Beta-Arrestin Recruitment Assays and Biased Agonism: BRET, Complementation, Imaging and Bias Factors

A beta-arrestin recruitment assay reports one physical event: the appearance of arrestin at, or close to, an activated receptor. Everything a reader usually wants from such an assay — whether a ligand is biased, by how much, and whether that number will survive a change of cell line — is inference layered on top of that single proximity measurement. This guide covers the three common measurement formats, the arithmetic that turns two concentration-response curves into a bias factor, and the reasons the resulting number describes an assay system at least as much as it describes a ligand.

What the readout actually is

Agonist occupancy of a receptor is followed by phosphorylation of its intracellular face by G protein-coupled receptor kinases, and arrestin then binds the phosphorylated, active receptor. A recruitment assay senses the consequence of that binding — two labeled proteins coming within molecular distance of each other, or one labeled protein arriving in a membrane compartment. It does not sense phosphorylation, does not sense downstream signaling, and does not distinguish a productive arrestin complex from a transient one.

Two configurations are in general use, and they answer slightly different questions. In the direct format, a tag is fused to the receptor's intracellular C-terminus and the sensor reports arrestin binding that specific receptor. In the translocation format, the receptor is left unmodified and the sensor reports arrestin arriving at the plasma membrane. Protocols for both, in resonance energy transfer and in nanoluciferase complementation versions, have been published as step-by-step procedures for live cells ^[1]. The direct format is receptor-specific but requires a modified receptor; the translocation format preserves the native receptor sequence but will also respond to arrestin recruited by any other activated receptor in the same cell.

Resonance energy transfer formats

In bioluminescence resonance energy transfer (BRET), a luciferase donor on one partner transfers energy non-radiatively to a fluorescent acceptor on the other. Transfer efficiency falls off very steeply with separation, so the ratio of acceptor to donor emission rises when the two labels approach within a few nanometers. Because the donor is enzymatic, no excitation light is applied, which removes photobleaching and excitation-driven autofluorescence as error sources. Those physical points are standard optics and are stated here without citation.

The acceptor need not be on the receptor. In enhanced bystander BRET, the luciferase donor is placed on beta-arrestin and a fluorescent acceptor is anchored in a defined compartment — the plasma membrane or the endosome — so the signal reports arrestin arriving in that compartment, and the same sensor design can follow receptor/arrestin complexes from the membrane into endosomes ^[2]. This buys compartmental resolution at the cost of receptor specificity: a bystander signal is a statement about where arrestin is, not about which receptor it is touching.

A third variant places donor and acceptor on the same arrestin molecule. Charest and colleagues built such an intramolecular sensor and reported that the BRET signal increased following receptor activation, which they interpreted as suggesting a rearrangement bringing the arrestin amino and carboxyl termini closer together; their kinetic analysis placed this conformational change after the initial arrestin/receptor engagement ^[3]. That ordering matters for interpretation. A conformational sensor and a proximity sensor are measuring different steps in the same sequence, so they can rank a set of ligands differently without either being wrong.

Complementation formats

Enzyme fragment complementation splits a reporter enzyme into an inactive small fragment and an inactive large fragment, and reconstitutes activity when the two fusion partners interact. In the beta-galactosidase implementation, the receptor carries a small optimized alpha-fragment peptide of roughly 4 kDa and beta-arrestin carries an N-terminal deletion mutant serving as enzyme acceptor; receptor activation drives complementation, and activation is quantified as restored beta-galactosidase activity, an approach validated on somatostatin receptor 2, with its potential application to high-throughput screening exemplified in both agonist and antagonist modes ^[4].

The enzymatic step is the trade-off. Turnover amplifies a small number of complementation events into a large signal, which is why these formats tolerate low expression and read well in a plate, but it also means the measurement integrates over the substrate incubation rather than reporting instantaneous occupancy. Split nanoluciferase formats take the opposite position, and published protocols describe them as monitoring reversible arrestin recruitment in living cells ^[1]; the design choice behind that reversibility — fragments engineered to have low intrinsic affinity for each other, so that complementation follows the interaction rather than driving it — is standard construct design rather than a result from any one study. If the comparison you care about is kinetic — onset rate, persistence, whether a signal decays — an integrating endpoint format cannot supply it, and the choice of read time becomes a hidden variable in any potency estimate derived from it. That last point is a consequence of the assay design rather than a finding from a specific study.

Imaging formats

Fluorescently tagged arrestin can simply be watched. Monitoring translocation of GFP-tagged arrestin from cytosol to activated receptors at the plasma membrane measures the pharmacological effect of a test compound, and because the readout is a whole-cell image it also carries high-content information about the compound's effects on cell processes ^[5]. Hudson and colleagues note that screening this way is best accomplished in clonal cell lines stably expressing an optimum level of the target receptor ^[5] — an explicit acknowledgment that the expression level is tuned to the assay, not inherited from biology.

Imaging supplies two things luminescence plate readouts cannot: spatial pattern, which separates diffuse membrane association from punctate vesicular structures, and per-cell distributions, which reveal whether a population mean is describing every cell or a responsive subpopulation. It costs throughput, and it moves a substantial part of the measurement into image segmentation and feature-selection choices that should be reported alongside the pharmacology.

From two curves to a bias factor

Bias is not read off a single curve. It is a comparison of comparisons. The standard treatment fits each concentration-response curve to the Black and Leff operational model, which separates affinity for the receptor from efficacy in activating a particular pathway, and combines them into a transduction coefficient, log(tau/KA). Kenakin and colleagues proposed this coefficient as a scale for evaluating selective agonist effects, and discussed both its independence with respect to receptor density and the calculation of statistical estimates of confidence in differences ^[6]. A stepwise practical account of fitting, computing and plotting these quantities, including limitations that may influence the conclusions drawn, has been published as a guide ^[7].

The arithmetic proceeds in two normalizations, and it is worth being explicit about which value sits in the denominator at each stage.

• Within a pathway. Subtract the reference ligand's transduction coefficient from the test ligand's: delta log(tau/KA) = log(tau/KA) test minus log(tau/KA) reference. The reference ligand is the denominator here. This step removes properties shared by every ligand in that pathway, such as coupling efficiency and receptor level.
• Between pathways. Subtract the two deltas: delta-delta log(tau/KA) = delta for pathway 1 minus delta for pathway 2. Pathway 2 is now the denominator.
• Back to a ratio. The bias factor is 10 raised to the delta-delta value.

A worked example, with values chosen purely for illustration and not taken from any dataset. Suppose a test ligand gives log(tau/KA) of 8.10 for cAMP and 6.90 for arrestin recruitment, while the reference agonist gives 8.60 and 8.20 in the same experiment. Then delta for cAMP is 8.10 minus 8.60, equal to minus 0.50; delta for arrestin is 6.90 minus 8.20, equal to minus 1.30. The delta-delta for cAMP over arrestin is minus 0.50 minus minus 1.30, equal to 0.80. The bias factor is 10 to the 0.80, equal to 6.3. Stated the other way round — arrestin over cAMP — the delta-delta is minus 0.80 and the factor is 10 to the minus 0.80, equal to 0.16. Same data, reciprocal number. A bias factor quoted without naming the reference ligand and the direction of the comparison is not interpretable.

Uncertainty travels with it, and it is not small. Continuing the illustration, if each delta carries a standard error of 0.15 log units, the standard error of their difference is the square root of 0.15 squared plus 0.15 squared, which is the square root of 0.045, or 0.212. A 95% interval on the delta-delta is 0.80 plus or minus 1.96 times 0.212, that is 0.384 to 1.216 — which back-transforms to a bias factor between 2.4 and 16.4. A point estimate of "roughly sixfold" carries an interval spanning almost an order of magnitude, and this is the routine case rather than a pathological one.

One structural caveat applies to the scale itself. Burgueño and colleagues point out that combining affinity and efficacy into a single parameter can yield compounds with a defined calculated bias that nonetheless display varying experimental preferences, and they propose an efficacy-driven log(tau) scale not as a replacement for log(tau/KA) but as a complement in situations where partial agonism is present ^[8].

Why bias is system-dependent

The normalizations above are designed to cancel system properties, and within a single experiment they largely do. They cannot cancel a system property that differs between the systems being compared. Several such properties are known.

Cell background. Cell type dependence is a hallmark of functional selectivity, and Kenakin and colleagues frame this as a particular problem for profiling agonists in whole-cell test systems removed from the system of interest ^[6]. This is the reason a bias factor should be reported with its cell line, not despite it.

Kinase complement. Arrestin recruitment depends on which GRK isoforms phosphorylate the receptor, and that routing is itself modifiable. Working in HEK293 cells with the beta2-adrenergic receptor and a C-tail mutant, Burghi and colleagues found that Galpha-s was not required for receptor and beta-arrestin conformational changes, arrestin recruitment or internalization, but that Galpha-s dictated which GRK isoforms were involved in arrestin recruitment ^[9]. A manipulation introduced as a G protein control can therefore alter the arrestin readout through kinase selection.

Expression of the machinery. The abundance of GRKs and beta-arrestins is not uniform. Matthees and colleagues analyzed GRK and beta-arrestin expression data across 61 tissues annotated in the Human Protein Atlas and argue that these expression profiles are underappreciated in most studies focused on the molecular mechanisms of receptor regulation ^[10]. A recombinant line's kinase and arrestin levels are a property of that line.

Other receptors in the same cell. Assessing 19 beta-adrenergic ligands across four pathways in a single cell background, van der Westhuizen and colleagues reported that in some cases apparent functional selectivity originated from off-target effects on other endogenously expressed adrenergic receptors, and emphasized assessing the selectivity of the responses before concluding that receptor-specific biased signaling had been observed ^[11].

The tag. A C-terminal fusion sits precisely where GRKs phosphorylate and where arrestin engages. Comparing the pharmacology of tagged and untagged receptor in a binding or functional assay is the conventional control for this, and it is convention rather than a result: the point is that a tagged construct is a different protein until shown otherwise.

What a recruitment assay does not establish

Recruitment is not evidence of arrestin-dependent signaling. Using HEK293 cells in which the Gs, Gq and G12 families of Galpha proteins were depleted by CRISPR/Cas9 and Gi/o was inactivated with pertussis toxin, Grundmann and colleagues reported that for the broad set of receptors they examined, arrestin recruitment and internalization still occurred in this "zero functional G" state, while ERK activation and label-free whole-cell responses failed ^[12]. Their finding is specific to that cellular system and that receptor set, and the authors present it as changing the understanding of how receptors activate ERK1/2 rather than as excluding arrestin-dependent events in general.

Nor does a bias factor identify a mechanism. It is an ordinal statement about two pathways, one reference ligand and one system. Michel and Charlton, reviewing the field, list the techniques for quantitative analysis of bias and the development of suitable screening assays among the technical challenges still facing the area ^[13]. Whether a bias factor measured in one recombinant background reproduces in another is an empirical question, settled by measuring it again, not by the calculation.

What to report

• Assay format and configuration: direct receptor-fused, bystander/translocation, complementation or imaging, and where each tag sits.
• Cell line, receptor expression level and how it was set, and whether GRKs or arrestins were co-expressed.
• Read time and whether the readout is instantaneous or integrated; for kinetic formats, the time window analyzed.
• The reference ligand, the direction of every ratio, and the pathway pairing.
• Fitted log(tau/KA) values with their standard errors for each pathway, not only the final bias factor.
• A confidence interval on the delta-delta value, back-transformed, alongside the point estimate.
• Evidence that the response is receptor-specific in the chosen background.

None of this makes a bias factor a stronger claim than it is. It makes the claim checkable, which is the only property that distinguishes a measurement from a number.