06 — RNA processing and splicing
Before this: Ch 05 — Transcription · Ch 01 — Chemistry and cell primer · Time: ~40 min
What you'll be able to do
- Trace a transcript from nascent RNA to exportable mRNA, say what each modification buys, and name the transcripts a poly(A)-selected library silently loses
- Derive the two-step chemistry of , explain from that derivation why the burns ATP although the reaction is energetically neutral, and say which of its components actually catalyses the reaction
- Read splice signals out of raw sequence, predict what a variant at each position does, and explain why a broken donor site causes exon skipping in humans and retention in yeast
- Predict how the local balance of SR proteins and hnRNPs — and the polymerase's elongation rate — sets an exon's percent-spliced-in, and say why a transcript catalogue bounds proteome diversity from above rather than measuring it
- Apply the 50–55 nt rule to decide whether a premature stop gives no protein or a truncated one — and why that changes the inheritance pattern
- Identify the splice-disrupting variants that standard pipelines label benign and silently discard
- Distinguish an RNA-editing site from a genuine variant in data, and say why calling variants from RNA without an editing filter yields thousands of spurious A→G "SNPs"
The core idea
In a bacterium, and happen in one compartment simultaneously: a latches onto the message while is still writing it. There is no room for an editing step, and there isn't one.
Eukaryotes put a membrane between the two. That creates a window in which a transcript exists, is not yet being translated, and can be worked on. Evolution filled the window completely. What comes off RNA polymerase II is not a message — it is raw material. Its 5′ end is a naked triphosphate that nucleases will chew. It has no defined 3′ end, because Pol II does not stop cleanly. And most of its length is : internal sequence that must be excised before the reading frame makes sense.
The mature mRNA is the output of a pipeline that adds a header, cuts and appends a trailer, and removes interior segments whose boundaries are specified — weakly and ambiguously — in the stream itself. That ambiguity is the point. If the boundaries were unambiguous, would be a compilation step and this chapter would be short. Which segments get removed depends on the cell. Splicing is not tidying-up bolted onto transcription; it is a second layer of coding, resolved at run time.
1. The pipeline
The arrows imply a sequence; in reality all four overlap. The enzymes are recruited to the C-terminal domain of Pol II, so the polymerase is not merely a producer — it is the scaffold that carries the processing machinery to the RNA it is making. Processing that fails leaves a transcript degraded in the nucleus rather than exported, so nuclear RNA and cytoplasmic mRNA are not the same population.
2. The 5′ cap
A 7-methylguanosine attached to the first transcribed nucleotide — backwards. A triphosphatase strips one phosphate from the RNA's 5′ triphosphate, a guanylyltransferase transfers GMP from GTP, a methyltransferase methylates N7 of the added guanine.
m7G first transcribed nucleotide
| |
5'-G ppp 5'-N ---- 3'-N ---- 3'-N ...
^^^^^^^^^^^
5'-5' triphosphate bridge — not the usual 5'-3' linkage
The inverted is the function. There is no free 5′ end and no 5′-monophosphate, so 5′→3′ exonucleases have nothing to grip. It is also a positive tag: the nuclear cap-binding complex reads it and licenses of the first , 3′-end processing and export; in the cytoplasm it is swapped for eIF4E, the rate-limiting initiation factor. And a further 2′-O-methyl on the first transcribed nucleotide (cap1) marks the RNA as self — unmethylated cap0 RNA is treated as viral, which is why synthetic mRNA must be capped correctly to avoid an interferon response.
3. The 3′ end: cleavage and polyadenylation
Pol II has no crisp terminator. The 3′ end of an mRNA is not where stopped — it is where the transcript was cut.
... A A U A A A ......(10–30 nt)...... C A ↓ ...... G/U-rich ...
^^^^^^^^^^^ ^ ^^^^^^^^^
hexamer, read by CPSF cleavage bound by CstF
AAUAAA is canonical, in roughly half of human 3′ ends; AUUAAA is the common second. After cleavage, poly(A) polymerase adds ~200–250 adenosines with no template — the tail is not encoded in the genome. Poly(A)-binding protein coats it, blocking 3′→5′ decay, and contacts eIF4G/eIF4E at the cap to loop the message into a closed circle, so a commits only to a message verified intact at both ends. Deadenylation is the first and usually rate-limiting step of turnover, which makes tail length the message's clock.
Two practical consequences. Most human genes have more than one usable poly(A) site, and the choice changes 3′ UTR length and with it the regulatory elements carried (Ch 24). And replication-dependent histone mRNAs have no poly(A) tail at all — they end in a stem-loop — so oligo-dT-selected reports near-zero histone expression in dividing cells. That is library prep, not biology (Ch 47).
The signal is mutable like anything else: AATAAA→AATAAG in HBA2 abolishes cleavage, the polymerase reads through, and the result is α-thalassaemia. One base, outside all coding sequence, and output collapses.
4. The size problem
| Quantity | Typical value |
|---|---|
| Internal exon length | median ~120 bp, mean ~170 bp |
| length | median ~1.5 kb, mean ~5.4 kb |
| Exons per protein-coding gene | ~8–9 (median) |
| Extremes | TTN, 363 exons in one transcript; DMD, ~2.2 Mb of DNA for ~14 kb of mRNA |
pre-mRNA [E1]~~~~~~~~~~~~[E2]~~~~~~~~~~~~~~~~~~~[E3]~~~~~~~~[E4]
120bp 3 kb 130bp 7 kb 110bp 2 kb 900bp
mRNA [E1][E2][E3][E4]
A typical pre-mRNA is 90–95% intron. Finding the coding islands is a search over megabases with weak signals, performed in real time by a machine that assembles fresh for every intron.
5. The signals — and why they are not enough
exon intron exon
... C A G │ G U R A G U .......... C U R A Y ......... Y Y Y Y Y Y Y Y N C A G │ G ...
│ +1 +2 ... +6 ▲ polypyrimidine -2 -1
│ branch point A tract (~10-20 nt)
5′ splice site 18-40 nt upstream 3′ splice site
(donor) of the 3′ SS (acceptor)
(R = A/G, Y = C/U, N = any.) The GU–AG rule — begin GU and end AG; GT…AG in DNA — holds for about 99% of human introns. Roughly 0.3–0.5%, several hundred introns, are U12-type, spliced by a separate minor built from U11, U12, U4atac and U6atac, plus U5 shared with the major spliceosome. They splice slowly and are rate-limiting in the genes containing them, which is why in minor-spliceosome components cause specific developmental syndromes despite affecting so few introns.
Now the part that matters if you have ever written a parser. A 5′ splice site carries roughly 8 bits across 9 degenerate positions; a 3.1 Gb genome contains on the order of 10⁸ GU dinucleotides. Eight bits means one position in 2⁸ = 256 matches by chance, so 3.1 × 10⁹ / 256 ≈ 1.2 × 10⁷ consensus-matching decoys per strand — against roughly 1.5 × 10⁵ real 5′ splice sites (19,442 coding genes × ~8 introns apiece). Eighty decoys for every true site. You cannot find splice sites by matching the consensus. It is necessary and nowhere near sufficient — exactly the specificity problem from Ch 01 §5. The true site is not uniquely recognisable; it is merely bound somewhat longer than the ~10⁷ decoys. Discrimination comes from context: auxiliary sequence elements, the proteins bound to them, the pairing geometry of §8, and the order in which the polymerase emits the sequence.
6. The chemistry: two transesterifications
Step 1 — branching. The 2′-hydroxyl of the branch-point adenosine attacks the phosphate at the 5′ splice site. The exon1– bond breaks; a bond forms between the intron's first nucleotide and the branch A's 2′ position. That 2′–5′ is a branch in an otherwise linear molecule, making the intron a lariat.
Step 2 — exon ligation. Exon 1 now has a free 3′-hydroxyl. It attacks the phosphate at the 3′ splice site. The exons join; the lariat is released, debranched, degraded.
substrate [--- exon 1 ---]G U ~~~~~~~~ A ~~~~~~~~ A G[--- exon 2 ---]
^ ^
5′ splice site 3′ splice site
after step 1 G
/ \ ← 2'-5' branch
[--- exon 1 ]-OH ~~~~~~~~ A ~~~~~~~~ A G[--- exon 2 ---]
after step 2 G
/ \
[--- exon 1 ][--- exon 2 ---] A ~~~~~~ A G-OH ← lariat, discarded
Count the bonds. Step 1 breaks one phosphodiester bond and makes one; so does step 2. The reaction conserves phosphodiester bonds and is therefore close to energetically neutral — the chemistry needs no ATP.
Which raises the question the derivation exists for: the hydrolyses ATP at eight distinct points, driven by eight RNA . If not for the chemistry, what for? Fidelity and order. Each ATPase drives an irreversible rearrangement and acts as a kinetic checkpoint — a correctly paired substrate advances before the helicase can reject it, a poorly paired one is discarded. The energy proofreads a near-neutral reaction rather than driving it: accuracy from imperfect filters in series, the same architecture as replication fidelity.
7. The spliceosome is a ribozyme
| snRNP | Job |
|---|---|
| U1 | Base-pairs the 5′ splice site. First commitment step |
| U2 | Base-pairs the branch region, bulging the branch A out of the duplex so its 2′-OH is exposed and aimed |
| U4/U6 | U4 delivers U6 pre-paired and inhibited — U6 is the catalyst, so it arrives muzzled |
| U5 | Holds the two exon ends together across the reaction, so exon 1 does not away between steps |
| U6 | Replaces U1 at the 5′ splice site and folds the catalytic core |
The load-bearing fact: the catalytic centre is RNA. U6 snRNA folds a triplex that coordinates the two magnesium ions doing the chemistry; the ~150 proteins are scaffolding, motors and proofreaders. That U6 core is superimposable on domain V of a group II self- , an RNA that performs the identical two transesterifications, through the identical lariat, with no at all. The economical reading is that the spliceosome is a group II intron broken into trans-acting pieces — and once the recognition elements were separable from the sequence they act on, they became regulatable, and became possible. The whole assembly is also rebuilt from scratch for every intron; it is not a persistent machine.
8. Exon definition, and why donor mutations skip exons
Which way does the machinery search when pairing a 5′ site with the correct 3′ site?
definition pairs across the intron — U1 at one end reaching U2AF at the other. Fine when introns are a few hundred nucleotides, as in yeast.
pairs across the exon: U1 on the downstream exon's 5′ site communicates with U2AF on that same exon's 3′ site over ~120 nt, and only later is this converted into a cross-intron complex. Humans — 120 bp exons inside 5.4 kb introns — use this.
Three predictions follow, all observed:
- A destroying a 5′ splice site removes the upstream exon, not just the intron. That exon has lost half the pair that defines it, so it is never recognised. Intron retention — the naive prediction — is the yeast answer.
- Exons are size-constrained. Internal exons much beyond ~300 nt are poorly defined.
- Exonic sequence carries information. If recognition spans the exon, the exon's interior is where activators and must bind — which is why a change altering no amino acid can delete the exon entirely.
9. Alternative splicing: classes and scale
cassette exon [1]---[2]---[3] most common class in mammals
(exon skipping) [1]---------[3]
mutually exclusive [1]---[2a]--[3] exactly one of a set is used
[1]---[2b]--[3]
alternative 5′ site [1===]------[2] donor moves
[1=]--------[2]
alternative 3′ site [1]------[===2] acceptor moves
[1]--------[=2]
intron retention [1]---[2] intron kept in the mature message
[1]~~~[2]
Add alternative first exons (alternative ) and alternative last exons (alternative polyadenylation) for the full inventory.
Scale. At least 95% of human multi-exon genes are alternatively spliced. GENCODE Release 50 annotates 19,442 protein-coding genes but 644,292 transcripts across 78,733 genes — about eight transcripts per gene (verified-facts). This is the standard answer to "how do ~19,000 genes build a human", and it is largely right. It is also routinely overstated: many isoforms are low-abundance, cell-type-restricted, or non-productive by design (§12), and mass spectrometry detects far fewer protein isoforms than the catalogue implies. Transcript count bounds proteome diversity from above; it does not measure it.
Pushed to an extreme it is spectacular: Dscam in Drosophila has four clusters of mutually exclusive exons (12 × 48 × 33 × 2), giving 38,016 possible mRNAs from one gene — more isoforms than the fly has genes.
10. How the choice is made
Splice sites are weak (§5), so inclusion of any exon is decided by proteins bound nearby.
| In the exon | In the | |
|---|---|---|
| Promotes inclusion | ESE — exonic | ISE |
| Represses inclusion | ESS — exonic splicing silencer | ISS |
SR proteins (SRSF1, SRSF2, …) carry RNA-recognition motifs plus an arginine/serine-rich domain; they bind ESEs and recruit U1 and U2AF to adjacent weak sites — broadly, they promote inclusion. hnRNPs (hnRNP A1, PTBP1, …) bind silencers and antagonise them.
The decision is quantitative and combinatorial, not switch-like. Exon inclusion is a continuous quantity — percent spliced in, PSI ∈ [0,1] — set by the local balance of activators and , hence by their concentrations, which differ by cell type. Tissue-specific splicing therefore needs no tissue-specific splice sites, only different mixing ratios of the same ubiquitous factors. Position matters as much as identity: a factor binding upstream of an exon may repress it while the same factor downstream activates, so regulators are described by splicing maps — effect as a function of binding position — not by a motif alone.
And because splicing is co-transcriptional, elongation rate is an input. A slow polymerase gives a weak upstream exon more time to be recognised before a competitor appears, favouring inclusion — so chromatin state, which sets elongation rate, modulates splicing (Ch 23).
11. Splicing mutations: the under-recognised disease class
Variants at the invariant GU/AG dinucleotides — ±1 and ±2 — are obvious, flagged by every pipeline, and account for roughly a tenth of catalogued pathogenic variants. They are the tip of the distribution. The rest look benign:
| Variant class | Standard | What it can actually do |
|---|---|---|
| Synonymous, in an exon | synonymous_variant | Destroy an ESE → whole exon skipped |
| , in an exon | missense_variant | The amino acid change is irrelevant; the real effect is loss of an ESE |
| Deep intronic, hundreds of nt from any exon | intron_variant | Create a new GU, activating a pseudoexon that is spliced in and frameshifts the message |
| Donor +3 to +6, acceptor −3 to −20 | splice_region_variant, or nothing | Weaken rather than abolish → partial skipping, dose-dependent severity |
A variant's annotation is a statement about a transcript model, not about a molecule.
synonymous_variantmeans "changes no amino acid in the canonical CDS". It does not mean silent. Filtering a variant list on consequence class — the first line of nearly every analysis pipeline ever written — discards deep intronic and synonymous variants before anyone looks at them. That filter is where a large share of undiagnosed Mendelian disease is hiding.
Two real cases. CEP290 c.2991+1655A>G sits 1,655 nt inside an , creates a cryptic donor, and activates a 128 bp pseudoexon carrying a premature stop; it is the commonest cause of Leber congenital amaurosis type 10, and every exome-only pipeline is blind to it. SMN2 c.840C>T is synonymous, yet by destroying an ESE it forces exon 7 skipping — which is why SMN2 cannot compensate for the loss of SMN1 in spinal muscular atrophy; see the worked example.
How much disease is this? Estimates run from ~15% to over 60% of pathogenic variants depending on gene and method; machine-learning surveys place roughly 16% of inherited-disease in the exonic-but-splice-disrupting class alone. The range is wide precisely because the class is under-ascertained: you find these only by predicting splice effects directly (SpliceAI and successors), by sequencing RNA from patient tissue, or by minigene assay — never by reading the consequence column. Ch 55 turns this into a procedure. The upside is that is one of the few disease mechanisms with a general fix: an base-paired to a silencer can redirect the decision.
12. Nonsense-mediated decay and the 50–55 nt rule
A truncated protein is often worse than none — it can poison a complex it half-assembles into. So the cell wants to destroy messages with premature stops. But a stop is a stop codon. How does the know a stop is premature? It has no reference. Position relative to what?
leaves marks. Each splicing event deposits an exon-junction complex (EJC) 20–24 nucleotides upstream of the junction, sequence-independently. The first round of sweeps EJCs off as the ribosome passes. A normal stop is in the last exon, so by termination every EJC is gone. Terminate with an EJC still downstream and the terminating complex contacts it via UPF1/UPF2/UPF3, UPF1 is phosphorylated, and the message is destroyed.
Now derive the threshold. The terminating ribosome protects roughly 25–30 nt downstream of the stop. Terminate too close to the junction and that footprint physically covers the EJC sitting 20–24 nt upstream of it, so no contact occurs.
...stop codon.................................|junction|...
└── ribosome footprint ~25-30 ──┘
└ EJC at 20-24 nt upstream ┘
stop within ~50-55 nt of the junction → EJC masked → NO decay
stop further than ~55 nt upstream → EJC exposed → DECAY
25–30 plus 20–24 puts the masking boundary somewhere in the mid-40s to mid-50s; the measured threshold is the empirical 50–55 nucleotide rule: a stop more than ~50–55 nt upstream of the final exon–exon junction triggers NMD; anything closer, or in the last exon, does not.
Three consequences. A nonsense variant early in a gene and one in the last exon are different — the first gives no protein (, usually ), the second gives a truncated protein that is actually made and can act -negatively. NMD is a regulatory circuit, not only quality control: a substantial fraction of events introduce premature stops deliberately, and SR proteins autoregulate by splicing "poison exons" into their own messages — negative feedback implemented in splicing, executed by NMD. And NMD targets are systematically under-observed in , having been degraded before you sequenced them.
13. RNA editing
Some changes are made to the RNA after . The genome does not change.
A-to-I, by ADAR. ADAR1 and ADAR2 deaminate adenosine to inosine in double-stranded RNA, and and reverse transcriptases read inosine as G. Millions of sites are catalogued in human, overwhelmingly in inverted Alu repeat pairs inside and 3′ UTRs. Most recode nothing; the bulk function is to mark endogenous double-stranded RNA as self so the innate sensor MDA5 ignores it — losing ADAR1 produces a spontaneous interferon response. Practically, edited sites appear in as A→G mismatches with no DNA support, so calling variants from RNA without filtering yields thousands of spurious A→G "SNPs" (Ch 46). Recoding sites are rare but can be essential: the GRIA2 Q/R site is genomically CAG (glutamine) and edited to read as CGG (arginine) at essentially 100% efficiency in brain, because the unedited channel is calcium-permeable and toxic.
C-to-U, by APOBEC1. In intestine, APOBEC1 deaminates one cytosine in APOB mRNA, turning the glutamine CAA into the stop UAA. Liver does not edit. Result: full-length ApoB100 from liver, truncated ApoB48 from intestine — two proteins, one gene, one DNA sequence, and a tissue-specific edit as the only difference. (APOBEC relatives that act on DNA are a major source of in cancer, Ch 56.)
14. Other RNAs, briefly
rRNA. Pol I makes one 47S precursor, cleaved and trimmed into 18S, 5.8S and 28S; 5S comes separately from Pol III. Over 200 nucleotides are chemically modified, each targeted by a snoRNA that base-pairs the site — an addressing scheme implemented in RNA. Many snoRNA genes live inside the of other genes, a useful corrective to "introns are waste".
tRNA. RNase P — itself a ribozyme — removes the 5′ leader, RNase Z the 3′ trailer, and CCA is added without a template. A minority of tRNAs contain introns, removed by a protein endonuclease and ligase unrelated to the : two mechanisms, both called .
Self-splicing introns. Group I introns use an external guanosine as and make no lariat. Group II introns use an internal branch adenosine and make a lariat by the identical chemistry of §6, with no spliceosome — and are the ancestors of the spliceosomal system.
Trans-splicing. Exons from two separate transcripts can be joined. In nematodes and trypanosomes it is routine: a short capped spliced-leader RNA donates its 5′ end to ~70% of C. elegans mRNAs and essentially all trypanosome mRNAs. Rare but real in humans.
Common misconceptions
| What people think | What's actually true |
|---|---|
| are junk the cell throws away | They carry branch points and regulatory elements, host snoRNA and genes, and their removal deposits the EJC marks used for quality control. The lariat is degraded; the sequence is functional |
| A synonymous variant is silent | It changes no amino acid. It can still destroy an exonic splicing and delete the whole exon. SMN2 c.840C>T changes no amino acid, yet it skips exon 7 — which is why SMN2 cannot compensate for loss of SMN1 in spinal muscular atrophy |
| GU–AG lets you find introns by pattern-matching | A 3.1 Gb genome has ~10⁸ GU dinucleotides. The consensus is weak and degenerate; real sites are distinguished by context and bound proteins. A learning problem, not a regex |
| Breaking a donor site causes the intron to be retained | In humans, means the upstream exon loses half its recognition pair and is skipped. Retention is the yeast prediction |
| The is a protein enzyme | The catalytic core is RNA — U6 snRNA positions the two catalytic magnesiums. The ~150 proteins scaffold, drive and proofread |
| Splicing needs ATP for the chemistry | The two transesterifications conserve phosphodiester bonds and are near-neutral. All eight ATP-dependent steps buy fidelity and irreversible ordering |
| ~644,000 transcripts means ~644,000 proteins | Many isoforms are low-abundance, cell-type-restricted, or deliberately non-productive NMD substrates. Transcript count bounds proteome diversity from above |
| A nonsense variant always means | Only if NMD fires. Within ~50 nt of the last junction, or in the last exon, the truncated protein is made — often , with a different inheritance pattern |
| Splicing happens after finishes | Most of it is co-transcriptional, and polymerase elongation rate is itself an input to splice-site choice |
Worked example: one synonymous base, and a disease
SMN1 and SMN2 are adjacent, near-identical paralogues on 5. Spinal muscular atrophy is caused by homozygous loss of SMN1. Everyone with the disease still has SMN2, which encodes an identical protein. Work out why that does not rescue them.
1 — The difference. SMN2 differs at position +6 of exon 7: c.840C>T. 280 goes TTC→TTT. Both are phenylalanine, so the change is synonymous; every predictor labels it synonymous_variant and every standard filter drops it.
2 — What the base was doing. Exon 7 is 54 nt, between a 5.76 kb 6 and a 444 nt intron 7. Its splice sites are weak and it depends on an exonic overlapping position +6. C→T disrupts that enhancer and strengthens hnRNP A1 binding. Under (§8), a short exon with weak sites and no net activator is not recognised.
3 — Consequence in RNA. Roughly 80–90% of SMN2 transcripts skip exon 7; only 10–20% are full-length. One synonymous base has cost ~85% of the gene's output.
4 — Frame arithmetic. 54 is divisible by 3, so skipping exon 7 does not shift the reading frame. Exon 6 joins exon 8 in frame, and — which normally terminates inside exon 7 — runs on into what is usually 3′ UTR, adding four residues (EMLA) before hitting a stop in exon 8. That stop is in the last exon, downstream of the final exon–exon junction, so §12's NMD test is not triggered: the SMNΔ7 message is stable and is translated.
5 — Why it still fails. The failure point is the protein, not the message. SMNΔ7 lacks the C-terminal region needed for oligomerisation, and the EMLA tail it picks up from exon 8 completes a degron, so the truncated protein is unstable and rapidly degraded. Note what NMD does not do here: exon 7 is the last coding exon, so even a frameshifting skip would have put its stop inside terminal exon 8 and escaped decay by the rule of §12. Divisibility decides which protein gets made, not whether the message survives — for that contrast you need an exon several junctions upstream of the last.
6 — Dosage. SMN2 copy number varies between people, 1 to 5+, and each copy contributes its 10–20% of full-length message. Copy number is the strongest known modifier of SMA severity: the disease is quantitative, and the quantity is a efficiency.
7 — The fix. If the problem is a splicing decision, change the decision. Nusinersen is an complementary to ISS-N1, an intronic silencer just inside intron 7. Blocking it prevents hnRNP A1 binding, exon 7 is included, full-length SMN rises. The drug touches no DNA, delivers no gene and encodes no protein — it occupies ~18 nucleotides of intron.
One base. Synonymous. Invisible to a consequence filter. It explains the disease, its severity gradient, and its treatment.
Connections
- Back to: Ch 05 — Transcription supplies the Pol II CTD that carries all this machinery; Ch 01 §5 set up weak signals and statistical specificity; Ch 02 for the 2′-OH that does the branching
- Forward to: Ch 07 reads the mature message; Ch 24 treats and 3′-UTR choice as regulation; Ch 44 must infer the exon structure described here from sequence and reads; Ch 47 quantifies isoforms and PSI; Ch 55 turns §11 and §12 into a classification procedure