A structure-activity relationship table is the most compressed document in peptide chemistry. Twenty analogs, four columns, one page. It reads like a ranking, and that is exactly the trouble: a ranking implies one underlying quantity that each row has more or less of. What the table actually holds is a set of measurements, each the output of a particular assay run under particular conditions, comparable only to the extent those conditions were held constant. Most of the work of reading a SAR table well is reconstructing what was held constant and what was not.

Start with the denominator of the fold-change column

A fold-change column is a ratio, and a ratio is meaningless until you can name what is on the bottom. Three questions settle it. First, which quantity is being divided — EC50, IC50, Ki, or a maximal response? Second, which row is the reference: the parent peptide, the endogenous ligand, or a standard agonist that does not appear in the table at all? Third, which direction is the column running — a value of 40 may mean forty times less potent than the reference or forty times more, and tables are inconsistent about it. It is a convention each laboratory sets for itself and not every paper states. Where the caption does not say, the column is uninterpretable, and no care applied downstream will repair that.

A fourth question is quieter but matters more. Was the reference measured alongside the analogs, on the same plate and on the same day, or was it a historical value carried forward? A same-plate reference makes the ratio a within-experiment comparison, in which shared errors — a mis-set dispenser, a slightly warm incubator, a cell passage drifting — partly cancel. A carried-forward reference makes the ratio a comparison between two experiments, and inherits the full variability of both.

Potency, affinity and efficacy are three different quantities

The columns of a SAR table often mix them, and readers often collapse them. They are not interchangeable.

  • Affinity describes how tightly a ligand occupies its target, and is expressed as an equilibrium dissociation constant. A larger constant means weaker binding.
  • Potency describes the concentration producing a defined fraction of a response — an EC50 or an IC50. It is a property of the ligand and the system together.
  • Efficacy describes how much response the occupied receptor produces, which is what separates a partial from a full agonist.

That potency and affinity can come apart is not a subtlety; it can invert a ranking. In guinea-pig isolated hearts, using atrial monophasic action potential duration as the readout and Furchgott's irreversible-antagonist method to obtain equilibrium dissociation constants, two A1-adenosine receptor agonists were compared: R-PIA gave an EC50 of 145 nM with a dissociation constant of 288 nM, while CCPA gave an EC50 of 44 nM with a dissociation constant of 734 nM, and the authors described CCPA as a high intrinsic efficacy, low affinity agonist and R-PIA as a low intrinsic efficacy, high affinity agonist [1]. The compound with the lower EC50 was the compound with the weaker binding. A potency column sorted alone would have put those two rows in the opposite order from an affinity column sorted alone.

Efficacy and potency can also move independently of one another when the system changes rather than the ligand. In PC12 cells expressing human alpha-2B adrenergic receptors at two densities, 3.8 and 1.3 pmol per mg protein, with cytosolic calcium as the readout, the maximal response to agonists was greater in the higher-expressing cells while the EC50 values of norepinephrine and oxymetazoline were reported as independent of receptor level [2]. Same ligands, same readout, one variable changed — and the efficacy column moved while the potency column did not.

Why the same peptide gets different numbers in different laboratories

Cross-laboratory reproducibility of these measurements has been estimated directly. An analysis of independent measurements deposited in ChEMBL, after filtering out unit-transcription errors, undifferentiated stereoisomers and repeated citations of the same measurement — which accounted for 90% of all pairs — put the experimental uncertainty at a mean error of 0.44 pKi units, a standard deviation of 0.54 pKi units and a median error of 0.34 pKi units [3]. Because pKi is a base-ten logarithm, a mean error of 0.44 units is a factor of about 2.8 in Ki itself. The same group's companion analysis of IC50 pairs from ChEMBL found the standard deviation of IC50 data to be about 25% larger than that of Ki data, concluding that mixing IC50 values from different assays adds a moderate amount of noise, and that for a broad database a Ki-to-IC50 conversion factor of 2 was the most reasonable [4]. Those authors carry that scaling through themselves: for heterogeneous biochemical IC50 data from that database they report a standard deviation of 0.68 pIC50 units, and state that 0.68 corresponds to a factor of 4.8 — 68.2% of independent IC50 measurements agreeing within a factor of 4.8, even when measured in different laboratories under potentially different assay conditions [4].

The practical consequence is a rule about which comparisons the noise floor applies to. It applies to a number you have lifted from one paper and set beside a number from another. It does not apply in the same way to two rows measured side by side in a single table, where much of that variance is shared and cancels. This is why a two-fold difference within one well-controlled series can be real while a two-fold difference between two publications usually is not worth discussing.

Assay-to-assay comparability: the system is part of the measurement

Even when the ligand and the target are nominally identical, the cellular background can move potency by orders of magnitude. A meta-analysis of published concentration-response data for the P2Y13 receptor pooled EC50 values by system: for signaling events induced by heterologously expressed P2Y13, ADP-like agonists gave an EC50 of 17.2 nM (95% CI 7.7-38.5), while for functional responses of endogenously expressed P2Y13 the same agonist class gave 1.76 micromolar (95% CI 0.3-10.06) [5]. Dividing 1.76 micromolar by 17.2 nanomolar gives about 100. Two conditions differ between those pooled groups, not one — the expression system and the class of response measured — so the number is not attributable to receptor density alone. It is, however, a fair measure of how far apart two pooled literature EC50 estimates for the same agonist class can sit while both are correct within their own scope. That same analysis reported a Hill coefficient of 4.4 (95% CI 3.3-5.4) for ADP-like agonists at heterologously expressed P2Y13; slopes far from unity are worth noticing, since a steep curve narrows the concentration range over which potency differences can be resolved — a point of standard curve-fitting reasoning, not a finding of that meta-analysis.

The amplification between receptor occupancy and the measured endpoint is the general mechanism here. A response measured far downstream of the receptor, through several amplifying steps, will report a lower EC50 than a response measured close to it, for the same ligand at the same receptor. Nothing about the ligand has changed. This is why an EC50 without its assay is not a property of a molecule, and why the operational model of agonism was built to separate the two: it parameterizes a concentration-response curve into an affinity term and an efficacy term, and the resulting transduction coefficient, log(tau/KA), was presented as a scale that is independent with respect to receptor density and that can be evaluated statistically for differences between ligands [6]. That scale was proposed precisely because cell-type dependence of functional selectivity makes profiling in test systems removed from the system of interest difficult to interpret [6].

The operational model is not a solvent for every problem. Where partial agonism is present, combining affinity and efficacy into a single transduction coefficient can yield compounds whose calculated bias points one way while their experimental preference points another, which is the argument for reporting an efficacy-driven log(tau) scale alongside rather than instead of log(tau/KA) [7]. A single-number summary of a two-parameter fit will always discard something.

When the readout changes, the series can reorder

This is the failure mode that turns a SAR table into a wrong conclusion rather than a noisy one. A series ranked by one readout is not guaranteed to keep its order under another.

An opioid analog series makes the point concretely. Concentration-response data were collected in an adenylate cyclase inhibition assay and a beta-arrestin2 recruitment assay, combined with a receptor depletion approach, and fitted to the operational model to separate intrinsic efficacy from affinity in each pathway. Relative to the reference agonist DAMGO, several of the fentanyl analogs were more efficacious at inhibiting cAMP production, whereas all of the fentanyl analogs were less efficacious at recruiting beta-arrestin2 [8]. One set of molecules, one receptor, two readouts, and the efficacy ordering relative to the reference does not survive the switch.

Worse, the ordering can depend on when you look. For a series of agonists at the dopamine D2 receptor, the authors showed that kinetic context — set by ligand-binding kinetics together with the temporal pattern of the receptor-signaling processes being measured — can profoundly influence apparent bias and can even lead to reversals in its direction [9]. The practical reading, which the authors do not themselves state, is that an assay read at one time point and an assay read at another are for these purposes different assays, and that a table which does not report its incubation time has withheld a variable capable of moving the ordering.

The underlying reason a table cannot settle this is that its columns are closed-system, equilibrium descriptions. IC50 and Kd quantify the effectiveness of a receptor-ligand interaction under equilibrium assumptions, and the duration of the complex — its residence time — is a separate quantity that those measures do not carry [10]. Two analogs with identical dissociation constants may occupy the receptor for very different lengths of time, and no column in a standard SAR table will show it.

What a peptide series looks like when these distinctions are kept

A well-reported peptide SAR study separates the quantities rather than merging them, and is explicit about which of its numbers it measured and which it inherited. In a study of GPR15L analogs, the C-terminal 11-mer GPR15L(71-81) is described as having full efficacy but about 40-fold lower potency than the 57-mer full-length GPR15L(25-81) [11]. That comparison is background the paper carries in, not a result it reports: the matching sentence in its introduction is referenced to earlier work by these authors and two other research groups. What the paper measures under its own conditions is the analog series, tested at GPR15 stably expressed in HEK293A cells using a homogeneous time-resolved FRET-based Gi cAMP functional assay; truncation, alanine scanning and terminal capping identified the C-terminal alpha-carboxyl group and residues Leu78, Pro75, Val74 and Trp72 as contributing significantly to potency. The full-length peptide appears in this paper in a separate bioluminescence resonance energy transfer assay of G protein activation, run on wild-type and mutant receptors [11].

Two things follow. The efficacy and potency columns tell different stories about the same truncation, and only reporting both makes the fragment interpretable. And the provenance of the 40-fold figure decides what may be done with it: carried between publications rather than established side by side, it is exactly the case the cross-laboratory noise floor governs. Forty-fold is 1.6 log units, about three times the 0.54-log-unit standard deviation estimated for Ki reproducibility across laboratories [3] — large enough to survive that check, which was derived for Ki and not EC50 values and is an order-of-magnitude comparison only.

Read the footnotes before the numbers

The information that determines whether a table is interpretable usually sits below it.

  • What n counts. Three wells on one plate and three independent cell preparations are both reported as n = 3 and mean very different things. Distinguishing between biological, experimental and observational units is what those authors argue clarifies where replication should occur; in a survey of published animal experiments from 2011 to 2016 in which interventions were applied to parents and effects examined in offspring, only 22% (95% CI 17-29%) replicated the correct entity-intervention pair, 46% (38-53%) showed pseudoreplication and 32% (26-39%) gave insufficient information to judge [12]. That survey is about animal designs, not cell-based SAR, but the unit question it formalizes is the same question a plate-based assay raises.
  • How the curve was fitted. The placement of dilutions along a concentration-response curve, whether a model was fitted directly or by a null method, and whether one-site and two-site fits were compared statistically all bear on what a reported EC50 means [13].
  • Whether the error is on the log scale. Potency estimates are approximately log-normal, so a mean and standard error should be computed on pEC50 values, not on EC50 values in nanomolar. A table reporting "EC50 12 +/- 8 nM" has probably averaged the wrong quantity.
  • What the compound actually was. A nominal concentration assumes the vial contents. Salt form, residual solvent and peptide content all sit between the mass weighed and the moles in the well, and any error there shifts every potency value for that analog by the same factor.

What is not known

There is no general conversion that makes potency values from different assays commensurable. The Ki-to-IC50 factor of 2 discussed above was derived for a broad, heterogeneous public database and offered as a reasonable aggregate correction [4]; it is not a per-target constant, and applying it to one series would be a misuse of its scope. Nor is there an agreed way to tell, from a table alone, whether a rank order established in one readout will hold in another — which is why series intended to distinguish pathways are profiled in more than one, and why authors who quantify bias increasingly report the model parameters rather than only the fitted curves [6, 7]. Much of what makes a table trustworthy is convention rather than evidence: running the reference on the same plate, keeping incubation times uniform across a series, reporting statistics on the log scale. These are practices that laboratories adopt because the alternatives have burned them, not because a controlled study established a threshold.

The one habit worth carrying into every table is the least technical. Before comparing any two rows, ask what differs between them besides the structure. If the answer is nothing, the comparison is about chemistry. If the answer is anything else, the comparison is about the assay.