The human genome holds ~3.2 billion base pairs across 46 chromosomes, yet only ~1.5% encodes protein. For the clinician this number reframes how we think about pathogenicity: the overwhelming majority of the genome is non-coding, so a variant landing outside an exon is the statistically expected case, not a reassuring one. The remaining ~98.5% is not inert filler — it contains promoters, enhancers, non-coding RNA genes (20,000–25,000), introns, and repetitive elements that govern when, where, and how much each gene is transcribed. Many neurogenetic disorders trace not to a broken protein but to mis-regulated dosage or timing of a perfectly normal one.

The genetic code is a degenerate triplet code: 64 codons specify 20 amino acids plus 3 stop signals. Degeneracy means most amino acids are encoded by several synonymous codons, so a nucleotide change — especially at the third 'wobble' position — frequently leaves the protein sequence untouched. This built-in redundancy buffers the proteome against the constant drizzle of point mutations, which is precisely why missense and nonsense changes are comparatively rare per base and why we should never assume a coding change is automatically tolerated or, conversely, that a 'silent' change is automatically safe.

The single most clinically useful feature of genome organization is the behavior of CpG dinucleotides. Cytosine in a CpG context is usually methylated to 5-methylcytosine, which spontaneously deaminates to thymine. Ordinary cytosine deaminates to uracil, a foreign base that repair enzymes excise efficiently; but 5-methylcytosine deaminates to thymine, a normal DNA base that the mismatch machinery often fails to recognize before the next replication locks in a C→T transition. The net effect is a mutation rate roughly 10× the genome average at CpG sites, which is why so many recurrent pathogenic missense and nonsense variants cluster at CGA, CGG, and other CpG-containing codons — a pattern worth recognizing when the same variant keeps reappearing in unrelated families.

Replication fidelity is layered, not absolute. Polymerase base selection alone is only modestly accurate; what makes replication trustworthy is a three-tier system. The polymerase first selects the correct nucleotide by Watson–Crick geometry, then its own 3'→5' exonuclease proofreads and removes the occasional misincorporated base, and finally the mismatch repair (MMR) system patrols the newly made strand to excise errors the first two steps missed. Each tier multiplies the accuracy of the one before it, driving the net error rate down to roughly 1 in 10⁹–10¹⁰ per base per division. The clinical payoff of understanding this hierarchy is what happens when the last tier fails: inherited MMR deficiency (Lynch syndrome) does not cause point-mutation chaos genome-wide but specifically destabilizes microsatellites — short tandem repeats where the polymerase is most prone to slip — producing the microsatellite instability that defines those tumors.

De novo variants are inevitable. Even with this fidelity, each child is born carrying roughly 60–70 new single-nucleotide variants absent from both parents. The dominant driver is paternal age: spermatogonia divide continuously throughout a man's life, so each replication is another chance to fix an error, whereas oocytes are largely arrested. The landmark whole-genome trio studies quantified this precisely — the de novo count rises by roughly 1.5 mutations per additional year of paternal age (Kong et al. 2012), about four times the maternal contribution. This is the molecular basis for the paternal age effect in de novo dominant conditions such as achondroplasia and several epileptic encephalopathies, and it explains why a striking new dominant phenotype in a child of older parents, with an unremarkable family history, is exactly what de novo disease looks like.

Repeat expansions break the fidelity rules in their own way. At tandem repeats the nascent and template strands can transiently misalign during replication or repair — replication slippage — adding or losing repeat units. Because longer tracts slip more readily, expansions tend to grow with each transmission, so the disease can present earlier and more severely in successive generations: the phenomenon of anticipation. This is why a parent's repeat length predicts, but does not equal, the child's, and why intergenerational instability must be counseled even when a parent is mildly affected or asymptomatic. Crucially, repeat length does not act through a single mechanism: an FMR1 full mutation (>200 CGG) is hypermethylated and transcriptionally silenced (loss of FMRP), whereas an intermediate premutation (55–200) stays active and over-produces toxic mRNA — the basis of the late-onset tremor/ataxia syndrome FXTAS first described in premutation carriers (Hagerman et al. 2001). The same locus can therefore cause opposite molecular lesions depending on tract length.

Splicing — excising introns and stitching exons together — is the step where genotype most often diverges from what a naive read of the coding sequence would predict, and it is therefore the part of mRNA processing a neurogeneticist must understand mechanistically. The spliceosome does not 'know' where exons are; it reads short consensus signals: the near-invariant GT at the 5' donor and AG at the 3' acceptor that bracket every intron, plus the branch point and polypyrimidine tract that position the catalytic chemistry. Variants at the donor/acceptor ±1/±2 positions almost always abolish a splice site, which is why they carry such heavy weight (PVS1) in ACMG classification — they predictably trigger exon skipping or intron retention rather than a subtle tweak.

The more counterintuitive lesson is that splicing depends on signals inside the exon too. Exonic splicing enhancers (ESEs) are short exonic motifs that recruit SR proteins to help the spliceosome recognize a nearby weak splice site. A nucleotide change that leaves the amino acid unchanged — a 'synonymous' variant — can nonetheless destroy an ESE and cause the whole exon to be skipped. This is the mechanistic reason a silent variant is never safe by inspection alone: SCN1A, for example, harbors synonymous and exonic changes that produce Dravet syndrome purely by disrupting splicing.

Finally, splicing is tissue-specific, which resolves one of the recurring puzzles in neurogenetics: how can a ubiquitously expressed gene cause a purely neurological phenotype? Alternative splicing generates brain-specific isoforms by including exons that other tissues skip. A variant that disrupts a brain-included exon affects only the neural transcript, sparing organs that use a different isoform — so the relevant transcript, not the canonical reference, is what must be modeled. Because roughly 10–15% of disease-causing variants act through splicing, in silico predictors (SpliceAI, MaxEntScan) should be run on every candidate, but they only flag a suspect; RNA studies (RT-PCR, RNA-seq) that show the aberrant transcript are the definitive functional evidence.

Once the mature mRNA reaches the cytoplasm, the ribosome reads it codon-by-codon in the 5'→3' direction, building the protein from N-terminus to C-terminus. Initiation is not as simple as 'find the first AUG': the 43S pre-initiation complex with Met-tRNA scans for an AUG sitting in a favorable Kozak context (the surrounding bases, especially −3 and +4). A poor Kozak context, or a variant that abolishes the start codon (p.Met1?), can leave the ribosome to skip past and initiate downstream, producing a truncated or absent protein — which is why initiation-codon variants are treated as likely loss-of-function even though they change only one residue's worth of sequence.

The protein's fate after the last peptide bond is just as consequential as its sequence. Signal peptides route nascent chains into the endoplasmic reticulum for secretion or membrane insertion; molecular chaperones (HSP70, HSP90) shepherd folding; and post-translational modifications — phosphorylation, glycosylation, ubiquitination — set localization, activity, and half-life. Misfolded products are tagged with ubiquitin and destroyed by the proteasome, so a missense variant can be 'pathogenic' not because the protein is non-functional but because the cell degrades it before it ever works.

The distinction that ultimately drives therapy is how a variant disrupts the protein. Loss-of-function (too little active protein, often via haploinsufficiency) is mechanistically opposite to gain-of-function or dominant-negative effects, where an abnormal product actively interferes — a mutant subunit poisoning a multimer, or a toxic conformation. This is not academic: a haploinsufficient gene is a candidate for boosting expression (e.g., antisense oligonucleotides that raise output, gene supplementation), whereas a toxic gain-of-function demands silencing the bad allele. Misjudging the mechanism points the entire therapeutic strategy in the wrong direction.

Naming a variant's class — missense, nonsense, frameshift, splice-site, synonymous — is only step one; the real work is predicting its molecular consequence, because that consequence drives which ACMG/AMP criteria apply and what the protein actually does. The central concept tying these classes together is nonsense-mediated decay (NMD), the surveillance pathway that distinguishes a 'broken protein gets made' variant from a 'no protein gets made' variant.

NMD works by reading the marks left over from splicing. After each intron is removed, an exon-junction complex (EJC) is deposited just upstream of the new exon–exon boundary; on the ribosome's first ('pioneer') round of translation, these EJCs are normally stripped off as the ribosome passes. If translation terminates at a premature termination codon (PTC) while an EJC still sits downstream — operationally, more than ~50–55 nucleotides upstream of the final exon–exon junction — the cell reads the transcript as defective and degrades it. This single rule explains the behavior of most truncating variants: a nonsense change (p.Arg100Ter) or a frameshift (a non-multiple-of-3 indel that almost always runs into a new stop) in an internal exon triggers NMD, eliminating the mRNA and yielding clean loss-of-function — the molecular justification for PVS1.

The instructive exceptions are where reasoning, not pattern-matching, is required. A PTC in the last exon (or the last ~55 nt of the penultimate exon) has no downstream junction to flag it, so it escapes NMD and a stable truncated protein is produced — which can act dominant-negatively and is therefore not interchangeable with a haploinsufficiency-causing variant elsewhere in the gene. Missense variants resist easy prediction altogether: impact depends on the residue's structural role and the chemical distance of the substitution, so a conservative swap in a flexible loop may be benign while a charge change in an active site is severe. And synonymous variants, despite leaving the protein untouched, can still disrupt splicing enhancers, mRNA stability, or translation efficiency. The unifying lesson is that the same nominal class can mean wildly different things — position and downstream context decide the consequence, not the label.