-driven and a phenotype that disappears is on-target — a discrimination a scrambled sequence cannot make, because it changes the seed and the target together (PLoS ONE 2012;7:e51942). Complex pools are the other route: siPools of up to 60 accurately defined siRNAs dilute each individual sequence far enough that transcriptome-wide off-target signatures fall below detection, while single-siRNA transfections severely perturb global gene expression (Nucleic Acids Research 2014;42:8049–8061).

In zebrafish the equivalent discipline is written down. The community guidelines state that “MO should be validated by comparison to a mutant, and if there is a discrepancy, by injection into embryos homozygous for a null allele or an allele lacking the MO-binding site”, and warn that “extra caution should be exercised when one has to inject more than 5 ng of a MO to cause a phenotype” (PLoS Genetics 2017;13:e1007000). Dose matters because morpholino injection carries sequence-independent toxicity of its own: activation of p53 and the apoptotic programme downstream of it is a documented off-target consequence, which is why a p53 knockdown control accompanies morphant work in this field (PLoS Genetics 2007;3:e78). A vascular or apoptotic phenotype at 10 ng is a dose report, not a gene function.

Acute versus chronic loss

A constitutive knockout has developed without the gene, and the organism or cell line has had every opportunity to adapt. An acute knockdown removes the protein from a system that was depending on it an hour ago.

Both are real biology and they answer different questions. Acute degradation systems — degron tags, auxin-inducible degradation — separate the two directly, and where the acute phenotype matches the knockdown, the discrepancy is adaptation rather than artefact.

Name the system and its kinetics in the methods. The auxin-inducible degron transplants the plant SCF–TIR1 pathway into non-plant cells and depletes a tagged protein rapidly and reversibly on auxin addition, in yeast and in chicken, mouse, hamster, monkey and human lines (Nature Methods 2009;6:917–922). AID2 replaces TIR1 with an OsTIR1(F74G) mutant paired with the bumped ligand 5-Ph-IAA, which removes the leaky basal degradation of the original system and works at a far lower ligand concentration, in cells and in mice (Nature Communications 2020;11:5701) — leak matters here because a leaky degron is a chronic partial knockdown wearing an acute label. The dTAG system pairs a degrader of FKBP12^F36V with in-frame knock-in of that tag at the endogenous locus, and was built explicitly to remove the delay between perturbation and measurement (Nature Chemical Biology 2018;14:431–441).

When tagging is unavailable — no validated knock-in, a functionally essential terminus, a protein with a half-life of days — the substitute is a time course rather than a single endpoint. An inducible shRNA, a Cas13d array or a gapmer antisense oligonucleotide sampled at several timepoints separates the first-order consequence from the adapted steady state, and a phenotype that appears at 24 hours and resolves by 96 hours is itself the answer. Acute rescue tests the same axis from the other direction: restore the protein in the established knockout line and ask whether the state reverses. If the chronic line has rewired, it will not, and that non-reversal is evidence for adaptation rather than a failed rescue.

Incomplete knockdown, and the residual

A knockdown removing 70% of a protein leaves 30%, which for many proteins is plenty. This produces the opposite pattern — knockout stronger than knockdown — and it is the easiest to check.

Always report knockdown efficiency at the protein level. Transcript knockdown is not protein knockdown, and a reviewer who sees only qPCR will ask.

Report it as a number with its method, not as a representative blot: percentage of control, mean and dispersion across the same biological replicates that produced the phenotype, the loading control, the antibody catalogue number and lot, and the exposure or detection range showing the signal is unsaturated. Densitometry from a single saturated film is the most common quantification defect in this class of figure — see image quantification reporting. A knockdown described as “greater than 80%” with no denominator and no replicate count is not a measurement.

Dose–response is the strongest form of the argument. If phenotype severity tracks residual protein across a series of siRNA concentrations or induction levels, the relationship is quantitative and the knockout sits at one end of it. If a 95% knockdown gives a strong phenotype and a complete null gives none, residual protein is not the explanation and you are back in compensation territory.

When the knockout is not a null

A CRISPR knockout line frequently retains functional protein, and the sequencing chromatogram will not tell you. Smits and colleagues combined RNA sequencing with mass spectrometry across a large panel of genetically verified deletion lines in HAP1 cells and observed residual protein expression for roughly a third of the quantified targets, at levels ranging from faint to indistinguishable from wild type (Nature Methods 2019;16:1087–1093). Two mechanisms accounted for it: translation reinitiation downstream of the lesion, giving an N-terminally truncated protein, and skipping of the edited exon, giving an isoform with an internal deletion. For BRD4, DNMT1 and NGLY1 the truncated products retained partial function.

Tuladhar and colleagues found the same class of failure at a similar rate by a different route, detecting foreign mRNA or protein products in approximately 50% of a panel of presumed knockout lines, arising from internal ribosome entry, conversion of alternatively spliced premature-termination-codon transcripts into coding molecules, and exon skipping caused by disruption of exon splicing enhancers (Nature Communications 2019;10:4056).

The detection consequence is concrete and frequently missed. An antibody raised against an N-terminal epitope reports “no protein” for a line expressing a reinitiated, N-terminally truncated product that still works. Blot with antibodies against two epitopes at opposite ends of the coding sequence, or run targeted mass spectrometry, and sequence the mRNA rather than only the genomic locus so that exon skipping is visible. A knockout validated by Sanger sequencing of genomic DNA plus one blot is validated against the wrong failure mode, which is also why a reviewer asking for more controls on a knockout line is usually asking a specific question rather than a ritual one.

Guide design follows from the same evidence. Target an early constitutive exon shared by all annotated transcripts, avoid placing the cut immediately upstream of a downstream in-frame ATG, and prefer a lesion that disrupts a functional domain over one that merely shifts the frame. Where the gene is small or the domain structure is unhelpful, a whole-exon or whole-locus deletion is cleaner than an indel, and has the second advantage of removing transcriptional adaptation from the picture.

Selection in the knockout line

Clonal knockout lines are selected. Cells that tolerate losing the gene are the ones that survived to become the line, which can select for a compensated or otherwise unusual state. Multiple independent clones address this; a single clone does not.

Three or more independently derived clones, ideally from at least two guide RNAs targeting different exons, separate gene loss from clone identity. Report each clone’s phenotype individually rather than pooling them into one bar, because between-clone variance is the quantity in question; treating wells from a single clone as independent replicates of gene loss is pseudoreplication, and reviewers name it as such. Where clonal derivation is impossible, a polyclonal edited population assayed early, before drift, carries less selection history than a line expanded for twenty passages.

Selection acts in whole animals too. A constitutive knockout with partial lethality leaves you analysing the survivors, and maternally deposited transcript or protein can mask a zygotic phenotype entirely in early development, which is why maternal-zygotic mutants exist as a category. Conditional and inducible alleles avoid both problems by removing the gene after the selective window has passed. For animal work, state the allele, the background and its generation of backcrossing, the sex distribution, and the littermate control structure — the ARRIVE guidelines items apply directly here, and in vivo rigor reporting is where this section becomes a methods paragraph.

Which pattern points at which explanation

Match the direction and shape of the discrepancy to a diagnostic before running anything expensive.

The table is a triage order, not a decision procedure. Two mechanisms frequently operate at once — a partially functional truncated protein in a line that has also upregulated a paralogue is a common combination — and the paralogue panel and the two-epitope blot are cheap enough to run together.

What reviewers say when the discrepancy is mishandled

Reviewers rarely write “your knockdown and knockout disagree”. They write the specific version, and these comments decide papers.

“The phenotype rests on a single siRNA; please include at least two independent sequences and a rescue with a silencing-resistant construct.” “Knockdown efficiency is shown by qPCR only; please quantify protein.” “The knockout line is a single clone; please confirm the phenotype in independent clones.” “Please verify that the knockout allele is a protein null rather than an in-frame deletion or a reinitiation product.” “The authors do not report whether paralogues are upregulated in the mutant.” “The morpholino dose exceeds the community guideline and no dose–response is shown.” “Figure 3 uses the knockdown and Figure 5 uses the knockout; the two are treated interchangeably in the discussion.” “The abstract states the gene is required for the process, but only the acute reagent produces the phenotype.”

The last two are the expensive ones, because they are about the claim rather than the experiment. A manuscript that switches reagent between figures without saying so, then generalises across both in the discussion, invites an overclaim objection that no additional experiment repairs — and a perturbation not reproduced by an independent method will not support the causal statement it is being asked to carry, which is the substance of correlative, not causal.

What to do

Run the cheap diagnostics first: paralogue expression in the knockout, protein-level knockdown efficiency, and a second independent siRNA. Between them they resolve most cases in a fortnight.

Then report the discrepancy rather than hiding it. A paper that shows both results and explains the difference is stronger than one that shows the convenient result — and a reviewer who discovers an unreported knockout that disagreed will treat everything else in the paper differently.

Sequence the remaining work by effort. The paralogue panel is a qPCR plate against the two or three closest sequence relatives of the target transcript. The second siRNA is a reagent order and one transfection. The two-epitope blot is one gel. Only then come the expensive moves: a promoter-deletion or whole-locus allele to remove transcriptional adaptation from the picture, a degron knock-in to separate acute from chronic loss, and rescue with a silencing-resistant construct, which answers the most objections at once because it re-establishes the causal link a failed replication broke. Where a phenotype survives none of these, the honest report is a negative result, and an experiment that did not replicate is a manuscript-shaping fact rather than a private one.

How to report the discrepancy

Report the discrepancy by naming the two reagents, the direction of the difference, and the mechanism you tested, in that order, in the results rather than the limitations.

State the reagents fully: siRNA or shRNA sequences with supplier catalogue numbers, morpholino sequence and dose in nanograms, guide RNA sequences with the targeted exon, and the exact allele of every mutant line including the indel and its predicted protein consequence. Give knockdown efficiency at the protein level with its denominator and the number of independent experiments. Show both phenotypes at the same scale in the same figure rather than putting the inconvenient one in a supplement.

Then state which mechanism you tested and what you found, including where the test was negative — “paralogue expression was unchanged in two independent mutant clones” is a result and belongs in the paper. Scope the conclusion to the reagent that supports it: an acute requirement demonstrated by degron depletion is an acute requirement, and extending it to a developmental requirement needs the constitutive allele. Where the mechanism remains unresolved, a bounded statement belongs in the limitations section — name the candidate explanations you could not exclude and the experiment that would exclude them, rather than writing that results should be interpreted with caution. If reviewers raised the discrepancy, the response letter should carry the same three-part structure: reagents, direction, mechanism tested.

Before submission

Reviewer wants more controls covers the specificity controls this situation demands, and reviewer says the results are correlative, not causal covers what a perturbation has to establish before a causal claim rests on it. If your manuscript reports one reagent and one clone, both objections are coming.

Related symptom: my replicates disagree. For preclinical manuscripts as a whole, preclinical review covers the reagent-validation and rigor items that reviewers of this literature apply as a set.

PerfectPaper reads the knockdown and the knockout figures against each other, flags a phenotype supported by a single reagent or a single clone, and checks that the claim in the discussion is scoped to the perturbation that produced it.

Review my manuscript

Frequently asked questions

Why do my knockdown and knockout phenotypes differ?

Five explanations account for most cases: transcriptional adaptation in the mutant, seed-mediated off-target silencing by the knockdown reagent, the difference between acute and chronic loss, residual protein in either the knockdown or a non-null knockout allele, and selection in a clonal line. Direction narrows the list immediately, because compensation and non-null alleles make the knockout milder while incomplete knockdown makes it stronger.

Why does my CRISPR knockout have no phenotype when the siRNA did?

Two mechanisms dominate. The mutant transcript may be triggering transcriptional adaptation, upregulating sequence-related paralogues that buffer the loss, which is testable by qPCR of those paralogues. Or the line may not be a protein null: residual protein was detected for roughly a third of the quantified targets in a large panel of verified HAP1 deletion lines, through translation reinitiation or exon skipping (Nature Methods 2019;16:1087–1093). Blotting two epitopes at opposite ends of the coding sequence distinguishes the two cases.

Is genetic compensation why my knockout is milder than my knockdown?

Genetic compensation is the first thing to test when the direction runs that way. Loss-of-function mutations that produce a degraded transcript can trigger upregulation of sequence-related genes, buffering the phenotype in the mutant while leaving a transient knockdown unbuffered. Measure paralogue expression in the knockout: if the closest sequence relatives are elevated, you have the explanation and a mechanistic result worth reporting.

Which result should I trust, my knockdown or my knockout?

Trust neither by default. The two reagents answer different questions — acute versus chronic loss — and the disagreement between them is information rather than noise. Work out which mechanism produced it before deciding which result supports your claim.

How many siRNAs do I need to rule out off-target effects?

Two or three independent sequences targeting different regions, ideally with a rescue using a resistant construct. A phenotype from a single siRNA is not separable from that reagent’s off-target profile. Check that the sequences do not share a seed, since off-targeting tracks with 3′ UTR matches to positions 2–7 or 2–8 of the antisense strand (Nature Methods 2006;3:199–204).

Should I report a knockout that contradicts my knockdown?

Yes. Reporting both and explaining the discrepancy is stronger science and safer publishing. Unreported contradictory data found later is a serious problem.

Does a degron system solve my knockdown-versus-knockout discrepancy?

A degron solves half of it. Acute depletion separates acute from chronic loss cleanly, which resolves the most common ambiguity, but it says nothing about off-target silencing by an siRNA — that still needs independent reagents with non-overlapping seeds and a rescue. Choose a low-leak system such as AID2, because a leaky degron is a chronic partial knockdown presented as an acute one.