A messenger RNA is a sequence of four kinds of base and a protein is a sequence of twenty kinds of amino acid, and no base has any chemical affinity for any amino acid.
That last clause is the crux, and it was appreciated early. There is no way to build a template on which amino acids line up against bases directly, because the shapes and chemistries do not match. Whatever connects them must be an intermediary object that recognises a base sequence at one end and holds an amino acid at the other, and the assignment it embodies is therefore arbitrary in the way a code is arbitrary rather than forced in the way base pairing is forced.
Why three bases
The arithmetic is the first thing anyone noticed, and it is decisive. One base per amino acid gives four possibilities, far too few. Two bases give , still short of twenty. Three give , comfortably enough, with a great deal left over. There is no way to get twenty from four except by using at least three positions, so the code must be at least a triplet code, and parsimony suggests exactly three.
Sixty-four for twenty is a large surplus, and in the 1950s the surplus was treated as a problem to be explained away. George Gamow proposed in 1954 that amino acids fit into diamond-shaped holes in the DNA helix, in an overlapping code where consecutive triplets shared bases, which reduces the effective number of possibilities and looked more economical. It was ruled out quickly by protein sequence data: an overlapping code constrains which amino acids can follow which, and real proteins showed no such constraint. Single amino acid substitutions in sickle cell haemoglobin also change only one residue, whereas in an overlapping code a single base change would alter three neighbouring residues.
The frameshift experiment
Francis Crick, Leslie Barnett, Sydney Brenner and Richard Watts-Tobin settled the structure of the code in 1961 with genetics alone, and without knowing a single codon.
They used proflavin, an acridine dye that causes insertions or deletions of single bases rather than substitutions, on the rII region of bacteriophage T4, where loss of function is easy to score. A single insertion destroyed the gene's function, as expected if everything downstream is read in the wrong frame. Combining two insertions in the same gene also destroyed function. Combining three insertions restored function to something close to normal.
The reasoning is complete and it is worth following. If the message is read in fixed-size groups from a fixed starting point, then adding one base shifts everything downstream by one and garbles it, adding two shifts by two and still garbles it, and adding a number of bases equal to the group size restores the original frame after a short scrambled stretch. Three insertions restoring function therefore says the group size is three. Three deletions did the same. One insertion combined with one deletion, close together, also restored function, which confirms the frame interpretation directly.
The same experiment shows the code is non-overlapping, since an overlapping code has no frame to restore, and that it has no punctuation between codons, since a comma-free code would not be disrupted by a shift in the way observed.
Poly-U
Marshall Nirenberg and Heinrich Matthaei broke the first codon later the same year, at the National Institutes of Health, using a cell-free system: an E. coli extract containing ribosomes, transfer RNA, enzymes and energy, from which the cell's own messenger RNA had been allowed to decay. Add a synthetic RNA and the system makes whatever protein that RNA specifies.
They added polyuridylic acid, an RNA of nothing but U, and the extract made a polypeptide of nothing but phenylalanine. UUU codes for phenylalanine. Nirenberg presented the result in August 1961 at a congress in Moscow, to a small audience, and Crick arranged for him to repeat it to a large one.
Poly-A gave polylysine and poly-C gave polyproline. Beyond the homopolymers the method becomes indirect, because a random copolymer of two bases produces a mixture of codons whose proportions can be calculated but whose order cannot be controlled, so the assignments come out statistically. Har Gobind Khorana solved that by synthesising RNAs of defined repeating sequence, and Nirenberg and Philip Leder solved it more directly in 1964 with a binding assay: a trinucleotide of known sequence, added to ribosomes, causes just one charged transfer RNA to bind, and the complex is large enough to stick to a nitrocellulose filter while free transfer RNA washes through. Sixty-four trinucleotides, twenty labelled amino acids, and a filter. All 64 codons were assigned by 1966.
Example. A cell-free extract is given a random copolymer of U and G in a 3:1 ratio. Assuming the bases are incorporated independently, work out the expected frequency of the codon UUG relative to UUU, and say why this method could not have completed the code on its own.
The probability of U at any position is 0.75 and of G is 0.25, so UUU has probability and UUG has , one third as common. Measuring the ratio of two amino acids in the product therefore tells you the base composition of their codons. What it cannot tell you is the order, because UUG, UGU and GUU all have the same composition and the same expected frequency, and they code for leucine, cysteine and valine respectively. A composition method can sort the 64 codons into ten composition classes and no further. That is why the completion of the code needed either defined repeating sequences, which fix the order, or the triplet binding assay, which tests one specific sequence at a time. The general lesson is about experimental resolution rather than about the code: a method that measures a summary statistic can only distinguish hypotheses that differ in that statistic, and no amount of extra data of the same kind will break the remaining degeneracy.
Now you. In the frameshift experiment, three insertions restored function but only if they were reasonably close together in the gene, and some combinations of three failed. Explain both observations.
Answer
Between the first insertion and the third, the message is read in the wrong frame, so that stretch encodes an essentially random sequence of amino acids. If the three insertions are close, the garbled stretch is a few residues long and the protein can often tolerate it, particularly outside its active site, so function is restored well enough to score as positive. If they are spread far apart, hundreds of scrambled residues intervene and the protein does not work no matter what the frame downstream is. The failures of particular close-together triples have a sharper cause: in a wrong reading frame, one of the three stop codons will appear on average once every twenty-one codons, and if a stop falls inside the shifted stretch translation ends there and no downstream protein is made at all. Both observations therefore support the model rather than complicating it, and the second is a nice indirect prediction of stop codons from an experiment that was not looking for them. It also shows something about how such experiments are read: "restores function" is a biological assay with a threshold, not a chemical measurement, and knowing where the threshold sits is part of interpreting it.
The code is not arbitrary in its arrangement
The assignments themselves could have been anything, but the pattern of assignments is strikingly non-random, and the pattern is the interesting part.
Sixty-one codons specify amino acids and three, UAA, UAG and UGA, specify stop. The redundancy is concentrated almost entirely in the third position: for eight of the twenty amino acids, all four codons beginning with the same two bases mean the same thing, so the third base is irrelevant. Where the third position does matter, it usually distinguishes only between the two purines and the two pyrimidines, so a change from A to G or from C to T is often silent.
Second, codons that differ in the first position often specify amino acids of similar character. All four codons with U in the middle position specify strongly hydrophobic residues.
Both features have the same effect: a random single base change is disproportionately likely either to change nothing or to substitute a chemically similar amino acid. Stephen Freeland and Laurence Hurst tested this in 1998 by generating alternative codes at random and scoring them for how much a point mutation changes the chemical character of the encoded residue. The natural code scored better than all but about one in a million randomly generated alternatives. The exact figure depends on which chemical property is scored and on what set of alternative codes is considered admissible, and it has been contested, but the qualitative conclusion is robust: the code is arranged so that the commonest kind of mistake does the least damage.
The code is also nearly universal, which is the strongest single piece of evidence that all life on Earth shares an ancestor. A human gene expressed in a bacterium produces human protein. The exceptions are informative and few: human mitochondria read UGA as tryptophan rather than stop and AGA and AGG as stop rather than arginine, some ciliates read UAA and UAG as glutamine, and a handful of organisms have made other reassignments. Every exception is a small deviation in a small genome, which is what one expects if changing a codon assignment is possible only when very few genes would be affected.
Example. A human gene is put into E. coli and the protein is made, but at a small fraction of the expected yield, with ribosomes stalling partway along the message. Nothing is wrong with the sequence of the protein. What is going wrong, and what would you change?
The problem is codon usage. Synonymous codons are not used equally, and each organism has its own preferences, matched to the abundances of its own transfer RNAs: a codon that is rare in E. coli is decoded by a transfer RNA that is scarce, so a ribosome reaching it waits. A human gene full of codons that are common in humans and rare in E. coli therefore translates slowly and stalls, and stalled ribosomes both reduce yield and trigger quality control that destroys the partial product. The fix is codon optimisation: rewrite the gene so that every codon is a synonym preferred by the host, changing not one amino acid. This is standard practice in any laboratory expressing a foreign protein, and it works. It is also a warning about the phrase "the code is universal", which is true and does not mean that a gene reads equally well everywhere: the dictionary is shared but the vocabulary frequencies are not, and the machinery is tuned to the frequencies.
Now you. Because the third position is often silent, a synonymous change alters the DNA and not the protein. Comparing the same gene in two related species, what would you expect the ratio of non-synonymous to synonymous substitutions to be if the protein is under no selection at all, and what would a ratio well below that tell you?
Answer
If the protein is under no selection, both kinds of site accumulate substitutions at the same underlying mutation rate, so once each is expressed per available site the ratio should be about one. A ratio well below one means non-synonymous changes have been removed by selection, which is evidence that the protein sequence matters and that the gene is functional and conserved. A ratio above one is the much rarer and more interesting case, since it means amino acid changes have been favoured, and it is the standard signature of positive selection: it is found in immune recognition genes, in surface proteins of pathogens evading immunity, and in genes involved in reproduction. This test, usually written as the ratio of dN to dS, is one of the workhorses of molecular evolution, and note what makes it possible. It exists only because the code's redundancy provides a built-in internal control, a class of mutation at the same locus, subject to the same mutational processes, that selection cannot see. Without third-position degeneracy there would be no way to separate mutation rate from selection.
Two amino acids beyond the twenty are inserted by special mechanisms that reinterpret a stop codon in a particular sequence context: selenocysteine at UGA and pyrrolysine at UAG. They are genuine exceptions and they show that the machinery has some flexibility, not that the code is loose.
The adaptor
Crick had argued in 1955, before any evidence, that the connection must be made by adaptor molecules: small nucleic acids that pair with the message and carry an amino acid. Transfer RNA is exactly that, about 76 nucleotides folded into an L shape, with the three-base anticodon at one end and the amino acid attached to the 3' end of the other arm, roughly 8 nm away.
The consequence is the important part, and it is easy to state and easy to underestimate. The ribosome reads the anticodon. It never inspects the amino acid. Once a transfer RNA has been charged, whatever is attached to it will be inserted wherever that anticodon pairs.
François Chapeville and colleagues demonstrated this in 1962 with an experiment of great directness. They took transfer RNA properly charged with cysteine, then chemically reduced the cysteine to alanine using Raney nickel while it was still attached, producing a molecule with a cysteine anticodon carrying alanine. Fed into a protein-synthesising system, it put alanine wherever the message called for cysteine. The code lives in the charging enzymes, not in the ribosome.
That makes the aminoacyl-tRNA synthetases the physical embodiment of the genetic code: twenty enzymes, each of which must recognise one amino acid and the set of transfer RNAs bearing the corresponding anticodons, and get both right. They activate the amino acid with ATP, splitting it to AMP and pyrophosphate, which is the irreversibility trick from the lesson on ATP, and then transfer it to the transfer RNA.
Their accuracy problem is severe. Isoleucine and valine differ by a single methylene group, worth only about 12 kJ/mol of binding energy, which by itself would allow discrimination of only about a hundredfold. The measured error rate for inserting valine at isoleucine positions is about one in 40,000, which would require 27 kJ/mol from a straightforward binding argument. The enzyme achieves it with a double sieve: the synthetic site is too small to admit anything larger than isoleucine, and a second, editing site is large enough to admit valine but too small to admit isoleucine, and hydrolyses whatever it can bind. Anything too big is rejected at the first sieve and anything too small is destroyed at the second. Two coarse filters in series, each cheap, give a precision neither could reach alone, which is the same architectural idea as the three-filter fidelity of DNA replication.
A machine made of RNA
The ribosome is two subunits, roughly 2.5 million daltons in bacteria, built from three or four RNA molecules and some fifty proteins. It has three sites for transfer RNA, conventionally A, P and E: a charged transfer RNA arrives at A, the growing chain is transferred onto it, and the now empty transfer RNA moves through P to E and leaves.
For decades the RNA was assumed to be scaffolding and the proteins to be the catalysts, because catalysis was what proteins did. When the atomic structures arrived, from Thomas Steitz, Venkatraman Ramakrishnan and Ada Yonath around 2000, the peptidyl transferase centre, where the peptide bond is actually made, turned out to have no protein within about 1.8 nanometres. The catalytic machine is ribosomal RNA, and the proteins sit on the outside, stabilising it. The three shared a Nobel Prize in 2009.
This is one of the strongest arguments for an early RNA world. The most conserved and most central machine in all of biology, the one that makes the proteins, is itself not made of protein, which is exactly what one would expect if RNA came first and protein synthesis was invented by RNA.
Rates, errors and cost
A bacterial ribosome adds 15 to 20 amino acids per second, so a 300-residue protein takes about 15 seconds. Eukaryotic ribosomes run at 3 to 8 per second.
The error rate is around one in per codon. For a 300-residue protein the chance of at least one wrong residue is per cent, and for a 1000-residue protein it is about 10 per cent. That is far worse than DNA replication, and it should be: a defective protein is degraded and the cell makes another, while a defective genome is inherited. The tolerable error rate of any process is set by how expensive and how permanent the mistake is.
Even one in is better than simple binding allows, since correct and incorrect codon-anticodon pairs differ by only a few kilojoules per mole. The extra accuracy comes from kinetic proofreading, a scheme John Hopfield described in 1974: the charged transfer RNA arrives bound to the factor EF-Tu, and an irreversible GTP hydrolysis divides the selection into two stages, with a delay between them during which an incorrectly paired transfer RNA is more likely to dissociate. The cell spends a GTP to create a second, independent chance to reject the wrong molecule. Accuracy beyond what equilibrium binding gives always has to be bought with energy, and this is where the price appears in the bill.
That bill is the largest single item in a cell's budget. Each peptide bond costs four high-energy phosphates: two in charging the transfer RNA, and two GTP in the elongation cycle. A bacterium making 2500 proteins per second at 300 residues each is spending three million ATP equivalents per second on translation alone, which is why the lesson on energetics found that protein synthesis consumes the majority of a fast-growing cell's income.
Example. A very large fraction of clinically used antibiotics target the bacterial ribosome: tetracyclines, aminoglycosides, macrolides, chloramphenicol, oxazolidinones. Why is a machine so ancient and so conserved a good drug target rather than a bad one?
Because conservation and identity are not the same thing. The ribosome is universal in function and ancient in origin, but bacterial and eukaryotic ribosomes have diverged enough in sequence and in surface detail that small molecules can bind one and not the other, while the parts that must be identical, the catalytic core, are largely not where these drugs bind. Tetracyclines block the A site, macrolides plug the exit tunnel, aminoglycosides distort the decoding centre so that the ribosome accepts wrong transfer RNAs. Being essential is what makes the target good: a cell cannot survive without translating, cannot easily do without a ribosome, and resistance requires changing a machine under heavy structural constraint. The prediction that follows from this reasoning is uncomfortable and correct. If selectivity comes only from the divergence between bacterial and eukaryotic ribosomes, then any human ribosome that resembles a bacterial one should be vulnerable, and human mitochondria contain exactly such ribosomes. That is the accepted explanation for the irreversible hearing loss caused by aminoglycosides, which is strongly associated with particular mitochondrial ribosomal RNA variants, and for the marrow suppression and lactic acidosis seen with prolonged linezolid. Why mitochondria have bacterial ribosomes at all is the subject of a later lesson.
Now you. A missense mutation changes one amino acid in a protein. A nonsense mutation changes a codon to a stop. A frameshift inserts or deletes one base. Rank these three by expected severity, and then say why the ranking has an important exception.
Answer
The usual ranking is missense least severe, nonsense more severe, frameshift most severe. A missense change substitutes one residue, and given the structure of the code the substitute is often chemically similar, so many missense mutations are tolerated or even silent in effect. A nonsense mutation truncates the protein at that point, losing everything downstream. A frameshift garbles everything downstream and then almost always hits a premature stop, so it combines the damage of both. The important exception is position. A missense mutation in an active site, at a residue that coordinates a metal ion or performs catalysis, destroys the protein completely, while a nonsense mutation in the last few codons removes a tail that may not matter. Severity is a property of the mutation and its location together, never of its class alone. There is a second effect worth knowing: cells possess nonsense-mediated decay, which detects a premature stop codon by its position relative to the marks left by splicing and destroys the message. That usually converts a truncation into a complete absence of product, which is more severe for a protein needed in two copies and less severe when a truncated protein would be actively harmful.
The flow from DNA to RNA to protein is now complete, and a cell that ran it on every gene at once would be bankrupt within minutes. Four thousand genes in a bacterium and twenty thousand in a human cannot all be expressed, and the ones that are must change with circumstances: a bacterium meeting a new sugar, a liver cell responding to a hormone, a cell in an embryo becoming a neuron rather than a muscle fibre. The next lesson is how a cell chooses, starting with a bacterial switch that was worked out in the late 1950s and that remains the clearest example of regulation anywhere in biology.