Learning the Nuisance, Not the Signal: Few-Label Analysis of Label-Free Biosensor Kinetics

Label-free biosensors detect a molecule directly from the physical change its binding causes, such as a shift in refractive index, capacitance, or current, with no fluorescent tag or reporter enzyme. That makes them cheap and fast. It also means they respond to everything else that changes those quantities: proteins sticking where they should not, temperature drift, and the sensor material slowly dissolving. Separating the target's signal from this background is a learning problem. It is also a small-data problem, because every labeled run (a known analyte at a known concentration on a fabricated chip) costs bench time, reagents, and often a fresh chip. This note proposes using a cheap and plentiful kind of data that current pipelines ignore: runs with no analyte at all. Blanks, buffer-only runs at different pH values, and runs with serum but no target cost almost nothing to collect and need no labels. They contain exactly the drift, dissolution, and fouling that hides the target.

The core idea fits in two sentences. Each sensor trace is modeled as a binding curve with only two or three physical parameters, plus a nuisance term drawn from a low-dimensional subspace learned from many analyte-free runs. The physical parameters are then classified from a handful of labeled runs. Because the expensive labels are spent only on the few parameters that carry identity and concentration, and the hard-to-model background is learned from blanks, the method should stay accurate in exactly the cases where closed-form physics fails: basic pH, where the porous sensor dissolves, and biological fluids, where fouling dominates. It should also adapt to a newly fabricated chip from blanks plus a few calibration standards.

Where this sits

The closest work is Ward et al.. They identify and quantify three proteins (BSA, ovalbumin, avidin) on porous silicon (PSi) chips etched at three current densities. They reduce each run to a single equilibrium spectral shift per chip, then apply linear discriminant analysis and a linear support vector machine. Their work motivates this proposal in two ways. First, it shows that capture-agent-free arrays can discriminate proteins. Second, it reports that at pH 10 the dissolving PSi skeleton produces negative optical shifts that compete with avidin infiltration, and the linear models could not separate the two. Their pipeline also discards the time course of infiltration. Our proposal keeps the full time course and treats dissolution as a nuisance learned from blank pH-10 runs.

The weak-value-amplification immunosensors of Guan et al. and Wang et al. record real-time binding. They handle drift in hardware with a self-reference differential channel, and quantify through a single linear calibration curve. They supply the simplest competing idea, subtracting a reference, which we take seriously below. Bi et al. show the opposite approach. An unconstrained one-dimensional convolutional network (1D-CNN) on raw voltammetric waveforms from a six-electrode tongue reached 94.0% vintage accuracy, beating hand-featured SVM (91.0%) and random forest (78.0%). They had 100 samples per class, from one production line, and did not test drift or transfer.

On the learning side, two kinds of prior work constrain the design. Dong et al. show that bounding a network with a deterministic geometric model preserves plausible output on only 45 training volumes. We reuse that structure, with a physical binding model playing the role of their deformable contour. Jia et al. found that recombining decomposed signal components hurt a raw time-series model by breaking phase continuity. Limitations of Synthetic Data Generation in Specialized Data-Scarce Domains warns that ungrounded generators invent features. Together these argue against generative augmentation of binding curves, which is why we use real recorded blanks rather than synthesized ones. Chen et al. found plain logistic regression beat RF, XGBoost, SVM, and k-NN on 80 breath samples, a reminder that low-capacity models win at these sample sizes. The proposal therefore keeps the labeled part of the model nearly linear.

Core Claim

For a sensor channel $c$ (one pore size or one electrode), the measured trace $y_c(t)$ is modeled as

$$ y_c(t) = A_c\left(1 - e^{-k_c t}\right) + \sum_{j=1}^{J} b_{cj}\,\phi_{cj}(t) + \varepsilon(t). $$

Here $A_c$ is the equilibrium binding amplitude and $k_c$ is the observed association rate, a pseudo-first-order (Langmuir) term whose rate depends on concentration and molecule size. The $\phi_{cj}$ are $J$ nuisance basis functions for that channel, learned from analyte-free runs. The $b_{cj}$ are run-specific nuisance weights, and $\varepsilon$ is measurement noise.

The fit penalizes overlap between the binding term and the nuisance span, so the nuisance cannot absorb the signal. Identity and concentration are predicted from the vector of $(A_c, k_c)$ across channels using a regularized linear classifier and regressor.

Run-specific weights $b_{cj}$ are what distinguish this from blank subtraction. Dissolution rate and fouling level vary from run to run, so a fixed subtracted blank leaves residue, while a subspace can track it.

The claim is falsifiable in a specific way. In clean pH-4 buffer, a plain kinetic fit should do as well as our model. The advantage should appear only, and grow, as the nuisance becomes stronger and more variable: pH 10, mixtures, diluted serum, and new fabrication batches.

Research Program

The build has three parts. The first is the data. We replicate the PSi protocol of Ward et al.: chips etched at 25, 40, and 55 mA cm⁻², the same three proteins, concentrations of 0.02, 0.2, and 2 g/L, and pH 4 and 10. We add time-resolved reflectance with phase tracked by their wavelet method, two-protein mixtures, and spiking into diluted serum. We fabricate at least four chip batches so that chip-to-chip variation is real rather than simulated. For every batch we record a large pool of blanks: buffer at both pH values, serum-only runs, and deliberate temperature steps. These cost a fraction of labeled runs.

The second part is the nuisance model. The baseline version is a per-channel principal subspace of blank traces, with $J$ chosen by held-out reconstruction error on blanks. A small nonlinear autoencoder is a variant to try only if the linear subspace leaves structured residuals.

The third part is the few-shot cross-batch procedure. On a new batch we re-estimate only the nuisance subspace from that batch's blanks, plus a per-channel gain from $K \le 5$ calibration standards. The labeled classifier is not retrained.

The analyses ask three questions. How does accuracy scale with the number of labeled runs per class (3 to 20) at a fixed blank budget? How does it scale with the blank budget at a fixed label count? How early in the transient can a decision be made? Kinetic fits allow pre-equilibrium calls, which endpoint methods cannot make. As a secondary, smaller study, we will apply the same decomposition to real-time weak-value traces from a CA125/HE4 sensor of the Wang et al. design run in spiked serum. Its reference channel then serves as an in-hardware baseline.

Baselines

The naive reference point is the Ward pipeline as published: equilibrium shift per channel, LDA, then a linear SVM, trained on the same labeled runs. No reported result was found on Lacuna for its leave-one-out or holdout accuracy, so we re-run it on our data under our splits.

The strongest physics-only competitor is a nonlinear least-squares fit of the same binding term without any nuisance subspace, with only a constant baseline. It feeds $(A_c, k_c)$ into the same regularized linear classifier. This isolates whether the learned nuisance adds anything.

The strongest hardware-style competitor is blank or reference subtraction. We subtract the mean blank of the matching batch, pH, and matrix, in the spirit of the self-reference channel of Wang et al., then apply the kinetic fit.

The strongest learned competitor is a 1D-CNN on raw multi-channel traces with the architecture class of Bi et al.. Their reported figure is 94.0% test accuracy against 91.0% for SVM, on 100 samples per class. That is a different task and a larger label budget, so it is context rather than a target number. We run this network twice: once plain, and once augmented by adding randomly scaled real blank traces to labeled traces. The augmented version gives the network access to the same blank pool our method uses, which makes it the fairest test of whether the structure matters or only the blanks do.

All methods see identical labeled and blank pools and identical batch-held-out splits. The CNN receives a tuning budget of at least as many configurations as our model's few hyperparameters ($J$ and the overlap penalty), selected on validation batches only.

Beating these baselines means three things. Identity accuracy and log-concentration error at pH 10 and in serum must be better than both the physics-only fit and blank subtraction, with non-overlapping confidence intervals across batch-held-out folds. On a new batch with $K \le 5$ standards, accuracy must be within a few points of training on that batch's full labeled set. The method must match or beat the blank-augmented CNN at 3 to 10 labeled runs per class. Two ablations isolate the key ingredient: replacing the learned subspace with a generic polynomial drift basis, and removing the overlap penalty.

Why not just…?

Why not just fit Langmuir kinetics directly? At pH 4 in clean buffer this may be all one needs, and we predict it ties with our method there. But Ward et al. report that dissolution at pH 10 creates competing shifts that equilibrium models could not separate. Bi et al. describe composite electrochemical signals that closed forms do not capture. The physics-only baseline measures the gap directly. If it closes the gap, the nuisance model is unnecessary and we will say so.

Why not just subtract a blank or reference channel? This is the most serious threat. It works in the weak-value sensors because their drift is largely shared between channels. Subtraction removes one fixed nuisance shape at fixed amplitude, while dissolution and fouling vary per run. The pH-10 and serum conditions test whether run-specific weights matter. If subtraction matches our model there, the contribution reduces to "record better blanks," which is still useful but much smaller.

Why not just train a CNN, adding blanks as augmentation? This is a real competitor because it uses the same cheap data. Prior small-sample evidence (Chen et al., Dong et al.) favors low-capacity or physically bounded models at tens of samples, which suggests the CNN will fall behind at 3 to 10 labels per class and catch up at 20 or more. The label-scaling curve will show where the crossover lies. A CNN that wins everywhere would refute the structural claim.

Why not just use high-affinity capture antibodies? Antibodies reduce but do not remove fouling. They also add cost, limit shelf life, and restrict detection to targets with an existing antibody, as Ward et al. note. The secondary weak-value study tests the antibody case directly: if our decomposition adds nothing over the reference channel on an antibody sensor in serum, the method's value is confined to capture-free arrays.

Decisive Checks

The experiment that settles the proposal is a two-way grid. One axis is nuisance severity: pH 4 buffer, pH 10 buffer, mixtures, then serum. The other is labeled runs per class: 3, 5, 10, 20. Results are evaluated with whole fabrication batches held out. The predicted signature has three parts. All kinetic methods should tie at pH 4. Our method should pull away from the physics-only fit and blank subtraction as severity rises. It should beat the blank-augmented CNN at low label counts, with the CNN converging only at the high end. A second check is cross-batch transfer with $K \in \{1, 3, 5\}$ standards, compared against fully retraining every baseline on the new batch. A third check plots decision accuracy against the fraction of the transient observed, which tests whether kinetic parameters allow calls before equilibrium. If the first check shows no widening gap, the central claim is wrong.