Pilot-Aided Kinetic Calibration for Receptor-Based Molecular Receivers

Wireless engineers know that a receiver designed around an exact channel model breaks when that model is slightly wrong. Mismatched channel knowledge leaves residual interference that no amount of extra signal power removes. Molecular communication is a setting where information travels as diffusing molecules that a cell-like receiver detects by binding them to surface receptors. It has the same weakness, but from a source the field has not yet treated as a channel-knowledge problem: the receiver's own chemistry. The best current channel-sensing scheme for such receivers estimates the concentrations of up to ten competing ligand types from how long a single receptor type stays bound and unbound. It does this only because it assumes the unbinding rate of every ligand is known exactly and hardwired into the estimator (Channel Sensing in Molecular Communications With Single Type of Ligand Receptors). Systems biology says such rate constants are rarely known and are highly consequential. Small errors in kinetic thresholds can flip the qualitative behavior of a model (Stability and reachability analysis of a hybrid model of luminescence in the marine bacterium Vibrio fischeri), and they drift with temperature, pH and ionic conditions.

This note proposes to treat uncertain receptor kinetics exactly as wireless treats imperfect channel state information (CSI), the receiver's estimate of how the medium distorts a signal. The receiver spends a fraction of its time observing "pilot" windows of known ligand composition, recalibrates its kinetic rates from them, and uses the recalibrated rates to sense unknown concentrations in the data windows that follow. The questions are when this pays off, how much pilot time it needs as a function of how fast the kinetics drift, and whether it beats the obvious alternative of fitting rates and concentrations jointly with no pilots at all.

Core Claim

Our first claim is that a receptor-based estimator that plugs in nominal unbinding rates has an error floor that longer observation cannot remove. This is the molecular analogue of the distortion ceiling that hardware impairments impose on relays, where outage saturates near 0.3 at impairment level 0.1 regardless of transmit power (Two-Way Relaying under the Presence of Relay Transceiver Hardware Impairments). The floor follows from the standard bias–variance split. If $\hat{c}$ is the estimated vector of ligand concentration ratios, $\Delta k$ is the error in the assumed unbinding rates, $b(\Delta k)$ is the resulting systematic bias, $\lambda$ is the rate of binding events, and $T$ is the observation time, then

$$ \mathrm{MSE}(T) \approx \|b(\Delta k)\|^2 + \frac{\mathrm{tr}\,\Sigma}{\lambda T}, $$

where $\Sigma$ is the per-event covariance of the known-kinetics estimator. The second term shrinks as more binding events accumulate; the first does not. Bias is largest when ligands have similar unbinding rates, because a small rate error then reassigns dwell times between ligand types.

Our second claim is that kinetic drift has a coherence time $\tau_k$, the interval over which the rates stay approximately constant. This gives the classic training trade-off. Spending a fraction $\rho$ of each coherence interval on pilots yields a total error of roughly

$$ \mathcal{E}(\rho) \approx \frac{A}{\rho\,\lambda \tau_k} + \frac{B}{(1-\rho)\,\lambda \tau_k}, $$

where $A$ measures how kinetic-estimation error propagates into concentration error and $B$ is the data-window variance. This has an interior optimum near $\rho^\star = \sqrt{A}/(\sqrt{A}+\sqrt{B})$. The prediction we test is that $\rho^\star$ grows as ligand rates become harder to distinguish, and that pilots lose their value when $\tau_k$ is long enough to calibrate offline once.

Research Program

The work has four parts.

The first part is analysis. We take the dwell-time model of the single-receptor channel-sensing paper and derive two things. One is the bias of its unbiased method-of-moments and maximum-likelihood estimators under misspecified rates. The other is the Cramér–Rao bound, the minimum variance any unbiased estimator can achieve, with unbinding rates treated as unknown nuisance parameters. Comparing that nuisance-parameter bound with the known-rate bound the original paper reports tells us how much information is lost by not knowing the chemistry. Where the two bounds diverge sharply, blind joint estimation is near-unidentifiable. This is the same structural non-identifiability that grey-box biological models hit (Current state and open problems in universal differential equations for systems biology).

The second part builds the pilot-aided estimator. Pilot windows with known composition give a maximum-likelihood estimate of the unbinding and binding rates. These enter the data-window estimator either as point values or as a posterior over rates that is integrated out. A simple random-walk model on the log-rates tracks drift across windows. Working in log-rates follows the positivity-preserving log-transformation used in the universal-differential-equation pipeline.

The third part attaches calibrated error bars. Following Conformal Robust Beamforming Via Generative Channel Models, we build concentration intervals by conformal prediction, a method that turns held-out residuals into intervals with guaranteed coverage. That paper builds guaranteed-coverage channel uncertainty sets from a held-out calibration set of 100 pairs. Here the residuals come from recent pilot windows. The open question is whether coverage survives drift, since conformal guarantees assume the calibration and test data are exchangeable.

The fourth part closes the loop to communication. The channel-sensing paper explicitly left the downstream effect of imperfect sensing on symbol error untested. We measure how the sensing error propagates into symbol error for concentration shift keying (information encoded in concentration level) and molecule shift keying (information encoded in which ligand type is sent).

Baselines

The naive reference point is the plug-in estimator. It runs the unbiased method-of-moments and maximum-likelihood sensing algorithms from Channel Sensing in Molecular Communications With Single Type of Ligand Receptors with nominal rates while the true rates are perturbed. It shows the cost of ignoring kinetic uncertainty and should exhibit the predicted floor.

The upper reference is the same estimator given the true rates (the oracle). That paper reports average NMSE "significantly below $10^{-2}$" for up to 10 ligand types, with its unbiased estimators close to the Cramér–Rao bound. We reproduce this regime first so our simulator matches theirs. No reported result was found on Lacuna for any estimator under uncertain rates, and none for end-to-end concentration-shift-keying symbol error with receptor-based sensing.

The strongest competitor is blind joint estimation. Bound dwell times from a mixture of ligands form a mixture of exponential distributions, so rates and mixture weights can be fit together by expectation–maximization. To avoid the local-minimum failures documented in the universal-differential-equation study, we run it with multi-start initialization, the remedy that study used, with up to 10,000 restarts. Blind estimation is the strongest competitor because it spends no time on pilots. For fairness, every method sees the same total number of binding events per coherence interval. Blind estimation gets all of them, while pilot-aided gets $(1-\rho)$ of them for data. Both receive the same restart budget and the same log-rate parametrization.

A further baseline is one-shot offline calibration: rates are estimated once at the start and never updated. It separates the value of calibration from the value of tracking.

Two ablations isolate the key ingredient. The first uses pilots without the drift-tracking prior. The second uses pilot-aided point estimates without conformal intervals. Pilot-aided calibration counts as winning if it achieves lower NMSE than blind estimation at equal event budget in the low-rate-separation, short-coherence regime. It must also stay within a stated factor of the oracle, fixed before running, while plug-in does not.

Why not just…?

Why not just observe for longer with nominal rates? Under the bias decomposition above, longer windows shrink only the variance term. The first experiment plots NMSE against $T$ for fixed rate perturbations and should show a plateau. If no plateau appears at realistic perturbation sizes, the problem is not worth solving, and we would report that.

Why not just fit everything blind with expectation–maximization? This is the real threat. Exponential mixtures are hard to separate when rates are close, which the channel-sensing paper already hints at in the error floors of its simplified estimator for similar unbinding rates. But when rates are well separated, blind fitting may match pilot-aided estimation at zero pilot cost. The nuisance-parameter Cramér–Rao comparison and the head-to-head experiment will map where each wins. If blind estimation wins everywhere, the contribution reduces to the identifiability analysis, and we would state that plainly.

Why not just calibrate once in the lab? One-shot calibration works if kinetics do not drift on the timescale of operation. Sweeping the drift rate, and therefore the coherence time $\tau_k$, shows where it fails. We expect a crossover, not a uniform win.

Why not just widen the error bars, conformally or with Gaussian rate noise? Wider intervals admit uncertainty without reducing it. In the wireless setting, Gaussian-error robust designs both missed outage targets at low SNR and were over-conservative at high SNR (Conformal Robust Beamforming Via Generative Channel Models). We keep conformal intervals as a layer on top of calibration and measure both their coverage and their width with and without recalibration.

Proposed Experiments

All experiments use the synthetic multi-ligand receptor simulation of the channel-sensing paper. Bound and unbound dwell times are generated by exact stochastic simulation for 2 to 10 ligand types. We extend it with multiplicative perturbations of each unbinding and binding rate, varied in magnitude, and with slow correlated drift that mimics a temperature-dependent shift in rates.

The first decisive check is the floor: plug-in NMSE against observation time at several perturbation sizes and rate spacings. The second is the information loss: the nuisance-parameter Cramér–Rao bound versus the known-rate bound, which predicts where blind estimation fails. The third is the trade-off: NMSE of pilot-aided estimation against pilot fraction $\rho$, testing whether the empirical optimum tracks the predicted $\rho^\star$ and moves as predicted with rate spacing and coherence time. The fourth is the head-to-head comparison against blind expectation–maximization, one-shot calibration and the oracle at equal event budget. The fifth is conformal coverage under drift. The last is symbol error rate for concentration and molecule shift keying under each sensing method.

As a sanity bridge to the wireless side, the same pilot-fraction analysis can be posed on the 3GPP urban macrocell setup used by the conformal beamforming work. There, the mobility-driven coherence time plays the role of $\tau_k$. The core contribution, however, remains the molecular receiver, where the "channel" error originates in the receiver's own reaction kinetics.