Small-angle X-ray scattering: what a solution curve licenses

A solution SAXS experiment returns one number per scattering angle. The azimuthally averaged excess intensity of the sample over its solvent, I(q), is the whole observable, and every structural statement made from it — size, shape, folded or disordered, one species or three — has to come out of that single curve. Solution scattering data are subtle in appearance, highly sensitive to sample quality and experimental error, and easily misinterpreted, and the structural information they carry is limited to roughly 10 Å resolution ^[1]. The value of the technique is that it works on the sample as prepared, with no crystal; the cost is that a one-dimensional curve underdetermines almost everything you would like to know.

Nothing structural exists until the solvent is removed. Every atom in the illuminated volume scatters, buffer components included, so an exactly matched solvent blank must be prepared and subtracted before the net macromolecular scattering can be recovered. Sample heterogeneity caused by contaminants, aggregates, mismatched solvents or radiation damage severely influences and complicates the analysis, which is why the sample has to be pure and monodisperse for the duration of the experiment, and why the workflow that establishes this — gel electrophoresis, size-exclusion chromatography, light scattering — happens before the beam is ever used ^[2].

Guinier analysis and the range in which it is true

At the smallest accessible angles, ln I(q) plotted against q² is linear for a monodisperse particle: the intercept gives the forward scattering I(0), the slope gives the radius of gyration R_g. The linearity is a truncated series expansion, not an identity, so it holds only over a bounded q range, and the conventional working limit for standard Guinier analysis is about qR_g = 1.1. Molecular dynamics simulations of disordered chains traced the deviations from that approximation to the polymer scaling exponent and produced an empirical correction term; tested on intrinsically disordered proteins and on experimental datasets, the corrected treatment inferred R_g to within about 3% of the true value and extended the valid range to roughly qR_g = 2 ^[3]. That extension was developed and validated on disordered protein chains. For an analyte whose conformational class you do not already know, the 1.1 limit remains the defensible default, and the extended treatment is a claim conditioned on a specific polymer model rather than a general licence to fit further out.

R_g from a Guinier fit is a single moment of the mass distribution. It says how mass is spread about the centroid and nothing about how that mass is arranged. Two peptides sharing an R_g can be a compact globule and a loose coil of quite different maximum extent, and the Guinier region is precisely the region in which that difference is invisible.

The pair-distance distribution and D_max

The real-space counterpart of I(q) is P(r), the distribution of all interatomic vector lengths within the particle, weighted by scattering contrast. It is recovered by an indirect Fourier transform of the scattering curve, and that transform requires you to supply the distance beyond which P(r) is forced to zero — the maximum dimension D_max. In a study of porcine pepsin across a pH series, low-resolution models were reconstructed inside a search volume set by D_max, itself calculated from the pair-distance distribution function p(r), with no restriction imposed on the symmetry or anisometry of the molecule ^[4].

D_max is the parameter most often reported as a measurement when it is in fact an input. It bounds the transform, and it then bounds the search volume of any shape reconstruction built on that transform ^[4], so an error in it propagates into both the real-space R_g and the model. The consistency check available to you is internal: the R_g obtained by integrating P(r) uses the whole curve, while the Guinier R_g uses only the first few points. They are not the same measurement, and a disagreement between them points either at the low-q data or at the D_max you chose.

Porod and Kratky: compactness and disorder

The decay of intensity at higher q reports on the interface between particle and solvent. The Porod–Debye region describes the nature of that decay in a way that captures the information needed to distinguish folded from flexible particles; applied to comparative experiments, it can separate a discrete conformational change from localised flexibility, and its authors present this as more robust and more conclusive than traditional Kratky analyses, with the particle density calculable by the Porod–Debye criteria providing, on their prototypic SAXS data, an objective quality-assurance parameter that they suggest may prove of general use for SAXS modelling and validation ^[5].

The Kratky representation, q²I(q) against q, remains a quick visual test. A compact folded particle gives a bell-shaped peak that returns towards the baseline; a chain without a persistent tertiary fold gives a curve that rises and does not come back. On the amyloid-β 1-40 and 1-42 peptides, in-line rapid-mixing SAXS found both to be monodisperse on timescales under 2 s in the absence of copper, and Kratky plots of those profiles indicated that immediately after mixing both were intrinsically disordered; ensemble optimisation modelling of the same data returned a wide range of structural conformers rather than one ^[6]. In the same study, at a copper-to-peptide molar ratio of 0.5, Aβ1-40 aggregated while Aβ1-42 adopted a relatively monodisperse cylindrical shape; at a ratio of 1.5 the assignments changed, Aβ1-40 adopting a monodisperse cylindrical shape and Aβ1-42 the shape of an ellipsoid of rotation ^[6].

That pair of results is the practical warning. The same peptide, at two ligand stoichiometries within one study, gave one profile the analysis called aggregated and one it called a well-behaved particle. A Kratky classification is a statement about the ensemble present in the cell at that moment, not a property of the peptide.

Where the analyte is a self-assembled peptide structure rather than a single molecule, the analysis leaves the model-free regime and becomes a form-factor fit. In SAXS on nanofibers formed by a designed β-peptide, a form factor describing solid nanofibers did not fit the data at high q; a hollow cylinder with radial polydispersity, combined with a structure factor computed from polymer reference interaction site model theory, described the intensity across the whole q range and gave a hollow core radius of 15 Å with 21% polydispersity and a shell thickness of 20 Å ^[7]. Note what changed: the conclusion now rests on a chosen geometric model, and the strength of the claim is the strength of the argument that no simpler model fits.

The concentration series

What you measure is the particle form factor multiplied by an interparticle structure factor. At low concentration that structure factor is near unity and can be ignored; as concentration rises it cannot be, and the distortion sits at exactly the low-q end where R_g and I(0) are read. SAXS on equine metmyoglobin at up to 0.4 g/mL and on bovine pancreatic trypsin inhibitor at up to 0.15 g/mL, in solutions also containing 50% D₂O and 1 M urea, showed significant effects from interference between X-rays scattered by different molecules, indicating non-ideal behaviour at high concentration ^[8]. Metmyoglobin, nearly isoelectric under those conditions, was well fitted by a hard-sphere model with an 18 Å radius, only slightly smaller than the 20 Å estimated from its three-dimensional structure. BPTI in phosphate buffer showed substantially less interparticle interference than the hard-sphere model predicted from its 15 Å structural radius; replacing phosphate with MOPS increased the interference, consistent with a larger effective radius and suggesting that phosphate ions may mediate attractive intermolecular interactions without the formation of stable oligomers ^[8].

Read that as a statement about buffers as much as about crowding. The same molecule at the same concentration behaved differently in two buffers, so a concentration series characterises the solution, not the particle alone. The routine remedy is a dilution series whose low-q region is extrapolated to infinite dilution, but that is not the only route: in one published treatment, scattering curves free of an inter-molecular structure factor were instead derived by dividing the curves of a dilution series by the curve from the most dilute sample in that series, rather than extrapolating to infinite dilution ^[9]. Either way, what a concentration series buys you is a curve without the concentration-dependent term. What it does not buy you is evidence that the same species is present at every concentration.

Radiation damage and its signature

At a high-brilliance source, the absorbed X-ray dose that gives good counting statistics is close to the dose that ruins the sample. In lysozyme solution the scattering did not change below 400 Gy and increased exponentially above it, the aggregation appearing only beyond a critical dose rate; mass spectrometry of the exposed solution showed no cleavage of the polypeptide chain, so the damage signature was a non-specific association of intact molecules without substantial alteration of folding topology rather than fragmentation ^[10]. Small amounts of cryoprotectant — glycerol, ethylene glycol or sucrose, similarly effective in the 100 mM concentration range — reduced the damage, and the reduction was attributed mainly to changes in protein–protein interactions rather than to slower diffusion of activated species ^[10].

Do not carry 400 Gy across to your own sample. When radiation damage was quantified with a common methodology on lysozyme, glucose isomerase and xylanase, no single metric described damage for all of them: R_g, molecular weight and the integrated intensity of the SAXS profile together formed the minimal set that captured every observed behaviour, and radiation sensitivities derived from those parameters varied by up to six orders of magnitude between the three proteins tested ^[11]. No peptide is covered by any of those figures. What transfers is the procedure — collect the exposure as a series of successive frames and track all three parameters frame to frame — not the threshold.

Monodispersity is a precondition, not a finding

Guinier analysis, P(r) and every reconstruction downstream of them assume a single species. A solution of monomer and dimer does not produce a curve that announces two populations; it produces an intensity-weighted average with an intermediate R_g, and that average will fit a plausible single-particle model. This is the reason the sample-preparation literature treats purity and monodispersity as conditions to be established before the measurement rather than parameters to be recovered from it ^[2].

In-line SEC-SAXS exists because that pre-characterisation expires the moment the sample sits in a capillary. Coupling the column to the beam means scattering is recorded on material that has just been separated, so an aggregate peak and a monomer peak arrive as different frames instead of one average. The engineering trade-offs are concrete. An optimised system at one beamline used a compact configuration to limit the sample dilution that costs sensitivity, a coflow sample presentation that allowed an 11-fold increase in flux on sample without capillary fouling, and multi-wavelength fibre-optic UV analysis close to the X-ray beam, which gives accurate concentration for samples of known extinction coefficient and hence an estimate of the molecular weight of the scattering particle from the forward scattering intensity ^[12]. The same work is explicit that fast-flow low-volume columns give throughput competitive with batch concentration-series measurements but at a concomitant reduction in potential resolution relative to lower flow rates and larger columns ^[12]. SEC-SAXS also displaces the batch dilution series that would most cleanly have exposed a concentration-dependent oligomer; you trade one diagnostic for another.

Ab initio shape and the ambiguity problem

Bead modelling builds a particle from densely packed dummy atoms, assigns each atom to a phase or to the solvent through a configuration vector, and uses simulated annealing to find a configuration that fits the data while minimising the interfacial area of the model ^[13]. That interfacial-area penalty is what makes the problem tractable; without a compactness constraint an enormous number of bead configurations reproduce any one curve.

The penalty does not make the answer unique. When scattering patterns were computed from geometrical bodies of known shape and anisometry — solid and hollow spheres, cylinders, prisms — and the shapes reconstructed ab initio, bead-modelling methods were found to deliver multiple solutions, requiring a described procedure to analyse them and to estimate the stability and reliability of the reconstruction; flat shapes proved more difficult to restore than elongated ones, and certain shape types were identified as needing additional information for reliable reconstruction ^[14]. A single envelope presented as the shape is therefore not a result. Multiple independent runs, an explicit estimate of their agreement, and a statement of which shape class the data can and cannot discriminate is a result.

The ambiguity can also be assessed before any reconstruction is attempted. A rescaled library of more than 14,000 shape topologies, built from up to seven beads closely packed on a hexagonal grid and reduced to topology rather than size information, forms a scattering map against which an experimental profile can be placed; the number of neighbours in that map yields a quantitative a-score reporting how many distinct shape categories are compatible with the data ^[15]. A high a-score tells you, without running anything, that the curve does not contain enough information to choose a shape.

The underlying constraint is information content. One-dimensional solution scattering profiles are generally restricted to low-resolution shape and size information that on its own cannot lead to unique three-dimensional structures comparable to all-atom models from crystallography or NMR ^[16]. That constraint is relieved only by adding orthogonal data. In one demonstration, contrast-matched scattering from a protein carrying specific heavy-atom labels in 65% aqueous sucrose — conditions under which the protein is effectively invisible and the heavy atoms dominate the intensity, allowing pairwise distances between them to be extracted directly — was combined with conventional aqueous SAXS/WAXS and NMR residual dipolar couplings to position domains relative to one another within 1 Å in a complex between a lead-substituted calmodulin and a target peptide, with uniqueness established by exhaustive search over all models compatible with the data and not achievable from aqueous SAXS and RDCs alone ^[16].

Guidelines for the critical evaluation of scattering data exist for this reason. The ease with which three-dimensional models can be calculated against inherently one-dimensional scattering data means that confidence in a SAXS result depends on the reader being able to assess sample quality, data quality and modelling assumptions ^[17].

What SAXS adds beyond DLS

Dynamic light scattering measures diffusion behaviour: the diffusion coefficient, and hence the hydrodynamic radius calculated from it, depends on the size and shape of the macromolecule, and DLS is a practical way to assess the homogeneity of protein and nucleic acid preparations and of complexes ^[18]. It is fast, needs little material, is non-destructive, and it is the right first instrument for asking whether a preparation is one population.

What it does not give you is a distance distribution. R_h is a single number derived from a transport property, and it folds together the particle, its hydration shell and its shape-dependent friction; a compact and an extended particle of equal friction return the same R_h. SAXS returns R_g, I(0), P(r) and D_max from the scattered intensity itself, separating the second moment of the mass distribution from the maximum extent, and reporting on compactness through the high-q decay. The two techniques are complementary in a specific and useful way: DLS has been used as a screening tool to validate solution scattering models against measured hydrodynamic radii ^[18]. A model that fits the scattering curve but implies an R_h the DLS does not see is a model with a problem.

The honest output of a SAXS session is a set of parameters with their conditions attached: R_g with the q range of the Guinier fit declared, D_max and P(r) with the chosen limit stated, a Porod–Debye or Kratky classification of compactness, and the concentration and exposure series showing that none of these drifted. Add an envelope only with its ambiguity scored. That set is not a structure, and the guidelines the field wrote exist to stop it being presented as one ^[17].