Quantifying Secondary Structure by Circular Dichroism: Reference Sets, Concentration Normalization and the Limits of Similarity Criteria

1 Introduction

Among higher-order structure methods, CD holds a special place: it was the first to be adopted, it is the easiest to run, and it is the easiest to over-interpret. With a CD instrument, a quartz cuvette and a nitrogen purge you can have a spectrum in half an hour; one click on "fit" in the software gives four percentages. Nothing in the process raises an error, so these four percentages are often written into reports as measurement results.

Yet at least four links in this chain introduce substantial uncertainty, and none of them raises an alarm:

• Reference set. Fitting assumes that "the unknown spectrum = a linear combination of several spectra of known structure". This holds roughly for globular proteins but less well for β-sheet-rich antibodies—reference sets lack spectra closely homologous to antibody variable and Fc domains.
• Algorithm. Different algorithms give different results for the same spectrum, commonly 5–10 percentage points apart. This is not a programming error; they constrain the "solution" differently.
• Concentration. Mean residue ellipticity uses concentration as its denominator: whatever the concentration error, the reported [θ]_MRW carries the same error in the same direction. Note, however, that the fitted percentages are determined mainly by spectral shape and are insensitive to overall scaling—so a concentration error will not make the four numbers look "absurd", but it shifts absolute values, absolute-intensity comparisons and cross-sample comparability as a whole, making it one of the hardest biases to detect.
• Instrument range. Light throughput in the far UV (especially 190–200 nm) is very low; with insufficient nitrogen purge or an ageing lamp, what is measured may be noise rather than signal. The symptom of this problem is that "the spectrum still looks smooth".

This article takes these four links apart and gives the basis for judgement and the reporting requirements for each.

2 Principles

2.1 Physical origin of circular dichroism

A chiral molecule absorbs left- and right-circularly polarized light differently; the difference ΔA = A_L − A_R is called circular dichroism. The instrument measures the intensity difference between the two beams and converts it to ellipticity θ (in millidegrees, mdeg): θ = 32980 × ΔA.

A CD signal appears only from chromophores in a chiral environment. In proteins these are mainly the peptide bond (transitions in the far UV) and aromatic side chains (transitions in the near UV). This physical origin gives CD two natural regions:

• Far UV (190–250 nm): the signal comes from the n→π* and π→π* transitions of the peptide bond and reflects the regular arrangement of the backbone, i.e. secondary structure.
• Near UV (250–320 nm): the signal comes from phenylalanine, tyrosine and tryptophan side chains and disulfide bonds and reflects the microenvironment of these residues, i.e. tertiary structure.

The two regions carry information of an entirely different nature and cannot substitute for each other.

2.2 Far UV: linear superposition of secondary structures

The three basic secondary structures each have a characteristic far-UV shape: α-helix shows two minima of similar intensity at 208 and 222 nm; β-sheet shows a minimum at 216–218 nm and turns positive near 195 nm (its shape varies considerably with parallel/antiparallel type and strand length, making it the least "standard" of the three); random coil has a strong minimum near 200 nm and turns positive at shorter wavelengths.

A protein's measured spectrum is a weighted superposition of these components, the weights being the residue fractions of each. This is the physical basis of the fitting inversion: given a set of reference spectra of known structure, find a set of non-negative weights whose weighted sum approximates the measured spectrum.

Linear superposition has a direct consequence, which is also CD's greatest limitation: it is sensitive only to changes in "amount", not to changes in "position". If a residue moves from the third to the fourth position of an α-helix, or a glycosylation site changes, the CD spectrum stays essentially the same as long as the overall secondary-structure proportions are unchanged. CD cannot see "conformational detail", only "conformational composition".

For α-helix there is an additional criterion: the ratio of ellipticity at 222 nm to that at 208 nm. A ratio below 1 (typically 0.8–0.9) suggests single-stranded helices; a ratio clearly above 1 (often above 1.1) suggests coiled-coil winding between helices. This ratio is useful when judging the structure of fusion proteins and peptides. Collagen-like proteins have their own criterion: an intact triple helix shows "a positive peak at 220–225 nm + a negative peak at 196–198 nm" in the far UV, and the positive-to-negative intensity ratio R_pn (the ratio of the absolute values of θ₂₂₁ and θ₁₉₈) characterizes triple-helix integrity (about 0.10–0.15 for intact collagen, falling to 0 or turning negative on denaturation)¹¹. This is one of the few cases where CD can give a binary "specific structure present/absent" judgement, and it is more robust than percentage fitting.

2.3 Near UV: a window on tertiary structure through aromatic residues and disulfides

Near-UV CD signals are one to two orders of magnitude weaker than far-UV signals, but the information is of a different kind: it reflects the polarity and flexibility of the environment around aromatic residues and their interactions with neighbouring groups.

The key feature is spectral shape, not intensity. When part of a protein unfolds and aromatic residues are exposed from the hydrophobic core to water, near-UV CD peak positions and relative intensities change, while the far UV may hardly change. In the HOS evidence chain, near-UV CD and intrinsic fluorescence therefore cover the same level (tertiary structure) and can corroborate each other.

The practical difficulty of near-UV CD is signal-to-noise: to obtain a usable signal, concentration must be raised to 0.5–1 mg/mL with a 10 mm path length, which in turn brings risks of aggregation and inner-filter effects. Near-UV data are therefore usually less reproducible than far-UV data and should carry correspondingly less weight in comparability studies.

2.4 Thermal denaturation curves and melting temperature

At a fixed wavelength (usually 222 nm, where α-helix gives its strongest signal) the temperature is raised continuously and ellipticity is recorded against temperature, giving a sigmoidal denaturation curve. The inflection point, or the extremum of the first derivative, is the melting temperature T_m.

Three points need attention. First, a single wavelength at 222 nm reflects only the loss of α-helix. For β-sheet-dominated proteins (such as IgG), the signal change at 222 nm is small and the precision of T_m drops accordingly; 208 nm or multi-wavelength principal component analysis of the full spectrum can be used instead. Second, T_m is a thermodynamic quantity, not the whole of stability—a high T_m means unfolding needs more energy, but it does not mean the protein will not aggregate at storage temperature; that is governed by kinetics and should be complemented by the aggregation onset temperature T_on from DLS. Third, heating rate affects the apparent T_m: at fast rates the system cannot reach equilibrium and the apparent T_m is higher; the method must fix and state the rate (usually 1 ℃/min), and results obtained at different rates are not comparable between laboratories.

3 Experimental method

3.1 Concentration, path length and nitrogen

Note: concentration and path length must be chosen together, not independently. The concentrations in the table correspond to a 1 mm path; with a 0.1 mm path, keeping the same absorbance requires roughly a ten-fold higher concentration (usually 1–2 mg/mL, where aggregation must be watched). This correspondence should be written into the method and verified experimentally during instrument qualification.

Nitrogen purging is not optional. Oxygen absorbs noticeably below 200 nm (especially at 185–190 nm); with insufficient purging, oxygen absorption mixes into the measured "protein signal", appearing as an abnormal negative offset and increased noise at 190–200 nm. The criterion is not "whether the spectrum is smooth"—this interference often shows no spikes or jumps and looks perfectly normal—but a comparison of the same sample before and after thorough purging: if the absolute ellipticity at 190–200 nm decreases and the noise falls after purging, oxygen interference was present.

Path length is coupled with concentration: 0.2 mg/mL in a 1 mm path and 0.1 mg/mL in a 2 mm path give the same protein absorbance and the same signal, but the latter doubles the absorbance of the buffer and dissolved oxygen as well, so the total absorbance is higher and stray-light error larger. Where the signal is sufficient, prefer a short path and higher concentration; the cost is a higher aggregation risk at high concentration, which should be checked by turbidity or DLS.

3.2 HT voltage: the usability criterion for far-UV data

This is the most important practical criterion in this section.

During a scan, the CD instrument records the high-tension (HT) voltage of the photomultiplier. With ample light throughput, HT stays low (commonly 300–600 V; some instruments display kV, which must be converted before comparison); when throughput is insufficient (weak lamp output at short wavelengths, strong sample absorption, cuvette absorption), the instrument raises HT automatically to maintain the signal, and noise is amplified with it¹².

This gives a method-design rule: scan HT first, then set the low-wavelength cut-off. If HT already exceeds the threshold at 200 nm, the fitting range should be 200–250 nm rather than a mechanically applied 190–250 nm.

3.3 Parameter settings and system suitability

Scan parameters (response time, bandwidth, step size, number of accumulations) follow the Chinese industry standard JY/T 0572—2020, General rules for circular dichroism spectroscopy². Bandwidth and step size must match: too wide a bandwidth smooths out narrow peaks, too large a step distorts peak positions.

System suitability is checked with a reference substance: ammonium camphorsulfonate (ACS) has a defined specific ellipticity at 290.5 nm (literature value [θ] ≈ +7 260 deg·cm²·dmol⁻¹) and is used to check the instrument's ellipticity scale; its far-UV features can also be used to check wavelength accuracy. Run it with every measurement and confirm that the deviation is within the permitted range. The specific reference substance, concentration and limits must be defined and validated in the method itself—the industry standard gives only general requirements and does not name a specific substance.

4 Data analysis and normalization

4.1 Calculating mean residue ellipticity and its dependence on concentration

The raw data are ellipticities θ (mdeg), which scale linearly with concentration and path length and so cannot be compared directly between samples. Normalization converts them to mean residue ellipticity:

[θ]_MRW = θ × MRW / (10 × c × l)

where MRW is the mean residue weight (molecular weight divided by number of residues), c is concentration (mg/mL) and l is path length (cm); [θ]_MRW calculated this way is in deg·cm²·dmol⁻¹. When θ is entered in mdeg, the number is in "mdeg·cm²·dmol⁻¹"; the unit must be converted or stated in reports—this is one of the most common sources of mismatch in inter-laboratory comparisons.

In this equation MRW is fixed by the sequence and l by the nominal cuvette value; only c has to be measured. And c usually comes from the A280 of the same sample divided by the extinction coefficient. This forms a chain:

CD spectrum → [θ]_MRW → c → A280 → extinction coefficient ε → measured amino-acid composition

What this chain means: the uncertainty of ε propagates 1:1 into the absolute value of [θ]_MRW—a 5% bias in ε gives a 5% bias in every normalized ellipticity, while the spectral shape is unchanged and nothing shows on the CD plot. It should be stressed that this bias is not a proportional bias in the fitted percentages: percentages are determined mainly by spectral shape and the relative amplitudes of the component reference spectra, and are insensitive to overall scaling of the spectrum (all three algorithms solve under the constraint "sum of components = 1", so the fit depends more on shape than on absolute amplitude); the effect of a concentration error on the component fractions is therefore partial and non-proportional, usually well below 1:1. Its harm lies precisely in "raising no alarm": distorted absolute values, distorted absolute-intensity criteria (θ₂₂₂/θ₂₀₈, R_pn, single-point ellipticity thresholds) and an overall shift in lot-to-lot and inter-laboratory comparability leave no trace in the spectrum or in fit-quality indicators.

In practice this gives three requirements: CD and UV must use the same dilution and the same concentration measurement; CD results must be reported together with the concentration and its source; for comparability studies that require high precision, first determine the extinction coefficient from a measured amino-acid composition and normalize with that concentration.

4.2 Choice of fitting algorithm and its uncertainty

Several algorithms are used for secondary-structure fitting, and the three mainstream ones differ in principle: SELCON3 is built on least-squares fitting to basis functions with a self-consistency check of the reference set; CONTINLL uses regularized ridge regression, is more robust to noise and tends to give smoother solutions; CDSSTR first reduces the reference set by singular value decomposition and then fits in the low-dimensional space, depending more heavily on reference-set coverage.

The difference among the three is not one of precision but that they constrain the "solution" differently: whether the sum of components may deviate from 1, whether unexplained residues are allowed, and how strong the regularization is. A spread of 5–10 percentage points in α-helix fraction for the same spectrum processed by all three is therefore within the normal range¹⁰. Reference-set coverage and version also affect results; Whitmore and Wallace systematically compared how different reference sets perform on similar proteins⁹, and Greenfield's protocol gives a complete operational framework from spectrum to secondary structure⁸.

This gives two methodological requirements. First, the method must fix the algorithm and reference set; you cannot "recalculate with another program to see"—changing the algorithm means changing the method. Second, reports must give fit-quality indicators: the fit residual (NRMSD), the deviation of the sum of components from 1, and the name of the reference set used. Without these indicators, four percentages alone do not let the reader judge how trustworthy the result is.

4.3 Quantifying similarity comparisons

A common practice in comparability studies is to overlay two CD spectra and see whether they "coincide". This needs to be quantified, or the conclusion rests on subjective judgement. Three kinds of quantitative indicator are available:

Spectral correlation coefficient—the Pearson correlation of the two spectra computed wavelength by wavelength; the closer to 1, the more similar. Its weakness is insensitivity to overall intensity: two spectra of identical shape but 10% different in amplitude still correlate close to 1.

Root mean square of the difference spectrum (RMSD)—point-by-point subtraction followed by the root mean square; it reflects amplitude differences but is sensitive to single outliers.

Difference in fitted components—fit each spectrum and compare the component fractions, giving the differences and their uncertainty. This is the indicator closest to "structural difference", but it is limited by the algorithm uncertainty.

In practice the three should be used together: the correlation coefficient confirms shape, RMSD confirms amplitude and component differences confirm structure. Any one alone is not enough to support a conclusion.

5 Results and discussion

5.1 How large is the disagreement between algorithms?

Taking a typical IgG far-UV spectrum fitted with the three algorithms, α-helix fractions commonly differ by 5–10 percentage points and β-sheet fractions differ less (IgG is β-sheet-dominated and well covered by the reference sets), whereas the allocation between turn and random coil often differs as much or more—these two structures are hard to distinguish clearly in the reference spectra, and algorithms have considerable freedom in allocating them. Their sum (turn + coil) is usually more stable than either alone, which is why some authors advocate reporting only three classes: "α-helix / β-sheet / other".

The methodological meaning of this fact: the figure "18% α-helix" has only one significant figure. Reporting "18.3%" mistakes algorithm uncertainty for measurement precision. Good practice is to report a range or "about 18%", and to state the algorithm and reference set in the method.

5.2 Why IgG CD spectra "all look the same"

This is the most commonly misused point of CD in antibody characterization. IgG secondary structure is dominated by β-sheet (CD fits commonly report 40%–45%, depending on algorithm and reference set; see 4.2), and this composition is quite stable across IgGs. Glycosylation changes the local environment around the two C_H2 domains of the Fc region, involving a few residues; their contribution to the far-UV spectrum is swamped by the overall β-sheet signal.

The result: two spectra overlay very well, but that does not prove the conformations are identical. Differences may exist in the near UV (aromatic microenvironment), in thermal stability (T_m differing by 1–2 ℃) or in local flexibility (hydrogen–deuterium exchange rates), none of which far-UV CD is sensitive to.

5.3 Orthogonal corroboration with FTIR

CD and FTIR both give secondary-structure fractions, but on different physical bases: CD on chiral differences in electronic transitions, FTIR on the vibrational dipoles of the amide bond. Their sensitive bands also differ—CD is very sensitive to the α-helix feature at 222 nm, FTIR more sensitive to β-sheet at 1620–1640 cm⁻¹.

Agreement between the two therefore has special value: when CD and FTIR give the same β-sheet fraction, the conclusion is stronger than from either method alone. Conversely, a clear disagreement suggests interference in one of them (CD may be affected by concentration normalization, FTIR by water vapour); look for the cause rather than averaging. Component fractions from a single method are in essence estimates, which is why many laboratories adopt "secondary-structure conclusions must agree between two methods before entering a report" as an internal methodological rule—it is a quality standard set by the company, not a regulatory requirement, and should be identified as such when written into documents.

5.4 Where CD data land in a filing, and how to word them

CD data appear in three places in a filing, each with different wording requirements. Structural characterization falls under ICH Q6B; comparability studies under ICH Q5E (comparability of biotechnological/biological products subject to manufacturing changes) and its implementation Q&A; and specifications under the identity test in each product monograph:

The most common wording error is applying the identity-test wording to comparability—"spectra conform" written as "structures are identical". The evidential strength of these two statements differs by an order of magnitude. Another common error is writing "no difference detected" as "no difference": the former reflects only the discriminating power at the range and signal-to-noise used, and the method's discriminating power (inter-spectrum RMSD of replicate measurements, lower bound of the correlation coefficient) should be given with it; otherwise "not detected" has no criterion behind it.

5.5 Method validation: design by "performance characteristics", not by "the full set of items"

The ICH Q2(R2) framework designs validation by performance characteristics rather than a fixed checklist: the intended use of the method determines what must be validated, rather than running every item⁴. CD suits this approach well—it is both a qualitative method (identity, spectral comparison) and one with semi-quantitative fitted output, and the two need different evidence.

A workable split: for identity use, assess specificity (forced-degradation or heat-denatured samples should give distinguishable spectra; negative controls and buffer should give no signal) and robustness (purge time, path length, concentration ±10%, heating rate, different cuvette lots); for semi-quantitative use, add repeatability (six consecutive measurements of one sample, reporting the RSD of θ and of each component fraction) and intermediate precision (different days, different analysts). Linearity needs care: ellipticity is indeed proportional to concentration (Beer–Lambert), but the method's output is the fitted fractions, which do not change with concentration—listing linearity as a validation item neither matches the method's actual use and invites questions about its rationale during inspection. As for accuracy, secondary-structure fractions have no accepted true value; they can be supported only indirectly with reference proteins or by comparison with FTIR or crystal structures, and recoveries should not be forced.

System suitability runs every time: ACS to check the ellipticity scale, the HT curve to confirm the low-wavelength cut-off, and baseline noise (RMS) within the value set by the method. All three are cheap and quick, and all can "raise an alarm" independently of the result—far more useful than inspecting the spectral shape afterwards.

6 Pharmacopoeias and standards

None of the three pharmacopoeias has a dedicated chapter for electronic CD^1,6,7. Two points to note: first, USP 2025 has chapter 〈782〉 on vibrational circular dichroism (VCD), but VCD measures vibrational transitions (useful for distinguishing stereoisomers and supporting conformational assignment) and is not the same technique as far-UV CD based on electronic transitions; the two cannot be cited for each other. Second, EP 2.2.42 is density of solids and has nothing to do with CD; when searching, it is easy to mistake that chapter number for a CD chapter.

The practical consequence: there is no pharmacopoeial chapter to cite as the basis of a CD method; only an industry standard or an in-house method can be cited, and a complete set of criteria must be built in the method validation. This differs fundamentally from citing "ChP 0401" as the basis for CD in a filing—0401 is UV-visible absorption and optically does not involve circularly polarized light; at most it can serve as a reference for shared elements such as instrument calibration, absorbance linearity and cuvette matching, but it cannot serve as a CD-specific basis. Citing 0401 as the basis of a CD method is a formal defect that reviewers readily point out.

7 Conclusions

The value of CD in the HOS evidence chain is clear, but its output is a fitted value, not a measured value, and this must be reflected in method design and in the wording of reports.

First, the reference set and algorithm determine the result. For the same spectrum, α-helix can differ by 5–10 percentage points between algorithms, so the algorithm and reference set must be fixed and reported with the result; "18% α-helix" has only one significant figure.

Second, concentration is the most hidden source of error. Mean residue ellipticity uses concentration as its denominator, and concentration comes via A280 and the extinction coefficient from the amino-acid composition—so the absolute value of [θ]_MRW inherits the uncertainty of UV and AAA, transferred approximately 1:1. Note that fitted percentages are insensitive to overall scaling of the spectrum, so this bias does not raise an alarm in the percentages or the spectral shape; what it contaminates are absolute-intensity criteria (θ₂₂₂/θ₂₀₈, R_pn, single-point ellipticity thresholds) and lot-to-lot and inter-laboratory comparability.

Third, HT voltage is the objective criterion for usable far-UV data, not spectral smoothness. Wavelengths where HT exceeds the threshold must not be used for quantitative fitting, and the cut-off wavelength should be set from the HT curve.

Fourth, similar spectra do not mean identical conformation. IgG CD spectra are highly similar, and local changes such as glycosylation fall within measurement error; overlaid CD spectra can support only "no significant difference detected", and HOS conclusions should combine methods such as FTIR, near-UV CD, intrinsic fluorescence and DSC.

References