A stability study answers a question about containers you never opened, at a time you have not yet reached. Every gap the design leaves — a concentration you did not put on test, a pull point you skipped, a temperature you never ran — is a place where that inference travels on an assumption instead of a measurement. Careful chromatography does not repair a design that made the wrong assumption cheap to make. What follows is about which assumptions a reduced design buys, what they cost in statistical power, and how far a 40 °C result can be pushed for a peptide whose degradation chemistry is not guaranteed to be the same reaction at 5 °C.
The three thermal arms and what each one answers
The framework almost every protocol inherits puts material on test at three thermal conditions. Long-term testing runs for a minimum of 12 months at 25 ± 2 °C / 60 ± 5% RH or at 30 ± 2 °C / 65 ± 5% RH; intermediate and accelerated testing each cover a minimum of 6 months, at 30 ± 2 °C / 65 ± 5% RH and 40 ± 2 °C / 75 ± 5% RH respectively, and the intermediate arm is unnecessary where the 30 °C condition is already serving as the long-term one [1].
The three arms are not three measurements of the same thing. The long-term arm is the only one that observes the quantity you actually want, and it observes it slowly. The accelerated arm is a rate amplifier: it exists to surface degradation early enough to act on, and to supply the second temperature that any kinetic extrapolation needs. The intermediate arm is the bridge — the leg positioned to say something when the accelerated arm has moved and the long-term arm has not yet accumulated the time to adjudicate. Compressing the schedule further is an active area: accelerated predictive stability studies carried out over three to four weeks, combining 40–90 °C with 10–90% RH, have been developed to predict long-term behaviour without the wait [1].
Two other conditions get confused with the accelerated arm and answer different questions. Forced degradation is degradation at conditions more severe than accelerated; its purpose is to demonstrate the specificity of a stability-indicating method and to give insight into degradation pathways and degradation products [2]. That is method qualification and chemistry mapping — the separate forced degradation guide covers it — and its output does not enter shelf-life arithmetic. Light is a separate design axis rather than a hotter chamber. In two peptides derived from the α-melanocyte-stimulating hormone sequence and mutated to isolate each residue, UV-A irradiation in the presence of a pterin photosensitiser, at wavelengths where the peptides themselves do not absorb, converted the tryptophan residue to N-formylkynurenine and hydroxytryptophan and drove tyrosine dimerisation and incorporation of oxygen atoms, the first step being electron transfer from the residue to the photosensitiser triplet excited state, with hydrogen peroxide released [3]. A dark chamber at 40 °C reports none of that; the separate photostability guide covers the design.
Bracketing: testing only the extremes
Bracketing puts only the extreme levels of a design factor on test. The canonical form is a container factor — of several bottle sizes, only the smallest and the largest are tested, an example that source gives under the older and broader heading of matrix designs [4]. Nothing is spent on the intermediate levels. What you buy instead is the assumption that the intermediate levels are bounded by the two you ran.
That assumption has a shape requirement, which is that the response be monotonic across the bracketed range, or at minimum free of an interior extremum. It also has a scope requirement that is easier to overlook: the direction has to be established for your material in its actual physical form, not for its compound class. Biosynthetic human insulin was used as the model protein in a comparison of crystalline solid against amorphous solid prepared by freeze-drying, with samples equilibrated at selected relative humidities between zero and 75% to give a range of water contents. Increasing water content increased the degradation rate of the crystalline solid, while rate constants for degradation in the amorphous solid were essentially independent of water content up to the maximum studied, approximately 15% [5]. One protein, two solid forms, the same design factor, and two entirely different response shapes. In the same work the amorphous form was far more stable than the crystalline one under all conditions investigated, which is the reverse of the generalisation established from small molecules [5].
Direction can be counterintuitive in solution too. For the hexapeptide Val-Tyr-Pro-Asn-Gly-Ala at pH 9 in 400 mM Tris, increasing the peptide concentration slows the rate of deamidation, a self-stabilisation the authors noted had not been reported previously [6]. A design that bracketed concentration on the assumption that the most concentrated sample is the worst case would have had the ranking backwards for that reaction.
So the circumstance in which bracketing is not defensible is specific. Where a factor is not expected to behave monotonically, or where the mechanism changes across its range, the extremes do not bound the middle, and no amount of replication at the extremes recovers what the middle was doing. The defensible version of the design states the expected direction, names the evidence for it, and puts a confirmatory intermediate level on test wherever that evidence is thin.
Matrixing: a fraction of the samples at each pull
Matrixing keeps every design combination in the study but tests only a fraction of the combinations of interest at any specified sampling time [7]. Bracketing removes levels; matrixing removes occasions. The design question is therefore two-dimensional — which fraction, and at which times — and it has been posed as an optimisation, choosing the time vectors so that the design is optimal in terms of maximum information per unit cost [7].
How much can come out has been tested against full designs. Across three parenteral products whose degradation kinetics differed, with degradation products taken as the critical parameter, 28 matrixing designs were each compared with the full design on root-mean-square error and on the resulting shelf life; all of the evaluated designs proved applicable whether the degradation product increased linearly or non-linearly, and a reduction of two time points per batch came out as the most appropriate [8]. Read that as what it is: a demonstration on three products whose degradation behaviour had already been characterised from full data, not a standing licence to delete two pulls from a study whose kinetics you have not yet seen.
The cost is statistical power, and it lands in two places. Bracketing and matrixing designs have been examined for their power to detect a significant difference between slopes and, through the mean square error, for the precision of the estimated shelf life, with the distributions of the two designs compared across 1000 simulations [9]; fractional-factorial-type designs for stability studies have been proposed and the statistical powers of various designs compared [4]. Losing power to detect a slope difference means losing the ability to show that a bracketed or matrixed factor did not matter, which is the very proposition the reduced design rests on. An underpowered test that returns no significant difference is not evidence that the factor is inert.
The precision loss is not abstract, because of how the reported number is built. Shelf life is calculated as the time point at which the lower 95% confidence limit about the fitted regression line crosses the lowest acceptable limit for content, frequently 90% of the labelled amount [10]. A wider confidence band crosses that limit earlier at an identical degradation rate, so a design that saves assays pays for the saving in the shelf life it can claim. The same paper sets out the second place power matters: preliminary testing for batch differences in both slope and intercept should precede pooling stability data across batches, and the significance level recommended for that preliminary test is 0.25 — deliberately large, so that the power of the batch-difference test stays high — alongside an alternative that fixes the power and computes from the data the significance level needed to achieve it [10]. A matrixed design thins the data feeding that preliminary test at the same moment it depends on pooling to recover precision.
Where the pull points go
Time-point spacing decides which curve shapes you can tell apart, and therefore what you are entitled to extrapolate. In the insulin solids work, degradation was found to be linear in the square root of time, and the rate constants were defined from the slopes of percentage degradation against square root of time [5]. A schedule with widely spaced early pulls cannot separate that behaviour from a linear-in-time law over the same window, and the two extrapolate to different places. Early points carry the curvature; late points anchor the extrapolation. Matrixing that removes early pulls and matrixing that removes late pulls are not the same decision, and describing both as a fractional reduction hides the difference.
This is also why the matrixing result above holds only in the direction it was demonstrated: those designs were checked against a full design, and that check is unavailable in a study whose full design was never run. The defensible sequence is full sampling until the shape of the degradation curve is established, then reduction argued from that shape.
What an accelerated result licenses
The step from a 40 °C dataset to a statement about 5 °C storage is an Arrhenius extrapolation, and the first thing to notice is that it does not produce a single number. For accelerated degradation studies of peptide and protein drugs, fitting all the data simultaneously rather than determining individual rate constants condition by condition was clearly preferable, combined with use of the reparameterised Arrhenius equation; across the procedures compared, the estimated shelf life at 5 °C varied between 2.2 and 4.0 years as a function of the temperature range and the procedure used, with the classic Arrhenius analysis giving the widest 95% confidence intervals, all of it evaluated by a bootstrap approach [11]. Close to a factor of two, on those two choices alone.
None of which makes the extrapolation hopeless. Peptide chemical degradation can be well behaved over a wide window: deamidation of the Asn residue in Val-Tyr-Pro-Asn-Gly-Ala at pH 9 followed Arrhenius behaviour over the range 5–65 °C, with an apparent activation energy of 13.3 kcal/mol [6]. In the solid state the picture is similar with a caveat attached. A review of the lyophile literature, covering amorphous small molecular weight drugs and proteins, concluded that the temperature dependence of degradation rate constants in lyophiles can be represented by the Arrhenius equation in most cases, with the majority of reported activation energies for the various degradation pathways falling in the range 8 to 25 kcal/mol, but noted that in some instances there is a break in the Arrhenius plot around the glass transition temperature or a related characteristic temperature, and that the equation is a reasonable empirical tool provided specific conditions are met [12]. For a lyophilised peptide that translates into a concrete design check: an accelerated arm sitting above the formulation's Tg may be on the far side of that break from the storage condition, in which case the two points were never on one line.
Which attribute you are extrapolating matters as much as which temperatures you ran. Temperature-induced protein aggregation is often non-Arrhenius over even relatively small temperature windows relevant for product development, a consequence of the complexity of the aggregation process and of temperature-dependent conformational stability, and this makes low-temperature extrapolation difficult on the basis of accelerated studies at high temperatures alone [13]. That review concerns protein aggregation rather than peptide covalent chemistry, but the structural point carries: an attribute whose rate depends on a temperature-dependent conformational equilibrium is a different extrapolation problem from a single covalent reaction with one activation energy. A study whose critical attribute is aggregate content needs its extrapolation defended separately from one tracking a chemical degradant.
Pathways can also swap outright when the condition changes. For the same hexapeptide, in the pH range 5 to 12 deamidation proceeded exclusively via a cyclic imide intermediate, forming both the Asp- and the isoAsp-hexapeptide, whereas at acidic pH the pathway involved direct hydrolysis of the Asn side-chain amide and produced only the Asp product, with buffer catalysis observed between pH 7 and 11 [14]. Different products, a different rate law, and no guarantee that the peak you are integrating is the same peak across conditions.
The consequence that most often catches a formulation screen is rank reversal. In a monoclonal antibody digested with papain and resolved into Fc and Fab fragments, with buffers prepared at room temperature and then placed at the stability temperatures and solution pH not adjusted to be the same at different temperatures, deamidation at 40 °C was faster in acidic buffers than in basic buffers, and that trend was reversed at 5 °C, attributed to the change in hydroxide ion concentration influenced by buffer and temperature; the apparent activation energy was higher for the rates generated in acidic buffer than in basic buffer [15]. Two conditions with different activation energies give two Arrhenius lines with different slopes, and lines with different slopes cross. The analyte there was an antibody rather than a peptide, but the arithmetic is indifferent to analyte class: wherever a design factor changes the activation energy rather than only the pre-exponential term, the ordering measured at the accelerated condition is not evidence about the ordering at the storage condition.
Because the extrapolation rests on a slope, the number of temperatures in the accelerated design is itself a design decision with a floor. In a study applying full and reduced designs to accelerated stability modelling of a parenteral product, with degradation products monitored across stress and long-term conditions, 13 models were assessed by coefficient of determination and predictive relevance and their predicted degradant levels compared against the actual long-term stability results; the full model and 11 of the reduced models were suitable, the two-temperature model was the exception, and the three-temperature model was identified as the most appropriate for that product [16]. Two points define a line whether or not a line is the right model. A third is the first opportunity to discover that it is not.
Writing the protocol so the conclusion survives it
Four questions decide whether the study you are about to start can carry the claim you intend to make from it. Which factors are bracketed, and what evidence — on this material, in this physical form — establishes that the response is monotonic across each bracketed range? Which pulls are matrixed away, and has the shape of the degradation curve been established from full sampling, or is it being assumed? Does any accelerated arm cross a physical or mechanistic boundary, a glass transition or a change in the dominant pathway, that puts it on a different line from the storage condition? And is the critical attribute a covalent reaction with one activation energy, or a physical process whose rate runs through a conformational equilibrium?
None of these are answered by the assay. They are answered before the first sample goes into the chamber, and each one of them narrows or widens what the finished dataset lets you say.