Expert-Coded Curve Shapes for Reading Immunosensors in Serum

Antibody biosensors, devices that report a binding event as a shift in an optical signal, now reach very low detection limits in clean buffer. A silicon ring resonator with chemically anchored antibodies detects SARS-CoV-2 spike protein at 100 fg/mL in 15 minutes (Ning et al.). These sensors are still read with one number: the final signal shift, after subtracting a reference channel. In real serum that number mixes true antigen capture with proteins that stick to the surface nonspecifically, and the reported limits have not been shown to survive the move from buffer to patient samples. The full time course of the signal, called a sensorgram, carries more than its endpoint. Specific capture and fouling rise with different curvature, saturate differently, and respond differently to a rinse. Experienced experimentalists read these shapes by eye. No instrument encodes what they see.

We propose to make that reading explicit and measurable. Thematic analysis is the qualitative method of building a vocabulary of recurring patterns, called codes, by close reading of material. We apply it to experts thinking aloud as they judge sensorgrams, and we keep a code only if independent coders apply it reliably. Each surviving code becomes a small detector trained on raw curves. The detector outputs, together with physical kinetic-fit parameters and a frozen neural embedding of the curve, feed a gradient-boosted tree classifier that decides whether target antigen is present. We run this on two attachment chemistries, randomly oriented amide coupling and upright Protein-G-oriented antibodies. That lets us also ask a scientific question: does controlling antibody orientation make the curve shape more informative about specific binding?

Core Claim

Our central claim is that, in serum, a classifier given curve-shape features detects antigen at lower concentrations than endpoint calibration at the same false-positive rate. These are features grounded in expert reading and checked for reliability. We also expect a large part of the gain to come from features that a standard one-to-one binding-model fit does not capture. Our second claim is that the gain is larger on oriented surfaces. Oriented antibodies give more uniform binding sites, and Wang et al. report 1.59 times more active sites with Protein G than with random attachment. With more uniform sites, specific capture should follow a cleaner, more distinctive shape against the less structured fouling signal. If the codes add nothing beyond kinetic-fit parameters, that is a clean negative result, and the proposal is built to report it.

Positioning Against Prior Work

The sensing side motivates the work. Ning et al. and Wang et al. build the covalent and oriented surface layers we reuse: silane and glutaraldehyde tethering on subwavelength-grating rings, and Protein G on a polydopamine layer. Both read concentration from univariate calibration with differential reference subtraction. Wang et al. note that their selectivity was checked only in buffer against isolated interfering proteins. That gap is exactly the serum regime we target. Marvi and Jafari explain why misoriented or adsorbed receptors degrade the signal. That mechanism predicts that orientation should change the curve shape, not only the amplitude.

The hybrid classification literature supplies the method and one important constraint. Several papers freeze or fine-tune a neural feature extractor and pass its features to a classical classifier. Reinders et al. used random forests on CNN patches with about 15 examples per class. Shayeed et al. reported expected calibration error below 0.009 from ConvNeXt embeddings with simple heads. Pang et al. used XGBoost on 1D-CNN protein-sequence features. In each case, the shift to a classical classifier helped with scarce data or calibration, and our setting has both problems: a few thousand curves, and a decision that needs calibrated confidence. None of these works touched physical binding kinetics. The constraint comes from MC-BEC. Adding dense text and radiology embeddings to a LightGBM model lowered near-term decompensation performance, an effect the authors call modality competition. We treat that as a live risk: a high-dimensional curve embedding could drown the few physically meaningful features. Accordingly, the embedding enters only as an ablation-controlled block.

Thematic analysis supplies the procedure for building the feature vocabulary. Dai et al. showed that a human-LLM coding loop can reach Cohen's κ of 0.81 to 0.87, against 0.77 between humans. Cohen's κ measures agreement between coders beyond what chance would produce. We borrow their loop only to code transcripts at scale. Yi et al. warn that machine themes rarely map one-to-one onto human codes and that evaluation in this area is often ungrounded. We answer that warning differently from prior thematic-analysis work. Each code must earn its place twice: once by inter-coder reliability, and once by measured predictive value against known spiked concentrations. Prior work stopped at text themes. LieXBerta fed predetermined emotion categories into XGBoost, whereas our categories are derived from the data and tied to a physical ground truth.

Research Program

The data come from a wet-lab campaign. Each recognized antibody is immobilized on ring-resonator chips by one of two routes: random amide coupling (silanization, then glutaraldehyde or EDC/NHS) or Protein-G-oriented attachment. Multiple rings per chip include an uncoated reference ring. The sample set covers several conditions:

  • buffer;
  • pooled human serum, both blank and spiked with SARS-CoV-2 spike protein at concentrations bracketing and below the buffer limit of detection;
  • serum spiked with H1N1 antigen, which does not match the antibody.

Every run records association, a buffer rinse, and dissociation. Ground truth is the spiked concentration, so labels need no expert judgment. Splits are held out by chip and by fabrication day, so that no chip contributes to both training and test.

The qualitative stage runs in parallel. Four to six experimentalists from different labs review randomized, de-identified sensorgrams while thinking aloud. Two analysts code the transcripts inductively following reflexive thematic analysis, with an LLM second coder in the style of Dai et al. to scale coverage. We expect codes for phenomena such as signal that reverses on rinse, a slow linear creep after an initial jump, and drift on the reference ring that matches the sample ring. Next, the experts annotate a curve subset with the codebook. A code is kept only if its κ is at least 0.6. For each kept code, a small 1D convolutional detector is trained to predict it from the raw curve.

The final classifier is a gradient-boosted tree model with three input blocks:

  • code-detector probabilities;
  • one-to-one binding-model fit parameters (on-rate, off-rate, saturation amplitude, fit residual);
  • a frozen 1D-CNN embedding.

The CNN is pretrained with a self-supervised objective on all curves, including unlabeled ones. Each block can be switched off.

Baselines

The naive reference point is the protocol used in the source papers: reference-subtracted endpoint shift, read through a calibration curve built in serum. Its threshold is set so that 95% of blank-serum runs are classified as negative. The file reports this approach's performance in buffer only: LOD 100 fg/mL with about 15 pm non-matching shift for Ning et al., and LOD 5.39 U/mL for CA125 and 1.79 ng/mL for HE4 for Wang et al.. No reported result was found on Lacuna for any method on these sensors in serum. Our first deliverable is therefore this baseline's serum LOD measured on our own data.

The strongest physical competitor is a kinetic-fit model: a GBDT on binding-model fit parameters plus endpoint shift. It encodes the textbook physics and is the alternative most likely to match the full system. The strongest learned competitor is an end-to-end 1D CNN with a softmax output, temperature-scaled on validation data. A further hybrid baseline uses the CNN embedding with a GBDT and no codes; it isolates whether reliable expert codes add anything to learned features. Two ablations isolate the qualitative ingredient. The first replaces the thematic-analysis codebook with a feature checklist that one author writes before seeing any transcripts. The second keeps all codes regardless of κ.

All learned models use the same chip-held-out splits and the same budget of 50 hyperparameter trials. The kinetic fit receives an equal budget for its initialization and model-choice search, so the physics baseline is not under-tuned.

Beating a baseline means one of two things. Either we obtain a lower effective serum LOD at 95% specificity, with chip-level bootstrap confidence intervals that do not overlap, or we obtain equal LOD at lower calibration error. The two-chemistry comparison counts as a positive result if the improvement over the kinetic-fit model is significantly larger on oriented surfaces.

Why not just…?

Why not just fit a one-to-one binding model and classify on its parameters? This is the real threat. Many fouling signatures may already appear as a poor fit or an implausible off-rate. We include it as the main baseline. The decisive test is the improvement of the full model over the kinetic-fit model, along with SHAP attributions that reveal whether code features carry signal the fit parameters do not. If the improvement is negligible on both chemistries, we will report that expert-coded shape adds nothing beyond kinetics. The remaining contribution would then be the orientation-by-learnability analysis and the serum benchmark itself.

Why not just rely on the reference channel? Wang et al. show that differential subtraction handles thermal drift. A reference ring, however, fouls differently from an antibody-coated ring, because nonspecific adsorption depends on the surface. Subtraction therefore cannot remove fouling that occurs only on the antibody-coated ring. The endpoint baseline measures exactly how much serum degrades the reference-subtracted readout.

Why not train an end-to-end CNN? With a few thousand curves from a few dozen chips, prior hybrid work predicts overfitting and overconfidence (Reinders et al.; Shayeed et al.). That prediction is testable, because the temperature-scaled CNN is a baseline and we measure its calibration error on held-out chips.

Why not skip thematic analysis and have one expert list features? This ablation exists for that reason. Thematic analysis costs expert hours. It earns them only if multi-expert, reliability-filtered codes outperform an a-priori checklist or transfer better to the held-out chemistry. If they do not, the cheaper checklist should be recommended.

Decisive Checks

Three results decide the project. The first is the effective serum LOD at 95% specificity for the endpoint, kinetic-fit, and full models, computed separately for each chemistry on held-out chips and days. The second is whether code features keep their value when the classifier is trained on one chemistry and tested on the other, which tests whether they capture physics or chip-specific quirks. The third is H1N1 cross-reactivity in serum, scored as the false-positive rate on non-matching antigen. We also report AUROC, sensitivity at 95% specificity, and expected calibration error.

Block-wise ablations test for modality competition. If adding the CNN embedding to the codes and kinetic features worsens held-out performance, we will treat it as a reproduction of the MC-BEC effect and drop the embedding.