An isothermal titration calorimeter measures one quantity: heat. Everything else on an ITC report — dissociation constant, binding enthalpy, entropy term, stoichiometry — is a parameter of a model fitted to a series of heats. It is unusual in yielding an enthalpy and a free energy from one experiment, which is why it is used for thermodynamic signatures rather than affinities alone, and unusual in how directly its answers depend on two numbers written on two tubes: the cell concentration and the syringe concentration.
This guide covers what the instrument records, why the Wiseman c parameter is a window rather than a rule, how buffer choice enters the measured enthalpy, why concentration accuracy sets the ceiling, and how enthalpy-entropy compensation can be manufactured.
What is recorded, and what is inferred
A titration calorimeter holds a sample cell and a reference cell at the same temperature and records the power needed to do so. Titrant is injected in aliquots, each injection deflects that power, and integrating the deflection gives one heat. The raw output is a list of heats indexed by molar ratio.
The instrument description that established the format reported data reduction and deconvolution giving least-squares estimates of n, ΔH°, ΔS° and the binding constant K, with binding constants as large as 10^8 M^-1 measurable under favorable conditions; on ribonuclease A and 2'-CMP, repeat determinations there suggested errors of only a few percent in n, ΔH° and K over the most favorable concentration range [1].
For a 1:1 model the regression returns three free parameters: K, ΔH° and n. The remaining entries are algebra, not measurement. ΔG° = -RT ln K, with K relative to a 1 M standard state, and ΔS° = (ΔH° - ΔG°)/T. The entropy term is a residual — what is left after subtracting a logarithmic function of K from a directly fitted enthalpy — so it inherits the error of both. That is a definitional relationship rather than an experimental finding, and is stated here uncited. The heat capacity change ΔCp is not available from one titration at all: it is the slope of ΔH° against temperature.
The c-value window is a proxy, not a rule
The Wiseman parameter c is the product of the binding constant and the site concentration in the cell, c = n·K·[M]0. It governs isotherm shape: at low c a shallow ramp with no discernible inflection, at high c a step whose position fixes n but whose steepness says almost nothing about K. The commonly taught window of roughly 1 to 1000 is a convention distilled from that geometry, not a threshold from any one experiment.
The stronger version has been tested. Turnbull and Daranas noted the widely held belief that accurate thermodynamic information cannot be derived when c is below about 10, and used experimental titrations with simulated data to argue this is false provided four conditions hold: enough of the isotherm is analyzed, the stoichiometry is known, both concentrations are known accurately, and signal-to-noise is adequate. Their conclusion is scoped: they support ITC for K and hence ΔG° at low c, while advocating greater caution for ΔH° [2].
A separate analysis reframes it. Tellinghuisen argued that feasibility is set less by c than by the total detectable heat q_tot and the extent to which titrand can be converted to complex, and gave a guideline: with 90% conversion, K can be estimated within 5% over the range 10 to 10^8 M^-1 when q_tot/σ_q is about 700, where σ_q is the standard deviation for estimating an injection heat, dropping to about 150 when n is known [3].
That guideline converts into a heat budget, though the example below deliberately pushes past the range the guideline was quoted for. Take Kd = 100 pM, so K = 10^10 M^-1 — two orders of magnitude above the 10^8 M^-1 upper end of the range over which the ratio of 700 was stated, so what follows extrapolates the criterion rather than applies it. Holding c at or below 10^3 caps the cell concentration at 10^3 / 10^10 = 10^-7 M. In a 200 µL cell with n = 1 that is 10^-7 mol/L × 2 × 10^-4 L = 2 × 10^-11 mol of sites. At ΔH° = -40 kJ/mol and 90% conversion the titration liberates 0.9 × 2 × 10^-11 × 4 × 10^4 = 7.2 × 10^-7 J, or 0.72 µJ. The ratio of 700 then requires σ_q below 7.2 × 10^-7 / 700 = 1.0 nJ, or 0.25 ncal at 4.184 J per calorie; fixing n relaxes that to 4.8 nJ, about 1.1 ncal. The binding is not too tight for the model; the experiment is too quiet to hear.
The same analysis criticizes the protocol most workers run: twenty to thirty injections to a final titrant/titrand mole ratio of about 2 is described as a scheme that can be far from optimal and can needlessly limit applicability [3]. An earlier optimization study prescribed instead 10 injections, a final stoichiometry range R_m = 6.4/c^0.2 + 13/c but no smaller than 1.1, and the largest cell concentration the instrument allows, subject to c < 10^3 for estimating K [4]. That formula shows the drift: at c = 10, R_m = 6.4/1.585 + 1.3 = 5.34; at c = 100, 6.4/2.512 + 0.13 = 2.68; at c = 1000, 6.4/3.981 + 0.013 = 1.62. The endpoint moves more than threefold while the standard protocol holds it at 2.
Concentration accuracy sets the ceiling
ITC has no internal concentration standard. The fit treats the stated concentrations as exact, so systematic error propagates into the parameters. Tellinghuisen's treatment states that systematic errors in the concentrations of titrant and titrand are fully compensated by the site number, or stoichiometry parameter, n [4] — compensation in the sense that the fit still describes the data, with n absorbing the discrepancy, rather than in the sense that every other fitted parameter is left untouched.
What that compensation protects is worth working out, since the two concentrations behave differently. Write f for the factor by which the cell concentration is overstated and g for the syringe. The physical quantities — moles of sites in the cell, moles injected, total heat evolved — are unchanged by a labeling error. Apparent stoichiometry is the assumed molar ratio at true equivalence, so it becomes n·g/f. Apparent enthalpy is total heat over assumed moles injected, so it becomes ΔH°/g. The apparent binding constant, from curvature at fixed physical c, becomes K/g. A cell-side error moves n alone; a syringe-side error moves n, ΔH° and K together. That derivation is carried out here from the 1:1 model, not taken from a source.
The consequences are asymmetric. Take a syringe concentration overstated by 10%, so g = 1.10. Apparent enthalpy is ΔH°/1.10 = 0.909 ΔH°, low by 9.1% in magnitude — note the denominator, since a 10% overstatement is not a 10% error. For a true ΔH° of -40 kJ/mol the reported value is -36.4 kJ/mol, off by 3.6 kJ/mol. The apparent Kd rises by that same 1.10, costing RT ln(1.10) = 2.479 × 0.0953 = 0.24 kJ/mol in ΔG° at 298.15 K. One error, roughly fifteen times larger in enthalpy than in free energy, because affinity enters logarithmically and enthalpy linearly.
A Monte Carlo and experimental study puts the magnitudes in range: injection heat noise was mainly proportional in character, with approximately 1% uncertainty at 27 °C and 3% at 65 °C, and concentration errors approximately 1%. Its authors concluded that uncertainty is in most cases minimized by more injections and by fixing N at its known value, and that with appropriate technique the uncertainty in binding enthalpies can be kept below about 2% under many conditions, including at low c [5]. Fixing n is not free: the optimization study warns that altering and freezing any fit parameter degrades fit quality and produces predictable changes in the others, and that at very small c, ΔH° may be statistically ill-defined unless n can be fixed from other information [4]. Which advice applies depends on how well the cell concentration is independently known.
Purity is the other half, and not symmetric either. A study using trypsin and tRNA-guanine transglycosylase with high-affinity ligands found that errors in protein concentration did not significantly change the thermodynamic properties obtained, whereas most ligand impurities led to pronounced changes in binding enthalpy; even after correcting the ligand concentration for impurities not expected to bind, differences of about 4 kJ/mol remained, which the authors considered significant [6]. They concluded that ligand purity is the critical parameter to monitor, and that artificially changing fitting parameters to obtain a sound stoichiometry in the presence of uncharacterized impurities may yield parameters deviating significantly from the true signature [6]. A tidy n is not evidence of a clean prep.
Buffer ionization: part of the enthalpy belongs to the buffer
If binding takes up or releases a proton, the buffer supplies or absorbs it, and the buffer's own ionization enthalpy is added to every measured heat. The observed enthalpy is ΔH°obs = ΔH°int + n_H · ΔH°ion, where n_H is the number of protons transferred from buffer to complex and ΔH°ion is the buffer's ionization enthalpy at that temperature.
The correction is not small: ionization enthalpies of common buffers span a wide range and are themselves temperature dependent. Fukada and Takahashi determined enthalpy and heat capacity changes for the deprotonation of 18 buffers calorimetrically in 0.1 M potassium chloride between 5 and 45 °C, and found the enthalpy change to be a linear function of temperature for every buffer except phosphate and glycerol 2-phosphate, whose second dissociations form a divalent anion and were better fitted with a second-order function of temperature [7]. Reading a value at the wrong temperature is a real source of error, and for phosphate linear interpolation is the wrong model.
The measurement that separates the two terms is a titration repeated at one pH in buffers of differing ionization enthalpy. Take buffers differing by 44 kJ/mol, with observed binding enthalpies of -34.0 kJ/mol in the low-ionization buffer and +10.0 kJ/mol in the high one. The slope gives n_H = (10.0 - (-34.0)) / 44 = 1.0, one proton taken up per complex formed. If the low-ionization buffer contributes +4 kJ/mol, the intrinsic enthalpy is -34.0 - (1.0 × 4) = -38.0 kJ/mol; the same arithmetic in the second gives 10.0 - (1.0 × 48) = -38.0 kJ/mol. Those ionization enthalpies are stipulated for the example, not literature values. Note the direction: the observed enthalpy changed sign between two buffers at one pH, and only one number describes the interaction.
Baker and Murphy set out the framework, showing that experiments across pH in buffers of varying ionization enthalpy allow determination of the pKa values of the group responsible for the proton linkage in the free and liganded states, the protonation enthalpy for that group in those states, and the intrinsic energetics of binding. They also showed that besides the variation of the binding constant with pH, the observed binding enthalpy and heat capacity change can undergo extreme deviations from their intrinsic values depending on pH and buffer conditions [8]. A later simulation study of such designs reported that global analysis across pH, buffer ionization enthalpy and temperature can yield reliable estimates of intrinsic binding and protonation parameters, and that in that context the exact component concentrations may not be required [9] — a real escape route from the concentration problem, bought with a much larger experiment.
Enthalpy-entropy compensation, real and manufactured
Plot ΔH° against -TΔS° for a ligand series and the points usually fall on a line of slope near -1, with affinity barely moving. Some of that is physics and some is bookkeeping, and separating them is the difficulty.
The bookkeeping follows from the concentration algebra above. ΔG° is nearly immune to a titrant concentration error while ΔH° is not, and ΔS° is computed as (ΔH° - ΔG°)/T. A ligand series prepared with varying, unrecognized concentration or purity errors will therefore give a scatter of ΔH° values, each almost exactly offset in -TΔS°, and a compensation line of slope near -1 that describes the weighing rather than the binding.
The statistical critique is older. Sharp asked whether observed compensation reflects anything beyond ordinary statistical thermodynamics, reanalyzed protein data purporting to show it and concluded other causes explained the data better, and found from a general statistical mechanical model that the most likely behavior is linear compensation over a rather limited range of perturbations, with a compensation temperature Tc = dΔH/dΔS within 20% of the experimental temperature — behavior insensitive to model details and revealing little causal information about the system [10].
The phenomenon nonetheless survives careful analysis. Olsson and colleagues modeled how experimental constraints affect ITC measurements, found those constraints obscure any compensation present in common representations such as ΔH versus -TΔS plots, and instead compared all pairs of ligands binding each protein. Across data from 32 diverse proteins they reported a significant and widespread tendency to compensation: strong compensation, meaning ΔΔH and -TΔΔS opposed and differing by less than 20% in magnitude, for 22% of modifications, twice what would be expected without compensation, and reinforcement, where the two share a sign, for 15% [11]. On their reading compensation is strong but imperfect, and prior theory had focused on a stronger tendency than the data support [11]. A review of molecular origins treats the effect as unquestionably real while emphasizing that characterizing it requires experimental care [12].
What the technique does not settle
ITC gives a model-dependent description of one solution equilibrium under one set of conditions. It does not identify the binding site, does not by itself distinguish competitive from allosteric interaction, and does not report kinetics: two systems with the same K and ΔH° may equilibrate on very different timescales, and the calorimeter sees only the endpoint of each injection. A fitted n near 1 is consistent with 1:1 binding, but is equally the product of a concentration error and an aggregation state.
Several routine practices are convention rather than validated procedure: matching cell and syringe buffers by dialysis or from a single stock, degassing before loading, subtracting a titrant-into-buffer blank, discarding the first injection, and reporting reference power and temperature with the parameters. They reduce known artifacts, no controlled study is cited here for any of them individually, and none corrects a concentration that was wrong before loading.
The honest summary of the error budget: the instrument is usually the most accurate part of the experiment. Injection heat noise and concentration error both run near 1% in that study at 27 °C, the heat noise rising to about 3% at 65 °C [5], but concentration error is systematic rather than random: it does not average out across replicates, and it enters the enthalpy linearly. A thermodynamic signature reported without a stated, independent method for both concentrations — extinction coefficient and its source, amino acid analysis, quantitative NMR — is a shape, not a measurement.