Discovery at the VICB
Detecting Biomolecular Interactions in Solution, Label-Free
By: Carol A. Rouzer, VICB Communications
Published: May 1, 2016
Computational modeling predicts the origin of the signal detected during protein-ligand interactions by back-scattering interferometry.
Back-scattering interferometry (BSI) is a highly sensitive label-free technique for monitoring the interactions of biomolecules in solution or in membrane vesicles. It is remarkable for its ability to detect protein-protein, protein-small molecule, and protein-ion interactions over a wide range of affinities and despite huge differences between the masses of the interacting species. In fact, BSI can often detect the binding of a ligand to a protein, even if the ligand is present at concentrations below its individual limit of detection. BSI delivers this remarkable performance in the cases of systems that exhibit no change in absorbance, total mass, or temperature, leading many to question the exact molecular properties detected by the interferometer. Now, Vanderbilt Institute of Chemical Biology members Darryl Bornhop and Jens Meiler provide new insight into this intriguing question [D. J. Bornhop, et al. (2016) Proc. Natl. Acad. Sci. U.S.A., 113, E1595].
BSI is based on the principle of light interference. Whenever light passes through matter, its speed is slowed, leading to a “bending” of the light. The refractive index of a medium, the ratio of the speed of light in a vacuum to its speed in the medium, is one measure of light interference. Refractive index is a unique quality of any form of matter, indicating that interference is highly dependent on the composition of a substance. The back-scattering interferometer (Figure 1) exploits this property of matter. The instrument comprises a microfluidic chip containing a narrow channel that serves as the sample chamber. A HeNe laser provides the light, which is directed to the chip by use of a mirror. The channel allows the light to be reflected multiple times as it passes through the sample, resulting in a long path length that magnifies the effects of any interference changes. After passing through the sample, the light is reflected back to the mirror in the form of a scattered fan pattern, actually a set of high contrast “interference fringes.” The reflected light is transferred by the mirror to a high-resolution linear charge-coupled device (CCD). Mathematical analysis of the fringe pattern reveals changes in the refractive index that result from interactions between components in the sample. The relationship between sample concentration and refractive index is a well accepted chemical principal. However, accumulating evidence garnered from studies of other techniques based on refractive index, such as surface plasmon resonance, have suggested that concentration effects alone cannot explain the signals measured by BSI during protein-ligand interactions. This led the investigators to postulate that alterations in protein conformation, as reflected in changes in properties, such as hydrodynamic radius and solvation, may be key.
FIGURE 1. Diagrammatic representation of a backscattering interferometer. A mirror directs the light from a HeNe laser to the sample placed in the channel of a microfluidic chip. The light leaves the chip in the form of a fringe pattern that is reflected by the mirror into a CCD array detector. Side, top, and front views illustrate the configuration of the light relative to the channel in the chip. An example of a fringe pattern is shown, illustrating the portion of the pattern selected for data analysis. Changes in the refractive index of the sample lead to shifts in the positions of the fringes. Figure reprinted with permission from D. J. Bornhop, et al. (2016) Proc. Natl. Acad. Sci. U.S.A., 113, E1595. Copyright 2016. D. J. Bornhop, et al.
To test their hypothesis, the investigators first focused on the interactions of calmodulin with various ligands. Structural data available from the Protein Data Bank illustrate the large conformational changes that calmodulin undergoes upon binding to Ca2+ and various other ligands (Figure 2). The Bornhop lab had already completed multiple binding studies of these interactions using BSI as outlined in Figure 3. The Meiler lab used the available structural data to calculate the radius of gyration (Rgyr) and the solvent accessible surface area (SASA) for each complex. The researchers then generated a linear equation that related the changes in Rgyr and SASA (ΔRgyr and ΔSASA, respectively) to the maximal BSI response observed for the formation of each complex. They obtained a remarkably good correlation between the BSI response predicted by the equation and the actual experimental values.
FIGURE 2. Ribbon drawings showing the structures of calmodulin unbound and bound to various ligands. Figure reprinted with permission from D. J. Bornhop, et al. (2016) Proc. Natl. Acad. Sci. U.S.A., 113, E1595. Copyright 2016 D. J. Bornhop, et al.
Their ability to mathematically model the calmodulin response encouraged the researchers to pursue their approach for a wider range of proteins. They hypothesized that the BSI Free solution response function (FreeSRF, ρ) could be described by the equation,
ρ = χβC,
where χ is the molar refractometry in refractive index units/mole/liter, β is an instrument response constant in radians per refractive index unit, and C is the molar concentration of the final complex.
To test the validity of this equation, which strongly resembles Beer’s law for absorption, required that values for each of the parameters be measured. The value of β is specific for each instrument. To determine it, the investigators measured the change in fringe pattern (radians) at various concentrations of glycerol for each instrument in their laboratory. Then, they used the literature value relating the refractive index of glycerol to its concentration in aqueous solution to convert the value for β to the desired units of radians per refractive index units.
To determine the value for C, the investigators decided to focus on the maximal value of ρ obtained at saturating concentrations of ligand. Thus, the value of Bmax (Figure 3) could be used as C in the equation. This approach provided values for ρ, β, and C for a given experiment in which a binding interaction was fully explored by BSI, enabling the determination of the corresponding experimental value for χ.
FIGURE 3. Outline of the design of a binding interaction experiment using BSI. Not only must the receptor be combined with multiple concentrations of the ligand, a set of reference samples must also be prepared that have the exact same concentration as the experimental samples except for the absence of the receptor. All samples are allowed to come to equilibrium, and then both experimental and reference samples are subjected to BSI. The difference in signal versus the experimental and reference samples is plotted versus ligand concentration to generate a hyperbolic curve from which a binding affinity constant (KD) and maximal receptor concentration (Bmax)may be derived. Figure reprinted with permission from D. J. Bornhop, et al. (2016) Proc. Natl. Acad. Sci. U.S.A., 113, E1595. Copyright 2016 D. J. Bornhop, et al.
Accumulated data from twenty binding interaction experiments collected over a period of three years using six different instruments operated by multiple investigators provided a training set for the purpose of developing a mathematical model to predict the value of χ. Expanding on the case of calmodulin, Michael Kammer and Amanda Kussrow of the Bornhop lab used structural data from the Protein Data Bank to calculate values for ΔRgyr, ΔSASA, aveRgyr (the average of the Rgyr values for the starting protein and the protein in the complex), and aveSASA (the average of the SASA values for the starting protein and the protein in the complex) for each of the binding interactions. They then derived a linear equation relating these structural characteristics to the experimental values of χ. In this initial attempt, they obtained an equation that exhibited a high correlation coefficient, but when they compared values of χ calculated from the equation with the actual experimental values, they observed large errors, particularly in the case of interactions that had produced small BSI signals. To address this problem, the researchers divided their training set into two, one including experiments in which the BSI response was large, and one in which a smaller response was observed. When data from these two sets were used separately to generate a mathematical relationship between protein structural changes and χ, excellent results were obtained in both cases (Figure 4).
FIGURE 4. Plot of experimental values for χ versus those predicted by the small response (A) and large response (B) mathematical models. As indicated in the key, the different symbols denote the various types of interactions evaluated in the training set, as well as the result of the predicted values for interaction not included in the training set. Figure reprinted with permission from D. J. Bornhop, et al. (2016) Proc. Natl. Acad. Sci. U.S.A., 113, E1595. Copyright 2016 D. J. Bornhop, et al.
To test the validity of their mathematical model, the investigators used it to predict χ values for two protein-ligand interactions that had not been included in the training set. They found that in both cases, only one of the models yielded a value of χ that was consistent with experimental observations. Use of that model produced excellent predictions of χ for both interactions. The predictions agreed with the experimental χ values to within 0.13% and 18% error (Figure 4).
The results support the hypothesis that BSI detects protein conformational changes that result from intermolecular interactions and that the BSI response can be predicted on the basis of structural data, when available. It is notable that the training set used to create the small and large response models included a wide range of interactions including protein-protein, protein-aptamer, protein-ion, and protein-small molecule binding, and that these interactions occurred both free in solution, and in membrane vesicles. However, the investigators note that it is possible that the model likely requires further refinement. In particular, the division of the training set into small and large responders was somewhat arbitrary. It is quite possible that other divisions would be better or that additional categories may be needed. It is also possible that additional terms may be necessary to mathematically predict χ for all types of intermolecular interactions. Such refinements will emerge from future experiments. However, these findings represent a major step forward in understanding how complex biomolecular interactions yield the changes in refractive index that can be observed with such great sensitivity by BSI.