Thermal Shift Assays: DSF, nanoDSF and What a Tm Shift Means

A thermal shift assay does not measure stability. It measures a fluorescence signal as a sample is heated, and reports the temperature at which that signal changes most steeply. Everything downstream — a melting temperature, a shift, a ranking of buffers, a claim that a compound binds — is an interpretation layered on top of one number extracted from one curve. The interpretations are not equally safe, and the least safe of them is the one people reach for most: that a bigger shift means tighter binding.

Two ways to watch a protein unfold

The two dominant formats differ in what produces the light.

Dye-based DSF, also called differential scanning fluorimetry or Thermofluor, adds an extrinsic reporter. In the protocol that made the method routine, the temperature at which a protein unfolds is measured by an increase in the fluorescence of a dye with affinity for hydrophobic parts of the protein, which are exposed as the protein unfolds; a simple fitting procedure gives the transition midpoint, and the assay is best performed on a conventional real-time PCR instrument ^[1]. That last detail is why the technique spread: most labs already owned the hardware. A subsequent methods survey confirmed the point from the other direction, describing which RT-PCR machines can be adapted to run Thermofluor assays and what conditions are needed to set a screen up ^[2].

nanoDSF removes the dye. It reads the protein's own aromatic residues — the measurements are based on the intrinsic aromatic amino acid residue fluorescence of the protein itself ^[3]. Tryptophan emission is sensitive to the polarity of its surroundings, so burying and exposing those residues changes the spectrum. Instruments in this class typically track emission at two wavelengths near 330 and 350 nm and report their ratio, which is an instrument convention rather than a finding, and many pair it with a light-scattering channel so unfolding and aggregation are recorded on the same heating ramp. One nanoDSF study of immunoglobulin G describes exactly that combination: a high-throughput technique that simultaneously monitors protein unfolding and aggregation properties ^[4].

The label-free format has a corresponding requirement that is easy to overlook: it needs aromatic residues to read. A sequence without tryptophan, and with little tyrosine, gives a weak intrinsic signal. That follows from what the measurement is and is stated here without citation.

What the dye limits, and what the dye cannot see

The extrinsic-dye format has a coverage problem that has been quantified. A 2025 dye-library paper states plainly that Sypro Orange is only compatible with a minority of proteins — under 30 percent — which limits the scope of the method ^[5]. The response has been to stop assuming one dye fits all. Protein-adaptive DSF pre-screens each protein against a chemically diverse dye collection and picks a working pair; the platform was successfully applied to 94 percent of the proteins tested, 66 of 70 ^[6]. The expanded collection reported in 2025 contains 517 structurally diverse molecules and was screened against a panel of about 100 proteins chosen in part to represent the most challenging targets, small size among them, reaching a 94 percent success rate; the authors read this as support for the idea that larger, more chemically diverse libraries improve the likelihood of detecting melting transitions across a wider range of proteins ^[5].

Note the direction of that failure. A protein incompatible with the dye does not usually produce a wrong Tm. It produces no usable transition, or a high, flat baseline — and the well looks like a failed experiment rather than a misleading one. The dangerous case is the reverse: anything else in the well that binds the dye or fluoresces at the same wavelengths contributes signal. Intrinsic-fluorescence formats have the analogous problem, since every tryptophan in the sample emits, not only the ones in your protein. A membrane-protein nanoDSF study leveraged a recently developed non-fluorescent membrane scaffold protein for this reason, so that the measurement could report the intrinsic tryptophan fluorescence of the membrane protein as it unfolds without signal interference from the surrounding "dark" nanodisc ^[7].

What a melting temperature actually reports

Tm is the midpoint of an observed transition: the temperature at which the reporter says half the population has gone from the folded-like state to the unfolded-like state. Three qualifications matter more than the number.

It is the midpoint of what the probe sees, not of the whole molecule. A protein can give more than one transition, and when it does, a single reported number is a summary of several events. The membrane-protein study above resolved a biphasic unfolding pattern with a minor transition followed by a major one, giving two distinct midpoint melting temperatures of 70.5 °C and 77.5 °C — 7.0 °C apart — which the authors read as consistent with a three-state unfolding model, and which they traced to a conserved disulfide bond between two cysteines driving the intermediate state ^[7]. Reporting "the Tm" for a protein like that discards most of the information.

It is not necessarily an equilibrium quantity. Thermal unfolding is often irreversible, and when it is, the apparent midpoint depends on how fast you heated. In a study of two full-length immunoglobulin light chains, the thermal unfolding of both proteins was irreversible and scan-rate dependent ^[8]. A Tm from an irreversible transition is a kinetic observable dressed as a thermodynamic one. It is still useful for ranking conditions measured the same way, and close to useless for comparison across papers that used different ramp rates.

It is conformational, not chemical. A melting temperature says something about the free-energy gap between folded and unfolded states under the conditions in the well. It says nothing about whether the peptide bonds are intact, whether asparagines have deamidated, or whether methionines have oxidized. Those are separate measurements with separate analytical signatures.

Comparison across techniques is also not free. A stability-probe evaluation describes differential scanning calorimetry as the "gold standard" for thermal stability assays while noting that the use of an external probe increases throughput, making it more suitable for screening, and that current probe-based methodologies suffer from relatively low sensitivity — adding that while DSF is an effective screening tool, interpretation and comparison of the results is often complicated ^[9].

Why a Tm shift is not an affinity

A ligand that binds the folded state preferentially will, other things equal, raise the melting temperature. That is the whole basis of the assay, and the original protocol states the relationship with a hedge that has been widely dropped since: the difference in midpoint temperature in the presence and absence of ligand is related to the binding affinity of the small molecule ^[1]. Related to is not equal to, and the gap has structure.

On the standard thermodynamic treatment the shift depends on the binding constant at the melting temperature and on the enthalpy of unfolding, and a fluorescence experiment measures neither of those directly — a consequence of the thermodynamic framework rather than a finding from any one study, and stated here without citation on that basis. The thermodynamic framework that converts a thermal transition into a binding constant was formulated for calorimetry, where the unfolding enthalpy is obtained from the same scan. Brandts and Lin developed models for interacting systems involving one thermal transition, protein–ligand, and two, protein–protein, and extracted binding constants and heats of binding by calculation or by simulation fitting. They reported that the two approaches agree when binding sites are completely saturated, but that substantial errors arise in the calculation method when site saturation is incomplete and the transition of liganded molecules overlaps that of unliganded molecules — traced primarily to an inability to determine Tm under weak-binding conditions ^[10]. Their simulations put the reach of the calorimetric method at binding constants up to perhaps 10^40 per molar or higher, while most equilibrium methods, they note, fail well below 10^10 per molar ^[10]. The direction is worth holding onto: the thermal approach extends toward tighter binding than equilibrium titration reaches, and degrades at the weak end — the opposite of the intuition that a small shift is simply a small affinity accurately measured.

For the fluorescence version, the honest summary comes from a group proposing a quantitative treatment of their own. Thermal shift data have traditionally yielded qualitative results, they write, and quantitative methods for dissociation-constant determination have relied either on empirical and inaccurate usage of Tm or on isothermal approaches; their alternative plots Tm against the logarithm of ligand concentration and fits a derived van 't Hoff expression, tested on maltose-binding protein and on an allosterically regulated homotetrameric enzyme ^[11]. The existence of that literature is itself the argument: if reading a Kd off a single ΔTm were sound, nobody would need a new model.

Two operational consequences. Rank-ordering compounds by ΔTm assumes their unfolding enthalpies and binding modes are comparable, which is an assumption, not a result. And hits need confirmation elsewhere: a review of DSF theory and applications concludes that although DSF is a potent tool in buffer optimization and large chemical library screens, orthogonal techniques are recommended when it comes to ligand-binding validation and optimization, as DSF is prone to false positives and negatives ^[12].

Buffer screening: the application that holds up

Where thermal shift work is on firmest ground is comparing the same protein across solution conditions. Nothing has to be converted into a binding constant; the readout is used as an ordinal one.

The EMBL Hamburg screens are the standard shape of this. Their two-step approach screens first for global parameters — designed so that stability trends can be discerned according to pH, salt concentration, buffer type and buffer concentration — then for protein-specific additives that can affect folding, aggregation state and solubility, including small molecules that specifically bind and stabilize. The screens were run on more than 200 different protein constructs at their facility, and were designed to evaluate and give hints for optimizing purification and storage protocols ^[2]. The DSF review makes the same case in the drug-discovery context, illustrating uses of the technique in protein buffer optimization for stability, refolding and crystallization purposes ^[12].

But even here the highest Tm is not automatically the answer. In the IgG nanoDSF study, across sodium acetate, sodium citrate and sodium phosphate from pH 4 to 8, the melting temperature tended to increase as pH increased in all three buffers — while aggregation propensity differed by buffer species, with the best stability against aggregation obtained in acetate buffers below pH 4.6; citrate at the same pH gave higher aggregation and viscosity than acetate, which a dynamic light scattering study of protein–protein interaction suggested reflected attractive intermolecular interaction in citrate and repulsive in acetate. Their conclusion was that sodium acetate at pH 4.6 is suitable for that formulation ^[4]. Read the two trends together: the pH region with the higher Tm values in their data is not the region they recommended. Conformational stability and colloidal stability are different axes, and a screen that ranks on Tm alone is only reading one of them.

That caution extends to shelf-life prediction, where the evidence is weaker than the practice. A nanoDSF comparison of two monoclonal antibody originators and their biosimilars, resolving domain-specific unfolding onsets and midpoints, found no clear correlation between formation of higher molecular weight species and the unfolding parameters — onset temperature and aggregation temperature — in accelerated stability studies ^[3]. A thermal ranking is a hypothesis about storage behavior, not a measurement of it.

Conventions, and what is not settled

Several routine choices are convention rather than evidence, and are stated here without citation because no primary source establishes them:

• Heating at roughly 1 °C per minute, and holding that rate constant across every plate you intend to compare.
• Replicate wells per condition, so that a shift smaller than the well-to-well spread is not reported as a shift.
• A protein-free control well containing dye, buffer and test compound, to catch intrinsic compound fluorescence and direct compound–dye interaction.
• Matching solvent concentration across every well when compounds are delivered in DMSO, since the solvent can shift Tm on its own.
• Reporting Tm as the maximum of the first derivative of the melt curve, or from a fitted sigmoid — stating which, because the two do not always agree to a tenth of a degree.

What is genuinely unresolved is more interesting. There is no general mapping from ΔTm to binding affinity that holds without knowing the unfolding enthalpy, which is why the fitting literature keeps growing ^[11]. There is no settled account of why a given dye works for one protein and not another, only the empirical finding that broader chemical diversity raises the hit rate ^{[5, 6]}. And for small, flexible peptides — the material most of this site is about — the premise may not hold at all: a molecule with no cooperatively folded state has no melting transition to shift, so a flat trace is the expected result rather than a failure to be troubleshot. That is a statement about what the method requires, not a measured claim.

The defensible use is narrow and genuinely useful. Compare like with like, on one instrument, at one ramp rate, in one session. Treat a shift as evidence that something changed and as a reason to run a second, different measurement. Do not convert it into a number it was never measuring.