English:Aging – Protein turnover beyond static proteomic signatures

aiMOOC

Aging – Protein turnover beyond static proteomic signatures

QR-Code

Introduction

Aging – Protein turnover beyond static proteomic signatures is an expert-level colloquium on how to infer protein dynamics from pulse–chase proteomics without confusing observed label loss with a single biological process. The central case study is the 2026 Nature paper by Guldner and colleagues, Ageing promotes microglial accumulation of slow-degrading synaptic proteins, which used cell-selective BONCAT in mice to follow neuronal proteins during ageing.

The key methodological question is not merely whether a labelled protein signal declines. You must ask what process generated that decline. A protein can disappear from the measured neuronal pool because it is proteolytically degraded, secreted, transferred to another cell, lost with a dying cell, diluted by changes in the sampled population, or shifted below the analytical detection limit. Conversely, labelled signal can persist because degradation is slow, because protein is physically sequestered, because labelled amino acid is recycled into new protein, or because the assay continues to recover labelled material after it has left the original neuron.

This course therefore treats protein turnover as a dynamic measurement-model problem. You will reconstruct the published analysis, separate latent biological fluxes, examine structural and practical identifiability, quantify uncertainty, test non-exponential alternatives, handle missing proteins explicitly, and design follow-up experiments that distinguish competing explanations.

The Guldner study compared young, middle-aged and aged mice and followed BONCAT-labelled proteins in several brain regions. The published experiment used four sacrifice time points during a two-week chase, with four BONCAT-labelled biological replicates per age and time point and wild-type background controls. BONCAT-enriched peptides were multiplexed with tandem mass tags and quantified by LC–MS. The authors reported an average increase in neuronal protein half-life with age and substantial regional heterogeneity.

The expert task is to determine which statements are directly identified by those observations and which depend on modelling assumptions.

Load video

YouTube

YouTube might collect personal data. Privacy Policy

Continue Dismiss

Learning Goals

By the end of the colloquium, you should be able to distinguish a static proteomic abundance difference from a kinetic turnover difference, reconstruct the published BONCAT pulse–chase measurement model, derive alternative observation equations, identify which kinetic parameters are structurally identifiable, evaluate practical identifiability in a short sampling window, propagate biological and analytical uncertainty, perform sensitivity analyses for non-exponential kinetics and missing proteins, and propose an experimental design that distinguishes degradation from synthesis, secretion, cell loss, intercellular transfer and label recycling.

From Static Proteomics to Fluxes

A static proteomic measurement gives a snapshot of the amount of protein present at the sampling time. That amount is the net result of multiple fluxes. In the simplest one-compartment model,

$\frac{dP}{dt}=S(t)-k_{\mathrm{d}\mathrm{e}\mathrm{g}}(t)P(t)$

where $P(t)$ is protein abundance, $S(t)$ is synthesis and $k_{\mathrm{d}\mathrm{e}\mathrm{g}}$ is a first-order degradation rate constant.

At steady state,

$S=k_{\mathrm{d}\mathrm{e}\mathrm{g}}P$.

This equation immediately shows why abundance alone does not identify turnover. The same abundance can arise from high synthesis plus high degradation or from low synthesis plus low degradation. Protein concentration and protein flux are different quantities.

In tissue, the model is usually more complicated. Removal from the measured pool can include proteolysis, secretion, export, intercellular transfer and cell loss. A static proteome therefore cannot determine which kinetic mechanism produced an age-associated abundance signature.

The image illustrates several routes by which proteins can leave a functional intracellular pool, including proteasomal and autophagic degradation. A pulse–chase experiment adds temporal information, but temporal information alone does not automatically make every underlying flux identifiable.

Why Half-Life Is a Model-Dependent Quantity

For a homogeneous population undergoing first-order decay with no new labelled input,

$L(t)=L(0)e^{-kt}$

and the half-life is

$t_{1/2}=\frac{\ln 2}{k}$.

This familiar relationship is exact only under the assumed model. If the hazard of removal changes with molecular age, if there are fast and slow subpopulations, if labelled input continues during the chase, or if the assay combines several compartments, a single exponential half-life becomes an effective summary, not necessarily the intrinsic degradation half-life of a protein molecule.

Pulse length also matters when degradation is non-exponential. A long pulse creates a labelled population containing molecules of different molecular ages at the start of the chase. For an age-dependent degradation process, those molecules can have different subsequent hazards. The resulting chase curve is therefore a convolution of the pulse history and the lifetime distribution.

The Guldner 2026 BONCAT Experiment

The case study used engineered phenylalanyl-tRNA synthetase activity to permit cell-selective incorporation of azido-phenylalanine into newly synthesized neuronal proteins. The azide handle enabled click-chemistry enrichment of the labelled proteome.

Young mice were approximately 4 months old, middle-aged mice approximately 12 months old and aged mice approximately 24 months old in the turnover experiment. Mice received AAV:Camk2a-PheRS* and a pulse–chase AzF administration scheme. Four groups were euthanized at four time points within the two-week chase. Brain regions were dissected, BONCAT-labelled proteins were enriched, peptides were TMT-multiplexed and analysed by LC–MS.

The published turnover analysis used non-normalized abundance data so that the expected decline in labelled protein was not removed by global normalization. At the first time point, proteins had to be sufficiently enriched over wild-type background. For comparisons across ages within a region, the analysis further focused on proteins detected across the compared groups. The first measured time point was set to 100% remaining and later points were expressed relative to it. Trajectories with substantial increases were excluded, while small increases of up to 5% between time points were tolerated.

The authors did not impute the turnover trajectories. Instead, they used stringent detection filters because most missingness occurred between TMT plexes rather than within a plex. This avoids inventing kinetic values for proteins absent from an entire comparison, but it also changes the estimand: the reported age comparisons apply to the subset of proteins that survive the detectability and completeness filters.

Published Half-Life Modelling

For half-life estimation, the mean trajectory for each protein was normalized so that the first measurement equalled one. The published methods fitted a one-level exponential model and a more complex two-level model by least squares. The one-level model was

${\Lambda }_1(t)=e^{-k_{A}t}$.

The two-level model represented a multi-stage kinetic process. Parameters were optimized numerically, and the model with the lower Akaike Information Criterion was selected. The authors also compared model-derived half-lives with direct interpolation for proteins whose measured trajectories actually crossed 50% remaining.

This is an important validation, but it does not eliminate extrapolation uncertainty for proteins that never approached 50% remaining. Direct interpolation is only possible after the trajectory spans the half-level. Model-based estimates can extend beyond the observed window, but their uncertainty can increase rapidly when the data contain little curvature or little total decline.

Reanalyzing the Measurement Model

The first reanalysis step is to separate the biological state model from the measurement model. A state model describes hidden biological quantities. A measurement model describes how the experiment converts those hidden quantities into observed intensities.

A Minimal Multi-Process State Model

Let $L_N(t)$ denote labelled protein in the targeted neuronal intracellular pool, $L_E(t)$ labelled protein in an extracellular or secreted pool, and $L_M(t)$ labelled neuron-derived protein residing in microglia or another recipient cell.

A useful starting model is

$\frac{d{L}_N}{dt}=u(t)-(k_{\mathrm{d}\mathrm{e}\mathrm{g}}+k_{\mathrm{s}\mathrm{e}\mathrm{c}}+k_{\mathrm{t}\mathrm{r}}+k_{\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}})L_N$

$\frac{d{L}_E}{dt}=k_{\mathrm{s}\mathrm{e}\mathrm{c}}{L}_N-k_{\mathrm{c}\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{r},\mathrm{E}}{L}_E$

$\frac{d{L}_M}{dt}=k_{\mathrm{t}\mathrm{r}}{L}_N-k_{\mathrm{c}\mathrm{l}\mathrm{e}\mathrm{a}\mathrm{r},\mathrm{M}}{L}_M$.

Here:

u(t) is new labelled protein synthesis during the nominal chase.

kdeg is intracellular proteolytic degradation.

ksec is secretion or export from the intracellular neuronal compartment.

ktr is transfer of labelled neuronal protein into another cell type.

kcell is loss associated with disappearance of labelled neurons from the sampled cellular pool.

kclear,E and kclear,M are clearance rates from the extracellular and recipient-cell pools.

The terms are conceptually distinct even when the original experiment cannot estimate them separately.

Cell Number Must Be Explicit

If the measured abundance is a tissue-level total, neuronal cell number can affect the signal independently of molecular degradation. Let $N(t)$ be the number of labelled neurons contributing to the sample and ${\ell }_N(t)$ the average labelled amount per surviving neuron. Then

$L_N(t)=N(t){\ell }_N(t)$.

A decline in $L_N$ can therefore arise from fewer labelled molecules per neuron, fewer contributing neurons, or both.

Cell loss is not always equivalent to immediate label loss from the tissue homogenate. If neuronal proteins released by dying cells remain extracellularly or are engulfed by microglia, a whole-region homogenate may continue to contain those labelled proteins. Therefore the effective influence of cell death depends on the sampling compartment and the fate of released material.

The Observation Equation

A regional tissue measurement can be written as

$Y_{\mathrm{r}\mathrm{e}\mathrm{g}\mathrm{i}\mathrm{o}\mathrm{n}}(t)=q(t)[L_N(t)+{\eta }_E{L}_E(t)+{\eta }_M{L}_M(t)]+B(t)+ϵ(t)$.

q(t) represents recovery, enrichment and MS response.

ηE and ηM represent how efficiently extracellular and microglial labelled material enters the assayed fraction.

B(t) is nonspecific BONCAT pull-down or analytical background.

ε(t) is measurement error.

This equation matters because BONCAT labels the protein molecule at synthesis. If the labelled protein later moves from a neuron into a microglial cell but remains within the dissected tissue, the label itself does not identify its current cellular owner. The molecule can still contribute to the regional BONCAT signal.

The Guldner study independently demonstrated accumulation of neuron-derived labelled proteins in microglia. That biological result is therefore directly relevant to the turnover measurement model: intercellular relocation is a plausible latent process that a whole-region signal cannot by itself separate from degradation or retention.

Label Recycling and Residual Precursor

During an ideal chase, $u(t)=0$. In practice, that condition must be demonstrated rather than assumed.

Let $Z(t)$ denote the intracellular free pool of label-capable AzF. Then

$u(t)=\alpha S(t)Z(t)$,

where $S(t)$ is the underlying protein synthesis flux and $\alpha$ converts precursor availability into labelled incorporation.

After administration stops, $Z(t)$ may decline because of clearance and dilution. It could also receive input from residual tissue stores or release of AzF-containing material after proteolysis. If recycled AzF remains competent for re-incorporation by PheRS*, newly synthesized proteins can carry the same label as the pulse cohort.

A simple precursor model is

$\frac{dZ}{dt}=-k_{Z}Z+\rho k_{\mathrm{d}\mathrm{e}\mathrm{g}}{L}_N$,

where $\rho$ is an effective recycling fraction. This does not assert that recycling materially affected the published experiment. It defines a competing explanation that should be tested experimentally.

If $u(t)>0$, the observed labelled signal is no longer a pure survival curve of the original pulse. Ignoring the input biases the inferred disappearance rate downward and the apparent half-life upward.

What Is Identifiable from the Published Chase?

Structural identifiability asks whether a parameter could be uniquely recovered from perfect, noise-free observations of the specified outputs.

Practical identifiability asks whether the finite, noisy experiment actually constrains that parameter with useful precision.

From a single whole-region BONCAT trajectory, the separate rates $k_{\mathrm{d}\mathrm{e}\mathrm{g}}$, $k_{\mathrm{s}\mathrm{e}\mathrm{c}}$, $k_{\mathrm{t}\mathrm{r}}$ and $k_{\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}}$ are generally not structurally identifiable. If all act as first-order losses from the same observed compartment, only their sum is directly visible:

$k_{\mathrm{o}\mathrm{u}\mathrm{t}}=k_{\mathrm{d}\mathrm{e}\mathrm{g}}+k_{\mathrm{s}\mathrm{e}\mathrm{c}}+k_{\mathrm{t}\mathrm{r}}+k_{\mathrm{c}\mathrm{e}\mathrm{l}\mathrm{l}}$.

If labelled input is also unknown, even $k_{\mathrm{o}\mathrm{u}\mathrm{t}}$ can trade off against $u(t)$. A smaller loss rate and a larger continued input can generate a trajectory similar to a larger loss rate and a smaller input.

The most defensible quantity from the original regional pulse–chase alone is therefore an effective disappearance trajectory from the assayed labelled-protein pool. Calling it a degradation trajectory requires assumptions about secretion, transfer, cell loss, recycling, recovery and background.

Effective Half-Life versus Degradation Half-Life

Under a mono-exponential no-input model,

$t_{1/2,\mathrm{e}\mathrm{f}\mathrm{f}}=\frac{\ln 2}{k_{\mathrm{o}\mathrm{u}\mathrm{t}}}$.

Only when non-proteolytic losses are negligible and labelled synthesis during the chase is negligible does

$t_{1/2,\mathrm{e}\mathrm{f}\mathrm{f}}\approx t_{1/2,\mathrm{d}\mathrm{e}\mathrm{g}}$.

This distinction should appear in tables, plots and conclusions. A useful reporting convention is to call the primary quantity BONCAT-labelled protein persistence half-life or effective regional disappearance half-life unless orthogonal measurements justify a mechanistic degradation interpretation.

Practical Identifiability in a Two-Week Window

The published design contains four time points within a two-week chase. That is enough to detect substantial differences in many trajectories, but it does not make every half-life equally estimable.

For a mono-exponential model, long half-lives produce shallow slopes. A protein with a 60-day half-life retains about 85% of its starting signal after 14 days. A protein with a 120-day half-life retains about 92%. When analytical and biological variation is comparable to those declines, the half-life is weakly constrained and the confidence interval can extend far beyond the observation window.

Very short half-lives create the opposite problem. If most labelled material disappears before the first informative post-pulse sample, later values cluster near the assay floor. The experiment then identifies only that turnover is fast, not the precise rate.

The most informative design places observations where candidate trajectories differ strongly and where the signal remains above the detection floor. For a target half-life, a sample near the expected 50% remaining point is especially valuable, but early points are also necessary to identify fast or multi-phase decay.

A Reporting Rule for Long and Short Half-Lives

Do not force every protein into a finite point estimate.

If the profile likelihood for the half-life is bounded on both sides, report the estimate and a 95% uncertainty interval.

If the likelihood is flat toward long half-lives, report a one-sided lower bound such as half-life greater than X days at 95% confidence.

If the protein is already near the detection floor at the earliest post-pulse sample, report an upper bound or a fast-turnover category.

If the curve is incompatible with monotonic no-input decay, report that the simple half-life is not estimable under the primary model and inspect recycling, background, missingness and non-exponential kinetics.

Reproducing the Published Analysis

A reanalysis should begin by reproducing the original result before changing the model.

The raw and processed mass-spectrometry resource is available through ProteomeXchange dataset PXD056701, titled Protein Degradation Dynamics Among Brain Regions and Ages. The journal article provides Supplementary Table 2 for the degradation analysis.

A baseline reconstruction should preserve the original age groups, brain-region definitions, background controls, TMT plex structure, protein identifiers, first-time-point normalization and trajectory filters.

Baseline Reconstruction Steps

1. ProteomeXchange: Retrieve PXD056701 metadata, raw files, search outputs and processed protein tables; record file names, sizes and checksums.
2. Experimental metadata: Build a machine-readable sample sheet containing age, brain region, time point, animal identifier, TMT channel, BONCAT status, background status and plex.
3. Background correction: Reproduce the first-time-point enrichment filter and background treatment used in the publication before introducing alternative background models.
4. Trajectory reconstruction: Calculate percent remaining relative to the first time point exactly as specified in the published method.
5. Half-life reconstruction: Refit the published one-level and two-level models and confirm that reconstructed half-lives agree with the reported Supplementary Table within a predefined numerical tolerance.
6. Quality control: Compare reconstructed distributions, age fold changes, direct-interpolation relationships and region-specific summaries with the published figures.

Only after the baseline result is reproduced should model extensions be introduced.

A Reproducible Reanalysis Protocol

The protocol below separates data reconstruction, model fitting, identifiability diagnostics and biological interpretation.

Step A: Freeze the Analysis Environment

Create a version-controlled repository with a plain-text environment specification, exact package versions, random seeds, operating-system information and a scripted workflow. Use a workflow manager or a single reproducible entry point so that every figure and table can be regenerated from the declared inputs.

Do not edit source tables manually. Store every transformation as code. Generate a manifest containing cryptographic checksums for downloaded source files and keep a read-only copy of the original inputs.

Step B: Preserve the Replicate Structure

The published half-life fit uses mean trajectories. The reanalysis should additionally fit replicate-level data.

Let $Y_{pirt}$ be the abundance for protein $p$, individual animal $i$, region $r$ and time $t$. Fit the observation model to replicate-level values rather than treating the time-point mean as error-free.

A hierarchical model can include animal-level residual variation, plex-level variation and protein-specific parameters. This prevents the uncertainty of the mean trajectory from disappearing during curve fitting.

Step C: Model Background Rather than Only Subtracting It

Background subtraction can create unstable ratios when the true signal approaches the background. An alternative is to fit raw positive intensities using

$Y(t)=B_{\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{x}}+Af(t;\theta )+ϵ$,

where $B_{\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{x}}$ is estimated from matched wild-type controls.

Fit at least two background specifications:

a fixed background estimated from controls;

a hierarchical background with uncertainty propagated into the half-life.

Compare results with the published subtraction approach. Proteins whose half-life changes strongly when the background model changes should be flagged as background-sensitive.

Step D: Fit a Predeclared Model Set

Use a small model set chosen for interpretable biological alternatives.

Model 1: Mono-exponential no-input decay

$f(t)=e^{-kt}$.

Model 2: Two-component mixture

$f(t)=\pi e^{-k_{1}t}+(1-\pi )e^{-k_{2}t}$.

This represents kinetically distinct subpopulations but requires more information than the mono-exponential model.

Model 3: Weibull survival

$f(t)=e^{-(t/\lambda {)}^{\beta }}$.

When $\beta =1$, this reduces to an exponential. Values below or above one represent decreasing or increasing hazard with molecular age.

Model 4: Continued labelled input

$\frac{dL}{dt}=u_0{e}^{-k_{Z}t}-k_{\mathrm{o}\mathrm{u}\mathrm{t}}L$.

This tests whether residual or recycled label could explain a shallow trajectory.

Model 5: Two-stage kinetic model

Use a pulse-aware two-stage model comparable to the class used in the publication and in prior analyses of non-exponential protein degradation.

With only four mean time points, do not treat a better fit by a multi-parameter model as proof of that mechanism. The model set is a sensitivity analysis, not a competition to maximize complexity.

Step E: Use Small-Sample Model Diagnostics

The published analysis used AIC to choose between kinetic models. With only four mean time points per protein, model-selection uncertainty is substantial.

Report residual plots, parameter boundaries and likelihood profiles in addition to AIC. Calculate AICc when mathematically defined, but note that for very small $n$ and multi-parameter models the AICc correction can become unstable or undefined.

Use cross-validation only when the replicate structure provides enough independent information. Do not pretend that leaving out one of four averaged time points gives a robust predictive assessment.

For each protein, report whether competing models are distinguishable over the observed window. If several models fit almost equally well but imply different long-term half-lives, the long-term half-life is model-dependent.

Step F: Quantify Parameter Identifiability

For each fitted model:

calculate the sensitivity of the predicted trajectory to each parameter;

compute a Fisher-information diagnostic or equivalent local sensitivity matrix;

generate profile likelihoods for the primary half-life or rate parameters;

flag parameters whose 95% profile-likelihood interval reaches the parameter boundary;

simulate replicate datasets at the fitted parameters and refit them to estimate empirical bias and coverage.

For the expanded mechanistic model, perform a structural identifiability analysis before fitting. Parameters that are structurally confounded should be fixed from external measurements, reparameterized into identifiable combinations, or omitted.

Step G: Propagate Uncertainty

Use a hierarchical bootstrap or model-based simulation that resamples biological replicates within age, time and region while preserving the experimental design.

For every protein, report at least:

the point estimate of the effective rate or half-life;

a 95% confidence interval or credible interval;

whether the interval is two-sided or one-sided;

the model used;

the observed fraction remaining at day 14;

a detectability or censoring flag;

a model-sensitivity flag.

For age contrasts, calculate uncertainty for the contrast itself, not only for the two separate half-lives.

For example,

${\Delta }_p={\log }_2\left(\frac{t_{1/2,\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{d},p}}{t_{1/2,\mathrm{y}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{g},p}}\right)$.

Bootstrap or posterior draws should propagate uncertainty into ${\Delta }_p$.

Sensitivity Analysis for Non-Exponential Kinetics

Non-exponential degradation is biologically plausible. Experimental work has shown that some proteins are especially unstable shortly after synthesis and become more stable after assembly into complexes. In such cases a single exponential can hide age-dependent molecular hazard.

Load video

YouTube

YouTube might collect personal data. Privacy Policy

Continue Dismiss

A robust sensitivity analysis should compare several summaries rather than only half-life.

Useful summaries include the fraction remaining at fixed times, the area under the labelled-survival curve over the observed window, an effective early slope, an effective late slope, the median lifetime when identified, and the mean lifetime for a fully specified survival model.

The area under the observed survival curve through day 14,

$AU{C}_{0,14}={\int }_0^{14}f(t)dt$,

is often less extrapolation-dependent than a half-life lying far outside the observation window.

Pulse-Length Sensitivity

For non-exponential decay, the labelled molecules present at chase start are not all the same molecular age. Therefore repeat the analysis under explicit pulse-history assumptions.

Simulate plausible synthesis histories during the pulse and propagate them into the chase-start age distribution. Ask whether the inferred age effect persists when the pulse duration or within-pulse synthesis profile changes.

A follow-up experiment can use different pulse lengths in matched animals. If the apparent half-life changes with pulse duration, a memoryless single-exponential model is inadequate.

Missing Proteins Are Part of the Model

Missing values in mass-spectrometry proteomics can arise from stochastic acquisition, low signal, left-censoring below the detection limit, failed identification, biological absence, or cross-plex incompatibility.

The original turnover analysis avoided imputation and restricted many comparisons to proteins reliably detected across the required groups. That is a defensible primary analysis, but it creates a selected protein set.

Proteins with high abundance, favorable peptides and persistent signal are more likely to survive all filters. Fast-decaying proteins can disappear at late time points. Low-abundance proteins can be missing in one age or region. Therefore complete-case filtering can alter the distribution of estimated half-lives.

Missingness Sensitivity Plan

Run at least four analyses.

Complete-case analysis: reproduce the publication's stringent common-detection criteria.

Censored-likelihood analysis: treat late missing observations with a known or estimated lower limit of quantification as left-censored rather than replacing them with a single small value.

Selection-model analysis: estimate detection probability as a function of latent intensity, plex and peptide characteristics.

MNAR sensitivity analysis: shift the assumed distribution of missing values across a plausible range and examine whether the age effect changes.

Do not impute a protein that is absent from an entire plex as though its abundance were a precisely measured low value. Such a protein may be below detection, technically absent or biologically different. The uncertainty must remain visible.

Protein-Level and Peptide-Level Missingness

Where peptide-level data are available, model peptides nested within proteins. A protein inferred from several consistently observed peptides carries different information from a protein represented by a single marginal peptide.

Repeat age comparisons after stratifying proteins by baseline abundance, peptide count and missingness rate. If the apparent age effect is restricted to highly detectable proteins, the conclusion should be qualified accordingly.

Competing Explanations for Slower Label Decline

Explanation	Predicted neuronal intracellular signal	Predicted extracellular or microglial signal	Predicted free AzF during chase	Discriminating measurement
Reduced intracellular proteolysis	Persists longer	No necessary increase	Near zero	Neuron-isolated labelled proteins plus proteasome and lysosome flux markers
Continued labelled synthesis from residual or recycled AzF	Persists longer or rebounds	Depends on trafficking	Detectable after nominal chase begins	Targeted free-AzF measurement and stronger precursor chase
Reduced secretion or export	Persists longer	Lower extracellular labelled protein	Near zero	Paired intracellular and CSF or interstitial-fluid BONCAT proteomics
Increased transfer into microglia with regional retention	May fall in neurons	Rises in microglia while whole-region signal persists	Near zero	Cell-type-resolved BONCAT measurement in neurons and microglia
Altered neuronal survival or sampled cell number	Changes with labelled-cell abundance	Depends on fate of released material	Near zero	Stereology, lineage counts and per-cell normalization
Aggregation or sequestration	Persists in insoluble fraction	May accumulate in recipient cells	Near zero	Parallel soluble and insoluble fractionation
Background or detection-floor artifact	Apparent shallow decline near floor	No biological prediction	No biological prediction	Replicate-level background model and dilution-series calibration

The purpose of this table is to design observations that give different predictions under the competing hypotheses.

Follow-Up Experimental Design

A stronger design should measure multiple outputs from the same conceptual system rather than adding more parameters to one output.

Time Sampling

Use dense early sampling and a longer late window. A candidate schedule is chase start, day 1, day 3, day 7, day 14, day 28, day 42 and day 56.

The final schedule should be chosen by simulation using the variance observed in the original study. Optimize the sampling times to reduce uncertainty for the half-life range of interest and to distinguish the mono-exponential, non-exponential and continued-input models.

If resources are limited, prioritize at least one early point before substantial fast-decay signal is lost and at least one late point near the expected half-level of long-lived proteins.

Measure the Label Precursor Pool

At every early chase point, quantify free AzF in plasma and the relevant brain regions using targeted mass spectrometry.

Measure the decline of free AzF directly rather than assuming an instantaneous switch to zero. A residual precursor curve supplies $Z(t)$ for the continued-input model.

Run a chase-strength experiment using increased unlabelled phenylalanine competition, provided the altered dosing is validated for safety and physiology. If stronger competition changes the apparent labelled-protein half-life, precursor persistence or recycling is contributing.

Separate Synthesis from Loss

Use a separate short-pulse cohort at each age and region to measure the rate of new neuronal protein labelling under standardized exposure. This estimates age-related changes in synthesis independently of the chase.

Where technically feasible, validate key proteins with an orthogonal nascent-protein method or a second chemically distinguishable pulse. The goal is to observe synthesis and persistence with labels that cannot be mistaken for one another.

A reduction in chase decay combined with unchanged synthesis supports increased persistence. A reduction in chase decay accompanied by persistent label-capable precursor or continued labelled synthesis requires a different interpretation.

Separate Cellular Compartments

At matched chase times, analyse:

the BONCAT-labelled neuronal fraction;

sorted microglia;

extracellular or interstitial fluid where feasible;

CSF and plasma for secreted labelled proteins;

soluble and detergent-insoluble brain fractions.

The sum of labelled material across compartments provides a mass-balance check. A label that disappears from neurons but appears in microglia is transfer, not immediate molecular destruction.

Quantify Cell Loss and Sampling Composition

Estimate the abundance of targeted Camk2a-positive neurons at each age and chase time using independent histological or lineage measures.

Report labelled protein both per tissue mass and, where feasible, per contributing neuron. This separates molecular turnover from changes in the number of labelled cells.

Include region volume and total protein recovery as covariates so that age-related anatomical or extraction differences do not masquerade as kinetic differences.

Design the Study by Simulation

Use the original replicate variance to simulate competing biological models.

For each candidate design, generate synthetic data under:

true degradation slowing;

unchanged degradation plus residual label input;

unchanged degradation plus reduced secretion;

unchanged degradation plus increased neuron-to-microglia transfer;

biexponential or Weibull decay;

age-dependent detection limits.

Fit the planned analysis model to every simulated dataset. Choose the design that recovers the correct mechanism most often and gives acceptable interval coverage for the primary kinetic parameters.

Decision Criteria for the Colloquium

A mechanistic conclusion should require three levels of evidence.

Trajectory evidence: the measured BONCAT persistence differs with age.

Model robustness: the difference remains under plausible background, missingness and kinetic alternatives.

Mechanistic separation: orthogonal compartment, precursor, synthesis and cell-number measurements distinguish degradation from competing explanations.

The first level supports a descriptive kinetic result. The second supports a robust effective-turnover result. The third is needed for a specific mechanistic degradation claim.

Recommended Output of the Reanalysis

The final reproducible report should contain a protein-level table with identifiers, region, age, observed abundances, percent remaining, fitted effective rate, half-life interval, model, background sensitivity, missingness class and identifiability flag.

Include region-level distributions of observed day-14 persistence as a low-extrapolation summary.

Include half-life distributions only for proteins meeting prespecified identifiability criteria.

Show profile-likelihood examples for well-identified, weakly identified and non-identifiable proteins.

Show how the age effect changes across the complete-case, censored, selection-model and MNAR-sensitivity analyses.

Provide model-comparison plots demonstrating where mono-exponential, multi-stage and continued-input models make distinguishable predictions.

Archive scripts, environment files, checksums and generated tables so that a second analyst can regenerate every result without manual editing.

Sources and Data

The central article is Guldner et al., Ageing promotes microglial accumulation of slow-degrading synaptic proteins, Nature 2026.

The turnover dataset is ProteomeXchange PXD056701.

The original BONCAT concept is described in Dieterich et al., Selective identification of newly synthesized proteins in mammalian cells using BONCAT.

Non-exponential pulse–chase analysis is developed in Sin, Chiarugi and Valleriani, Degradation Parameters from Pulse-Chase Experiments.

Age-dependent protein degradation kinetics are demonstrated in McShane et al., Kinetic Analysis of Protein Stability Reveals Age-Dependent Degradation.

Experimental and modelling considerations for intact-animal turnover studies are reviewed in Harmonizing Labeling and Analytical Strategies to Obtain Protein Turnover Rates in Intact Adult Animals.

Structural identifiability concepts for systems-biology models are reviewed in Structural Identifiability of Dynamic Systems Biology Models.

Recent proteomics missing-value methodology is discussed in msBayesImpute as a versatile framework for addressing missing values in biomedical mass spectrometry proteomics data.

BONCAT analysis of intracellular and secreted proteins is described in Bioorthogonal Non-Canonical Amino Acid Tagging to detect newly synthesized proteins in cells and their secretome.

Interactive Tasks

Quiz: Test Your Knowledge

What does a decline in regional BONCAT-labelled protein most directly measure? (Disappearance from the assayed labelled protein pool) (!Intracellular proteolysis only) (!Protein synthesis only) (!Neuron death only)

Under a single exponential no-input model, how is half-life related to the decay constant? (Half-life equals ln 2 divided by the decay constant) (!Half-life equals the decay constant divided by ln 2) (!Half-life equals twice the decay constant) (!Half-life is independent of the decay constant)

Why can secretion confound an intracellular degradation estimate? (Secreted labelled protein leaves the intracellular pool without being proteolytically destroyed there) (!Secretion always creates new label) (!Secretion makes mass spectrometry impossible) (!Secretion converts every protein into RNA)

What is the main identifiability problem when degradation, secretion and cell loss are observed only through one regional decay curve? (Only a combined effective loss rate may be identifiable) (!Every loss process is automatically identifiable) (!No kinetic parameter can ever be estimated) (!Cell loss has no effect on tissue measurements)

Why can residual or recycled AzF bias an apparent half-life upward? (It can create new labelled protein during the nominal chase) (!It forces every protein to aggregate) (!It removes all analytical background) (!It guarantees exponential kinetics)

Why are very long half-lives difficult to estimate in a short observation window? (The measured trajectory may show too little decline to constrain the half-life) (!Long-lived proteins cannot be measured by mass spectrometry) (!Long half-lives always produce increasing trajectories) (!The half-life formula changes with mouse age)

What is a major risk of complete-case filtering in turnover proteomics? (It can select for proteins that are easier to detect across all samples) (!It guarantees unbiased population estimates) (!It removes all biological variation) (!It creates additional time points)

Which measurement most directly tests whether label-capable precursor persists during the chase? (Targeted quantification of free AzF) (!Static RNA sequencing) (!Total brain weight) (!A single endpoint Western blot)

Which result would most strongly support neuron-to-microglia transfer rather than immediate destruction? (Neuronal labelled signal falls while matching labelled proteins rise in microglia) (!Every compartment declines at the same rate) (!Free AzF is undetectable) (!Total RNA remains constant)

What should be reported when a half-life profile likelihood remains open toward very long values? (A one-sided lower bound or non-identifiable long half-life) (!A precise finite half-life without an interval) (!A zero half-life) (!The arithmetic mean of all proteins)

Memory Game

Effective half-life	Time for the assayed labelled pool to fall by half under a specified model
Structural identifiability	Whether ideal observations uniquely determine a model parameter
Practical identifiability	Whether finite noisy data constrain a parameter precisely enough to use
Label recycling	Re-entry of label-capable precursor into newly synthesized protein
Left censoring	Observation known only to lie below an analytical detection threshold
Secretion	Export of protein from the intracellular compartment
Profile likelihood	Parameter uncertainty analysis based on constrained refitting
Pulse history	Time-dependent labelling input that determines the molecular-age distribution at chase start

Drag and Drop

Match the correct terms.	Topic
Free AzF measurement	Tests residual or recycled label precursor
Sorted microglial proteomics	Tests intercellular transfer of neuron-derived labelled proteins
Neuron count	Tests whether changing cell abundance contributes to tissue-level signal loss
Censored likelihood	Handles observations below the quantification limit without fixed-value imputation
Profile likelihood	Tests whether a kinetic parameter is practically constrained

...

Crossword Puzzle

Proteostasis	What term describes maintenance of protein homeostasis?
Identifiability	What property asks whether model parameters can be inferred uniquely?
Recycling	What process can reintroduce label into newly synthesized protein?
Secretion	What process exports protein from a cell?
Censoring	What statistical concept describes a value known only to lie beyond a detection boundary?
Bootstrap	What resampling method can propagate biological-replicate uncertainty?

Cloze Text

Complete the text.

A regional BONCAT decay curve directly measures persistence of the . A single exponential converts its fitted rate into an effective . Separate degradation and secretion rates are not identifiable from one output unless additional are added. Continued incorporation during the chase can occur if label-capable precursor is . Long half-lives are weakly constrained when little decline occurs within the observation . Proteins that disappear below the mass-spectrometry limit should be treated as potentially . A profile likelihood can reveal whether a kinetic parameter is practically . A design that samples neurons, microglia and extracellular compartments can separate degradation from intercellular .

Open-Ended Tasks

Easy

1. Measurement model sketch: Draw a compartment diagram containing labelled neurons, extracellular space, microglia and free AzF, and label every flux that could change the measured BONCAT signal.
2. Half-life window calculation: Calculate the fraction remaining after 14 days for proteins with half-lives of 3, 14, 60 and 120 days, then explain which are easiest to estimate.
3. Paper figure audit: Select one turnover figure from the Guldner study and list every transformation between the biological sample and the plotted quantity.
4. Missingness map: Create a visual matrix that distinguishes proteins missing within a plex from proteins missing across entire plexes.

Standard

1. Replicate-level reanalysis: Refit several protein trajectories using replicate-level data and compare uncertainty with fits to the time-point means.
2. Model comparison notebook: Fit mono-exponential, biexponential and Weibull models to simulated trajectories and document when the models become indistinguishable over 14 days.
3. Background sensitivity experiment: Recalculate half-lives under fixed subtraction, fitted additive background and a higher hypothetical background floor.
4. Interview a proteomics researcher: Conduct a structured interview about detection limits, TMT missingness and how analysts decide whether an absent peptide is biological or technical.

Advanced

1. Structural identifiability analysis: Analyse the neuron–extracellular–microglia model and determine which parameter combinations are identifiable from regional signal alone and after adding compartment-resolved outputs.
2. MNAR simulation study: Simulate intensity-dependent dropout and compare complete-case, fixed-value imputation, censored likelihood and selection-model estimates of age-related half-life change.
3. Optimal sampling design: Use Fisher-information or simulation-based design to choose a limited set of chase times that discriminates degradation slowing from residual labelled input.
4. Competing-mechanism protocol: Write and preregister a complete in vivo follow-up protocol that measures precursor, neuronal, microglial, extracellular, aggregate and cell-number readouts with predefined decision rules.

Learning Assessment

1. Mechanistic interpretation: Given an aged trajectory with slower regional BONCAT decline and increased microglial labelled protein, explain at least three compatible mechanisms and identify the measurement needed to distinguish them.
2. Identifiability proof: Show mathematically why degradation and secretion cannot be estimated separately from a single intracellular first-order decay curve when both act only through their summed loss rate.
3. Uncertainty critique: Evaluate a protein whose fitted half-life is 80 days from a 14-day experiment and decide what additional interval or bound should accompany the point estimate.
4. Missing-data transfer: Compare how an age effect could change when late missing values are treated as censored observations rather than excluded complete cases.
5. Experimental redesign: Design a sampling schedule and set of orthogonal measurements that can separate true degradation slowing from label recycling.
6. Model robustness: Interpret a case in which exponential and Weibull models fit equally well within 14 days but predict very different 60-day persistence.
7. Reproducibility audit: Inspect a hypothetical analysis repository and determine whether another laboratory could regenerate its protein-level half-life table from raw inputs without manual intervention.

Evidence of Learning

Strong evidence of learning includes the ability to reconstruct the published Guldner turnover analysis from archived data, explain every term in an explicit state-and-observation model, distinguish effective disappearance from intracellular degradation, identify structurally confounded parameters, diagnose weak practical identifiability, produce uncertainty intervals rather than unsupported point estimates, compare non-exponential models, analyse missingness mechanisms, and design measurements that separate competing biological explanations.

A high-quality product includes a version-controlled analysis repository, data manifest, executable environment specification, scripted preprocessing, protein-level results table, identifiability diagnostics, sensitivity-analysis matrix, uncertainty plots, model-comparison figures, missingness audit, simulation-based design report and a concise statement of which mechanistic conclusions are supported by which observations.

Transfer is demonstrated when you can apply the same reasoning to other pulse–chase systems, such as stable-isotope proteomics, RNA decay, secretome turnover, organelle turnover or tracer studies in which observed disappearance is a mixture of biological transport and destruction.

OERs on the Topic

Useful open resources include the English Wikipedia articles on Protein turnover, Proteostasis, Pulse-chase analysis, Mass spectrometry, Ubiquitin, Proteasome and Microglia, together with the openly accessible ProteomeXchange dataset PXD056701.

Linked Learning Areas

Protein turnover Proteostasis Pulse-chase analysis BONCAT Mass spectrometry Systems biology Structural identifiability Missing data Neuroscience Aging