XSci

Kinetic Proofreading Activation Times Beyond the Weibull Limit

Thomas Mourier, Cameron Hopkinson, Kenny Wong

Submitted by Thomas Mourier

Published November 4, 2026 · Version v1, November 4, 2026 · DOI 10.66977/xsci.2611.000i

Computational Biology

Abstract

T cells can detect a few foreign peptide-major histocompatibility complexes (pMHCs) among a far larger number of self pMHCs. Kinetic proofreading explains this discrimination: a receptor that has bound its ligand must complete nKP signaling steps, each at rate kp, before it activates, and dissociation at any step resets its progress. When many receptors are engaged at once, the time to the first activation has been approximated by a Weibull law taken from the short-time limit of this process. The Weibull law is convenient, but it does not contain the dissociation rate koff, even though koff is what proofreading uses to tell ligands apart. Using the exact first-passage law of the underlying continuous-time Markov chain together with stochastic simulations, we show that the relative error of the Weibull law is set by a single dimensionless number ΛΛ, the Weibull-predicted activation time measured in units of nKP + 1 bound-state lifetimes. For ΛΛ < 1 the first receptor to activate does so before its ligand has ever dissociated, and the Weibull law holds. As ΛΛ grows, activation relies increasingly on repeated unbinding and rebinding. The true activation time then exceeds the Weibull prediction by up to 2900-fold on our parameter grid, and the effective Weibull shape falls to one, so activation becomes memoryless instead of timer-like. We derive a closed form that keeps koff by treating the successful pass through the proofreading chain separately from the time lost to failed attempts. It matches exact medians to within 2% in 19 of 25 parameter sets, with a worst case of about 6%, and reproduces the loss of timer-like behavior. Because ΛΛ is computed from the Weibull prediction itself, it shows in advance how far that formula can be trusted; where it cannot, the new expression requires only a one-dimensional root find.

View PDF

Review reports

Reviewer3 (AI review)

November 4, 2026

AI Review (reviewer3.com)

Overview

Weibull Limit Validity Criterion

Contribution

The authors establish a single dimensionless parameter, Λ, that predicts the relative error of the Weibull short-time approximation by comparing activation times against bound-state lifetimes.

Off-Rate-Dependent Activation Formula

Contribution

The study derives a closed-form expression for first-passage activation times that explicitly retains the ligand dissociation rate by separating initial uninterrupted attempts from subsequent retry cycles.

Transition to Memoryless Dynamics

Contribution

The authors demonstrate that as dissociation and rebinding dominate over initial attempts, kinetic proofreading activation shifts from a sharp timer-like distribution to a memoryless exponential process.

Comments (3)

1. Ambiguous units in the definition of Λ — Major

so the relative error of the Eq. (1) quantile is Λ = (kp + koff)tq / (nKP + 1) = (1 + ω) × tq / (nKP + 1), (2) with tq in units of 1/kp as above.

In Equation (2), the first expression writes (k_p + k_off)t_q / (n_KP + 1), which requires dimensional time t_q, whereas the second expression writes (1 + ω) × t_q / (n_KP + 1) and the text states that t_q is in units of 1/k_p. Using the same symbol t_q for both dimensional time and dimensionless time (k_p t_q) in the same equation is inconsistent and could cause a factor-of-k_p error when applying the diagnostic whenever k_p ≠ 1 s^-1 (such as in Section 6 where k_p = 1.25 s^-1).

2. Condition λ < μ in Eq. (3) is not universally guaranteed — Major

The single assumption we add is that the geometric sum W is exponential apart from its atom at the origin, W ≈ σ δ 0 + (1 − σ ) Exp(

Equation (3) states that λ = σ/τ_c < μ without qualification, and Eq. (4) relies on 1 - λ/μ > 0 in the prefactor (1 - λ/μ)^(-n_KP). However, as ω -> 0 (so σ -> 1 and τ_c -> 1/k_1 + n_KP/(2 k_p)) and k_1/k_p >> 1 (for instance, k_1/k_p = 10^4 as considered in Table 1 and Section 6), λ approaches 2 k_p / n_KP, which exceeds μ ≈ k_p whenever n_KP < 2 (and approaches μ at n_KP = 2); the parameter condition required for λ < μ to hold should be stated explicitly.

3. Upper bound t_q^U overlooks the first-attempt term — Major

nine of the 25 cells. An upper bound follows from writing the retry term of Eq. (5) as

The derivation of the upper bound t_q <= t_q^U applies Jensen's inequality exclusively to the retry term of Eq. (5) while omitting the positive first-attempt term σ P(n_KP, μt). While t_q <= t_q^U still holds because σ P(n_KP, μt) >= 0, this makes the upper bracket needlessly loose by τ_c / (1 - (1 - q)^(1/n)) in the retry-dominated regime; retaining σ P(n_KP, μt) -> σ in the Jensen bound yields the much tighter bound n_KP/μ + τ_c [(1 - (1 - q)^(1/n))/σ - 1], matching t_q^E in Eq. (7) whenever the retry term is positive.

XSci (AI review)

November 4, 2026

Summary

The manuscript asks when the Weibull law for the first activation time among n engaged TCR/pMHC pairs — the large-n limit of the short-time behaviour of a McKeithan proofreading chain — can be used. Because that limit keeps only the leading term of the single-pair first-passage law, it contains k_p and n_KP but not k_off, the quantity proofreading discriminates on. The authors show that the leading-order relative error of the Weibull quantile is one dimensionless group, Λ = (k_p + k_off)t_q/(n_KP + 1), and that on a 5 × 5 grid in (n_KP, ω = k_off/k_p) the true median exceeds the Weibull prediction by up to 2.9 × 10^3 while the effective shape falls from n_KP to 1. Splitting the first-passage time into the successful pass and the time lost to failed cycles, and replacing the geometric sum of failures by its exponential limit, they obtain a closed form, Eq. (5), accurate to 2% in 19 of 25 cells.

Contribution and Significance

The core contribution is a usable domain-of-validity criterion for an asymptotic formula, with a closed form that follows the activation time across the crossover that criterion identifies. Λ is attractive because it is computed from the Weibull prediction itself, so a user can decide in advance whether to trust that prediction; and the closed form yields an explicit trigger rate, κ = σ/τ_c with σ = (1 + ω)^{-n_KP} [Eq. (8)], which makes the retry-dominated regime interpretable rather than merely numerical. The derivations are correct: I rederived the expansion behind Eq. (2), τ_c, Eqs. (4), (5), (7) and (8), and the bracket [t_q^W, t_q^U], and the reported numbers reproduce.

What is genuinely new is narrower than the framing suggests. The chain is McKeithan's [12]; the extreme statistic and the Weibull limit are Morgan and Lindsay's [17]; the deterministic-versus-exponential dichotomy of completion times is Bel, Munsky and Nemenman's [18, 19]; fast-versus-slow criteria for extreme first-passage times are Lawley's [22, 23]. The authors say as much in Sec. 1. The advance is the specialisation of that programme to this chain: worth publishing, but it refines the reach of one prior result rather than changing what is known about discrimination.

Strengths

  • The analysis from Eq. (2) to Eq. (8) is correct at every step I verified, including the identity that the retry term of Eq. (5) equals σE[(t − G)^+]/τ_c and the result E[bound time per attempt] = (1 − σ)/k_off behind τ_c.
  • The quantile root of Eq. (5) is rigorously bracketed (Sec. 5): t_q^W is a lower bound because forward jumps are no faster than a rate-k_p Poisson clock, and t_q^U follows from Jensen's inequality.
  • Sec. 4 is a result in its own right, not a restatement of the scale error: Table 3 shows the effective shape falling from 5.4 to 1.0 across the n_KP = 10 row while Eq. (1) asserts 10.
  • The attempt-level sampler of Sec. 7 costs O(n_KP) draws per pair however many attempts it makes, reaching regimes where a Gillespie run would need some 3 × 10^8 events per replicate.
  • The numerical self-discipline is unusual: a derived R^2 ceiling of 0.999667 for an unfitted model (Sec. 7) and an analytic Monte Carlo standard error of 0.019/k on the median (Fig. 1).

Recommendation

Major Revision for Physical Biology. The technical core is sound and reusable, and Λ and κ are genuine contributions. But two of the three central claims are stated more strongly than the evidence supports: that Λ sets the Weibull error rests on a collapse partly forced by a bounded error measure, and that Eq. (5) is the replacement to use is never weighed against the exact law. Both are repairable within the manuscript's own approach, largely with evidence the authors have.

Major Concerns

  1. The collapse onto min(Λ, 1) is partly an artifact of the error measure (Sec. 3, Fig. 2). The error is E = 1 − t_Eq(1)/t_exact, and Sec. 5 proves t_q^W lies below the exact quantile, so E is confined to [0, 1) by construction. Every sweep point with Λ > 1 must then sit at E ≈ 1, and the sweep has many: at n_KP = 10, t_q/(n_KP + 1) = 0.158, so ω = 20 gives Λ = 3.3. The informative branch is Λ ≲ 0.3, and there the Fig. 2 caption concedes the measured error exceeds Λ "from under 2% for Λ < 0.05 to roughly 24% at Λ = 0.32". The quantitative content of the claim is thus the first-order Taylor coefficient of F(t), good to a few percent only while Λ is a few percent; the large-Λ agreement does not test it. The caption notes the ceiling; the abstract and Sec. 8 do not. Restating the result as the tolerance statement already in Sec. 8, and re-plotting Fig. 2 with an unbounded statistic such as ln(t_exact/t_Eq(1)), would settle it.

  2. No case is made for Eq. (5) over the exact law it approximates (Sec. 5; Abstract; Sec. 8). The exact object is a matrix exponential on n_KP + 2 ≤ 12 states, computed routinely by uniformization (Reproducibility), and Eq. (6) applies to it and to Eq. (5) alike, so both need the same one-dimensional root find for a quantile. Eq. (5) therefore appears to cost the same while carrying 0.01–6.12% error (Table 4); the abstract's "the new expression requires only a one-dimensional root find" implies a saving never demonstrated, and no timing comparison appears.

    Scope compounds this. The headline accuracy is one slice at n = 10^4 and k_1/k_p = 1; the wider 1080-set sweep, where the worst error rises to 12.5%, is a single unillustrated sentence. And one structural choice is never varied: all pairs start in C_0, although Sec. 2 states that pairs starting unbound "raises the Weibull shape to n_KP + 1" — the baseline Table 3 measures the collapse against — and a contact begins with unbound receptors. What would resolve this: a cost and accuracy comparison against uniformization, the wider sweep shown, and the unbound-start case reported.

  3. The motivating failure is one the manuscript constructs, and the biological margin is fragile (Sec. 1, Sec. 6). Both sections say so, Sec. 6 in these words: Ref. [17] "invokes Eq. (1) only to identify the n → ∞ limit, so the accuracy reported there is unaffected by what follows". No other published analysis using Eq. (1) outside its window is identified. The consequences advertised — Γ_max = 0.500018, self pMHCs triggering "in under two seconds", an energetic cost understated by "about 3000" — are therefore the cost of a substitution the authors themselves introduce, yet the abstract and Sec. 1 present them without that qualification. The positive counterpart is weaker than stated too: "roughly an hour" for self at n_KP = 10, ω = 5 holds at k_1 = k_p, while the authors' own sweep gives 11 min at k_1 = 10^4 k_p, the rebinding rate they attribute to Ref. [17] for self — well inside a phase-II conjugate, so the margin largely disappears at the fast end. Sec. 6 should be reframed as a prospective caution, the hour quoted with its k_1 dependence, and the chance-level framing dropped from the abstract.

Minor Comments

  1. In Eq. (7), t_q^E = n_KP/μ + τ_c(L/(nσ) − 1) mixes orders: the exponential-limit quantile is n_KP/μ + τ_c L/(nσ), and the extra −τ_c is the mean E[M − 1], which is what drives t_q^E negative in nine of the 25 cells.
  2. Sec. 3 states that at n_KP = 2 "ω shifts the median by only 1.6% across the whole row", but that row of Table 2 spans 0.0118–0.0121 non-monotonically; printing the exact grid beside the simulated one would remove the appeal to Monte Carlo scatter.
  3. The abstract and Sec. 8 should quote the 12.5% worst error of the 1080-set sweep beside the 2%/19-of-25 figure from the 5 × 5 grid.
  4. The per-pair rebinding rates attributed to Ref. [17] in Sec. 6 ("1 s^-1 for the agonist and 10^4 s^-1 for the self population") are asserted rather than derived from that model's coupled binding kinetics, yet the 67 min / 11 min contrast turns on them.
  5. A relevant reference on rebinding is absent: O. Dushek, R. Das and D. Coombs, "A role for rebinding in rapid and reliable T cell responses to antigen", PLoS Comput. Biol. 5(11):e1000578, 2009, which treats a proofreading chain where rebinding partially retains progress — the alternative to the complete reset assumed here.
  6. A Kolmogorov–Smirnov statistic would test the simulator against the exact law more sharply than the R^2 of Table 6.
  7. The departure of Ref. [17]'s coupled-pair law from the independent-pair reduction (Sec. 2) is acknowledged but never quantified.
  8. Table 5 places genuine self ligands at ω = 5.6–28, outside the grid's ω ≤ 5; one or two extra columns would let Sec. 6's claims be read off the grid directly.
  9. Notation needs unifying: n_KPR and τ_KPR in Sec. 6 and Table 5 against n_KP elsewhere; "Eq. (1)'s median scales as n^{-1}/n_KP" in Sec. 6 should read n^{-1/n_KP}; and "T_{1,n} → exp(nκ)" in Sec. 5 should be Exp(nκ).