Surface plasmon resonance does not measure binding. It measures the refractive index in a thin layer above a gold film and reports it as a response unit which, for most biological material, tracks mass at the surface. Everything else — an association rate constant, a dissociation rate constant, an affinity — is inferred from that trace under a model you chose. The model is where the errors live.

What the sensorgram is

One partner, the ligand, is attached to the sensor surface. The other, the analyte, flows over it. Response is recorded continuously: a baseline in running buffer, a rise during injection as complex forms, a decay once injection ends. The rise is not association alone — it is the net of association and dissociation. Only the decay isolates dissociation, and only if rebinding is suppressed.

Two conventions, stated without citation because they are instrument practice, not findings. Response units are scaled so a given number is roughly the same surface mass density whichever protein produced it, which is what makes the arithmetic below possible. And every trace should be double-referenced: subtract a reference channel carrying the same chemistry without ligand, then a buffer-only injection, removing bulk refractive index shifts rather than fitting them.

Immobilization chemistry and orientation

Amine coupling — activating surface carboxyls and letting them react with lysine side chains and the N-terminus — is the default because it needs no modified reagent. It is also the least controlled. Coupling happens wherever an amine reacts first, so the surface population is a mixture of orientations, some with the binding site facing the matrix or blocked by the bond itself. A review of the technique's artifacts placed immobilization of the binding partner alongside steric hindrance between adjacent sites at the sensor surface and the finite rate of mass transport among the identified sources of distortion in measured kinetics and equilibria, each with suggested control experiments [1].

Directed attachment is the alternative: an engineered cysteine for thiol coupling, a site-specific biotin captured on streptavidin, an affinity tag captured by an immobilized partner. Comparing strategies for two single-domain antibodies, direct covalent attachment and a randomly biotinylated monomer on NeutrAvidin were the poorest performers, while a single-domain antibody fused to a streptavidin core tetramer gave the highest density of active molecules, by at least a factor of two over the next construct [2]. Single-domain antibodies are not peptides and the ranking should not be assumed to transfer; what transfers is that active surface density, not total immobilized mass, is what matters.

The surface itself has chemistry. Carboxylated sensor surfaces are anionic at working pH and basic analytes adsorb to them — so reliably that one group turned the liability into a method, using the electrostatic interaction between naturally positively charged basic proteins and a negatively charged carboxylated chip to immobilize ligand with no modification at all, with sodium dodecyl sulfate for complete regeneration [3]. Many research peptides carry net positive charge at neutral pH. If yours does, the same electrostatics appear as nonspecific binding on the reference channel — and a reference channel that binds analyte is not a reference.

Which partner goes on the chip

Two things decide it. The first is valency. In a study of a monoclonal Fab binding influenza N9 neuraminidase, the neuraminidases — homotetramers of 190 kDa — were immobilized specifically to avoid multivalent binding, leaving the monovalent Fab as analyte [4]. Putting the multivalent partner on the surface turns avidity into a surface-density effect you control, not an apparent affinity you cannot interpret.

The second is signal, and it is arithmetic:

Rmax = (RU of ligand immobilized) × (MW of analyte ÷ MW of ligand) × binding stoichiometry

The denominator is the molecular weight of the immobilized ligand, not the analyte and not the complex. Getting it backwards inverts the conclusion. Work it with reagents from a benchmark: acetazolamide at 222 g/mol as analyte, carbonic anhydrase II at 30,000 Da as ligand [5]. Suppose 3,000 RU of enzyme and 1:1 binding. Then Rmax = 3,000 × (222 ÷ 30,000) = 22.2 RU. Flow a 30,000 Da protein over the same surface and Rmax = 3,000 RU — a 135-fold difference in dynamic range, purely from analyte size.

For a peptide the choice can go either way. A 2,500 Da peptide as analyte over a 50,000 Da receptor construct immobilized at 2,000 RU gives Rmax = 2,000 × (2,500 ÷ 50,000) = 100 RU. Reverse it — peptide immobilized at 30 RU, 50,000 Da protein as analyte — and Rmax = 30 × (50,000 ÷ 2,500) = 600 RU: six times the signal from about one sixty-seventh of the immobilized mass, 30 RU against 2,000 RU. In molar terms the peptide surface carries roughly 3.3-fold fewer binding sites, each of which reports a 20-fold heavier analyte. That is the argument for putting the peptide on the chip, against the risk that a peptide tethered at an uncontrolled point no longer folds or binds.

The two rate constants, and the one derived from them

The association rate constant ka is second order, in units of per molar per second. The dissociation rate constant kd is first order, per second, and independent of concentration. The equilibrium dissociation constant is their quotient, KD = kd ÷ ka. Only two of the three are independent.

Because kd is first order it converts directly into a complex half-life, ln(2) ÷ kd. Take the averaged constants from a benchmark in which 36 investigators measured acetazolamide binding to carbonic anhydrase II under similar conditions: ka of 3.1 ± 1.6 × 10^(6) per molar per second and kd of 6.7 ± 2.5 × 10^(−2) per second, corresponding to an average KD of 2.6 ± 1.4 × 10^(−8) M [5]. That kd gives a half-life of 0.693 ÷ 0.067 = 10.3 s, and falling to 5% remaining takes ln(20) ÷ 0.067 = 44.7 s, so a 60 s dissociation phase is ample here. Make kd a hundredfold smaller, 6.7 × 10^(−4) per second, and 5% remaining takes 2.996 ÷ 0.00067 = 4,471 s, about 75 minutes; the same 60 s phase would then report little but baseline.

Note too that 6.7 × 10^(−2) ÷ 3.1 × 10^(6) = 2.2 × 10^(−8) M, not the 2.6 × 10^(−8) M reported: the mean of a set of ratios is not the ratio of the means, and a report should say which it quotes.

Kinetics carry information affinity does not. Comparing two arylsulfonamides against carbonic anhydrase II by surface plasmon resonance, isothermal titration calorimetry and stopped-flow fluorescence, one group found the enzyme bound one compound with twofold higher affinity, yet kinetic analysis revealed the lower-affinity compound had a fourfold slower dissociation rate; they also reported that when the biosensor experiments were performed with care, the constants matched those acquired in solution [6]. Two compounds can rank one way on KD and the opposite way on complex lifetime.

Mass transport limitation

Analyte must diffuse from the flowing stream to the surface before it can bind. If binding is fast enough the layer next to the surface is depleted, and the observed rate reports delivery rather than chemistry. Conventional analyses may yield arbitrary apparent rate constants lying between the mass transport rate constant and the true intrinsic chemical rate constants, depending on the concentrations used; a two-compartment description of transport was introduced to indicate how far a curve is transport controlled and, in favorable cases, to recover correct rate constants 10- to 100-fold larger than earlier analyses could evaluate [7].

The experimental controls are ligand density and flow rate. In the neuraminidase work a contribution of mass transport to the kinetic constants was demonstrated at higher surface densities and low flow rates, and minimized at low ligand densities and relatively high flow rates, up to 100 microliters per minute; a transport-corrected model fitted to transport-affected data gave constants similar to a plain 1:1 model fitted to data in which transport had been minimized experimentally [4]. Separately, both the immobilization matrix and the flow rate were experimentally observed to influence the effective transport rate constant in antibody-antigen models [8].

Of the two levers, density is the stronger. Under the standard laminar-flow scaling for a rectangular flow cell — textbook transport theory, stated without citation — the transport rate constant rises with roughly the cube root of volumetric flow, so a tenfold flow increase buys only 10^(1÷3) = 2.15-fold more delivery, while halving ligand density halves demand directly. The diagnostic follows: run the same concentrations over two or three surfaces of different density. If fitted rate constants move with density, they are not rate constants.

Corrective models have limits. The compartment approximation was found to describe moderately transport-limited data well but to break down for strongly transport-limited binding, and in that regime the authors reported experimental evidence for gradients forming within the sensing volume of the evanescent field biosensor they used [8]. A transport term in the software is no substitute for a surface that does not need one.

Heterogeneity, and why a 1:1 fit can still look good

Mass transport limitation and heterogeneity of the surface sites are described as the two most common sources of deviation from ideal pseudo-first-order bimolecular kinetics, and as problems intrinsic to using a biosensor surface rather than avoidable mistakes [9]. Randomly coupled ligand is a population, not a species: different attachment points, environments, accessibility.

The uncomfortable part is that a heterogeneous surface under transport limitation can produce curves a 1:1 model fits with small residuals and confident-looking parameters. Computational tools exist to determine a distribution of affinity and kinetic rate constants from surface binding progress curves rather than assume one class of sites [8]; that they were needed is the point. A good fit to a single-site model is weak evidence that the surface has a single class of sites. The honest tests are external: does the fit hold across densities, across flow rates, and across a series wide enough to constrain the plateau?

Regeneration, and the option of not regenerating

Between cycles the surface must return to baseline. A regeneration buffer has to break every interaction between ligand and analyte while maintaining the ligand's active binding site — two requirements pulling opposite ways, which is why regeneration is scouted empirically, not predicted. One approach combined a continuous flow microspotter with surface plasmon resonance imaging to test 48 regeneration buffers simultaneously on one biosensor, finding optimal conditions within hours; the authors describe the workflow as applicable to any ligand coupled through amine, thiol or streptavidin immobilization [10]. The control is always the same: repeated identical injections should give repeated identical responses. Downward drift means the surface is being consumed, and every later cycle is a different experiment.

Some surfaces cannot be regenerated without being destroyed. Kinetic titration — sequentially injecting an analyte concentration series with no regeneration steps — was examined by a combination of simulation and experimentation, and the authors report that the method can be as robust as the classical one, that kinetic titrations can be more efficient than the conventional data collection method, and that they allow full characterization of analyte binding to ligand surfaces that are difficult to regenerate [11]. It is the natural first choice for a fragile immobilized peptide.

Why a low-molecular-weight analyte is hard

Steady-state analysis — fitting an isotherm to plateau response against concentration — is the conventional fallback when kinetics will not resolve. It is convention rather than a cited finding, and it keeps the affinity while losing the residence time. Even the plateau has to be visible, though. Return to that 22.2 RU Rmax. Half-saturation is 11.1 RU, and a 5% error on that reading is 0.56 RU. Every refractive index mismatch, temperature excursion and trace of nonspecific binding competes with a signal that size. Nothing about the chemistry is harder for a small analyte; the signal tracks mass, and there is less of it.

Two benchmarks bound the consequence. When 26 investigators characterized 10 sulfonamide inhibitors binding carbonic anhydrase II, the experimental errors in ka, kd and KD averaged 34%, 24% and 37% respectively, the greatest variation being reported for compounds of exceptionally weak affinity, fast association rates, or both; the constants nonetheless correlated well with solution-based titration calorimetry [12]. The first two figures are worth combining. Were those errors independent, the propagated error on their quotient would be the square root of 0.34 squared plus 0.24 squared, or 42%. The reported 37% is smaller — consistent with the two constants being positively correlated in a global fit, as parameters from one curve generally are.

The acetazolamide study makes the design point. Each of the 36 participants created three different density surfaces and analyzed an identical dilution series from 4.1 to 1000 nM, fitted globally to a 1:1 model including a mass transport term; the greatest variability arose at the immobilization step, since each investigator provided their own surface activating reagents [5]. The habit is to span roughly a tenth to ten times KD — convention, not a cited requirement. Against the reported average KD of 26 nM, that series spans about 0.16-fold to 38-fold KD, stopping short of the habit at the bottom and running well past it at the top — the bracket a low-signal analyte needs to fix both slope and plateau.

Size is not the only source of scatter. In a global benchmark where 150 participants from 20 countries received the same protein samples — a 50 kDa Fab and a 60 kDa glutathione S-transferase antigen — most collected responses that could be fit, and once a few outlying data sets were removed the average reported affinity across the remaining panel was 620 pM with a standard deviation of 980 pM [13]. A standard deviation larger than the mean of a strictly positive quantity means the distribution is right-skewed, so that summary is not a symmetric interval. The authors concluded that when the assay was designed and executed appropriately, the rate constants were consistent and independent of which protein was immobilized and which instrument was used.

What is convention and what is unresolved

Several statements above are practice rather than evidence, and are flagged where they appear: the response-unit-to-mass scaling, double referencing, the cube-root flow scaling, steady-state analysis as fallback, and spanning a tenth to ten times KD. Every arithmetic example is worked from stated suppositions and published molecular weights; no supposed immobilization level is a specification.

What remains unsettled is more basic. There is no general method for knowing the true distribution of orientations and activities on a randomly coupled surface — it is inferred from binding behavior, not observed. There is no accepted criterion for how much drift across densities is tolerable before a fit must be rejected. And a good fit to a 1:1 model is weak evidence for a 1:1 mechanism: transport limitation and heterogeneity distort curves in partly compensating directions.

A defensible report is not a KD alone. It states:

  • the immobilization chemistry, the level in RU, and the theoretical Rmax
  • which partner was ligand, the valency reason, and what the reference channel carried
  • densities and flow rates tested, and whether fitted constants moved between them
  • the regeneration condition and its cycle-to-cycle stability, or kinetic titration
  • the concentration series as multiples of fitted KD
  • overlaid experimental and fitted curves with residuals

None of that guarantees a correct number. It makes an incorrect one visible.