Thermal unfolding methods in drug discovery

Abstract

Thermal unfolding methods, applied in both isolated protein and cell-based settings, are increasingly used to identify and characterize hits during early drug discovery. Technical developments over recent years have facilitated their application in high-throughput approaches, and they now are used more frequently for primary screening. Widespread access to instrumentation and automation, the ability to miniaturize, as well as the capability and capacity to generate the appropriate scale and quality of protein and cell reagents have all played a part in these advances. As the nature of drug targets and approaches to their modulation have evolved, these methods have broadened our ability to provide useful chemical start points. Target proteins without catalytic function, or those that may be difficult to express and purify, are amenable to these methods. Here, we provide a review of the applications of thermal unfolding methods applied in hit finding during early drug discovery.

INTRODUCTION TO PROTEIN THERMAL UNFOLDING

Thermal unfolding methods have been employed increasingly as primary screening approaches during early drug discovery. These methods detect compounds that stabilize target proteins either in isolated protein assays or in the more physiological setting of the cell. As these broadly different settings suggest, the nature of the different detection methods is also broad with fluorescence, aggregation, antibody binding, mass spectrometry, and luminescence methods, all being employed to monitor the thermal unfolding process. However, the principle of the method is simple and common to all these detection approaches, the ability to discriminate between folded and unfolded states of a protein, in this case brought about by heating. The principles underlying protein thermal unfolding and denaturation have been studied since the late 19th century when Hardy and Whetham first described the processes of denaturation and agglutination of albumen.¹ Since then, we have recognized that temperature plays a central role in the balance between the different conformational states of proteins and thermal unfolding measurements of proteins began to be studied more widely following the introduction of the first adiabatic differential scanning calorimeter, applicable for studying protein unfolding by Peter Privalov in 1964.² Differential scanning calorimetry (DSC) follows protein unfolding by measuring the heat capacity of protein states and the excess heat associated with transitions that are induced by a change in temperature. This breakthrough allowed us to design experiments focused on understanding the nature of the forces that control the formation and stabilization of the three-dimensional structure of proteins and their functions. It also allowed the effects of ligand binding equilibria on macromolecular unfolding equilibria to be considered and studied.^3–5 Subsequently, DSC was demonstrated to be valuable in determining binding constants for very tightly binding ligands, which could not be accessed using traditional methods.⁶ Whilst DSC was revolutionary in delivering thermodynamic studies on protein thermal unfolding, and measurements of ligand binding could be accessed using the technique, the throughput of the instrument limited wide application of thermal unfolding in hit identification. This situation changed in 2001 with the description of a fluorescence-based thermal shift assay applicable to screening in 384-well plates.⁷ The Thermofluor™ approach, now also known as thermal shift assay (TSA) or differential scanning fluorimetry (DSF), provided a route by which high-throughput protein stabilization experiments could be undertaken without the need for a calorimeter and with reduced protein consumption per interaction. The method involves monitoring the unfolding process utilizing the change in fluorescence as an environmentally sensitive dye partitions itself into the unfolded protein. The decrease in the dielectric constant of the unfolded protein environment, compared to aqueous solution, results in a large increase in fluorescence of the dye as a function of protein unfolding with temperature. This method soon became widely adopted with the implementation of real-time polymerase chain reaction (PCR) systems coupled with the use of Sypro^® Orange dye to monitor the unfolding event. A modification to the method, whereby the requirement for the dye was removed, with detection relying on tryptophan or tyrosine fluorescence to monitor protein unfolding, was introduced and termed nanoDSF. Addressing two common issues facing DSF and other isolated protein assays in early drug discovery: the requirement for purified protein and the question of the physiological relevance of the target protein outside of its native cellular environment, a further development in thermal unfolding studies, the cellular thermal shift assay (CETSA), was presented in 2013.⁸ For the first time, this allowed the use of thermal unfolding to detect compound-target engagement in cell lysates. The method has also been applied to intact cells, and the Western blot methodology described in the original publication has been extended to incorporate bead-based,⁹ mass spectrometry-based,¹⁰ and luminescence-based detection methods.¹¹ Thermal unfolding methods clearly represent powerful approaches to screen compounds for potential target engagement, and there is a wide choice of approaches dependent upon the nature of the target protein and the available reagents. Their common feature is that they exploit the thermodynamic stabilization of the protein occurring upon ligand binding,¹² allowing binding events to be indirectly detected. This review presents the principles of the different thermal unfolding methods that are employed in high-throughput screening, provides examples of screening outcomes, and highlights opportunities for future development.

Two potential issues that need consideration with thermal unfolding methods are (1) the potential change in pH of buffers used during thermal unfolding studies due to the enthalpy of ionization of the buffers used, and (2) the choice of scan rate. For buffers, such as Tris, with a reasonably large enthalpy of ionization (for Tris dpH/dT = −0.031), the pH can change by around 1.5 units over a 50 °C temperature range. Thus, the use of buffers with lower enthalpy of ionization (such as phosphate, dpH/dT = −0.0022, where the change over a 50 °C temperature range is only around 0.1 pH unit) is recommended, to avoid complications in interpreting the unfolding temperature and its stabilization by ligand binding. Scan rate dependence, which is the change in observed T_m when the rate of temperature increase is varied, can indicate the presence of non-equilibrium processes, such as irreversible unfolding. Typical scan rates are often around 1 °C/min, although modern instruments, such as nanoDSF, can have much faster rates. These fast scan rates can be problematic as the temperature change may be faster than the equilibration of the system, leading to a correlation between the scan rate and the melting temperature. It is also possible that ligand binding may alter the kinetics of the relaxation process complicating interpretation. The observation of scan rate dependence precludes the interpretation of ligand binding and stabilization in terms of equilibrium thermodynamics.

DIFFERENTIAL SCANNING CALORIMETRY

Differential scanning calorimetry (DSC)^13,14 is a powerful method able to study thermally induced transitions between conformational states of biological macromolecules and is the only method to directly measure the enthalpy of this transition. Usually, this involves monitoring the folded to unfolded transition of proteins or nucleic acids. The differential scanning calorimeter measures the excess heat capacity (C_p) of a solution of the macromolecule of interest as a function of temperature. The transition is signified by an endothermic peak, where the value of C_p is a maximum, which is centered on the melting temperature (T_m). The melting temperature is defined as the temperature at which 50% of the protein molecules are unfolded or, under ideal dynamic reversible two-state equilibrium conditions, the temperature at which any one molecule spends 50% of its time in the folded and 50% of its time in the unfolded state. Calculating the area under the thermogram yields the transition enthalpy (ΔH°), and the difference between the pre- and post-transition baseline yields the ΔC_p. Analyzing the shape of the peak, by fitting a two-state model to enable determination of the van't Hoff enthalpy (ΔH°_vH) gives an indication of the size of the cooperative unit of the transition from native to denatured state or may indicate the presence of intermediates in the unfolding pathway. For a two-state reversible equilibrium, we expect ΔH°_vH = ΔH°_cal. Where ΔH°_vH > ΔH°_cal, this may indicate that the size of the cooperative unit may be larger than the monomeric form and could indicate that an oligomeric state is the cooperative unit. Observation of ΔH°_vH < ΔH°_cal may indicate that the transition involves unfolding intermediates or independent domains. Thus, DSC provides a low throughput, but thorough, thermodynamic characterization of macromolecular thermal unfolding. The method may be extended to evaluate the effect of ligands on macromolecular unfolding, by including, usually saturating concentrations of ligand in the thermal unfolding experiment. A shift of the transition peak to higher temperatures in the presence of ligand provides an indication of macromolecule-ligand binding, although there is not a simple relationship between the magnitude of the shift in T_m and the ligand binding dissociation constant, K_d (the relationship depends upon the contribution of the entropic and enthalpic contributions to the K_d value and the concentrations of the protein and added ligand and the unfolding enthalpy).^6,15 The ability to fully characterize macromolecular thermal unfolding and the resulting effect of ligands, directly (without the need for additional equilibria associated with measuring the degree of unfolding) thus positions DSC as the benchmark method for comparison to other methods employed to study the thermal unfolding of isolated, purified macromolecules, such as differential scanning fluorimetry or nanoDSF (Fig. 1).

FIG. 1.

Technologies that can be utilized to monitor the thermal unfolding of isolated protein samples. Several technologies can be utilized to monitor the thermal unfolding of isolated protein samples. Differential scanning calorimetry (DSC) detects the change in enthalpy of the transition between folded and unfolded states and is the most direct method; however, it is only capable of low throughput analysis. nanoDSF can analyze up to 48 samples in parallel and measures the intrinsic fluorescence of tryptophan residues within the protein structure as their exposure to the environment changes during protein unfolding. Differential scanning fluorimetry (DSF), while less direct, is capable of higher sample throughput and can be used to analyze samples prepared in a multi-well plate format. This dye-based method can be used to simultaneously analyze up to 384 or 1536 samples using multi-well plate-compatible thermocycler/PCR instruments. This form of DSF measures the extrinsic fluorescence generated by the inclusion of environmentally sensitive dyes such as Sypro^® Orange, which bind to exposed hydrophobic regions of the protein structure during unfolding.

Several technologies can be utilized to monitor the thermal unfolding of isolated protein samples. Differential scanning calorimetry detects the change in enthalpy of the transition between folded and unfolded states and is the most direct method; however, it is only capable of low throughput analysis. nanoDSF can analyze up to 48 samples in parallel and measures the intrinsic fluorescence of tryptophan residues within the protein structure as their exposure to the environment changes during protein unfolding. Differential scanning fluorimetry, while less direct, is capable of higher sample throughput and can be used to analyze samples prepared in a multi-well plate format. This dye-based method can be used to simultaneously analyze up to 384 or 1536 samples using multi-well plate-compatible thermocycler/PCR instruments. This form of DSF measures the extrinsic fluorescence generated by the inclusion of environmentally sensitive dyes such as Sypro^® Orange, which bind to exposed hydrophobic regions of the protein structure during unfolding.

PRINCIPLES OF DSF

Differential scanning fluorimetry is a convenient, cost-effective, and robust method used to measure the thermal stability of isolated proteins under a range of different conditions.⁷ As with DSC, the assay is founded upon principles of protein thermodynamics, whereby proteins exist in equilibrium between their native (folded) state, favored under ambient conditions, and their denatured (unfolded) state.¹⁶ Gradually increasing the energy of the environment through heating in a controlled manner acts to shift this equilibrium toward the unfolded state, generating a thermal denaturation profile that can be used to derive a melting temperature (T_m) [Fig. 2(a)]. The T_m value is susceptible to changes in environment; changes in pH, buffer composition, or ligand binding can result in the formation of new molecular interactions which thermally stabilize a protein, leading to an increased T_m. The ability to monitor changes in T_m (ΔT_m) under different conditions has enabled DSF to be applied for multiple purposes within early drug discovery, from buffer screening to determine optimal conditions for protein stability^17–19 and crystallization²⁰ to ascertaining relative stabilities of isolated protein mutants or batch purifications.^21,22 Commonly, DSF is utilized for small molecule screening to indicate complex formation between a protein of interest and a compound, the occurrence of which usually manifests as a positive ΔT_m value compared to screening of protein alone [Fig. 2(c)].^23–26 It is possible for DSF traces to display greater than one unfolding transition, and therefore, more than one T_m value. This can be due to the presence of multiple domains within a protein structure [Fig. 2(b)]. While one domain may exhibit stabilization upon ligand binding, the other may remain the same as in the native protein, offering information about potential binding site location. Knowledge of the oligomeric state of the protein can be advantageous before performing a DSF experiment as the presence of multiple states can also result in greater than one thermal transition.

FIG. 2.

(a) Typical dye-based DSF thermal denaturation curve for protein sample. At the starting temperature, the protein is present in its folded, native state. Fluorescence emission increases with temperature as the protein unfolds and SYPRO^® Orange co-operatively binds to exposed hydrophobic residues. The melting temperature (Tm) (dotted line) can be defined as the midpoint of the transition, or the temperature at which 50% of the protein is unfolded. Fluorescence decreases after the peak is largely attributed to protein aggregation causing shielding and exclusion of dye from hydrophobic residues. (b) Thermal denaturation profile of protein showing two distinct transitions (T_m1 and T_m2) indicating the presence of multiple structural domains. While T_m1 remains the same as in the native protein, T_m2 shifts to a higher temperature in the presence of a ligand, offering information about potential binding site location. (c) Thermal denaturation profiles of a protein in the absence and presence of a ligand. Ligand presence causes the Tm to shift positively (positive ΔT_m) indicative of thermal stabilization due to binding. Raw fluorescence emission is plotted as a function of temperature. (d) First derivative of thermal denaturation curve plotted as a function of temperature. The T_m (dotted line) can also be defined as the inflection point, or peak, of the first derivative trace of the thermal shift curve, the temperature at which the steepest increase in fluorescence signal occurs.

Rather than observing the heat capacity of the protein (as in DSC), DSF monitors fluorescent signal as a function of temperature. Utilizing a fluorescence end point provides a precise, sensitive readout that is amenable to miniaturization within a multi-well plate format. The assay also benefits from much lower protein requirements than DSC.²⁷ These factors, coupled with the increased availability of lab instrumentation and software capable of monitoring unfolding across a temperature gradient (for example, instruments such as the LightCycler^®480 from Roche), have driven increased application of DSF as an accessible method for high throughput screening within early drug discovery.^28–31 The source of fluorescence monitored can either be extrinsic, as in dye-based DSF methods, or intrinsic to protein structure, as in the label-free DSF approach.

DYE-BASED DSF

Dye-based DSF involves monitoring the extrinsic fluorescence of environmentally sensitive dyes which are incubated with the protein sample. Under aqueous conditions, dye fluorescence is quenched. However, during thermal unfolding, hydrophobic residues within the protein’s core are exposed and can interact non-covalently with dye molecules, increasing the quantum yield of the dye and producing a fluorescent signal. This typically occurs in a cooperative manner resulting in a sigmoidal thermal denaturation profile, or “melt transition,” across a defined temperature range. Most commonly, a temperature gradient ranging from room temperature to ∼95 °C is applied, with a ramp rate of approximately ∼1 °C/min. However, parallelisable and sensitive isothermal screening approaches have been described whereby the stability of a protein is compared at a constant temperature mimicking physiological conditions.³² Although there are many commercially available extrinsic dyes, Sypro^® Orange (Thermo Fisher Scientific) is widely favored for DSF applications due to its high signal to background fluorescence in the presence of native protein. One caveat is the sensitivity of Sypro^® Orange dye to detergents; hydrophobic buffers and detergent presence cause high background fluorescence which can mask the assay signal and any resultant T_m shift in the presence of ligand. Likewise, the presence of EDTA aggregates at basic pH has also been shown to induce interference.³³ Fluorescent molecular rotor dyes, such as probe 4-(dicyanovinyl)julolidine (DCVJ) and the commercially available dye Proteostat^® (Enzo Life Sciences), however, have been applied for monitoring unfolding under high-detergent conditions.^34,35 A comprehensive summary of other extrinsic dyes is given by Hawe et al.³⁶ and more recently Gao et al.²² Sypro^® Orange dye has an excitation maximum of λ = ∼470–490 nm and an emission maximum of λ = ∼570–590 nm and is typically utilized at 1:1000–1:5000 of stock concentration diluted in DMSO.^37,38 However, extrinsic dye concentration is often optimized prior to each specific DSF assay undertaken.

As well as protein studies, dye-based DSF has been successfully applied for investigation into RNA stability. While intercalating dyes, such as Sybr^® Green, can affect RNA stability, sensitive analysis of structured RNA in the presence of stabilizing additives can be carried out through use of single-strand specific fluorescent dyes, such as Quant-iT™ RiboGreen^® (Thermo Fisher Scientific).³⁹

LABEL-FREE DSF

The label-free DSF approach removes the requirement for extrinsic dye molecules to be equilibrated with a protein sample. The technique instead relies upon monitoring the intrinsic fluorescence of tryptophan and/or tyrosine residues present within the protein structure, which alter as a result of changes in microenvironment polarity during protein unfolding.^40,41 Observing the ratio of fluorescence emission at λ = 330 and 350 nm as a function of temperature allows determination of a T_m value for the protein sample of interest. Development of instrumentation such as the Prometheus NT.48 instrument (NanoTemper Technologies) has facilitated miniaturization of the label-free DSF assay for higher throughput application; nanoDSF allows for monitoring of up to 48 samples in parallel, each requiring only 10 μl total volume.⁴² While lower throughput and less sensitive than dye-based DSF, nanoDSF is applicable to a wider range of protein samples, such as those containing detergents which cause interference in the former technique. nanoDSF is thus suitable for use in monitoring the thermal stability of membrane proteins for optimal detergent selection.⁴³ Use of nanoDSF for the reliable determination of protein T_m directly from crude cell lysates has also been described.⁴⁴ The label-free nature of nanoDSF also can be advantageous given that inclusion of extrinsic dyes such as Sypro^® Orange introduce a second binding equilibrium to the analysis. Dye binding can cause slight destabilization of the protein being monitored, and therefore underestimation of T_m value. This effect is thought to be due to dye molecules selectively associating with aromatic residues, biasing the thermodynamic equilibrium toward the unfolded state; the extent of destabilization has been shown to be more pronounced for protein sequences containing a high proportion of charged or hydrophobic aromatic residues (e.g., Arg, Glu, Trp, and Phe).⁴² Inclusion of excess extrinsic dye may also lead to competition with small molecule ligands for binding to the protein of interest, a factor that can be avoided through use of label-free DSF. Utilizing intrinsic fluorescence also offers the potential to follow refolding of a protein (i.e., transition from an unfolded to folded state), something that cannot be monitored using extrinsic dye.^45,46

ASSUMPTIONS AND LIMITATIONS OF DSF

While dye-based and label-free DSF techniques can be used to obtain valuable information regarding the thermal stability of isolated proteins, it is important to consider that successful implementation of both relies upon certain assumptions and that both are susceptible to their own limitations. The recent review by Gao et al.²² provides an extensive summary of the key assumptions relating to both DSF techniques, some of which are discussed below.

First, dye-based DSF experiments rely upon the assumption that proteins being studied possess minimal exposed hydrophobic residues in their native state and that such proteins are stable and non-aggregated prior to unfolding. Proteins that are unstable, aggregated, or those that possess significant hydrophobic surface area can cause elevated background fluorescence at low temperature. Likewise, specific binding interactions between the protein and the extrinsic dye could prevent dye exposure to the aqueous environment, resulting in high background prior to unfolding. For small molecule screening, experiments are performed under the assumption that there is minimal interaction between compounds or other assay components and the extrinsic dye. Furthermore, screening collections can be specifically designed to minimize inclusion of compounds that interact with or sequester the dye, as presence of these can lead to either high background or low fluorescence, respectively.

A key assumption for nanoDSF small-molecule screening experiments is that compounds included are not intrinsically fluorescent with emission spectra overlapping that of the protein. This assumption also applies to dye-based DSF methods; however, such interference is less likely in Sypro^® Orange assays, due to a greater assay signal and high emission maximum of the dye. In nanoDSF, the assay signal correlates with the number of tryptophan residues within the protein sequence, and thus, this should be considered prior to setup. Equally, as signal sensitivity is lower than dye-based assays, it should be considered that the presence of multiple protein domains with different unfolding transitions may be difficult to characterize using a nanoDSF approach.

Despite the availability of automated DSF analysis software, Simeonov⁴⁷ discusses the fact that there is typically a requirement to conduct manual analysis of thermal melt transitions. Inspection of individual traces within DSF screening data is often required as noise, signal interferences, or secondary transitions can complicate T_m determination leading to false positives and negatives. While greater ΔT_m shifts are often associated with higher affinity binding, this relationship does not always apply. The extent of the thermal shift can be influenced by multiple factors, including the net interplay between both entropic and enthalpic contributions during binding. T_m shift values are also influenced by additional equilibria present within the assay, for example, binding of a small molecule ligand to both the folded and unfolded protein states or dye itself binding to the protein. In this way, ligands with equal affinity can produce varying ΔT_m values.⁴⁷ Although large-scale DSF screening can be conducted at a single ligand concentration to increase throughput, follow-up of suspected binders in concentration-response format is often applied in order to determine a “true ΔT_m,” for which the presence of the ligand under saturating conditions is a prerequisite,⁴⁷ although it should be recognized that the T_m value increases with ligand concentration even beyond saturation, due to a positive entropy of mixing effect.

The observation of negative thermal shifts (i.e., decrease in T_m) in the presence of small molecules has also been shown by DSF. While ligand binding to a protein form is always associated with stabilization of that form, due to the lower free energy of the protein-ligand complex, negative thermal shifts or observed destabilization can occur in several scenarios. For example, ligand binding preferentially to the unfolded form of a protein over the native form can cause an observed decrease in T_m relative to that of the protein in the absence of ligand.⁴⁸ Equally, if analysis is conducted on a heterogenous sample containing multiple protein populations, binding of a ligand preferentially to a minority population with a lower T_m value could perturb the equilibrium leading to an observed decrease in T_m. Other factors that have been associated with negative thermal shifts include covalent modification of the protein due to ligand binding, and altered ionic strength, depleting ions that would otherwise promote protein stabilization.⁴⁷ Inclusion of binding partners within a DSF assay, or their physiological presence in the case of cellular thermal shift assays, can complicate analysis further due to potential for additional interactions to occur. Due to difficulty in interpretation of negative thermal shift data, along with observations of nonspecific binding leading to decrease in T_m, to date focus has largely been placed on ligands producing typical positive shifts. Comprehensive analysis into negative shift data and larger scale validation of negative shifting ligands would be beneficial to understanding the true value of this observation for future studies.

PRINCIPLES OF CETSA

The Cellular Thermal Shift Assay (CETSA) was developed following a collaboration between the Karolinska Institutet, Sweden and the University of Nanyang, Singapore and is the most recent of the thermal unfolding methods described here to be adapted to drug discovery.⁸ CETSA, like other thermal shift assays, works on the principle that compound binding to a target protein alters the thermal stability of the target protein.^12,49 Unlike thermal unfolding methods such as DSC and DSF, CETSA is advantaged by its sensitivity to factors such as cell permeability of a compound, intracellular protein complex formation, and the impact of the presence/absence of cellular cofactors.⁵⁰ The general method for CETSA varies significantly depending on the application and the detection technology. Broadly, cells are co-incubated with compounds of interest for a short period. The cells are then heat-treated, usually using a thermal cycler or a heat block, and are subsequently lysed. Following cell lysis, the samples must usually be centrifuged, to separate the folded, soluble target of interest in the supernatant from the aggregated, melted target in the pellet. Compounds that alter the quantity of the folded soluble target of interest are posited to bind to the target of interest.⁸

DETECTION METHODS LEVERAGED FOR CETSA

In the first CETSA experiments, Molina Martinez et al. used Western blot to detect the amount of soluble folded protein that remained following thermal destabilization of whole cells, lysed cells, or lysed tissues, in the presence and absence of compounds.⁸ To generate a CETSA melt curve for a particular target protein, CETSA must be run across a range of temperatures, which means that target detection by Western blot is relatively low in terms of compound throughput. Therefore, the use of Western blot was rapidly followed by the development of other methods (Fig. 3). Multiplexing of samples allows for higher throughput protein detection by mass spectrometry,⁵⁰ whilst AlphaScreen antibody/bead or NanoLuciferase protein detection reagents have enabled CETSA to be miniaturized into 384-well plate-based formats termed “AlphaCETSA” by PerkinElmer^51,52 and “BiTSA,”¹¹ “SplitLuc CETSA,”⁵³ or “NaLTSA.”⁵⁴ Though mass spectrometry remains lower throughput than AlphaScreen or NanoLuciferase-based detection technologies, mass spectrometry has the advantage of being totally label-free, requiring neither specific antibodies nor the introduction of a protein tag to the target of interest.

FIG. 3.

A range of technologies can be used to detect unfolding of the protein target(s) of interest for CETSA. Detection of the endogenous protein is enabled by mass spectrometry or antibody-based methods, whilst detection of genetically tagged target is most often based on NanoLuciferase technologies. Left to right images represent the target detection by (1) Western blot,⁸ (2) AlphaScreen anti-target antibody pairs,⁵¹ (3) immunofluorescence microscopy,^56,57 (4) mass spectrometry,⁵⁰ (5) acoustically dispensed reverse phase protein array,⁵⁵ (6) TIRF of anti-target antibody, bound to immobilized target, released following single cell lysis within a microfluidic device,⁶¹ (7) monitoring of NanoLuciferase formed following the binding of LgBiT to the HiBiT-tagged target,⁵³ and (8) monitoring of NanoLuciferase engineered to be thermostable using a Gly-Ser linker, and thus suitable for real time CETSA.⁵⁸ (b) When applying CETSA to drug discovery, the concentration of the compound, the temperature of the melt event, or the treatment of the cells may all be varied to elucidate further information about the compound's mechanism of action.^8,49,62–64

The detection method for the target of interest continues to be developed to drive higher throughput compound screening in drug discovery. Acoustic droplet ejection technology has been used to enable 384-well plate-based reverse phase protein array (RPPA)-CETSA experiments, effectively providing a miniaturization of the Western blot approach with an imaging end point (Fig. 3).⁵⁵ However, this method requires that the CETSA is performed on cell lysates; thus, information about compound membrane permeability is lost, along with all cellular protein that is not in the soluble fraction. Other groups have recently developed single-cell high content imaging (HCI)-CETSA where antibody detection is achieved by single-cell immunofluorescence (Fig. 3).^56,57 Conversely to RPPA-CETSA, HCI-CETSA requires the CETSA experiment to be undertaken using intact cells. HCI-CETSA has been miniaturized to a 384-well plate-based format, and indeed Axelsson et al.⁵⁶ screened a library of >1000 compounds as proof-of-principle.

Most recently, the Henderson group, which originally implemented split NanoLuciferase HiBiT-CETSA, developed a 384-well plate thermal cycler capable of capturing luminescence (Fig. 3).^53,58 They used this to detect their target of interest, which had been labeled with a NanoLuciferase engineered to be thermostable (termed “ThermLuc”). This allowed for the first instance of live cell real-time CETSA experiments and thus enabled the detection of melt curves for 384 samples within a 4-minute total melt/read cycle. Further miniaturization to 1536-well plates for high-throughput compound screening has been attempted, though remains challenging due to the requirement to separate the supernatant of the cell lysate from the pellet in all but real-time ThermLuc CETSA and a few target-dependent exceptions.^53,58 Research into CETSA method development over the last 10 years has enabled the first truly high-throughput drug discovery effort driven by CETSA, in which Rowlands et al. have undertaken a library screen of >300 000 compounds using the AlphaCETSA method.⁵⁹

Conversely, in drug discovery applications where compound throughput is not a priority, CETSA can be used to help to elucidate our understanding of the compound/target interaction. CETSA is particularly useful in proving cellular target engagement soon after high-throughput biochemical screens and alongside low-throughput phenotypic cellular screens. The use of mass spectrometry enables the detection of the wider proteome, and therefore also provides information about how a compound might affect the melting temperature (T_m) of thousands of different proteins. This method, known as thermal proteome profiling (TPP), mass spectrometry (MS)-CETSA or CETSA^®Explore can provide information about direct binding partners that appear to be co-stabilized or destabilized by the compound/target complex.^50,60 Another method that also aims to facilitate cellular target engagement studies, but does not improve compound throughput, is single-cell total internal reflection fluorescence microscopy (TIRF)-CETSA (Fig. 3).⁶¹ Relative to other CETSA methods that often require many cells per sample, this method requires only hundreds of cells per conditions, and thus can be used on rare patient-derived cells. The authors captured and lysed one cell at a time using a microfluidic chip and laser-based optical trapping, and then used TIRF microscopy to enable single-molecule antibody-based target detection.

There are very few publications that directly compare the findings of CETSA using different detection methods. Martinez et al.⁵³ have shown that 18 different HiBiT-tagged targets melt at similar T_m values to the endogenous protein; meanwhile, Sanchez et al.⁵⁸ have benchmarked ThermLuc CETSA against SplitLuc CETSA and have shown strong agreement between the two methods across a panel of 29 previously identified binders to LDHA. On a smaller scale, Herledan et al.⁵⁵ have demonstrated that Western blot CETSA and RPPA CETSA produced concordant thermal shifts for three inhibitors of insulin-degrading enzyme (IDE), whilst Massey et al.⁵⁷ have shown that AlphaCETSA and HCI CETSA produced concordant thermal shifts for five inhibitors of p38a. Whilst these studies are encouraging, there are few studies to date that compare many binders or targets across different CETSA detection methods. It seems likely that different CETSA detection methods have different liabilities that might result in different putative binders being identified as false positives or false negatives. We therefore recommend that this be a focus of future drug discovery efforts to maximize the sensitivity and specificity of CETSA assays.

VARIATIONS TO THE CETSA FORMAT

CETSA is an extremely flexible method in its application to drug discovery which can be run in many different experimental formats to provide different information (Fig. 3). To generate a classical CETSA melt curve, the temperature of the heat shock is varied whilst the concentration of the compound is fixed in excess.⁸ As is the case for DSF, compounds that bind and stabilize the target will cause the T_m of the target to increase, though the degree to which a compound (de-)stabilizes a target of interest does not necessarily correlate with the potency with which it binds to the target.⁴⁷ To obtain potency, the compound concentration is instead varied at a fixed temperature usually close to the target T_m.⁸ The concentration at which a particular compound stabilizes the target by 50% of the maximal possible stabilization effect that the compound can cause is used to determine the EC₅₀ and thereby provides an estimate of compound potency. This method variation is known as isothermal dose-response CETSA (ITDR-CETSA). If compound throughput is a particular priority, then isothermal CETSA, where the temperature of the heat shock and the compound concentration are both fixed, has been used to screen larger compound libraries of up to 1120 compounds.⁵⁶ We have since screened libraries of >300 000 compounds using this approach.⁵⁹

In 2D-CETSA, ITDR-CETSA and melt curves are combined such that the temperature of the heat shock and the compound concentration are varied in tandem, to access even more information.^64,65 For instance, some proteins might alter in abundance in response to increasing concentration of a particular compound, across all temperatures. These proteins are likely to be changing in expression or extractability due to compound treatment, rather than be direct binders that cause thermal stabilization, and might instead be involved in signaling pathways downstream of compound binding. Meanwhile, proteins that are thermally stabilized by the compound of interest even at very low compound concentrations might represent previously unknown off-target binding at high potency.⁶² A variant of 2D-CETSA, termed compressed CETSA, has also been developed.⁶⁶ Samples melted at a range of different temperatures are pooled to increase the throughput relative to 2D-CETSA, whilst the output remains more robust than a truly isothermal approach.

The final variable that may be altered in a CETSA experiment is the cellular state. There are studies that have applied TPP/MS-CETSA across cells synchronized to different stages of the cell cycle, cells with different genetic knock out backgrounds, or to a cellular time series post-treatment.^49,62–64 Although to date these studies have not been undertaken with a direct drug discovery application, they highlight that unlike the other biophysical methods described here, CETSA enables compound potency to be interrogated in a label-free manner across a range of cellular states/conditions; therefore, CETSA holds great potential for impact in drug discovery.

Often, melting temperatures with and without ligand agree between DSF and CETSA experiments, for example.^67,68 However, it is possible that the additional complexity of the cellular system may not lead to correlations between CETSA experiments and those conducted on isolated proteins. In these cases, understanding of the protein system and its potential for interactions in a cellular context should be attempted to rationalize any differences in T_m and ΔT_m measurements.

The different methods used for monitoring thermal unfolding discussed above are applied at different stages and employ different reagents during the early part of drug discovery. They have a range of strengths and weaknesses, and the choice may depend upon the access to reagents and the different types of information required. Often it may be useful to combine methods as this may provide increased confidence in the engagement of the compound with the target. Table I provides a brief description of the pros and cons of the methods described.

TABLE I.

Overview of the advantages and disadvantages of the different thermal unfolding methods applied in drug discovery.

Technique	Advantages	Disadvantages
DSC	No dyes, labels, Abs required. Direct measurement of heat capacity and enthalpy of unfolding	High protein consumption, low throughput.
DSF	High throughput, lower sample consumption. Can be automated	Requires dye binding. Potential issues with dye binding perturbing the equilibrium. Temperature dependence of fluorescence signal. May detect aggregation of dye.
nanoDSF	No dye required. Ability to take the ratio of 2 wavelengths to remove temperature dependence of fluorescence effect	Requires tryptophan / tyrosine content. More sensitive to compound fluorescence.
CETSA	Measurement in a physiological setting. A range of detection methodologies available	Complexity of interpretation.

Within AstraZeneca, thermal unfolding approaches are often utilized as primary screening approaches, either employing DSF at the 100k compound scale or CETSA at around 500k compound scale. Typically, following confirmation in the primary assay, hit compounds with the largest degree of stabilization are followed up with concentration response curves (in DSF) and concentration response curves and/or full melt curves in CETSA. Recognizing that apparent affinities measured in this way provide values at the melting temperature, applying biophysical approaches such as SPR provides additional confidence in the hits identified, before chemistry is applied. The degree of agreement between the thermal methods and the downstream biophysical methods is often target dependent.

ANALYSIS OF THERMAL UNFOLDING DATA

Most high-throughput thermal shift assays (DSC, DSF, and CETSA) focus on determination of the melting temperature of proteins (T_m) and often more specifically the changes in T_m induced upon compound binding.⁷ All three methods in their original form quantify protein stability either as protein in solution (CETSA), dye fluorescence (DSF), or the excess heat capacity (DSC) as a function of increasing temperature. Figure 2(c) illustrates the melting curve for a protein in an example of a DSF experiment in the absence and presence of a ligand. The ligand stabilizes the protein and increases its melting temperate as shown by the dashed lines.

For DSF, there are three general data analysis approaches to determining the T_m and ΔT_m upon binding. First, the derivative method [Fig. 2(d)] uses the smoothed first derivative of the melting curve signal. The T_m is defined as the temperature at the inflection point, with the steepest increase in signal with temperature. Usually, a smoothing function such as a Savitzky–Golay filter is required to reduce experimental noise and produce a reliable T_m estimate. The main advantages of the derivative method are that it is simple, quick, robust, and has fewer assumptions compared to the “thermodynamic” methods below. It can also help uncover proteins with multiple transitions manifesting as multiple peaks in the derivative signal. However, if those transitions are very close to each other, the T_m estimates of the derivative may become more variable. In those cases, an even simpler model using the middle point between the bottom and the top of the curve may be more robust. The midpoint model approximates the T_m as the intersection of the arithmetic mean of the top and bottom fluorescence plateaus with the melting curve. However, it is based on a linear relationship in the region between the bottom and the top, which is rarely seen in experimental data. Therefore, the midpoint method is generally recommended only when all other methods fail. The third group of methods relies on equilibrium and thermodynamic models of protein unfolding. It encompasses a wide continuum of methods with increasing levels of complexity and number of parameters.

Generally, when a protein is heated in thermal stability experiments, the relationship of folded (F) and unfolded (U) protein during heating can be represented as a reversible reaction which reaches equilibrium faster than the heat ramp rate.

$$\mathrm{F}\rightleftharpoons \mathrm{U}.$$

The equilibrium constant between the folded and unfolded states is defined as

$$K_u=\frac{\left[U\right]}{\left[F\right]},$$

where $\left[U\right]+\left[F\right]=\left[P_{\mathrm{tot}}\right]$ are the concentrations of the folded and unfolded protein, and $P_{\mathrm{tot}}$ is the total concentration of all protein species. With that in mind, the fraction folded can be expressed as

$$f_{\left(F\right)}=\frac{1}{1+K_u}.$$

$K_u$ is related to the difference in Gibbs free energy such that

$$\Delta G_u=-RT\mathrm{\hspace{0.17em}ln}{K}_u,$$

where R is the universal gas constant, and T is the absolute temperature. Substitution leads to the Boltzmann equation:

$$f_{\left(F\right)}=\frac{1}{1+e^{-\Delta G_u/RT}}$$

In thermal stability experiments for CETSA and DSF, this is often expressed as

$$f_{\left(F\right)}=\frac{1}{1+e^{\frac{T_m-T}{\mathrm{Slope}}}}$$

This equation is a convenient simplification used to determine the melting point, T_m of the protein unfolding transition, i.e., the inflection point of the curve. It can incorporate a further constant or linear dependency for the start (S₀) and the end (S_inf) of the curve:

$$\mathrm{Measured}\hspace{0.17em}\mathrm{Signal}=Y\left(T\right)=S_0+\frac{S_{\inf }-S_0}{1+e^{\frac{Tm-T}{\mathit{Slope}}}}$$

This sigmoidal fit, known as the Boltzmann fit, is in good agreement with the behavior of proteins in CETSA experiments where the melt-curves have a sigmoidal shape. That is why it is the most used model fitted to data from cellular thermal shift assays.

In contrast, biophysical techniques, such as DSC and DSF, applied to isolated protein systems can determine the thermodynamics of thermal unfolding. For example, the expression above for ΔG can be recast as

$$\Delta G_u=\Delta H_u-T\hspace{0.17em}*\hspace{0.17em}\Delta S_u,$$

where ΔH_u and ΔS_u are the differences in entropy and enthalpy at the temperature for ΔG_u. This relationship and the deviation from the simplistic sigmoid fit is readily observed in differential scanning fluorimetry and calorimetry studies (DSF and DSC).

As described by Robertson and Murphy in 1997,⁶⁹ the temperature dependence of ΔH_u and ΔS_u is governed by the change in heat capacity ΔC_pu between the folded and unfolded states. In DSC experiments, ΔC_pu is in principle simply the difference in heat capacities (C_p) (baselines) before and after unfolding. Due to solvent restructuring, it takes more heat to raise the temperature of a solution of unfolded protein compared to a folded protein with the same concentration. If the, usually small, temperature dependence of ΔC_pu is ignored, the relationship of ΔG_u on temperature ΔG_u(T) can be expressed as

$$\Delta G_u\left(T\right)=\Delta H_u\left(T_R\right)-T\Delta S_u\left(T_R\right)+\Delta C_p\left[\left(T-T_R\right)-T\ln \left(\frac{T}{T_R}\right)\right].$$

When the reference temperature T_R = T_m, the melting temperature of the protein, then ΔG_u = 0 and $\Delta S_u=\frac{\Delta H_u}{T_m}$, and the relationship simplifies to the modified Gibbs−Helmholtz equation:

$$\Delta G_u\left(T\right)=\Delta H_m\left(1-\frac{T}{T_m}\right)+\Delta C_p\left[\left(T-T_m\right)-T\ln \left(\frac{T}{T_m}\right)\right].$$

The experimentally observed equilibrium constant can then be expressed as

$$K_{\mathit{obs}}\left(T\right)=e^{-\frac{\Delta G_u}{RT}}=e^{-\frac{\Delta H_m(1-\frac{T}{T_m})+\Delta C_p\left[\right(T-T_m)-\mathit{Tln}(\frac{T}{T_m}\left)\right]}{RT}}.$$

In the case of DSF data, the observed change in fluorescence can then be expressed as

$$Y\left(T\right)=S_0-\frac{S_{\inf }-S_0}{1+K_{\mathrm{obs}}\left(T\right)}=S_0-\frac{S_{\inf }-S_0}{1+e^{-\frac{\Delta H_m(1-\frac{T}{T_m})+\Delta C_p\left[\right(T-T_m)-\mathit{Tln}(\frac{T}{T_m}\left)\right]}{RT}}}$$

In practice, the S₀ and S_inf in DSF experiments often change with temperature, and the relationship can be modeled as linear:⁷⁰

$$S_0=Y_{f\left(T\right)}=Y_{f,T_m}+m_f(T-T_m),$$

$$S_{\inf }=Y_{u\left(T\right)}=Y_{u,T_m}+m_u\left(T-T_m\right),$$

where $Y_{f,T_m},Y_{u,T_m}$ are fluorescence intensities of the dye in the presence of folded and unfolded protein at the melting temperature T_m, and m_f and m_u are the slopes describing the linear temperature dependency of dye fluorescence for the folded and unfolded states. This fit with linear dependency for the top and the bottom is the expression that is most commonly referred to as a “thermodynamic” fit or more confusingly “Boltzmann” fit.

This model has seven parameters: $Y_{f,T_m},\hspace{0.17em}{Y}_{u,T_m},\hspace{0.17em}{m}_f,\hspace{0.17em}{m}_u,\hspace{0.17em}\Delta H_u,\Delta C_{pu},\hspace{0.17em}\text{and}\hspace{0.17em}{T}_m$, which can result in overfitting to experimental data and poses problems with identifiability of individual parameters. For example, ΔC_pu can often be poorly defined, and approaches have been suggested to measure it using separate DSC data⁷⁰ or to calculate it based on the number of amino acid residues (N_res) and area of the protein.^69,71

$$\Delta C_P=13.88\hspace{0.17em}{N}_{\mathrm{res}}\frac{\mathrm{cal}}{\mathrm{mol}\hspace{0.17em}\mathrm{K}}=58\hspace{0.17em}{N}_{\mathrm{res}}\frac{\mathrm{J}}{\mathrm{mol}\hspace{0.17em}\mathrm{K}}.$$

Robertson and Murphy⁶⁹ and Rees and Robertson⁷¹ also report estimates for ΔH_u and ΔS_u which can be used as starting points to aid convergence during nonlinear regression analysis.

All the above considerations have focused on an equilibrium model of protein stability with a reversible transition between folded and unfolded protein and no intermediate states. The models above require the assumptions that ΔC_pu does not change with temperature, that $\Delta S_u=\frac{\Delta H_u}{T_m}$ along the whole temperature range, when that is only valid at the T_m and that there are linear relationships for the temperature-dependent fluorescence signals.⁷⁰ Often, one or more of these assumptions may not hold, requiring more complex models be fitted to the experimentally observed data.

For example, it has been suggested, where Sypro^® Orange dye is used that a nonlinear model may better describe the post-transition behavior.⁷² A common issue may be irreversible aggregation of the protein, following the initial reversible unfolding transition. Lumry–Eyring models have been proposed by Sanchez-Ruiz⁷³ to account for this irreversible aggregation process, and useful tools have recently become readily available in open-source code (https://github.com/taiawu/dsfworld_package) and applications (https://gestwickilab.shinyapps.io/dsfworld/) to analyze these more complex situations.⁷⁴

In addition, it is important to bear in mind when fitting these more complex models that the magnitude and steepness of the pre- and post-unfolding baselines affect the entire fit and the values of the thermodynamic parameters both in simple linear models⁶⁹ and in the exponential decay models.⁷⁴ Therefore, melting points determined using the thermodynamic set of methods would often differ from those using the derivative or the midpoint methods.

High throughput thermal shift experiments also can provide information on the affinity of ligands for the target. The equation for the ligand binding equilibrium association constant at the T_m is given in the original⁷ publication for high-throughput DSF

$$K_L^{T_m}=\frac{\hspace{0.17em}\text{exp}(-\frac{\Delta H_u^{T_0}}{R}(\frac{1}{T_m}-\frac{1}{T_0})+\frac{\Delta C_{pu}^{T_0}}{R}(ln\left(\frac{T_m}{T_o}\right)+\frac{T_0}{T_m}-1)}{\left[L_{Tm}\right]},$$

where $K_L^{T_m}$ is the ligand association constant at $T_m,\hspace{0.17em}{T}_m$ is the midpoint of the unfolding transition in the presence of ligand, $T_0$ is the midpoint of the unfolding transition in the absence of ligand, $\Delta H_u^{T_0}$ is the enthalpy of protein unfolding in the absence of ligand at $T_0$, $\Delta C_{pu}^{T_0}$ is the change in heat capacity upon protein unfolding in the absence of ligand, $\left[L_{Tm}\right]$ is the free ligand concentration at $T_m$ which is approximately $L_{\mathrm{total}}$ when $L_{\mathrm{total}}\gg {\left[\mathrm{Protein}\right]}_{\mathrm{total}},$ R is the gas constant.

More recent work employing an isothermal approach⁷⁵ close to the T_m has described a further simplified method for determination of the ligand association constant at temperatures close to T_m

$$K_L^{Tm}=(1-f_{u0}^T)*(E{C}_{50}-\frac{{\left[P\right]}_T}{2}),$$

where [P_T] is the total protein concentration, and EC₅₀ is the ligand concentration where the fraction of unfolded protein is half of that observed in the absence of ligand. Note that the EC₅₀ in this case is not defined by the ligand concentration at which half of the protein is unfolded. $f_{u0}^T$ is the fraction of protein that is unfolded at the T ≈ T_m in the absence of ligand, and it can be determined directly from the thermal unfolding curve for the protein in the absence of ligand. The authors provide open-source code (https://sourceforge.net/projects/dsf-fitting/) to perform the suggested isothermal analysis.

Of course, to be physiologically relevant, association constants need to be transposed to values at physiological temperatures. This may be accomplished using the expression

$$K_L^T=K_L^{Tm}*\text{exp}(-\frac{\Delta H_L^T}{R}(\frac{1}{T}-\frac{1}{T_m})+\frac{\Delta C_{pL}^T}{R}(\ln \left(\frac{T}{{\mathrm{T}}_m}\right)+1+\frac{T}{T_m}).$$

The binding constants at physiological temperatures can then be recalculated provided that $\Delta C_{pL}^T$ can be measured, for example, using isothermal titration calorimetry, or assuming a low value, so that the second term can be ignored. These assumptions simplify the estimate for $K_L^T$ to

$$K_L^T=K_L^{T_m}\text{exp}(-\frac{\Delta H_L^T}{R}(\frac{1}{T}-\frac{1}{T_m}\left)\right).$$

While Pantoliano et al.⁷ suggest informed estimates of −10 to −15 kcal/moL for $\Delta H_L^T$, if real enthalpies of binding differ substantially between ligands the ranking based on the association constants and their corrected estimates at physiological temperature would be unreliable.⁷⁶ Therefore, the binding enthalpy of the ligand (ΔH_L^T) must be measured using calorimetric or van't Hoff methods.

Recent attempts have been made to quantify binding affinities using DSF data based on the isothermal method in nanoDSF experiments⁷⁷ and the authors provide readily available open tools for others to model and analyze their data.

Overall, with a plethora of commercial packages (such as GeneData Screener) and open-source tools like DSF World,⁷⁴ IsothermalFitting,⁷⁵ FoldAffinity,⁷⁷ MeltDown,⁷⁸ and SimpleDSFViewer,⁷⁹ scientists can understand, model, and fit DSF data in a high-throughput manner to find, rank, and characterize compounds which interact with the target of interest.

CETSA MELT CURVE ANALYSIS

As discussed above, the CETSA melt curve data fits the Boltzmann equation well, and thus allows for an accurate determination of T_m.^50,80 However, whilst for DSC and DSF experiments the ΔT_m is a sufficient metric for identifying binding events, it has been suggested that for CETSA experiments, the ΔT_m method may be too reductionist. This is because in DSF and DSC, the melt curves are determined from a definite amount of uncomplexed protein, but in CETSA experiments, the cellular environment could be driving changes in factors including target expression, localization, or complex formation. The result is that the melt curves report more than a simple thermal unfolding event, and therefore may vary in T_m, slope and plateau position.⁸¹ It is also possible that the T_m of a compound might lie beyond the assayed temperature range.⁸² As with DSC and DSF, there is a desire to be able to determine the degree of thermal shift at scale, i.e., for many compounds. Thus, there are several analysis approaches that have been developed to assess the degree of stabilization in an automated manner, for example, see recent developments in software packages such as TPP (https://www.bioconductor.org/packages/release/bioc/html/TPP.html).

FUTURE PERSPECTIVE

Thermal unfolding methods have had an important impact in early drug discovery, providing an understanding of protein stability, the thermodynamic detail around changes in stability upon ligand binding and by providing rapid and high-throughput approaches for identifying potential ligands. This impact is set to continue. However, we may expect to observe a shift from isolated protein primary screening approaches to those that make use of cell-based methods. The combination with CETSA, which does not rely on producing large quantities of purified protein, often for targets that are difficult to work with, opens the possibility for new screening cascades. For example, HiBIT CETSA methods will be used to identify potential hit molecules with follow-up using additional thermal unfolding methods but applied to far fewer compounds. This reduces the need for screening quantities of purified protein but ensures that rigorous characterization of actives from the primary screen is undertaken. We are already using thermal unfolding methods in this way, employing DSF to verify CETSA actives, using both single concentration and concentration response analysis, and bringing in additional biophysical methods once the confidence in those actives has been sufficiently demonstrated. The ability to undertake isothermal studies as well as thermal melting profiles and to combine different CETSA methods (SplitLuc and Alpha) alongside the isolated protein work in DSF and DSC is very powerful and forms the basis of a robust and rigorous downstream cascade. We expect that the MS-based methods will also feature strongly as thermal unfolding continues to have an important impact in hit identification and characterization.

AUTHOR DECLARATIONS

Conflict of Interest

The authors have no conflicts to disclose.

Author Contributions

Poppy Llowarch and Laura Usselmann contributed equally to this work.

Poppy Llowarch: Writing – original draft (equal); Writing – review & editing (equal). Laura Usselmann: Writing – original draft (equal); Writing – review & editing (equal). Delyan Ivanov: Writing – original draft (equal); Writing – review & editing (equal). Geoffrey Allan Holdgate: Conceptualization (lead); Supervision (lead); Writing – original draft (equal); Writing – review & editing (equal).

DATA AVAILABILITY

Data sharing is not applicable to this article as no new data were created or analyzed in this study.