# miRNA vs siRNA and RNA interference

> A miRNA is encoded in the genome and pairs partially; an siRNA arrives as double-stranded RNA and pairs completely. Both use the same Dicer and Argonaute machinery, and a perfectly matched siRNA still behaves like a miRNA on other mRNAs through its seed. Learn how the two differ, what duplex free energy can and cannot tell you, and which siRNA drugs are approved.

## How do miRNA and siRNA differ?

A microRNA (miRNA) is a short [non-coding RNA](/guides/how-many-types-of-rna-are-there) from the cell's own genome that pairs partly with many of its mRNAs and turns them down. A small interfering RNA (siRNA) comes from double-stranded RNA made by a virus, an experiment or a drug, and pairs completely with one mRNA, which is then cut. Both silence their targets through RNA interference, held inside the protein Argonaute.

The textbook split says miRNAs pair imperfectly and repress, while siRNAs pair perfectly and cut. Doench, Petersen and Sharp showed in 2003 that the split belongs to the target, not to the RNA. In mammalian cells, an siRNA given only partially matching sites in the 3' untranslated region (UTR) of a target mRNA repressed it without cutting it, the way an endogenous miRNA does. The same 21 nucleotides acted as either, depending on how many bases paired.

Origin and precursor separate the two RNAs on paper (Table 1). Inside the cell both use the same enzymes and the same Argonaute protein, and the extent of pairing decides whether the target is cut or merely turned down. That one fact explains why an siRNA has miRNA-like off-targets, and why a free-energy calculation says something useful about both and still misses the point.

  <caption><strong>Table 1. miRNA and siRNA compared.</strong> Animal miRNAs are described; plant miRNAs pair almost perfectly and cleave their targets. Sources: Bartel (2009, 2018), Lee et al. (2003), Elbashir et al. (2001), Liu et al. (2004).</caption>
  <thead>
    <tr>
      <th scope="col">Feature</th>
      <th scope="col">miRNA</th>
      <th scope="col">siRNA</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Origin</td>
      <td>Genes in the cell's own genome</td>
      <td>Long double-stranded RNA from a virus or an experiment, or a synthetic duplex</td>
    </tr>
    <tr>
      <td>Precursor</td>
      <td>Hairpin of about 70 nucleotides (pre-miRNA), cut from a longer primary transcript</td>
      <td>Long double-stranded RNA, or a ready-made 21-nucleotide duplex</td>
    </tr>
    <tr>
      <td>Processing</td>
      <td>Drosha in the nucleus, then Dicer in the cytoplasm</td>
      <td>Dicer only, or none for a synthetic duplex</td>
    </tr>
    <tr>
      <td>Length</td>
      <td>About 22 nucleotides</td>
      <td>21 to 23 nucleotides</td>
    </tr>
    <tr>
      <td>Pairing with the target</td>
      <td>Partial in animals, anchored on the seed; near-perfect in plants</td>
      <td>Complete across the guide strand by design</td>
    </tr>
    <tr>
      <td>Targets</td>
      <td>Hundreds of mRNAs per miRNA</td>
      <td>One intended mRNA, plus off-targets through seed matches</td>
    </tr>
    <tr>
      <td>Effect on the target</td>
      <td>Translational repression and mRNA decay in animals; cleavage is unusual</td>
      <td>Cleavage by Argonaute2</td>
    </tr>
    <tr>
      <td>Typical use</td>
      <td>Natural gene regulation; a target for anti-miR drugs</td>
      <td>Gene knockdown in the lab; approved drugs</td>
    </tr>
  </tbody>

## Where RNA interference came from

Two lines of work converged. In 1993 the Ambros lab reported that the worm gene lin-4 encodes a small RNA complementary to sequences in the lin-14 mRNA, the first miRNA.

In 1998 Fire, Mello and colleagues injected double-stranded RNA into the nematode Caenorhabditis elegans and found it far more effective at silencing a gene than either strand alone. Only a few molecules per cell were needed, which suggested a catalytic or amplifying step rather than simple one-to-one binding. The pair received the 2006 Nobel Prize in Physiology or Medicine.

The lines met on size. Both small RNAs and the products of double-stranded RNA silencing turned out to be about 22 nucleotides long, and both were produced by the same family of enzyme, Dicer. In 2001 Elbashir and colleagues showed that synthetic 21-nucleotide duplexes silence genes in human embryonic kidney and HeLa cells, which turned RNA interference into a tool for mammalian cells.

## How miRNAs and siRNAs are made and loaded

A miRNA gene is transcribed as a long primary transcript containing a hairpin. In the nucleus the enzyme Drosha cuts out a stem-loop of about 70 nucleotides, the pre-miRNA, and in the cytoplasm Dicer cuts that into a duplex of about 22 nucleotides. An siRNA skips the first step. Dicer cuts a long double-stranded RNA into 21 and 22 nucleotide pieces, or the duplex is synthesized directly and delivered into the cell.

Either duplex is then handed to Argonaute, which keeps one strand, the guide, and discards the other, the passenger. Schwarz and colleagues showed that the choice follows base-pair stability: the stability of the pairs at each strand's 5' end decides which strand enters the silencing complex. That is the first place where duplex free energy decides a biological outcome, for miRNAs and siRNAs alike.

![Two lanes. In the miRNA lane a primary transcript with a hairpin is cut by Drosha in the nucleus into a pre-miRNA of about 70 nucleotides, Dicer cuts it into a duplex of about 22 nucleotides, and Argonaute keeps the guide strand. In the siRNA lane a long double-stranded RNA, or a synthetic 21 nucleotide duplex, gives a 21 to 22 nucleotide duplex that Argonaute loads the same way.](/images/guides/mirna-vs-sirna-and-rna-interference/biogenesis.webp "**Figure 1. Two routes to the same loaded Argonaute.** Hairpin and strand lengths are schematic. The miRNA route has an extra cut by Drosha in the nucleus; both end with Argonaute holding the guide and discarding the passenger.")

The loaded protein sets a limit on what happens next. In mammals only Argonaute2 cuts RNA, and cells lacking it cannot respond to an siRNA. Cutting also needs extensive pairing between guide and target, which a seed match alone does not provide.

## How a small RNA recognizes its target

In animals the guide recognizes its target mostly through the seed, nucleotides 2 to 7 or 2 to 8 counted from the 5' end. Requiring Watson-Crick pairing to the seed sharply cuts false positives in target prediction.

Canonical sites match the seed plus one or two neighboring positions, 7 or 8 nucleotides in all, while 6-nucleotide seed matches are marginal and usually less effective. Extra pairing near the 3' end of the guide can support a site but rarely rescues a seed mismatch; in the TargetScan analysis it accounted for under 2% of conserved sites.

Seed pairing is cheap and common, so a single miRNA has many targets. Friedman and colleagues estimated that more than 60% of human protein-coding genes have been under selective pressure to keep pairing to miRNAs, based on over 45,000 conserved sites in human 3' untranslated regions. Animal miRNAs most often act by translational repression, by destabilizing the mRNA (which shortens its [half-life](/guides/how-long-does-mrna-last-in-a-cell)), or both.

Extensive pairing changes the outcome. When the match runs through the center of the guide, Argonaute can adopt a conformation that positions its active site for cleavage. In plants that is the normal mode, so many plant miRNA targets can be found by searching for near-complementary messages. In animals it happens at a few endogenous sites and at the intended site of an siRNA.

## How many miRNAs miRBase counts

The miRNA registry miRBase counts 46,603 hairpin precursors and 74,463 mature miRNAs from 323 species in release 23.1, dated September 2026. The previous full release, version 22 of March 2018, held 38,589 hairpins and 48,860 mature miRNAs from 271 species. We counted the human entries in the release 23.1 FASTA files on November 11, 2026: 1,861 hairpins and 2,789 mature sequences with the hsa prefix.

Those totals count annotations, not validated functional miRNAs. When Kozomara and colleagues mapped 5.5 billion small-RNA sequencing reads to miRBase 22, the data supported between 20% and 65% of annotations in well-studied animal genomes, and more than 200 sequences were removed.

Relative to release 23, version 23.1 deleted 589 hairpins and 1,078 mature sequences as misannotated. Before trusting the predicted targets of a miRBase entry, check that sequencing reads support it in the tissue you study.

## Why an siRNA also acts like a miRNA

The seed does not care about the rest of the guide. Jackson and colleagues profiled human cells treated with siRNAs and found signatures specific to each siRNA, not to its intended target, including silencing of genes that shared as few as 11 contiguous nucleotides with the siRNA.

Birmingham and colleagues then showed that overall sequence identity contributed little to off-targeting except for near-perfect matches. Off-targeted genes instead carried one or more perfect 3' UTR matches to the hexamer or heptamer seed of the antisense strand, positions 2 to 7 or 2 to 8.

So a single siRNA guide behaves differently on different mRNAs, and Figure 2 follows one guide strand across the three kinds of target it meets.

![A 21 nucleotide guide strand drawn three times with its seed, positions 2 to 8, in green. Against the intended mRNA all positions pair and the target is cut. Against an mRNA with partial sites in its 3' UTR the seed and a short stretch near the 3' end pair, and the target is repressed without cutting. Against an unrelated mRNA only the seed pairs, which reduces expression like a miRNA.](/images/guides/mirna-vs-sirna-and-rna-interference/guide-three-targets.webp "**Figure 2. One siRNA guide strand, three kinds of target.** The effect depends on the pairing at each site, not on the identity of the small RNA. Pairing patterns are schematic. Sources: Elbashir et al. (2001), Doench et al. (2003), Birmingham et al. (2006).")

Against the intended mRNA, Argonaute2 cuts the target, and one loaded complex can act on many copies. Against an mRNA with several partial sites in its 3' UTR, the pairing is interrupted by bulges and mismatches, and the target is repressed without being cut, by a cooperative mechanism distinct from cleavage.

Against an unrelated mRNA with a seed match of six or seven nucleotides at positions 2 to 8, expression drops like it does for a miRNA target, which is the off-target signature seen in profiling. A clean phenotype from one siRNA therefore does not show that it hit only one gene. Check for seed matches in the 3' UTRs of other expressed genes.

## What duplex free energy can tell you

Base-pair stability can be estimated. The Turner nearest-neighbor model assigns a free-energy change to each stacked pair, loop and mismatch, and the 2004 parameter set is the standard in the ViennaRNA Package. For two RNAs it gives the energy of the best duplex: a more negative number means a more stable helix at the chosen temperature. Strand selection by 5' end stability, described above, is a statement about the same base-pair stability.

[RNAduplex](/app/rnaduplex) applies this to two sequences. It finds the lowest-energy set of intermolecular base pairs between a miRNA or siRNA and a candidate target, and reports the binding site and the hybridization free energy. It ignores any folding inside either strand, which keeps it fast enough to screen many pairs and suits a first pass over candidate sites.

The catch is that a duplex energy describes two free strands in solution. In a cell the guide sits inside Argonaute, and the target is folded and bound by proteins. Because RNAduplex allows only intermolecular pairs, a site buried in the target's own [secondary structure](/guides/rna-secondary-structure) can look better than it is. Nor can a two-strand calculation represent Argonaute itself reshaping to allow extensive pairing, as in Bartel's model.

Each missing term has its own ViennaRNA program. Table 2 lists what each one adds and what it still cannot see.

  <caption><strong>Table 2. Choosing a duplex calculation for a miRNA or siRNA question.</strong> Each program computes a different quantity and misses something the next one adds. Source: ViennaRNA RNAduplex documentation.</caption>
  <thead>
    <tr>
      <th scope="col">Question</th>
      <th scope="col">Start with</th>
      <th scope="col">What it cannot see</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Where does this guide bind best on this target, and how strongly?</td>
      <td>[RNAduplex](/app/rnaduplex)</td>
      <td>Structure inside either RNA, and the Argonaute protein</td>
    </tr>
    <tr>
      <td>Is the site open, or does the target have to unfold first?</td>
      <td>[RNAup](/app/rnaup)</td>
      <td>Proteins on the mRNA, and Argonaute</td>
    </tr>
    <tr>
      <td>Where along a long mRNA does a short guide bind, given accessibility?</td>
      <td>[RNAplex](/app/rnaplex)</td>
      <td>Proteins on the mRNA, and Argonaute</td>
    </tr>
    <tr>
      <td>Does the guide fold on itself or pair with its own duplex partner?</td>
      <td>[RNAcofold](/app/rnacofold)</td>
      <td>Whether a stable duplex represses anything in a cell</td>
    </tr>
  </tbody>

Treat the energy as a ranking aid. A site that scores well and sits in an accessible 3' UTR deserves an experiment; a site that scores well alone does not predict repression.

## Approved siRNA drugs

As of October 29, 2026, eight of the 25 [oligonucleotide drugs the FDA has approved](/guides/approved-oligonucleotide-drugs) are siRNAs, from patisiran in August 2018 to plozasiran in November 2025 (Table 3). Each is a duplex built for full complementarity to a single liver mRNA, so each works the siRNA way: Argonaute2 cuts the target. Patisiran uses a lipid nanoparticle for delivery; the other seven carry a GalNAc sugar cluster that pulls them into liver cells.

  <caption><strong>Table 3. FDA-approved siRNA drugs.</strong> Approval dates are the first FDA approval. List checked October 29, 2026. Sources: Drugs@FDA data files, NIHS table of nucleic acid drugs.</caption>
  <thead>
    <tr>
      <th scope="col">Drug</th>
      <th scope="col">Target</th>
      <th scope="col">Condition</th>
      <th scope="col">FDA approval</th>
    </tr>
  </thead>
  <tbody>
    <tr><td>Patisiran (Onpattro)</td><td>TTR</td><td>Hereditary transthyretin amyloidosis</td><td>August 2018</td></tr>
    <tr><td>Givosiran (Givlaari)</td><td>ALAS1</td><td>Acute hepatic porphyria</td><td>November 2019</td></tr>
    <tr><td>Lumasiran (Oxlumo)</td><td>HAO1</td><td>Primary hyperoxaluria type 1</td><td>November 2020</td></tr>
    <tr><td>Inclisiran (Leqvio)</td><td>PCSK9</td><td>High LDL cholesterol</td><td>December 2021</td></tr>
    <tr><td>Vutrisiran (Amvuttra)</td><td>TTR</td><td>Hereditary transthyretin amyloidosis</td><td>June 2022</td></tr>
    <tr><td>Nedosiran (Rivfloza)</td><td>LDHA</td><td>Primary hyperoxaluria type 1</td><td>September 2023</td></tr>
    <tr><td>Fitusiran (Qfitlia)</td><td>SERPINC1</td><td>Hemophilia A or B</td><td>March 2025</td></tr>
    <tr><td>Plozasiran (Redemplo)</td><td>APOC3</td><td>Familial chylomicronemia syndrome</td><td>November 2025</td></tr>
  </tbody>

The drugs aimed at miRNAs are a different story. Miravirsen, an antisense oligonucleotide that sequesters miR-122, reduced hepatitis C virus RNA in a 36-patient phase 2a trial, with a mean maximum reduction of 3.0 log10 IU/mL at 7 mg/kg, but it is an inhibitor of a miRNA, not a miRNA or an siRNA.

## miRNA and siRNA analysis tools on ProteinIQ

Fold the precursor first to see whether it forms a hairpin, then check how a guide [pairs with a target RNA](/use-cases/rna-rna-interaction-prediction), then add the structure and accessibility terms that the quick screen leaves out.

- [RNAfold](/app/rnafold): predict the hairpin fold of a pre-miRNA or the structure of a target region
- [RNAduplex](/app/rnaduplex): find the best binding site and free energy between a guide and a target
- [RNAup](/app/rnaup): add the cost of opening structure at the target site
- [RNAplex](/app/rnaplex): scan a short guide along a long mRNA
- [RNAcofold](/app/rnacofold): fold two strands together as a dimer
- [Reverse complement](/app/reverse-complement): write out the sequence a guide pairs with

## Frequently asked questions

### What is the main difference between miRNA and siRNA?

A miRNA is encoded in the cell's own genome and processed from a hairpin by Drosha and Dicer; an siRNA comes from long double-stranded RNA or a synthetic duplex and needs Dicer only. In animals miRNAs pair partially through the seed and repress many targets, while siRNAs are designed to pair completely with one target and cut it.

### Who discovered RNA interference?

Andrew Fire and Craig Mello reported in 1998 that double-stranded RNA silences genes in C. elegans far more effectively than either strand alone, and they shared the 2006 Nobel Prize in Physiology or Medicine. The first miRNA, lin-4, had been described in 1993.

### How long are miRNAs and siRNAs?

A mature miRNA is about 22 nucleotides long, cut from a pre-miRNA hairpin of about 70. Dicer produces siRNAs 21 to 22 nucleotides long, and the synthetic duplexes used in cells are 21 nucleotides.

### Can an siRNA act like a miRNA?

Yes. With only partial matches in a 3' UTR, an siRNA represses its target without cleaving it, and most siRNA off-target effects trace to seed matches at positions 2 to 7 or 2 to 8.

### How many siRNA drugs are approved?

The FDA had approved eight siRNA drugs as of October 29, 2026, the first being patisiran in August 2018 and the latest plozasiran in November 2025. Seven of the eight carry a GalNAc sugar cluster that delivers them to liver cells; patisiran uses a lipid nanoparticle.

### How many miRNAs do humans have?

miRBase release 23.1 lists 1,861 human hairpins and 2,789 mature sequences by our count of its files, but sequencing data support only part of such annotations, so the count of functional human miRNAs is lower.
