Standard Curve Guide: How to Build One, Read an Unknown and Know When to Doubt It

A standard curve turns an instrument reading into a concentration: you measure samples of known concentration, fit a model to the responses, and read unknowns from that model. A flawed curve produces a wrong result that still looks precise, and a high R² does not protect you from that.

This guide shows how to build a calibration curve, works through the arithmetic, and sets out how to judge the result: residuals, linear range, detection limits, matrix effects, four-parameter fits and the peptide-specific problem of whether your “known” concentration is really known. The example’s numbers are illustrative.

What a standard curve is, and what it assumes

IUPAC describes calibration as establishing the functional relationship between signal and analyte amount, including choosing a model and estimating its errors and parameters [3]. Reading an unknown uses that relationship in reverse, estimating amount from signal [3].

Every curve has testable assumptions. The standard must behave like the analyte, the sample’s surroundings must not change the response, the model must describe the data, and the reading must fall inside the calibrated range. A curve is one tool within the wider process of analytical method validation.

How to make a standard curve

1. Choose the standard. Ideally it is the analyte itself. ICH M10 says the reference standard should be identical to the analyte and of assured purity and identity [2].
2. Bracket your samples. IUPAC suggests covering 0–150% or 50–150% of the concentration you expect, with evenly spaced levels [4]. Include a blank.
3. Use enough levels. ICH Q2(R2) recommends at least five concentrations [1]. ICH M10 and IUPAC’s single-laboratory guidance ask for six or more [2,4].
4. Match the diluent. Standards should preferably be diluted in the same diluent as the samples [9].
5. Replicate and randomize. IUPAC advises at least duplicates, preferably triplicates or more, in random order [4].
6. Run it with the samples. A curve describes that run’s reagents, instrument and conditions, so measure standards and unknowns together unless a validated method says otherwise.
7. Correct for the blank. Subtracting the response of the zero standard is common practice [9].
8. Fit, then interrogate. Use the simplest model that describes the data [2], and look at residuals and back-calculated standards before accepting it.

Our peptide calculator handles molar conversions.

A worked example with the standard curve equation

Suppose a peptide standard is measured at seven levels from 5 to 100 µM. The blank reads 0.052 absorbance units, and every value below has had it subtracted. Least-squares fitting of the “clean” column gives:

y = 0.01202x + 0.0011, with R² = 0.9999 (y in absorbance units, x in µM).

An unknown is diluted 1 in 5 (20 µL into 100 µL final volume) and reads 0.472, or 0.420 after blank correction.

• Solve for x: (0.420 − 0.0011) / 0.01202 = 34.9 µM in the well.
• Apply the dilution: 34.9 µM × 5 = about 174 µM in the original sample.
• Check the range: 34.9 µM sits between the 20 and 40 µM standards, so the reading is supported. Undiluted, at about 174 µM, it would not have been.

Why R squared cannot prove a standard curve is good

R² measures how much of the variation in y tracks the fitted line. It does not test whether a line is the right shape. Anscombe’s well-known 1973 data sets share the same correlation and regression line yet look entirely different when plotted [6]. For calibration the warnings are explicit: IUPAC’s 1998 recommendations say the correlation coefficient has no meaning in calibration under their conditions, because x values are set by the analyst and are not random [3], and its 2002 validation guidelines call it misleading and inappropriate as a test of linearity [4]. ICH Q2(R2) still asks for the coefficient to be reported, but alongside a plot, the slope and intercept, and an assessment of any nonrandom pattern in the residuals [1].

The table shows why. The “curved” series has an R² of 0.990, yet its standards do not come back to their nominal values.

The curved series under-reads its lowest standard by 79% while R² stays at 0.990. In published cadmium atomic absorption curves, every curve had r above 0.997, yet lack-of-fit and Mandel’s tests rejected the straight line in every case, and the linear model gave systematically biased results [5].

Three checks that do more than R²

• Residual plot. Random scatter around zero is what you want. IUPAC notes that a systematic or curved pattern suggests nonlinearity and a funnel shape suggests unequal variance, which calls for weighted regression [3,4].
• Back-calculated standards. Run each standard through the equation as if it were an unknown. For regulated chromatographic bioanalysis, ICH M10 expects results within ±15% of nominal, or ±20% at the lower limit, for at least 75% of standards [2]. These are a benchmark, not a universal rule.
• Lack-of-fit test. With true replicates, this compares the means’ distance from the line with the replicate scatter [3,4].

Linear range, LOD, LOQ and why you never extrapolate

The range is the interval over which the curve has been shown to work. ICH Q2(R2) defines the detection limit as the lowest amount that can be detected but not necessarily quantitated, and the quantitation limit as the lowest amount that can be determined with suitable precision and accuracy [1]. It offers these estimates, where σ is the standard deviation of the response and S is the slope:

• Detection limit (LOD): 3.3σ / S, or a signal-to-noise ratio of about 3:1.
• Quantitation limit (LOQ): 10σ / S, or a signal-to-noise ratio of at least 10:1.

σ can come from replicate blanks or from the curve’s residual standard deviation [1]. If our illustrative blank has σ = 0.004, the LOD is 3.3 × 0.004 / 0.01202 = about 1.1 µM and the LOQ is about 3.3 µM, so a 5 µM lowest standard is reasonable. These are estimates, and IUPAC cautions that a simple detection limit can be in error by a factor of two [4].

Never read beyond your highest or lowest accepted standard. At the top, response flattens as the slope approaches zero, so tiny absorbance differences become large concentration errors [9]. In some immunoassays, a hook effect can suppress the signal above the upper limit [8], so a concentrated sample can look moderate. IUPAC’s interval formulas also show uncertainty growing with distance from the centre of the calibration range [3]. ICH M10 handles over-range samples by diluting and re-measuring them [2]; readings below the lowest standard should be reported as below the limit, not given a calculated value.

Matrix effects, parallelism and the wrong standard

A curve in clean buffer may not describe a sample in culture medium or another complex matrix. Two tests catch this:

• Spike recovery. Add known analyte to the sample matrix and measure how much comes back. One practical guide cites a usual range of 80–120% [8].
• Dilution linearity and parallelism. ICH M10 defines parallelism as a parallel relationship between the calibration curve and serially diluted samples [2]. Results that shift after dilution correction mean the sample is not behaving like the standard.

A standard unlike the analyte

Assays that respond differently to different molecules make standard choice decisive. The reference is ideally a purified form of the main species being measured [9]. A BSA standard curve is convenient [9], but it expresses a result in BSA equivalents. That matters even more for short peptides: Thermo Fisher notes that a peptide must reach about 3000 daltons to be detected by Coomassie dye reagents, and that single amino acids and dipeptides do not give the biuret reaction [10]. Our comparison of dye-binding and copper-based methods covers those chemistries. For a specific short peptide, a same-peptide standard on a method that separates and detects it, such as HPLC or LC-MS, is usually the more defensible design.

Linear fit or four-parameter logistic?

ICH Q2(R2) states that linearity of the concentration-response relationship is not required when a suitable model describes it, and names four- and five-parameter logistic functions for S-shaped curves [1].

In ligand-binding assays, Findlay and Dillard describe a sigmoidal response and call the 4PL the accepted reference model; a fifth parameter can help with asymmetric curves, and weighting generally improves performance [7]. A common form of the equation is y = D + (A − D) / [1 + (x/C)^B], where A and D are the lower and upper plateaus, C is the inflection point and B is the slope factor [13].

Net peptide content: how known is your known?

A curve can only be as accurate as the concentrations assigned to its standards. This is the peptide-specific trap. Lyophilized powder is not all peptide, as our guide to net peptide content and TFA or acetate salts explains. Purity does not settle it either, because HPLC purity describes the share of peptide material that is the target and does not state how much of the powder is peptide.

Published work shows the scale of the issue. One proteomics study warned that water and salts lead to overestimating peptide amounts, and reported that about 30% of peptide concentrations were inaccurate when based on weight [11]. A multi-laboratory oxytocin study, comparing HPLC, quantitative NMR and amino acid analysis, noted that peptides are very hygroscopic and that bulk standards need humidity control and water correction [12].

No fit statistic can detect this error. It takes an independent assignment of the stock’s concentration, such as amino acid analysis or a characterized reference material. A standard can also change with age, so follow good peptide storage practice and record when each stock was made.

What’s established and what isn’t

Sourcing peptide standards for a Canadian lab

For anyone buying research peptides in Canada to use as an in-house standard, ask these questions first:

• Does the documentation report content or only purity? A purity and identity result supports using a material, but it does not assign a concentration. Our guide to reading an analysis report shows what each field does and does not tell you.
• What is the salt form? Counter-ions change the mass-to-mole conversion.
• Which lot are you calibrating with? Record the lot of the standard with each curve, and repeat any concentration assignment when it changes.
• Do you need an assigned value? If absolute concentration matters, plan an independent content measurement.

GHRP-2Six residues, about 818 g/mol: far below the size dye-binding reagents detect.View product SNAP-8Eight residues, about 1075 g/mol: another case for a same-peptide standard.View product LL-3737 residues, about 4493 g/mol; cationic, so low standards can be lost to plastics.View product

Frequently asked questions