mRNA Display and Ribosome Display: Covalent Linkage, Cell-Free Libraries, and What Each Buys Over Phage

Every display method has to solve the same problem. A selection step throws away almost everything, and afterward you have to be able to read what survived. That requires the peptide being tested and the nucleic acid encoding it to stay together through every wash. Phage solves it by packaging both inside one particle. mRNA display and ribosome display solve it without cells at all — one by forming a covalent bond, the other by holding an assembly together with buffer conditions.

The difference is not academic. It determines how long a wash can run, how much library you can build, and which building blocks the peptide can contain. This guide is about the linkage itself.

Puromycin, and why the linkage is covalent

Roberts and Szostak reported that in vitro translation of synthetic mRNAs carrying puromycin — which they described as a peptidyl acceptor antibiotic — at their 3' end generates covalent fusions between an mRNA and the peptide or protein it encodes, and that the stable linkage between the informational and functional domains of those molecules allows a specific mRNA to be enriched from a complex mixture on the properties of its encoded peptide. They demonstrated this by enriching fusions carrying a myc epitope peptide from a pool of random-sequence mRNA-peptide fusions by immunoprecipitation ^[1].

The mechanism is that the ribosome, arriving at the 3' end of the message, accepts puromycin into the A site as though it were an incoming aminoacyl-tRNA and transfers the nascent chain onto it. Because the puromycin is attached to the message that encoded the chain, the product is one molecule containing both. Roberts and Szostak describe that linkage as stable and show the fusions survive immunoprecipitation from a random-sequence pool ^[1]; the further reading — that a single covalently joined molecule tolerates handling a non-covalent assembly would not — is this guide's inference from the chemistry rather than a comparison those authors ran.

One half of the fusion is still RNA, which is the format's weak point. Yamaguchi and colleagues addressed it with a puromycin-linker DNA carrying a ligation site, a biotin site for solid-phase handling, a reverse transcription primer site for converting the unstable mRNA-protein fusion into a stable mRNA/cDNA-protein fusion whose cDNA is covalently linked to its encoded protein, and a restriction site for release from the support. They reported that this arrangement also promotes protein folding and disulfide shuffling, demonstrated an enrichment efficiency of 20-fold per selection round in their model systems, and selected a 32-residue random library against interleukin-6 receptor, obtaining peptides with multiple disulfide bonds that bound in the low nanomolar range ^[7].

Twenty-fold per round is worth converting. If that rate held unchanged across four rounds — which those authors did not demonstrate, having reported a per-round figure from model systems — the compounded enrichment would be 20 to the fourth power, or 160,000-fold. The denominator in each round is the pool entering that round, not the original library, which is why the factors multiply rather than add.

Ribosome display: a complex held together by conditions

Ribosome display never forms a bond. Hanes and Pluckthun reported a system in which a correctly folded complete protein and its encoding mRNA both remain attached to the ribosome and can be enriched for the ligand-binding properties of the native protein; they selected a single-chain antibody fragment 100-million-fold by five cycles of transcription, translation, antigen-affinity selection and PCR, and reported that the selected fragments all mutated in vitro, acquiring up to four unrelated amino acid exchanges over the five generations while remaining compatible with antigen binding ^[2].

That total is a five-round product, so the geometric mean is 10 to the power 8/5, about 40-fold per cycle. The authors did not report per-cycle factors and this average should not be read as one: rounds are rarely equal.

Because the complex is an assembly rather than a molecule, its integrity is a buffer property. The mRNA must lack a stop codon, so that release factors never terminate translation and the ribosome stalls with the chain still attached. Hanes and colleagues, comparing an optimized Escherichia coli system with rabbit reticulocyte systems using antibody single-chain fragments, reported that with all systems enrichment was always dependent on the absence of a stop codon and the presence of cognate antigen; that the E. coli system was 100-fold more efficient than an optimized uncoupled rabbit reticulocyte system, which was in turn several-fold more efficient than a previously reported coupled system; and that they could detect no intrinsic advantage to a eukaryotic translation system here ^[3].

Two consequences follow from the assembly being non-covalent, both convention across published protocols rather than results this guide can point to: selections are run cold and in buffers that stabilize the ribosome, and ribonuclease control is part of the selection rather than optional hygiene.

Where the library size actually comes from

The size advantage of cell-free formats is a consequence of one omitted step. Lipovsek and Pluckthun described the two combined advantages of in vitro display: by obviating the need to transform cells in order to generate and select libraries, these methods allow a much higher library diversity, and by including PCR as an integral step they make PCR-based mutagenesis strategies convenient ^[5]. In a phage or bacterial library, every distinct member has to get inside a cell at least once, and that transformation step is what the cell-free formats delete.

For mRNA display, Umemoto and colleagues describe the technology as generating libraries with over 10 to the twelfth to 10 to the thirteenth diversity ^[6]. Goto and Suga describe the RaPID platform as conducting affinity-based selection of libraries of more than 10 to the twelfth unique sequences of in vitro expressed macrocyclic peptides ^[9].

Those numbers are large and still sparse, and the sparseness worsens with length. A random 15-mer drawn from 20 proteinogenic amino acids has 20 to the fifteenth possible sequences, which is 3.2768 times 10 to the nineteenth. A library of 10 to the thirteenth members therefore covers at most about 3 times 10 to the minus seventh of that space — roughly three parts in ten million. The denominator is the full sequence space, and this is an upper bound, since it assumes every library member is distinct. A display selection is not an exhaustive search; it samples, and the fraction sampled falls steeply as the randomized region lengthens.

The yield term that does not appear in the library size

A designed library size is not a delivered one, because not every mRNA in the pool actually forms a fusion. Umemoto and colleagues examined this directly, noting that the protein-puromycin linker/mRNA complex formation yield is determining for preparing libraries, and that how mRNA sequences affect it had remained unclear. They translated puromycin-attached mRNAs containing three random codons after the start codon (32,768 sequences) or seven random bases next to the amber codon (6,480 sequences), and calculated enrichment scores by dividing the appearance rate of every sequence in the protein-linker/mRNA complexes by its appearance rate in total mRNAs. They reported a range of 0.09 to 2.10 for the N-terminal coding sequences and 0.30 to 4.23 for the C-terminal ones, and read that wide range as showing that both regions strongly affected complex formation yield; using the C-terminal GGC-CGA-UAG-U sequence, which gave the highest enrichment scores, they constructed highly diverse libraries of monobodies and macrocyclic peptides ^[6].

Take the ratios with their denominators named. Within the N-terminal set, best over worst is 2.10 divided by 0.09, about 23-fold; within the C-terminal set, 4.23 divided by 0.30, about 14-fold. Both ratios are internal to that study's translation system and those two sequence sets, and are not a general constant. What does travel is structural: the constant regions flanking the randomized one are part of the experiment, and a designed diversity quoted without a measured fusion yield is the wrong number.

Building blocks the genetic code does not supply

Cell-free translation is a defined mixture, and that is what opens the door to non-proteinogenic residues. Katoh and Suga explain that because aminoacyl-tRNA synthetases are highly specific for their respective amino acid-tRNA pairs, natural synthetases cannot be used to prepare tRNAs charged with non-proteinogenic amino acids, whereas flexizymes are adaptable aminoacylation ribozymes that can prepare diverse aminoacyl-tRNAs using amino acids activated with suitable leaving groups; as recognition elements they require only an aromatic ring in the leaving group or the side chain of the activated amino acid, plus the conserved 3'-end CCA of the tRNA, so the system handles both side-chain-modified and backbone-modified substrates ^[8].

What that permits in practice is described by Goto and Suga, who report that the RaPID system integrates genetic code reprogramming by the flexible in vitro translation system, facilitated by flexizymes, with mRNA display; that thioether-closed macrocyclic peptides obtained this way have generally shown high affinity and specificity for their targets; and that beyond the ring-closing thioether, the system has allowed incorporation of diverse non-proteinogenic amino acids, hydroxy acids giving an ester linkage, amino carbothioic acid giving a thioamide linkage, and abiotic foldamer units ^[9].

The macrocyclization chemistry is unusually clean. Sako and colleagues developed a non-proteinogenic amino acid bearing a chloroacetyl group on its side chain, which forms a thioether bond by intramolecular reaction with the sulfhydryl group of a cysteine residue in the same chain upon translation; they reported that this takes place spontaneously in the translation solution and gives the corresponding cyclic peptides independent of ring size, using flexizyme and reconstituted cell-free translation ^[11]. No separate cyclization step is needed, so the displayed species is already macrocyclic when selection begins.

Not every exotic residue is equally easy, and the reasons are mechanical. Iwane and colleagues noted that proteinogenic aminoacyl-tRNAs are fine-tuned to bind elongation factor Tu with roughly uniform affinity, and that incorporation of some non-proteinogenic residues such as N-methyl-amino acids is less efficient, especially when they appear frequently or consecutively. Using flexizymes they experimentally verified that N-methyl-aminoacyl-tRNA affinities for EF-Tu are indeed weaker than those of proteinogenic aminoacyl-tRNAs, engineered the tRNA T-stem to tune those affinities, and reported that uniform affinity-tuning enhanced incorporation, achieving nine distinct N-methyl-amino acids in both linear and thioether-macrocyclic peptide scaffolds ^[10].

Goto and Suga scope the remaining difficulty to a particular system: they describe multiple or consecutive incorporation of highly exotic amino acids, such as D- and beta-amino acids, as still challenging in the conventional FIT system, and the synthesis of peptides bearing non-carbonyl backbone structures as a particularly demanding task there — and in the same account report their answer: semilogical engineering of tRNA body sequences gave a suppressor tRNA, tRNAPro1E2, that effectively recruits EF-Tu and EF-P and allows not only single but also consecutive and multiple elongation of D-, beta- and gamma-amino acids as well as aminobenzoic acids, while integrating the FIT system with chemical or enzymatic post-translational modifications expands the accessible backbones to non-carbonyl moieties such as azolines, azoles and ring-closing pyridines ^[9].

What each format buys over phage

The clearest gain is the removal of a growth step. Bakhshinejad and colleagues, reviewing phage biopanning, describe target-unrelated peptides as arising two ways: selection-related ones from binding to components of the screening system rather than the target, and propagation-related ones from the faster growth rate of some phage clones, which lets them outcompete slow-propagating clones; library amplification between rounds, they write, contributes significantly to selecting clones with a propagation advantage ^[12]. A cell-free selection has no propagation step, so that second class cannot arise by that route.

It does not follow that cell-free selections are artifact-free. The first class — binding to beads, capture reagents, blocking agents — is a property of the selection surface rather than of the display format, and survives the change. Reverse transcription and PCR still sit between rounds, and whether that amplification imposes its own sequence bias is a question to test rather than assume.

The second gain is in-selection evolution: because PCR is integral, mutations introduced during amplification are carried forward. Hanes, Schaffitzel and colleagues applied ribosome display to single-chain antibody fragments from a synthetic human library against bovine insulin and reported that three independent experiments selected different clusters of closely related fragments, all binding the antigen with high affinity and specificity; that all selected fragments had affinity-matured up to 40-fold compared with their library progenitors by accumulating point mutations during the ribosome display cycles; and that dissociation constants of the isolated fragments were as low as 82 picomolar ^[4].

The formats are also combined rather than treated as rivals. Zimmermann and colleagues describe a protocol that generates synthetic nanobodies against any purified protein or protein complex within a three-week period: a single round of ribosome display using libraries encoded by mRNA, followed by two rounds of phage display and binder identification by ELISA. They report that the protocol is optimized to avoid undesired reduction in binder diversity and enrichment of non-specific binders, and that it typically gives rise to 10 to 30 unique binders with affinities in the range of 500 picomolar to 500 nanomolar ^[13]. Those authors report the sequence and the outcome, not a rationale for it; reading the two steps as a division of labor is this guide's interpretation.

What is not settled

• Designed and delivered library diversity are not the same quantity. Complex formation yield varies with the constant regions flanking the randomized one, at least in that study's translation system ^[6], so a designed diversity reported without a measured yield leaves the delivered figure unknown.
• Whether those enrichment-score ranges generalize beyond that study's translation system and sequence sets is untested ^[6].
• Whether reverse transcription and PCR between rounds bias sequence recovery in cell-free selections is unresolved by the work cited here.
• How far consecutive and multiple incorporation of the most exotic building blocks now extends is unsettled: the same authors call it challenging in the conventional FIT system while reporting that an engineered suppressor tRNA, and post-translational backbone conversions, extend the accessible range, without stating where the new limits lie ^[9].
• A selection reports binding under one buffer, on one surface, against one immobilized preparation of a target; nothing about the display format tells you what a recovered sequence does elsewhere, which is why resynthesis away from the display context and re-measurement in an independent assay is standard practice.

Both formats trade a biological constraint for a physical one. Phage asks that every library member survive a cell. mRNA and ribosome display ask instead that the linkage survive the wash — covalently in the first case, by buffer conditions in the second — and hand back a larger library and a translation reaction whose composition is yours to set.