Specificity is a shape, not a switch
The textbook antibody binds one antigen. It is the lock to a single key, and everything else is “cross-reactivity” — a defect, a footnote, noise to be subtracted. This framing is convenient and, for many purposes, wrong. Real antibodies bind a spectrum of epitopes with a spectrum of affinities, and polyreactivity is the rule the lock-and-key picture was built to ignore1,2.
This note argues for a different primitive object. Instead of asking what does this antibody bind?, we ask how is this antibody’s binding free energy distributed across the space of all epitopes? Specificity then stops being a label and becomes a property of a distribution’s shape.
The one equation you need — and the term it usually hides
An antibody in serum is not a system that occupies one epitope. It sits in contact with a reservoir of epitopes at finite concentration, so the honest starting point is grand‑canonical . For a single paratope exchanging ligand with that reservoir, the occupancy of epitope \(x\) is
\[\theta(x)=\frac{1}{1+e^{\beta(\varepsilon(x)-\mu)}},\qquad \mu=\mu^{\circ}+k_BT\ln\frac{c}{c^{\circ}},\qquad \beta=\frac{1}{k_BT},\]
the Langmuir isotherm in Fermi form. The chemical potential \(\mu\) is the free-energy cost (or availability) of taking one ligand molecule from the surrounding solution. It determines how strongly bulk ligand abundance favors occupancy of a binding site. Equivalently, for one site facing mutually competing ligand species with activities \(\require{action}\mathtip{\color{#EE00EE}{a_x}}{\color{#b24c00}{molar\ concetration\times\ activity\ coefficient\ \gamma}}\) and standard binding free energies \(\Delta G^{\circ}_x\), the binding polynomial is \[\Xi=1+\sum_x a_xe^{-\beta\Delta G^{\circ}_x} \], whose leading \(1\) is the unbound state and whose terms give \[\theta_x=a_xe^{-\beta\Delta G^{\circ}_x}/\Xi\]. Thus, a plentiful weak ligand can occupy more antibody than a scarce strong one.
Now take the limit an antibody microarray is normally run in — far from saturation, \[\beta(\varepsilon(x)-\mu)\gg0\] for every spot. Then \[\theta(x)\approx e^{\beta\mu}e^{-\beta\varepsilon(x)}\], and if we condition on the paratope being engaged at all, the factor \[e^{\beta\mu}\] cancels:
\[p(x)=\frac{c(x)\,e^{-\beta\varepsilon(x)}}{Z_{ep}},\qquad Z_{ep}=\sum_x c(x)e^{-\beta\varepsilon(x)}=\int d\varepsilon\,g_c(\varepsilon)e^{-\beta\varepsilon}.\]
This is the equation this notebook is built on, and it is worth being exact about what it is. \(p(x)\) is not the probability that the antibody is bound to \(x\). It is the probability that the epitope it is bound to is \(x\), given that it is bound, in the linear regime of the isotherm. \(Z_{ep}\) is not a thermodynamic partition function of any physical system — it is the normalizing constant of that conditional law, and derivatives of \(\ln Z_{ep}\) describe the conditional law, not the antibody’s free energy in serum. Use \(p(x)\) for questions about the shape of a specificity, and \(\theta(x)\) for questions about what an experiment will read out.
And note which density of states appears: \[g_c(\varepsilon)=\int dx\,c(x)\,\delta(\varepsilon-\varepsilon(x))\] is concentration‑weighted. This is normally the fatal complication, and it is why a Boltzmann reading of binding is usually a dead end for natural antigens — you never know \(c(x)\). It is also the deepest methodological argument for the platform used here. A near‑uniform random‑peptide phage library and an equimolar peptide microarray are physical realizations of \(c(x)=\mathrm{const}\). On those platforms, and essentially only on those platforms, the measured spectrum estimates \(g(\varepsilon)\) itself rather than \(g_c(\varepsilon)\). The mimotope library is not merely a convenient sample of epitope space; it is an instrument for measuring a density of states.
Being explicit about the conditioning buys the framework predictions rather than costing it anything. Two follow immediately. Specificity is a shape at a given chemical potential. As \(\mu\) rises, epitopes saturate from the strongest downward, marginal weight moves to weaker epitopes, and the effective number of epitopes engaged increases: the same monoclonal is sharply specific when scarce and visibly polyreactive when abundant, with the crossover near \(\mu\approx\varepsilon_{min}\). An antibody does not have a specificity; it has a titration curve, and a dilution series on one array measures it. Affinity and abundance are formally interchangeable. Clone \(i\) enters with \[\mu_i=\mu_i^{\circ}+k_BT\ln c_i\], so a hundred‑fold abundance advantage substitutes for \[\ln 100\approx 4.6\,k_BT\] of binding energy. That is the quantitative reason abundant polyreactive IgM can dominate an array over rare high‑affinity IgG, and it predicts where the two isotypes should cross over: at a concentration ratio of \[e^{\beta\Delta\varepsilon}\].
For intact antibodies this remains a baseline rather than a complete model: bivalency, IgM multivalency, epitope density, rebinding and steric geometry can make apparent affinity a nonlinear avidity effect. And \[\Delta G^{\circ}=\Delta H^{\circ}-T\Delta S^{\circ}\] already contains conformational and solvent entropy, so a flexible paratope should not be given a separate “fuzziness” parameter unless that parameter is experimentally identified.
On priority: the chemical‑potential and activity formulation of serum antibody binding is due to József Prechl, who defines chemical potential as the capacity of the system to generate antibody–antigen complexes and decomposes it into standard affinity, concentration and an activity coefficient3,4. On the concentration axis, that work is ahead of this notebook and should be credited as such. What is not in it is a temperature analogue; the effective‑temperature reading developed on the following pages is this notebook’s own, and it is the part most exposed to being wrong. The thermodynamic‑model machinery being borrowed here is standard in the physics of gene regulation, where the chemical‑potential term is treated as non‑optional5.
Specificity is a spectrum, not a statistic
It is tempting to summarise the shape of \(p(x)\) with one number. Two are usually quoted:
\[S=-\sum_xp(x)\ln p(x),\qquad N_{eff}=\Big(\sum_xp(x)^2\Big)^{-1}=\frac{1}{Y_2}.\]
Quoting both, separately, is like describing a spectrum by naming two wavelengths. They are two points on one curve, and the curve has a physical meaning that neither point does. Write the Rényi entropies \[R_q=\frac{1}{1-q}\ln\sum_xp(x)^q\], or equivalently the Hill numbers \[D_q=(\sum_xp(x)^q)^{1/(1-q)}=e^{R_q}\]. For a Boltzmann \(p\) there is an exact identity,
\[\sum_xp(x)^q=\frac{Z(q\beta)}{Z(\beta)^q}\quad\Longrightarrow\quad R_q=\frac{q\ln Z(\beta)-\ln Z(q\beta)}{q-1},\]
so the Rényi spectrum of an antibody’s specificity is that antibody’s free energy at rescaled temperatures \(q\beta\). Small \(q\) weights the broad, weak, rare part of the landscape; large \(q\) weights the few dominant epitopes. The landmarks are \(D_0\), the observed number of reactivities; \(D_1=e^{S}\); \(D_2=N_{eff}\), the participation ratio; and \(D_\infty=1/\max_xp(x)\), set by the strongest binder. Two antibodies with the same Shannon entropy can have sharply different \(D_q\) curves, so the spectrum distinguishes “one peak plus a broad haze” from “several comparable peaks”. A monoclonal and a natural polyreactive IgM differ not in a scalar but in the slope of \(R_q\) — and the slope is the landscape’s response to stringency. Diversity profiles of exactly this form are already standard in repertoire immunology[6]7; here they acquire a thermodynamic reading.
Recognising \(S\) and \(N_{eff}\) as Hill numbers of order 1 and 2 costs nothing and buys a great deal, because the entire diversity‑estimation literature then applies verbatim — including its warnings. Plug‑in estimates from an incomplete sample are severely and asymmetrically biased downward, the bias grows as coverage falls and is worst at small \(q\), and two samples of different completeness cannot be compared even after rarefying to equal read depth [8]9 . An antibody will look more specific than it is simply because you sequenced it less deeply — a lesson the T‑cell repertoire field learned the hard way10. The fix is off the shelf: estimate sample coverage as \(\hat C=1-f_1/n\), with \(n\) the total read count and \(f_1\) the number of mimotopes seen exactly once, and report coverage‑standardized profiles \(^qD(\hat C)\) at a common \(\hat C\).
The third axis needs replacing rather than standardizing. The gap \[\Delta=\varepsilon_{min}-\langle\varepsilon\rangle\] is intuitive but is the least defensible of the three, because \(\varepsilon_{min}\) is a sample minimum whose expectation drifts like \(\sqrt{2\ln n}\) with sequencing depth even when the antibody is unchanged. Since selected mimotopes are by construction a peaks‑over‑threshold sample, the identifiable summary of the tail is the generalized Pareto shape parameter \(\hat\xi\) — negative for a bounded landscape, zero for a Gumbel‑type tail, positive for a heavy tail. One should report \(\hat\xi\) with its threshold‑stability plot, and \(\Delta\) only as a descriptive footnote.
A monoclonal with one dominant epitope and a natural polyreactive IgM occupy opposite ends of these axes — not different categories, but different points on a continuum of distribution shapes. The point of writing the axes down carefully is that positions on a continuum have standard errors, and categories do not.
Condensation, and the one prediction that could kill this framework
In a random energy model with \(M\) epitopes and Gaussian energies of width \(\sigma\), the Boltzmann measure condenses onto \(O(1)\) states below \[\beta_c=\frac{\sqrt{2\ln M}}{\sigma},\]
above which weight is spread over exponentially many epitopes. (The algebra is right: for \(M=2^N\) and \(\sigma=\sqrt{N/2}\) this recovers the textbook \(\beta_c=2\sqrt{\ln 2}\).) It is tempting to read this as a phase transition an antibody undergoes between monospecificity and cross‑reactivity. That reading is stronger than the mathematics allows, in three ways worth stating plainly.
First, freezing in the REM is a thermodynamic‑limit statement — it exists because \(\ln M\) is extensive. One antibody facing a fixed finite library has no thermodynamic limit and hence no sharp transition, only a crossover of width \(O(1/\ln M)\). Second, \(\beta_c\) depends on \(M\), so “the freezing temperature of this antibody” changes when you buy a bigger library; a quantity that moves when you change your assay is not a property of the antibody. Third, every quantity in the derivation is averaged over the disorder — over the random draw of the \(M\) energies. A single monoclonal is one realization, and in the condensed phase the participation ratio is famously not self‑averaging: \(Y_2\) has order‑one fluctuations between realisations that do not shrink as \(M\) grows. “The freezing temperature of this antibody” is therefore a distribution, not a number. And \(\beta\) cannot be tuned across \(\beta_c\) anyway: literal temperature buys a factor of about \(1.12\) between 4 and 37 °C, and the effective temperature obeys no zeroth law. A transition you cannot cross is not an observable. The approach is also soft — \(Y_2\) goes to zero linearly in \(T_c-T\) — so even at ensemble level there is no sharp dichotomy, which is what the rest of this notebook argues anyway. Condensation is a property of the repertoire ensemble, not an event in the life of a molecule.
What the same theory does give, and this is worth much more, is an exact prediction with one free parameter that is already measured. Bouchaud and Mézard computed all the participation ratios of a condensed random‑energy measure. With \[\mu=T/T_c=\beta_c/\beta\] and weights \[w_x=e^{-\beta\varepsilon(x)}/Z_{ep}\],
\[P(w)\;=\;CM(1-w)^{\mu-1}w^{-1-\mu},\qquad Y_k=\sum_xw_x^k=\frac{\Gamma(k-\mu)}{\Gamma(k)\,\Gamma(1-\mu)},\qquad k>\mu,\]
11. Setting \(k=2\) gives \(Y_2=1-\mu\), the familiar \[\mathbb{E}[Y]=1-\beta_c/\beta\] of the random energy model — note in passing that \(Y_2\) is the Simpson index, the probability that two independent draws land on the same epitope. So one measurement fixes everything:
\[\widehat{\mu}=1-Y_2,\qquad Y_3=\tfrac{1}{2}Y_2(1+Y_2),\qquad Y_4=\tfrac{1}{6}Y_2(1+Y_2)(2+Y_2),\qquad Y_k=\frac{Y_2\prod_{j=1}^{k-2}(j+Y_2)}{(k-1)!},\qquad N_{eff}=\frac{1}{1-\mu}.\]
Estimate \(Y_2\) from a normalised reactivity profile, predict \(Y_3,Y_4,\dots\), and plot predicted against observed. The model is heavily over‑identified: one parameter \(\mu\) must simultaneously account for the whole rank‑abundance curve of a monoclonal’s mimotopes, its Shannon entropy, its \(N_{eff}\) and every higher moment. If the points sit on the diagonal, the reactivity distribution behaves like a condensed random‑energy measure, \(\mu\) is a legitimate reduced temperature, and “effective temperature” has stopped being a metaphor — we would have measured it. If they deviate, the random energy model is rejected for that antibody, and the direction of the deviation is informative: excess \(Y_k\) at large \(k\) points to correlated, clustered low‑energy epitopes — a genuine structural motif, the generalised random energy model rather than the plain one — while a deficit points to a heavier tail than Gaussian. The shape statistics of the previous section stop being descriptive summaries and become the second and first members of a predicted family.
These are quenched averages, valid for \(\mu<1\); intensities must be converted to weights on a common scale; and finite libraries bias every \(Y_k\), so the test must be run on rarefied subsamples with the extrapolation shown.
The dimensionless statement survives unchanged: specificity is governed by \[\beta\sigma=\sigma/k_BT\], the spread of binding energies in units of \(k_BT\). A “cold”, specific antibody is not one at low temperature but one whose landscape has large \(\sigma/k_BT\). What is new is that the claim now has a residual plot attached.
One caveat stated honestly: the REM assumes uncorrelated energies, and this notebook’s own 5‑of‑7 mimotope graph is a claim that energies are strongly correlated along sequence neighbourhoods. The weight law above should therefore be read as the null against which correlation is detected — deviations are informative, not embarrassing. See A hierarchy in the landscape below.
A hierarchy in the landscape
The caveat above deserves more than a caveat, because the violation of REM’s uncorrelated‑energy assumption is not a distant worry from the physics literature — it is the operating premise of our own antiphospholipid project, which joins mimotopes sharing 5 of 7 residues and validates that cutoff against mimotopes grouped by the panning monoclonal. That construction asserts, and confirms against data, that peptides at small Hamming‑type distance have correlated binding energies for the same antibody. REM asserts the opposite. We hold, on adjacent pages, a model and its empirical refutation, and the refutation is the better‑evidenced of the two.
The right response is replacement, not apology, and the replacement is standard. Derrida’s generalised random energy model imposes hierarchical correlations between configuration energies while retaining exact solvability12; rigorous treatment in Bovier & Kurkova (2004)13. For sequence spaces specifically, the Rough Mount Fuji model tunes between a smooth additive landscape and an uncorrelated one with a single parameter and has been fitted to real mutational data14, and Stadler’s amplitude spectra give a complete Fourier decomposition of landscape ruggedness on Hamming graphs15.
The immediately actionable point is that the energy correlation function
\[C(d)=\mathrm{cov}\big(\varepsilon(x),\varepsilon(y)\big)\Big|_{d_H(x,y)=d}\]
is directly estimable from any monoclonal’s mimotope set with measured intensities — it is a variogram over peptide space — and it is the single number that decides whether REM is even approximately admissible.
If the landscape is hierarchical, the mimotope similarity graph should be approximately ultrametric, measurable as Gromov \(\delta\)‑hyperbolicity or as the treeness of the correlation dendrogram — plain REM predicts \(\delta\) indistinguishable from a composition‑matched null, GREM predicts significant hyperbolicity. And GREM freezing is a cascade, one transition per level, so under increasing stringency the measure should condense hierarchically — first onto motif families, then onto individual sequences — visible as a staged collapse of community structure across panning rounds.
A further honest correction to what is written elsewhere on this site: the Gaussian \(g(\varepsilon)\) is justified by a central‑limit argument over \(N\) additive contact energies. Real paratope contacts are neither independent (packing, electrostatics, water) nor identically distributed (hotspot residues dominate), and the number of contacts is small — a median of three to four interaction motifs per variable domain, with 89% of paratope‑interacting residues in CDRs16. A central limit theorem with \(N\approx10\) strongly correlated, heavy‑tailed terms is not a theorem; it is a hope. One should rather fit the extreme‑value shape parameter \(\xi\), rather than assume \(\xi=0\).
Affinity and kinetics are independent inputs
A point the equilibrium picture alone hides: the energy gap \(\Delta\varepsilon = \varepsilon_{\text{bound}} - \varepsilon_{\text{unbound}} < 0\) fixes the dissociation constant,
\[ K_D \;=\; c^\circ\, e^{+\beta\,\Delta\varepsilon}, \qquad \Delta G^\circ \;=\; +k_B T \ln K_D \;=\; -k_B T \ln K_a, \]
so stronger binding (more negative \(\Delta\varepsilon\)) gives a smaller \(K_D\) — the direction most easily gotten wrong. But the rates are set by activation barriers, not by the gap:
\[ k_{\text{on}} \propto e^{-E_a^{\text{on}}/k_B T}, \qquad k_{\text{off}} \propto e^{-E_a^{\text{off}}/k_B T}, \qquad K_D \;=\; \frac{k_{\text{off}}}{k_{\text{on}}}. \]
Detailed balance pins only the difference of the barriers, \(E_a^{\text{off}} - E_a^{\text{on}} = -\Delta\varepsilon\), so the same \(K_D\) is consistent with many \((k_{\text{on}}, k_{\text{off}})\) pairs. Two antibodies of identical affinity can have wildly different residence times \(1/k_{\text{off}}\) — kinetics carries information that affinity alone does not, and affinity maturation is in part the sculpting of slower-off-rate, deeper wells17. \(K_D\) is therefore an equilibrium quantity related to kinetics through \(K_D = k_{\text{off}}/k_{\text{on}}\); it is not itself a kinetic metric.
A further consequence worth stating because it is so often mis-pictured: association in solution is diffusive, not ballistic. A ligand does not fly over its barrier with Maxwell–Boltzmann velocity; it random-walks across it under heavy solvent friction (Kramers’ overdamped regime), with a diffusion-limited ceiling on \(k_{\text{on}}\) of order \(10^{9}\)–\(10^{10}\,\text{M}^{-1}\text{s}^{-1}\) set by Smoluchowski17. Any reported on-rate above that ceiling is suspect.
What a microarray actually measures is \[k_{off}\] — and the wash is the temperature
Everything above assumes spot intensity is a monotone stand‑in for \(-\varepsilon\). It is not, and the three reasons are worth separating because the third is useful.
Valency. Serum IgM is decavalent, and array spots present peptides at high, uneven local density. Multivalent attachment converts a monovalent \(K_D\) into an avidity that depends on spot coverage in a way that has been measured and shown to distort binding constants extracted from arrays18. An IgM intensity behaves less like \(e^{-\beta\varepsilon}\) than like \(e^{-n\beta\varepsilon}\) with \(n\) an unknown, spot‑dependent effective valency — formally indistinguishable from a spot‑dependent temperature. Since IgM is the isotype of particular intrest in our studies, this is the largest single threat to the quantitative claims here, and the control is to titrate spot density.
\(\beta\) is not a clean lever. Physical temperature is weak (a factor of about \(1.12\) in \(\beta\) between 4 and 37 °C), but worse, it is not a lever at fixed landscape: antibody binding shows large negative \(\Delta C_p\) and strong enthalpy–entropy compensation, so van’t Hoff plots are curved and \(\varepsilon(x)\) is itself temperature‑dependent (e.g. - anti‑DNP measurements from −3 to 67 °C)19. Thought experiments of the form “cool the system and specificity sharpens” are not available.
Washing. A washed assay does not report equilibrium occupancy; it reports occupancy that survived a finite dissociation window \(t_w\). To first order the intensity on spot \(x\) is
\[I(x)\;\propto\;\theta^{eq}(x)\Big(1-e^{-k_{on}(x)[\mathrm{Ab}]t_i}\Big)e^{-k_{off}(x)t_w},\]
and since \(\ln k_{off}\) is roughly affine in the binding energy, the measured weights take the form \(e^{-\beta_s\varepsilon(x)}\) with an assay inverse temperature \(\beta_s>\beta\) that the experimenter sets by wash duration, buffer and flow. This is the honest version of what earlier notes here called “detection stringency”, and it is much better than a metaphor because it is dialable over a range physical temperature cannot reach: the same serum on the same library at \(t_w\in\{1,5,25\}\) minutes is a scan in \(\beta_s\). It also explains why successive panning rounds sharpen a mimotope set — each round is another dissociation window, so round number is a stringency schedule. The corollary that must be stated: at fixed \(K_D\) the signal still varies over orders of magnitude along an iso‑affinity line, so the empirically reconstructed “\(g(\varepsilon)\)” is a projection of a two‑dimensional \((\varepsilon,k_{off})\) landscape.
The wash‑time scan and the \(q\)‑scan of the Rényi spectrum are the same scan, which gives a consistency check that substitutes for the zeroth law this framework does not have. Because \(\sum p^q\) at stringency \(\beta_s\) equals\(\sum p^{q'}\) at stringency \(q\beta_s/q'\), a Rényi profile measured at one wash time predicts the ordinary \(Y_2\) measured at another. The claim is not that everything equilibrates to a common effective temperature — it does not — but that two independent routes to the same reduced temperature should agree. If they do, the effective‑temperature language is earned locally. If they do not, the size of the disagreement is the size of the non‑equilibrium correction, which is worth knowing either way.
One assumption in the above is not established and should be checked before the two‑route test is trusted: that \(\ln k_{off}\) is affine in \(\varepsilon\) for antibody–peptide pairs. That is a linear free‑energy relationship, plausible but not verified here.
A note on “temperature”
Because the formalism is borrowed, “temperature” is used in three different senses, and conflating them is the main way to abuse the analogy.
- Literal physical \(T\) is real but a weak lever: across 4 °C–37 °C the accessible range is a factor of \(\sim 1.12\), far too small to move \(\beta_c\).
- Effective temperature \(T_{\text{eff}}\) — fitting a measured signal as if \(p \propto e^{-\varepsilon/k_B T_{\text{eff}}}\) — is a width parameter, not a thermodynamic temperature. An antibody’s breadth is set by the chemistry of its CDR loops, not thermal agitation20,21, so two antibodies in the same tube can have very different \(T_{\text{eff}}\). Prefer reporting \(S\), \(N_{\text{eff}}\), \(\Delta\) directly.
- Selection / detection stringency (wash stringency, panning rounds, the positivity threshold) acts like an inverse temperature on the recovered distribution — and it is the strongest knob, but it is the temperature of a non-equilibrium process where detailed balance fails. A sharpened profile may reflect real biology or merely turned-up stringency; the two are confounded unless stringency is fixed and reported.
The cleaner, dimensionless statement is that specificity is governed by
\[ \beta\sigma \;=\; \frac{\sigma}{k_B T}, \]
the spread of binding energies in units of \(k_B T\). A “cold,” specific antibody is not one at low temperature — it is one whose landscape has large \(\sigma/k_B T\), with a few epitopes sitting many \(k_B T\) below the rest.
Why this is more than a reframing
Three things follow that the binary picture cannot give:
- Polyreactivity becomes a measurable quantity — an entropy \(S\) or an effective count \(N_{\text{eff}}\), not an error bar — and one with a known biochemical basis: polyreactive Fabs bind diverse epitopes with uniformly low affinity and characteristic CDR signatures1,20, while even “promiscuous” binding is built from specific hydrogen bonds across multiple discrete binding modes rather than nonspecific stickiness22.
- Mimotope and microarray data become samples from \(p(x)\), with an explicit model of when the sample is faithful (sub-saturation) and when it lies (saturation, or non-equilibrium selection)23.
- The repertoire becomes an ensemble of distributions, opening genuinely statistical-mechanical questions about the antibody population as a whole — and connecting to our observation that the autoimmune repertoire is restricted rather than simply redirected24.
None of the framing layer is settled. The epitope space is not obviously enumerable, \(\beta\) is a modelling choice rather than a measured constant, the REM’s uncorrelated-energy assumption is violated by real landscapes, and whether the equilibrium reading is the right one in a dynamic immune system is open. The mapping — affinity distribution as a canonical distribution over epitopes, with specificity as its shape — is sound and computable; the effective-temperature and freezing pictures are framing-level hypotheses. But as a way to organise mimotope and microarray data, and as a bridge to the repertoire-physics programme, treating specificity as a shape has been more productive than treating it as a switch.
The physics each step rests on is derived in full in the companion tutorial, <em>From the Boltzmann distribution to receptor–ligand kinetics</em>.