Resonance-Measured Concepts for Auditable Cell Classifiers

Interpretable classifiers of the "concept bottleneck" kind first predict a few human-readable quantities, then make the final decision from those quantities alone. They are expensive because a person must label every concept on every training example. Physical sensors avoid that cost. A vibrating micro-cantilever that a cell touches or flows past shifts its resonant frequency in proportion to the cell's added mass, and it loses sharpness in proportion to how much energy the cell dissipates. Those two numbers are concepts that come with units, need no annotator, and a biologist already understands. We propose to build single-cell classifiers whose bottleneck is made of resonance-measured physical quantities. The central question is whether the usual accuracy cost of interpretability in this setting is mostly a measurement-noise problem that mechanical resonance amplification can remove.

The claim is concrete and falsifiable. When a sensor runs off resonance, the physical concepts are noisy. A black-box network reading the raw sensor trace should then beat the interpretable model, partly by exploiting nuisance cues such as flow position or chip identity. As amplification rises, the gap should shrink, and under a change of chip or flow condition the interpretable model should win outright. If this holds, then in physical sensing "how interpretable can we afford to be" becomes an engineering quantity you can tune, not a fixed trade-off. The outcome is also a tool a bench biologist can audit in physical units, which is the accessibility goal that current platforms leave unmet.

Core Claim

The proposal builds on prior work in four ways. Concept Bottleneck Models route predictions through labeled concepts and let experts correct them at test time. Its authors report a joint model at 19.9% error on bird-species classification against 17.5% for an unconstrained network. Under shifted image backgrounds, the unconstrained network's error rose to 62.7% while the bottleneck model's was roughly 48–49%. That paper motivates the design and also names its main weakness: dense concept annotation, and accuracy loss when the concept vocabulary is incomplete. We keep the bottleneck but replace annotation with measurement.

Resonance supplies the measurement. Towards Optimal Kinetic Energy Harvesting for the Batteryless IoT tunes a piezoelectric cantilever's tip mass to its excitation frequency and reports a roughly 100x power gain over off-resonance operation. High quality factor silicon cantilever driven by PZT actuator for resonant based mass detection uses frequency shifts of a high-sharpness cantilever to detect small biological masses. The first paper gives us the amplification lever; the second gives the sensing principle. Neither connects the signal to a classifier.

For a resonator of effective mass $m$, a small added mass $\delta m$ shifts the natural frequency $f_0$ by

$$\frac{\delta f}{f_0} \approx -\frac{1}{2}\frac{\delta m}{m},$$

which follows from $f_0 \propto \sqrt{k/m}$ for stiffness $k$. The standard estimate of how precisely $f_0$ can be read is roughly the resonance linewidth $f_0/Q$ divided by the signal-to-noise ratio. Here $Q$, the quality factor, measures how sharp the resonance peak is. So every measured concept carries an uncertainty predicted from physics, and resonance amplification shrinks that uncertainty directly. The component that is new relative to both lines of work is a bottleneck whose concepts are supervised by physical fits and whose concept uncertainties are set by the resonator's measured $Q$ and noise floor, not learned freely.

Research Program

The work has three parts.

The sensing task. The biological task is taken from IMPEDANCE-MASS DUAL FLOW CYTOMETRY FOR MULTI-MODAL SINGLE CELL ANALYSIS. That work pushed mouse colon cancer cells (MC-38), with and without the growth factor TGFB, through a passive constriction channel with impedance electrodes. TGFB treatment induces a shift to a more invasive, mesenchymal cell state. The authors reported a 15.60% drop in stiffness and a 66.43% rise in fluidity, but performed no predictive modeling. Their passive channel is our "no amplification" arm. We add a PZT-driven cantilever resonator downstream, tuned to its drive frequency by tip-mass calibration in the manner of Sandhu et al. The resonator records per-cell frequency-response traces. From each trace a physics fit yields frequency shift (converted to buoyant mass), change in $Q$ (dissipation), transit time, and the impedance peak and deformability valley already produced by the Zhu et al. design. We collect data at controlled detunings, from on-resonance to well off it. Detuning is the knob that sets concept noise.

A simulation benchmark first. Hardware is slow and the true concepts are unknown, so we start with a lumped mass-spring-damper simulation. Synthetic cells with sampled mass and viscoelasticity pass the resonator, with nuisance variables including flow-position jitter, flow-speed drift, and chip-to-chip stiffness spread. Class labels depend mostly on mechanics plus a controlled non-mechanical component, so that concept incompleteness is known exactly. Autonomous Sensors Powered by Energy Harvesting from von Karman Vortices in Airflow is a warning here. That paper's simulated Strouhal number (0.23) differed from its measured value (0.17), and its harvester detuned sharply once flow passed about 4 m/s. We therefore treat the simulation as a hypothesis generator, recalibrate its damping and detuning against bench sweeps, and base claims only on hardware results.

The model. A small network $g$ maps the raw trace to concept estimates and their variances. The variances are fixed by the physics-derived noise floor for the current $Q$ and signal-to-noise ratio. The network is trained against the physics fits, not human labels. A logistic head $f$ then predicts cell state from concepts in physical units, and it accounts for measurement error in its inputs by down-weighting noisy concepts. Two things make $g$ necessary rather than decorative. Fitting a resonance curve per cell is too slow for in-line use, whereas a tiny network runs within the microsecond-to-millisecond budget of flow. This addresses the latency mismatch that makes Deep generative classification of blood cell morphology, at about 1.8 s per image, unusable in-line. Second, test-time intervention becomes literal: a user can replace a suspect predicted concept with the full physics fit or a repeated measurement.

The tool. We package the model as a point-and-click analysis tool in the spirit of PIVOT and WebMeV, both of which wrap statistical pipelines so non-programmers can run them. We also expose it as a callable tool for agent systems like CellAgent. CellAgent runs sequencing workflows from natural language, but its toolkit has no biophysical signal modules and it executes black-box models. The tool shows each cell's concepts in grams and dissipation units, lets the user override a concept, and reports the head's coefficients as odds changes per physical unit.

Baselines

Naive reference. A single threshold on transit time or impedance amplitude, tuned on the training set. This is the reading a biologist would make from the Zhu et al. passive channel. It tells us how much any learning adds.

Classical biophysics pipeline. Physics-fitted concepts fed to logistic regression or gradient-boosted trees, with no learned $g$ and no uncertainty handling. This is the most dangerous interpretable competitor, because it is effectively a sequential bottleneck with a hand-made concept extractor.

Strongest accuracy competitor. A 1D convolutional or transformer network on the raw resonator and impedance traces, with post-hoc attribution for explanation. Its sensor-domain analogue in the literature is the interactive time-series detectors surveyed by InTimeAD.

Fair comparison. All learned models get identical data splits, the same hyperparameter search budget, and matched parameter counts where possible. Splits are held out by chip and by day, not by cell.

Reported numbers. No reported result was found on Lacuna for single-cell classification accuracy on TGFB-treated versus untreated MC-38 cells, nor for any resonant-sensor cell classifier. The nearest reported comparison is the bird-species gap above: 19.9% versus 17.5% error in distribution, and 48–49% versus 62.7% under background shift, from Concept Bottleneck Models.

What counts as beating the black box. Held-out-chip AUROC within 2 points of the black box in distribution, and higher under chip or flow shift.

Ablations. Three ablations isolate the key ingredients:

  • a bottleneck with learned rather than physics-set concept variances;
  • a joint bottleneck with an extra unconstrained residual channel, which measures how much label information escapes the physical concepts;
  • the full model using only passive-channel concepts, which isolates what resonance adds.

A prototype-matching variant after This Looks Like That ("this trace looks like that training trace") is a secondary interpretable arm. We watch it for the same shortcut failure its authors reported with image backgrounds.

Why not just…?

Why not just fit the resonance curves and run logistic regression? This alternative could genuinely match our accuracy. If it does, the learned extractor adds only speed and calibrated uncertainty. We report that outcome directly rather than treating it as a loss. The study's main claim, that amplification controls the size of the interpretability gap, holds regardless of which interpretable model sits on the concepts. Two measurements will decide whether $g$ earns its place: per-cell latency, and accuracy at low signal-to-noise, where individual curve fits become unstable.

Why not just train a black box on raw traces and explain it with attribution maps? Attribution maps cannot be edited, and they can point at plausible regions while the model relies on shortcuts. This Looks Like That documents this, and the background-shift result in Concept Bottleneck Models shows how large the resulting cost under shift can be. In our setting the likely shortcuts are chip identity and flow position. The held-out-chip split and the planted-fault user study below test directly whether attribution lets users catch those shortcuts.

Why not just use passive deformability cytometry, which already measures stiffness? It may suffice. The Zhu et al. stiffness and fluidity shifts are large. The passive-only ablation measures exactly what resonance adds. If passive concepts already close the gap, the amplification hypothesis is refuted for this task, and we would recommend the simpler device.

Why not just add a signal module to an agent system? An agent that runs a black-box classifier inherits that classifier's opacity. The agent evaluates results by inspecting plots, not by exposing auditable decision variables. Our tool is meant to be one such module, not a replacement for the agent.

Proposed Experiments and Decisive Checks

Simulation gap curve. In simulation, we plot the AUROC gap between the black box and the bottleneck as a function of $Q$ and detuning, at several levels of non-mechanical label content. The prediction is that the gap shrinks steadily with amplification when labels are mostly mechanical, and levels off at a floor set by the non-mechanical share. A floor that does not track that share would signal a modeling error.

Hardware gap curve. On the bench we repeat the curve with MC-38 cells, with and without TGFB, at on-resonance and detuned settings, holding out chips and days.

Kill criterion. The claim fails if, at best tuning, the black box still leads by more than 5 AUROC points and the residual-channel ablation shows most of the label information bypassing the physical concepts. That result would mean mechanics does not adequately describe this cell-state shift.

Planted-fault user study. Following CellAgent's human-in-the-loop design of 10 experts and 10 novices, we recruit biologists without machine-learning training. We inject sensor faults such as a drifted frequency baseline or a clogged transit. Participants use either the concept panel or the attribution panel, and we measure how often they detect the fault and how often their interventions correct the prediction.

Latency and power. We measure the extractor's per-cell latency and energy on a microcontroller-class device. The target is the microwatt-scale budget that Sandhu et al. show resonance-tracking electronics must meet. This establishes whether the auditable model can run in-line.