Report 30 September 2026

Dissociation rate drives binding affinity differences among optimized cetuximab-derived EGFR binders

We went back to the raw surface plasmon resonance data behind our results in the 2026 EGFR/cetuximab binder competition, where 3 of our 10 designs surpassed the previous world best. 7 designs bind more tightly than the starting antibody, the 10 split into a strong group and a weak group, and our top design lets go of its target most slowly. The exact binding constants, the order among the best designs and how much tighter they bind than the parent or the published designs need a follow-up experiment.

Interlit Team 30 September 2026 17 min read

Ink and oil bubbles drifting through blue and violet liquid

In the 2026 EGFR/cetuximab binder optimization competition, 3 of our 10 designs surpassed the previous world best. We described the result in an earlier post: our top design bound its target 43.9 times more tightly than cetuximab, the approved antibody we started from, and 1.6 times more tightly than the previous best design, which bound 26.9 times more tightly than cetuximab. Because a headline number is only as reliable as the measurement behind it, we went back to the raw measurements to see which conclusions survive a close reading. This post reports what we found, including what still needs checking.

Three conclusions hold: 7 of our 10 designs bind the target more tightly than cetuximab, the 10 split into a strong group of 7 and a weak group of 3 that are 21-fold apart, and our top design lets go of its target more slowly than any other design.

Three questions are still open: the exact binding constants of the tightest designs, which of the 7 strong designs binds most tightly, and how much tighter our designs bind than the parent or the published designs, which were measured on other instruments or in other labs. The last question covers the 43.9-fold headline figure and the 1.6-fold lead over the previous best design, which is about the size of the 1.46-fold replicate reproducibility, the typical spread in $K_{D}$ between repeat measurements of the same design. A follow-up experiment would answer all three, and the last section describes it.

A lab measured how tightly each of our ten designs binds its target

Each design is a 241-residue single-chain variable fragment (scFv), a small antibody fragment made of a heavy-chain variable domain, a (GGGGS)3 linker and a kappa light-chain variable domain, and it binds domain III of the outer part of the EGFR receptor. The competition required every design to differ from the cetuximab scFv by at least 5 residues, so none is a copy of the starting antibody. We refer to the designs by the rank our pipeline gave them before any measurement, so rank #1 is the design we expected to perform best.

Adaptyv Biosystems, the competition's measurement provider, measured each design by surface plasmon resonance (SPR), which reports how tightly a design holds its target and how fast it attaches and lets go. Each design was measured in 2 replicate concentration series (repeated sets of measurements at increasing concentrations), which gives 20 series across 3 instrument runs, where a run is one session on the instrument. The same provider reported the cetuximab scFv we started from, which we call the parent, at 9.94 nM using a different method, biolayer interferometry (BLI).

We report tightness as the equilibrium dissociation constant $K_{D}$ , in nanomolar (nM). $K_{D}$ is the target concentration at which half of the binding sites are occupied, so a smaller $K_{D}$ means tighter binding.

SPR measures how fast a design attaches to its target and how fast it lets go

In SPR, one binding partner is fixed to a sensor surface while the other, the analyte, flows over it, and the instrument records the response $R (t)$ in response units (RU). Here the design is fixed and EGFR is the analyte. The response rises as complexes form and falls when the flow switches to buffer, and the standard model for this behavior is a 1:1 Langmuir interaction:

\frac{d R}{d t} = k_{on} \cdot C \cdot (R_{\max} - R) - k_{off} \cdot R

Here $C$ is the analyte concentration, $k_{on}$ is the association rate constant (how fast binding happens), $k_{off}$ is the dissociation rate constant (how fast complexes fall apart), and $R_{\max}$ is the response when every surface site is occupied. The equilibrium dissociation constant is $K_{D} = k_{off} / k_{on}$ .

We took the fitted $k_{on}$ , $k_{off}$ and $R_{\max}$ from the provider's own fits. The other quantities in this post come from our own analysis of those fits and of the raw response curves, which the provider's complete data package supplies for all 20 series, together with the blank series of all 3 runs. With these quantities in hand, the first thing to check is how each design compares with the parent.

Seven designs bound more tightly than the parent and three bound less tightly

The 7 improved designs bind 6.9 to 44 times more tightly than the parent, while the 3 weaker designs bind 3.1, 14.1 and 14.2 times less tightly. In energy terms, the improvements correspond to a change in binding free energy of −1.14 to −2.24 kcal/mol, where a more negative number means tighter binding. This comparison rests on a BLI measurement of the parent and SPR measurements of our designs, a limit we describe below. The plot below shows the $K_{D}$ of both replicates of every design against the parent and 2 published scFv values (taken up in a later section), so you can see how closely the replicates agree and which designs sit below each reference.

The designs split into a strong group and a weak group, 21-fold apart

The 7 strong designs have a $K_{D}$ between 0.23 nM and 1.45 nM, and the 3 weak designs lie between 31 nM and 141 nM, a 21-fold gap between the 2 groups. That gap is larger than every effect we could quantify. The largest disagreement between 2 repeats of the same design was 2.5-fold, the amount of immobilized design on the sensor surface changed from run to run and shifts the apparent $k_{on}$ by 2.65-fold across the observed range (95% confidence interval 1.97 to 3.57), and switching between binding models changed the fitted values by at most 65%. Two sources of error remain unquantified: the sensor and buffer conditions, which the provider did not record, and the background signal in runs 1 and 3, which the checks below describe. Because all of our designs were measured by SPR, the split into 2 groups does not depend on the BLI measurement of the parent.

Our top design lets go of its target more slowly than any other design

Most of the difference between designs comes from how slowly they release the target, not from how fast they attach. Across the 20 replicates, $k_{off}$ spans 393-fold while $k_{on}$ spans only 4.8-fold. Decomposing the variance of $\ln K_{D}$ , 81% comes from $k_{off}$ , 3% from $k_{on}$ , and the remaining 16% from the way the two vary together.

Rank #1 has the slowest release of the 10. Its 2 $k_{off}$ measurements agree to 1%, and both lie below every value recorded for any other design. After 600 seconds of buffer flow, its response still holds 77% (first replicate) and 76% (second replicate) of its plateau, the level reached at the end of the binding phase. The plot below places each replicate by both rate constants, so you can compare the two on one set of axes and find rank #1 at the slowest-release end.

Six of the seven improved designs bind more tightly than both published scFv values, but the comparison is only indicative

The 2 published scFv values we compare against are monovalent, meaning each molecule has one binding site, and both were measured by BLI: 1.21 nM for a framework-only variant from Cradle, and 0.86 nM for a variant from Synthyra, whose model card reports the BLI result obtained through Adaptyv. Sensor chemistry, immobilization density, concentration series, association and dissociation times and fitting strategy all shape a kinetic result, and we do not hold those conditions for the published measurements.

The Converge result, 0.315 nM, comes from a full-length IgG1 antibody captured on the sensor, with EGFR-His as the analyte, on a different SPR instrument. An IgG has 2 binding arms, and whether they bind together depends on the geometry and dynamics of the measurement, such as the local density of the target, the distance between binding sites and the analyte concentration. Its value therefore cannot be placed next to a monovalent scFv without a matched format.

Ranking the seven strong designs needs a follow-up experiment

The 7 strong designs span 6.4-fold in $K_{D}$ , whereas the replicate reproducibility, the geometric standard deviation of $K_{D}$ between the 2 replicates of a design, pooled over the 7, is 1.46-fold (95% confidence interval 1.29 to 2.17; the standard deviation of $\log_{10} K_{D}$ is 0.165). The designs at the 2 ends of the group are therefore separated by more than that, while designs close together in the group may not be. Rank #1 has the lowest fitted $K_{D}$ in both of its runs, but $K_{D}$ also depends on $k_{on}$ , the less certain of the 2 constants because it shifts with the amount of immobilized design on the surface. We therefore claim the slowest release for rank #1, not a resolved lead in $K_{D}$ over the designs behind it.

Exact binding constants below 1 nM need an experiment designed for them

The model-based analyses we ran put all 7 strong designs below 1.5 nM and the 6 tightest at roughly 0.2 nM to 0.7 nM under the main fit. The steady-state check cannot resolve values below 1 nM and gives estimates between 0.84 nM and 4.2 nM. Because the analyses share assumptions, we treat the absolute binding constants as provisional. One shared assumption is that the slow decay reflects the intrinsic $k_{off}$ , which can fail when released molecules bind again before leaving the flow cell, a process called rebinding.

Comparing with the parent and the published designs needs a head-to-head experiment

The parent and both published scFv values were measured by BLI and our designs by SPR, and we do not know the offset between the 2 methods for this system. We did not assume one. Instead, we report how large the offset would have to be to erase each result: for rank #1 against the parent, SPR would have to read 43.9 times tighter than BLI, and for the weakest improved design the figure is 6.9 times. The larger the required offset, the more securely the conclusion holds.

The 1.6-fold lead of rank #1 over the previous best design has the same limit because the 2 values come from different measurements, and it is about as large as the 1.46-fold replicate reproducibility, so we treat it as a lead we have not established.

Two measurement effects could distort the numbers, and we checked both

The first effect is that the amount of immobilized design on the sensor surface, the surface density, changes the apparent $k_{on}$ . Apparent $k_{on}$ fell in the later run by a factor of 0.69 across the 9 designs with replicates in different runs (p = 0.027), a drop that could be chance once we correct for testing 4 quantities at once. $R_{\max}$ , the binding capacity of the surface, rose by a factor of 2.24 (p = 0.0049), which survives the correction and shows that the surface capacity grew with run order.

Among the 14 series of the 7 strong designs, $\log k_{on}$ falls as $\log R_{\max}$ rises. Within run 2 alone the correlation is −0.92 (95% confidence interval −0.99 to −0.45, 6 series), and within run 3 alone it is −0.65 (95% confidence interval −0.96 to +0.34), which cannot be told apart from no correlation. After removing the differences between designs, adding $R_{\max}$ explains 94% of the remaining variance in $\log k_{on}$ , whereas adding run identity explains 80%. With 14 observations and 7 design terms these tests have few degrees of freedom, and $R_{\max}$ is confounded with run order. The plot below shows both rate constants of those 14 series against $R_{\max}$ , with a switch between them, so you can compare how each depends on surface density; the $k_{on}$ view carries the mass-transport curve described next.

The pattern is consistent with partial mass-transport limitation, a known SPR effect in which analyte near the surface is used up faster than it can be replaced. The data do not establish that mechanism, because they cannot separate the amount of immobilized design from the run itself or from designs that behaved differently in different runs.

The second effect is that mass transport and rebinding push $K_{D}$ in opposite directions, so the net bias is unknown. Transport limitation lowers the apparent $k_{on}$ and raises the apparent $K_{D}$ , and it also promotes rebinding, which slows the apparent decay and so lowers both the fitted $k_{off}$ and the apparent $K_{D}$ . These data cannot say which effect wins.

To test whether $k_{off}$ depends on surface density, we regressed $\log k_{off}$ on $\log R_{\max}$ with design identity removed. The fitted coefficient is −0.085 (95% confidence interval −0.254 to +0.083, p = 0.26), against −0.519 for $k_{on}$ (p < 0.001), so the data are consistent with no dependence of the release rate and do not exclude a modest one.

Correcting each series for transport reduces the pooled within-design spread of $\log_{10} K_{D}$ , the typical disagreement between repeats, from 0.165 to 0.132, a 20% reduction. The correction addresses $k_{on}$ only. It narrows the disagreement for some designs, such as rank #9 (from 1.52-fold to 1.01-fold), barely changes rank #6 (from 2.46-fold to 2.29-fold), widens it for rank #1 (from 1.78-fold to 1.93-fold), and does not change the ranking of the top 2 designs. We treat the corrected values as a sensitivity analysis and keep the reported values.

The fitted intervals describe how tightly one curve constrains a parameter given the model. They run 1% to 3% of the fitted value and so understate what repeating the measurement would show, which makes the 1.46-fold spread between repeats the realistic uncertainty when comparing designs.

We repeated the key result under five different assumptions, and each check has a limit

Two checks refit the decay phase of the raw curves under a different assumption, a third examines the background signal, a fourth uses the equilibrium form of the same 1:1 model, and a fifth fits a different binding model. Each rests on assumptions of its own, and none is free of a model. Both decay-phase checks begin 10 s after the nominal end of the injection, because the response of the weak designs first falls between 306 s and 315 s (27 reads, median 311 s). A window that starts at 300 s changes both results, giving a slow share of 9.7% to 98.6% in the two-exponential fit and a floating-baseline $k_{off}$ of up to $3.1 \times 10^{- 3}$ per second.

Allowing 2 release speeds does not change the result. Fitting each dissociation phase with 2 exponentials, one fast and one slow, assigns 93% to 99.8% of the decay amplitude to the slow component across the 12 series of the 6 tightest designs, so a small tightly bound fraction does not dominate the tail. The check cannot tell an intrinsic slow $k_{off}$ from a slow tail produced by rebinding or by an uneven surface.
Letting the baseline float makes the designs look somewhat weaker but not weak. Fitting the dissociation phase with a floating baseline raises the median $k_{off}$ of each series to between $1.1 \times 10^{- 3}$ and $3.0 \times 10^{- 3}$ per second, yet even under this most pessimistic reading of the decay phase, $K_{D}$ stays at or below 1.4 nM for every series of the 7 strong designs.
Background signal is small in run 2 and in run 3 up to 316.2 nM, but not in run 1 or at 1000 nM in run 3. Each run has a blank series, a measurement with buffer only, and the blanks peak at 15.2 RU in run 1, 3.5 RU in run 2 and 14.4 RU in run 3, against maximum sample signals of 55 to 847 RU across the 20 series. The largest rise of the blank within a read is at most 17% of the association response of any series in run 2, and at most 7% in run 3 from 10 nM to 316.2 nM, but it reaches 49% at 1000 nM in run 3. In run 1 it is 27% to 61% of the smallest association response at 100 nM, 316.2 nM and 1000 nM, and at 10 nM and 31.6 nM it exceeds the response of the rank #7 series measured in that run (8.3 RU and 4.0 RU against 3.8 RU and 3.7 RU). The data package does not say whether the sample curves were corrected for the blank, so we used them as supplied.
A steady-state estimate agrees in range but cannot resolve values below 1 nM. Plateau ratios at the 2 lowest concentrations give estimates between 0.84 nM and 4.2 nM. This is the equilibrium form of the same 1:1 model, and the lowest concentration already occupies 87% to 98% of sites for the strong designs, so the check confirms that equilibrium was reached and does not test the model.
One design fits better when one molecule binds 2 surface sites, and the data cannot settle whether it does. Rank #6 fits best with a model in which one analyte binds 2 sites, at 2.16-fold lower residual at both of its surface densities, but its fitted first-step affinity differs by 90-fold between the 2 replicates, so the curves do not determine the model. Whether such crosslinking occurs also depends on the spacing of the immobilized designs, the concentration and the geometry, and 2 surface densities cannot test it. We neither confirm nor exclude an avidity contribution, which would be extra apparent tightness from binding 2 sites at once.

Our pre-submission ranking did not tell the working designs from the failing ones

The ranking our pipeline produced before any measurement did not anticipate which designs would fail: the 3 weaker designs sit at ranks 2, 5 and 7, and failed and successful designs do not separate by rank (Mann–Whitney p = 0.67). Among the 7 designs that succeeded, the rank correlation with measured affinity is +0.50 (p = 0.25), a weak association that chance could produce, and across all 10 designs it is +0.152 (p = 0.676). We have written about what our ranking got right and wrong in an earlier post.

The next measurement should change five things

Immobilize the design at low density, with $R_{\max}$ below 50 RU, which the fitted transport model predicts recovers $k_{on}$ to within 13% of its transport-free value, and measure the leading designs at 2 or more densities within one run. This separates density from run and tests directly whether $k_{off}$ depends on density.
Use a concentration series from 0.05 nM to 5 nM, so that the plateau is informative for sub-nanomolar designs.
Record dissociation for at least 1,800 seconds, so that the slow decay of the tightest designs is observed over more than one half-life (about 21 minutes for rank #1).
Include the parent and the published variants in the same run, on the same surface, in the same format, because that is the only way to establish how much tighter our designs bind than they do.
Confirm the top designs with a solution-phase method, which has no surface-density effects, and confirm the epitope, the exact site on EGFR where each design binds, by competition against cetuximab.