Deep Inquiry🧊 Stability, formulation & solution chemistry

Beyond the lag phase: defending nonlinear aggregation models without inventing shelf life

β€œHow should we statistically model and defend non-linear, autocatalytic aggregation profiles (e.g., amyloidogenic peptides) to an auditor when ICH Q1E assumes linearity?”

The short answer

The premise needs correcting: ICH Q1E does not require every stability relationship to be linear. Section 2.6 explicitly permits linear or nonlinear relationships and says that nonlinear regression can sometimes better reflect the true relationship.[1] The difficult part is demonstrating that a model predicts the relevant quality attribute beyond the observed interval, not obtaining permission to draw a curve. I would separate mechanism discovery from shelf-life estimation, use a fit-for-purpose measurement and predeclared model/uncertainty analysis, and retain real-time evidence as the anchor.[1,2] A precise sigmoid fitted mostly to a lag phase can provide an unjustifiably confident expiry even when its equation sounds mechanistic.

The actual residual gap after later-course review

PDQC831 already contains sections on lag phases, nonlinear behavior, censored observations, residual diagnostics, and conservative restriction to observed real-time data when a linear model is untrustworthy. Its statistical lesson also distinguishes confidence limits on a mean from a point-estimate crossing. The provisional screen's β€œopen” label therefore overstates the absence of instruction. What is missing is an explicit bridge from those warnings to model identification, uncertainty in a threshold-crossing time, and a defensible audit record.

PDQC824 separately explains movement between soluble assemblies and particles, including why particle counts and aggregate mass can move differently. I inspected the later PDQC892/893 sections returned by broader aggregation/nonlinearity searches: they address microbial-test interference, facility requirements, or queueing nonlinearity, not a missing aggregation shelf-life framework. This is a bounded coverage finding, not a claim that a keyword search proves universal absence.

The earlier inquiry When Arrhenius Breaks already develops temperature extrapolation, phase transitions, and real-time confirmation. I build beyond it here. I also qualify its categorical statement that Arrhenius behavior requires exactly one elementary mechanism: several pathways can yield an approximately straight empirical relationship over a restricted range. That observation does not justify extrapolation through an untested transition; an apparently straight plot is weaker evidence than a validated prediction.[5]

What Q1E actually permits, and what is still draft

Q1E section 2.6 states: β€œThe relationship can be represented by a linear or non-linear function on an arithmetic or logarithmic scale.” It immediately adds: β€œIn some cases, a non-linear regression can better reflect the true relationship.”[1] Thus an auditor-facing defense should cite the actual text rather than apologize for departing from an imaginary linear-only rule. Appendix B's regression examples do not exhaust the methods permitted by the guidance.

The same section describes the earliest intersection of the appropriate 95% confidence limit for the mean with the proposed acceptance criterion. For an increasing attribute, such as a quantitatively meaningful aggregate measure, the upper one-sided limit is relevant; a direction-unknown attribute requires the corresponding two-sided treatment.[1] This is not a promise that 95% of individual containers will remain acceptable. The scientific and regulatory context must determine whether the mean is a sufficient estimand for the actual decision.

I checked the live agency records during this inquiry. EMA's current-version entry identifies the consolidated Q1 document as Draft, Step 2b, consultation closed, with consultation dates in 2025. FDA identifies its June 2025 Q1 guidance as Draftβ€”Not for implementation.[3,4] Although EMA's descriptive overview uses β€œsupersedes,” its current-version label is still draft; that overview sentence does not establish completed regional implementation. I therefore do not replace adopted Q1E with proposed consolidated provisions or invent new refrigerated extrapolation allowances.

Q5C is also a scope-sensitive source. Its primary text calls for long-term, real-time, real-condition data and covers specified biological/biotechnological proteins and polypeptides.[2] That is scientifically relevant to aggregation-sensitive materials but does not automatically make every synthetic RUO peptide a Q5C-regulated product. PP needs a truthful research-material stability claim, not a borrowed assertion of approved expiry, compounding beyond-use date, or regulatory acceptance of a particular model.

Start with the observable, not the fitting function

β€œAggregation percent” can mean SEC high-molecular-weight area, loss of soluble monomer, fluorescence, or a calculated aggregate mass fraction. These are not interchangeable observations. Particle number can increase by fragmentation without an equal increase in total aggregate mass; SEC can lose particles or dissociate reversible assemblies. Before modeling, I would define the attribute, sampling boundary, response calibration, reporting units, recovery, and acceptance criterion.

Meisl and colleagues' kinetic-analysis framework starts from aggregate-mass accumulation and emphasizes data quality, concentration dependence, global fitting, and model verification.[6] That matters because a fluorescence time series is not automatically an absolute mass assay. A change in dye binding, optical interference, or assembly morphology may change signal without proportional material conversion. An empirical fluorescence endpoint could still be useful for a defined research application, but should not be relabeled β€œall aggregates” without supporting calibration and orthogonal evidence.

A second issue is handling-induced aggregation. The state in a stored sealed vial need not be the state after opening, dilution, transfer, or analytical separation. Analytical preparation can create, destroy, or selectively exclude an assembly. Repeated readings from one prepared aliquot characterize that aliquot and analytical repeatability; they do not replicate independent storage containers. The statistical unit must match the inference.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ STORED MATERIAL + PACKAGEβ”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
              β–Ό
       SAMPLING / RECOVERY
              β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ DEFINED QUALITY ATTRIBUTEβ”‚ ← ORTHOGONAL CHECKS
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
              β–Ό
      CANDIDATE MODEL FAMILY
        β”‚             β”‚
        β–Ό             β–Ό
   EMPIRICAL FIT   MECHANISM TESTS
        β””β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”˜
               β–Ό
     IDENTIFIABILITY + DIAGNOSTICS
               β–Ό
       UNCERTAINTY IN CROSSING
               β–Ό
 REAL-TIME SUPPORT + EXTRAPOLATION DOMAIN
               β–Ό
    LIMITED CLAIM / MORE DATA / NO CLAIM

The feedback loop is essential: poor measurement recovery cannot be repaired by a more elaborate likelihood. Likewise, disagreement between physical and chemical attributes is evidence to investigate, not a reason to select the one giving the longest shelf life.

Three model classes answer different questions

Local empirical models

A straight line, transformed line, or low-order curve can summarize a restricted observation range without identifying a molecular mechanism. A first-order loss process is not linear in untransformed amount, but may be linear in its logarithm under appropriate assumptions. A logistic or Gompertz curve can describe a sigmoid. The curve family should be selected because it plausibly represents the observed attribute and intended prediction domain, not because a software menu makes it available.

A logistic midpoint is not necessarily a physical nucleation time. It is a parameter describing the location of a fitted transition. A lag duration extracted from a fluorescence tangent depends on how lag is defined and measured. Calling either quantity a mechanistic rate constant and fitting its logarithm against inverse temperature can combine two unjustified transformations.

Mechanistic aggregation models

Cohen and colleagues used kinetic studies, selective radiolabeling, and additional assays to identify a fibril-catalyzed secondary-nucleation pathway for AΞ²42.[7] That is evidence for positive feedback in that experimental system, not proof that every lipidated peptide follows the same network. A sigmoid alone cannot distinguish secondary nucleation from fragmentation, a conformational pre-equilibrium, heterogeneous nucleation, or changing detector response.

A schematic model might track soluble monomer, fibril number, and fibril mass. Primary nucleation creates new growth-competent structures, elongation consumes monomer, and secondary processes increase the number of growing structures. Different terms can generate nearly indistinguishable mass curves. Global analysis across independently varied concentrations and suitable mechanistic perturbations can constrain models more strongly than fitting each trace separately; that is the central experimental logic of Meisl and colleagues.[6]

I would not import a published AΞ² rate constant into a PP shelf-life model. Sequence, formulation, concentration, surfaces, package, and experimental history change the system being modeled. A mechanistic model can be useful for designing experiments even when its individual rate constants are not identifiable enough for extrapolation. The distinction between explanatory usefulness and predictive validation is crucial.

Statistical variation across units and batches

An aggregate trajectory fitted to a mean may conceal variable onset across vials. A hierarchical model could allow batch effects on baseline, growth, or transition timing, and container-level variation within batch. If the relevant endpoint is the first time a container exceeds a defined threshold, interval-censored time-to-event modeling may better match the data than pretending each container follows one deterministic average curve. These are my proposed modeling options, not requirements attributed to Q1E.

With very few batches, complex random-effect distributions are themselves weakly estimated. A Bayesian model can express external knowledge and uncertainty, but a narrow prior can also manufacture precision. I would report prior sensitivity and predictive checks, preserve the distinction between observed and future batches, and avoid interpreting lack of a detected batch difference as proof of exchangeability. Repeated injections do not supply the missing batch information.

Worked example: apparently stable early data, incompatible futures

This is an analyst-constructed hypothetical calculation, not a fit to laboratory measurements. Define an aggregate attribute in percentage points by A(t) = 0.1 + 19.9/[1 + exp(-k(t - t50))], with time in months. Model A uses k = 0.8/month and t50 = 12 months; model B uses k = 0.4/month and t50 = 18 months. The upper asymptote is twenty percent in both cases, chosen only to illustrate model dependence.

Month Model A (%) Model B (%)
0 0.1013 0.1148
3 0.1148 0.1492
6 0.2624 0.2624
9 1.7551 0.6293
12 10.0500 1.7551
15 18.3449 4.7064

Over the first six months the curves differ by only a few hundredths of a percentage point at these selected observations. Whether those differences are distinguishable depends on independently established method and container variability. They are plainly not enough to claim that the future transition is identified merely because the early trace is smooth. The next observations become much more discriminating.

A least-squares line fitted only to model A's values at months zero, three, and six is 0.078999 + 0.026848t. Its point-estimate crossing of an illustrative two-percent limit is 71.55 months. Solving the actual logistic equations gives 9.19 months for A and 12.38 months for B. These are mathematical crossing times, not shelf-life claims, and no extrapolation rule is being applied to authorize them.

The example reveals two separate failures. The early straight line is the wrong long-horizon shape, while choosing a sigmoid does not by itself determine which sigmoid is credible. Both curves cross the two-percent criterion before their respective inflection points, so β€œset expiry before the inflection” is not a sufficient rule either. The criterion and its uncertainty must be evaluated directly.

Quantify uncertainty in the prediction, not only the parameters

Raue and colleagues distinguish structural non-identifiability from practical non-identifiability due to limited observations, and describe profile likelihood as a tool for diagnosing both and obtaining confidence intervals.[8] That distinction applies here. A model may be intrinsically unable to separate two rate constants from one observable, or the model may become identifiable only after observing enough of the transition. A small optimizer-reported standard error does not resolve either problem automatically.

For a nonlinear mean function f(t, ΞΈ), define the threshold-crossing quantity explicitly: T(ΞΈ) = inf{t: f(t, ΞΈ) β‰₯ L} for an increasing attribute with upper limit L. Some parameter sets may not cross within the valid model domain. Those cases must not be silently dropped from bootstrap output or replaced with an arbitrary large time. If the credible parameter region includes very early crossings, an attractive median crossing is not an adequate conservative claim.

One analysis option is a profile likelihood for T itself: constrain a candidate crossing time, optimize nuisance parameters, and examine how fit quality changes. Another is a design-aware parametric bootstrap, simulating data from the fitted observation/error model and refitting each replicate. Resampling independent storage units or batches may be appropriate when enough such units exist; resampling every injection independently is not. These are statistical proposals grounded in identifiability logic, not a claim that the cited paper validates a PP shelf-life workflow.[8]

The final uncertainty must include observation noise, finite parameter information, relevant batch/container variation, and sensitivity to defensible alternative model families. Bootstrap intervals conditional on one wrong sigmoid omit model-form uncertainty. I would compare plausible models, restrict prediction to a justified domain, and show sensitivity to error structure and censoring. If useful conservative bounds cannot be established, the result is β€œinsufficient evidence for extrapolation,” not an expiry rounded down from an unstable estimate.

Mean confidence, individual prediction, and population assurance

Consider a separate hypothetical linear calculation, used only to expose the statistical distinction. At months 0, 3, 6, 9, 12, let measured attribute values be 0.10, 0.20, 0.34, 0.43, 0.62%. Ordinary least squares gives intercept 0.0840, slope 0.042333 percentage points/month, and residual standard deviation 0.028225. Assume independent normally distributed residuals, a valid linear mean, and a one-percent upper limit; none is asserted for actual aggregation data.

The fitted mean crosses one percent at 21.64 months. Using the one-sided 95% Student-t factor for three residual degrees of freedom, 2.35336, the upper confidence limit on the mean crosses at 19.32 months. Adding the individual-observation residual variance to form a one-sided prediction limit yields 18.90 months. These calculations demonstrate different estimands; they do not authorize extrapolation beyond twelve months of observations.[1]

A tolerance bound adds a different requirement: a stated fraction of a population is covered with a stated confidence level. Neither the mean bound nor a single-future-observation prediction interval is automatically that tolerance bound. With multiple quality attributes and models chosen after looking at the results, nominal confidence also does not automatically supply simultaneous product-level assurance. I would name the inferential target in the statistical analysis plan and test the implemented procedure's behavior, rather than write β€œ95% safe.”

For nonlinear models, a first-order covariance approximation can be useful locally but can fail near weakly identified transitions or constraints. A pointwise upper confidence curve and the distribution of crossing times are related but not interchangeable under every construction. Their coverage and interpretation must be justified for the chosen design. I would preserve the numerical algorithm, horizon, convergence diagnostics, failed fits, and simulation checks in the audit package.

Censoring, residuals, and experimental leverage

PDQC831 already warns against replacing every below-quantitation result with zero or half the threshold. To make that operational, suppose a validated observation model is normal with mean ΞΌ and standard deviation Οƒ, and the laboratory reports only Y < L. Its likelihood contribution is Ξ¦[(L - ΞΌ)/Οƒ], not a fabricated exact value at L/2. A different error distribution or reporting mechanism requires a different likelihood.

For an illustrative threshold L = 0.05% and Οƒ = 0.02 percentage points, the probability of a below-threshold result is 0.9332 when ΞΌ is 0.02%, but only 0.0668 when ΞΌ is 0.08%. Thus censored observations contain information without revealing an exact amount. Logging a substituted zero is impossible; logging a substituted small constant can impose arbitrary apparent growth. Raw instrument estimates below a validated quantitative range may aid investigation but should not be promoted to precise measurements merely to simplify fitting.

Diagnostics should examine residuals against time, fitted value, batch, analyst, and run, as well as recovery and reference-standard trends. Heteroscedasticity may require a validated variance model; serial correlation may arise from shared preparations or instrument runs. A plateau can reflect saturation of a detector rather than depletion of monomer. High RΒ² says little about any of these problems.

The most valuable next pull is often near where credible models begin to disagree, not merely another measurement deep inside the lag. Independent containers and representative batches are more informative about onset variability than additional injections of one vial. Stress arms can help reveal plausible mechanisms and measurement failure modes. They do not necessarily place the same transition at the same scaled time under intended storage conditions.[5,6]

Temperature and phase changes are a domain boundary

Wang and Roberts review why temperature-induced aggregation can be non-Arrhenius even over practically relevant windows.[5] This extends beyond a bad straight-line fit: temperature may change conformational populations, association equilibria, or the rate-controlling route. Solid-state moisture changes or phase transitions can introduce a different relationship entirely. I would not share one temperature coefficient across states without evidence that the fitted model remains valid.

Conversely, statistically significant curvature over time does not prove non-Arrhenius temperature dependence. One concerns the trajectory at a given condition; the other concerns how kinetic behavior changes across temperatures. A nonlinear time course can have temperature-dependent parameters that follow Arrhenius behavior over a validated domain, and a nearly linear time course can have non-Arrhenius temperature dependence. Separating these questions prevents the earlier inquiry's mechanistic warning from becoming a blanket rejection of every accelerated experiment.

An excursion may also create seeds whose effects persist after cooling. In that case, subsequent behavior depends on history and state, not just an accumulated temperature number. Mean kinetic temperature cannot infer whether such an event occurred. Real-time confirmation should use the intended material, package, concentration, and condition, while accelerated information supports a clearly bounded question rather than a universal time-conversion factor.[2,5]

What this means for Peptide Partners

I would give an auditor a compact argument with linked evidence. First define the attribute and the claim: mean quantitative trend, future-container behavior, or another explicitly justified endpoint. Then document measurement selectivity and recovery, independent units and batches, the candidate model family, censoring rules, variance assumptions, prespecified diagnostics, and failure criteria. Cite Q1E section 2.6 for methodological flexibility, while separately explaining applicability and the actual evidence supporting the proposed duration.[1]

The report should show fitted observations and uncertainty together with withheld or subsequently acquired real-time checks. Include alternative plausible trajectories, parameter profiles or equivalent diagnostics, and the reasons for restricting the extrapolation horizon. A model validated only before an aggregation transition has not validated its prediction of that transition. A disagreement in a new batch reopens the model, not just the batch's data processing.

PP's practical options are to support a limited observed storage interval, collect more discriminating data, improve the assay, change the material/package, or narrow the claim. Better analytical precision can reduce uncertainty conditional on a valid model; it cannot repair an incorrect mechanism or missing transition. Neither this inquiry nor the agency texts establish a product-specific shelf life for PP's untested peptides. RUO labeling does not justify an unsupported stability promise.

Questions to bring to the lab

  1. What does our aggregate result actually measure, and can we demonstrate recovery of material outside the soluble SEC window?
  2. Are independent storage containers, preparations, and injections distinguished in both the dataset and the variance model?
  3. Which plausible trajectories still fit the early data, and at what next time point would their predictions diverge enough to test?
  4. Do the parameters or crossing time have bounded likelihood profiles, or is the apparent precision an optimizer artifact?
  5. Is our intended assurance about a mean, a future container, or a stated fraction of future batches, and why?
  6. What evidence supports sharing a kinetic model across temperatures, moisture states, packages, and stressed versus unstressed material?
  7. What prospectively defined failure would stop extrapolation, trigger more real-time observations, or require revising the research-storage claim?

What I still don't know

I do not have PP aggregation trajectories, independent-container variability, suitable aggregate reference materials, or proof of a secondary-nucleation pathway for any catalog peptide. The calculated logistic and linear examples are teaching counterexamples, not validation datasets, simulated PP results, or universal expiry estimates. I cannot select a kinetic model or an acceptable stability interval without those experimental inputs.

My source inspection establishes Q1E's explicit allowance for nonlinear regression and the draft labels on current agency Q1 pages. It does not establish future adoption dates or guarantee acceptance of a particular analysis. The mechanistic and identifiability papers were verified by matching DOI metadata and read at abstract scope; detailed equations and worked numbers here are my stated mathematical constructions. Their limitations are part of the conclusion, not a footnote to a shelf-life promise.

References

  1. International Conference on Harmonisation. Evaluation for Stability Data Q1E. Current Step 4 version, 6 February 2003. https://database.ich.org/sites/default/files/Q1E_Guideline.pdf
  2. International Conference on Harmonisation. Quality of Biotechnological Products: Stability Testing of Biotechnological/Biological Products Q5C. 30 November 1995. https://database.ich.org/sites/default/files/Q5C_Guideline.pdf
  3. European Medicines Agency. ICH Q1 guideline on stability testing of drug substances and drug products β€” Scientific guideline. Current-version entry: Step 2b draft, consultation closed. Accessed 15 October 2026 (America/New_York). https://www.ema.europa.eu/en/scientific-guidelines/ich-q1-guideline-stability-testing-drug-substances-drug-products
  4. U.S. Food and Drug Administration. Q1 Stability Testing of Drug Substances and Drug Products. Draft Guidance for Industry. June 2025. Accessed 15 October 2026 (America/New_York). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/q1-stability-testing-drug-substances-and-drug-products
  5. Wang W, Roberts CJ. Non-Arrhenius Protein Aggregation. The AAPS Journal. 2013;15(3):840-851. doi:10.1208/s12248-013-9485-3
  6. Meisl G, Kirkegaard JB, Arosio P et al. Molecular mechanisms of protein aggregation from global fitting of kinetic models. Nature Protocols. 2016;11(2):252-272. doi:10.1038/nprot.2016.010
  7. Cohen SIA, Linse S, Luheshi LM et al. Proliferation of amyloid-Ξ²42 aggregates occurs through a secondary nucleation mechanism. Proceedings of the National Academy of Sciences. 2013;110(24):9758-9763. doi:10.1073/pnas.1218402110
  8. Raue A, Kreutz C, Maiwald T et al. Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood. Bioinformatics. 2009;25(15):1923-1929. doi:10.1093/bioinformatics/btp358