Cancer cells that turn invasive often pass through epithelial-mesenchymal transition (EMT). In this program, tightly joined sheet-forming cells loosen their junctions, change shape and start to crawl. Drugs that reverse it are judged by slow, destructive assays: a scratch in a cell layer is watched for a day to see whether it closes, or cells are fixed and stained for junction proteins such as E-cadherin and mesenchymal proteins such as vimentin. Before a mesenchymal cell moves anywhere, though, its edge is restless. It extends and retracts thin sheet-like protrusions on a timescale of seconds. An epithelial cell held by its neighbours is much quieter. We propose to measure that restlessness directly with an event camera, a sensor whose pixels report only when brightness changes and stay silent otherwise. The sensor's output is passed through a bank of resonate-and-fire neurons, spiking units that act as frequency-tuned filters and need only additions. The result is a per-well "fluctuation spectrum" of cell-edge activity, computed at a cost proportional to how much the cells move.
The core claim is biological and testable. Fast edge-fluctuation spectra, sampled in short bursts, separate epithelial, intermediate and mesenchymal states. They also detect an EMT inhibitor's effect hours before scratch closure or cell-track statistics do. The neuromorphic side is not the headline. It is a constraint: the readout must be cheap enough to run continuously on every well of a plate inside an incubator. That constraint makes event-proportional sensing and computation the natural design, and the energy budget is measured as an outcome rather than assumed.
What exists and where this sits
The closest work is Neuromorphic-enabled video-activated cell sorting. It pairs an event camera with spiking networks on a Lynxi HP201 chip to sort suspended cells flowing past at 1,000 cells/s. Its multi-view event sequences cut red-blood-cell versus spherocyte error from 19.93% (single frame) to 0.99%. That paper motivates us: temporal event information beat snapshots. Our setting differs in three ways. Its cells are detached and classified by gross three-dimensional shape during microsecond transit. Ours are adherent and nearly stationary, and are classified by the spectrum of their edge fluctuations. It does no drug-response measurement, which is our whole purpose.
EAP4EMSIG is "event-driven" only in the software sense. It runs closed-loop microscopy automation on ordinary frame cameras for bacteria. It constrains our design in a useful way: its shallow MLP autofocus, with mean absolute error 0.105 µm in 87 ms, outperformed foundation models. That warns against heavy learned components where cheap features suffice, and we adopt its autofocus approach to suppress drift-induced events.
Two papers supply the strongest motion-analysis competitors. TRACKASTRA is a transformer tracker that links cells from coordinates, area and intensity alone, reaching an error score (acyclic oriented graph matching, AOGM) of 7.9 on DynamicNuclearNet against 18.1 for Caliban. The Cell Behavior Video Classification Challenge found that trajectory-based classifiers beat end-to-end video networks on small live-cell datasets; the winning trajectory 1D CNN with attention scored 0.922 with AUC 0.944. Both papers say slow trajectories carry most of the usable signal at the frame rates they studied. Our claim is that a separate, faster signal exists that trajectories at those rates cannot see. Such a signal would have to exist for this proposal to matter.
On the biology side, Intermediate cell states in epithelial-to-mesenchymal transition argues that EMT is a continuum, so a continuous live readout has something to report beyond binary labels. Two studies show endpoint molecular assays can mislead. In Systems analysis reveals a transcriptional reversal of the mesenchymal phenotype induced by SNAIL-inhibitor GN-25, the inhibitor reversed the mesenchymal program across 2,737 genes, yet SNAI1 mRNA rebounded. In Impact of Nischarin on EMT regulators in breast cancer cell lines, seven markers shifted at the mRNA level but only three were confirmed at the protein level. These results motivate a non-destructive, time-resolved phenotypic readout, and they supply our perturbation systems.
Core Claim
Each region of interest yields an event stream $e(t)$: the count of brightness-change events in a short time bin. We filter it with a bank of $K$ resonate-and-fire units with characteristic frequencies $\omega_k$ and decay rates $\gamma_k$. Unit $k$ holds a complex state
$$z_k(t+\Delta) = e^{(-\gamma_k + i\omega_k)\Delta}\, z_k(t) + e(t),$$
and emits a spike when its amplitude $|z_k|$ crosses a threshold. Here $\Delta$ is the bin width. Because $e(t)$ is a count, each update needs one addition per incoming event plus a fixed decay between events. Compute therefore scales with event count, not with pixels times frames. The spike rate of each unit gives a $K$-dimensional fluctuation spectrum per region. This is the resonate-and-fire mechanism of A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series. That paper uses heterogeneous resonators as a learned backbone for general time series. We repurpose them as an interpretable spectral front end whose frequency bands correspond to protrusion-retraction cycles.
The hypothesis is that mesenchymal cells shift spectral power upward and increase total edge activity relative to epithelial cells. We further expect effective EMT inhibitors to reverse this shift within hours, well before gap closure differs between conditions.
Research Program
We build an event-camera module on a standard incubator microscope, using the same EVK4-class sensor as the cell-sorting work. A synchronized frame camera shares the light path through a beam splitter. Every well therefore receives short event bursts (minutes, repeated hourly) and, simultaneously, conventional low-rate time-lapse and high-frame-rate frame bursts. Thermal and illumination drift create spurious events, so we use fixed-cell wells and medium-only wells as noise floors. We also apply MLP autofocus in the style of EAP4EMSIG between bursts.
We use three perturbation systems from the file, chosen because each has a known direction of EMT change. The first is SNAIL-overexpressing HMLE cells treated with GN-25 versus vehicle, which tests reversal and the reported SNAI1 rebound. The second is Hs578T cells with and without Nischarin overexpression, plus MDA-MB-231 cells with and without NISCH knockdown, which tests bidirectional control. The third is hypoxic MRC-5 fibroblasts with and without 25 µM curcumin, which tests a fibrosis-type transition with reported E-cadherin and N-cadherin changes. Ground truth at each endpoint comes from per-cell immunostaining for E-cadherin and vimentin, scratch closure at 24 h, and Transwell invasion. Fixing a fraction of wells at intermediate time points gives staggered labels.
The analysis has three parts. First, we test whether spectra classify state and whether they track per-cell marker ratios as a continuous score for intermediate states. Second, we test whether spectra add information beyond slow trajectories, using a nested-model comparison. Third, we test how early a drug effect becomes detectable. We measure energy on neuromorphic hardware where available and with a wall-power meter on an embedded GPU. We report synaptic operations alongside measured joules. We do not import the cited orders-of-magnitude figures as if they applied here.
Baselines
The naive reference point is the assay biologists already run: scratch closure fraction and Transwell counts at 24 h, read from conventional time-lapse. It defines the time-to-call our method must beat. There is no reported result on Lacuna for time-to-detection of EMT inhibition by this or any other method, so we establish it in our own plates.
The strongest existing competitor is trajectory-based classification. Cells are tracked with Trackastra on the low-rate frame time-lapse, and a 1D CNN with attention, the CBVCC-winning design, classifies each track. Trajectory models won CBVCC precisely because they encode the right prior on small biological datasets. The closest reported numbers are CBVCC score 0.922 and AUC 0.944 on T-cell turning, from Cell Behavior Video Classification Challenge. That task is different from ours, so the number is context, not a target. We retrain this baseline on our labels with the same wells, splits, and hyperparameter search budget as our method. Beating it means a statistically higher AUC for EMT state at matched time points. Alternatively, it means a significantly earlier time point at which treated and untreated wells separate, with no loss of final accuracy.
The frame-camera fluctuation baseline uses high-frame-rate frame bursts acquired simultaneously from the same field. We compute frame differences thresholded to mimic events and pass them through the identical resonator bank and through an FFT. This isolates whether the sensor matters or only the timescale. To beat it, the event camera must match its accuracy at markedly lower data volume and compute, or exceed it at equal bandwidth.
The ablations separate the components. A fixed, hand-tuned resonator bank is compared with an FFT plus logistic regression on the same event counts, with a learned spiking network trained by surrogate gradients on raw event streams, and with a small dense CNN on accumulated event frames. If the FFT or the fixed bank matches the learned spiking network, we report that the spiking network is unnecessary. The deployment then becomes the cheapest variant that holds accuracy.
Why not just…?
Why not just run low-frame-rate time-lapse and track cells? This is the main threat. Trajectory models are strong, and one frame every 15 minutes costs almost nothing. If edge fluctuations carry no information beyond position and speed, the proposal fails. The nested-model test settles it: we add spectral features to the trajectory classifier and ask whether held-out AUC and time-to-call improve. If they do not, we report a negative result. The GN-25 rebound is a specific case where we expect an advantage. A rebound in protrusion activity could precede any renewed displacement, which tracks would miss.
Why not just use a fast ordinary camera in bursts and difference the frames? This alternative is real. A burst of frame differences emulates an event camera. The simultaneous dual-camera recording makes the comparison exact. If differencing matches the event camera at comparable cost, the conclusion becomes "fast fluctuation spectra, sensor-agnostic." The resonator front end still applies, and the event camera is dropped. The event sensor's expected advantages are microsecond timing and silence on static background. These matter only if useful signal lies above practical frame rates or if per-well bandwidth limits plate-scale monitoring. We measure both.
Why not just use endpoint Western blots and qPCR? They remain the label source, not the competition. The cited rebound and mRNA-protein gaps show that endpoint snapshots can disagree with each other. They also destroy the sample and cannot show when arrest begins or whether it persists. Our method is valuable only if it predicts those endpoints earlier and non-destructively. Correlation with per-cell marker staining is therefore a primary metric.
Why not just run a dense network on an edge GPU? Accuracy-wise it might be fine, and the dense CNN ablation tests exactly that. The case for event-proportional computation depends on scale: always-on monitoring across many wells where most pixels are static. We report measured energy per well-hour for both. If the GPU cost turns out negligible at realistic plate sizes, we state that the neuromorphic hardware is optional.
Proposed Experiments and Decisive Checks
We run a pilot on GN-25-treated HMLE-SNAIL cells and vehicle controls, roughly a few dozen wells per condition across independent passages, with splits made by biological replicate rather than by field. The first decisive check is that the fluctuation spectrum separates vehicle from treated wells above the fixed-cell noise floor. It must also do so before scratch closure diverges. If this fails, we stop.
The second check is the nested comparison against Trackastra plus the trajectory classifier across all three cell systems. We report AUC, the correlation of the continuous score with per-cell E-cadherin/vimentin ratio, and the earliest separating time point with confidence intervals. The third check is the matched event-versus-frame-differencing comparison at equal bytes per well. The fourth is the accuracy-versus-measured-energy curve across the resonator bank, FFT, learned spiking network and dense CNN.
A positive outcome is an early, non-destructive EMT readout that generalizes across all three cell systems and adds information beyond tracks. A negative outcome on the second check would show that, for adherent EMT models, slow trajectories already contain the signal. That result is worth publishing because it bounds where neuromorphic sensing helps in cell biology.
References
- Neuromorphic-enabled video-activated cell sorting
- EAP4EMSIG - Enhancing Event-Driven Microscopy for Microfluidic Single-Cell Analysis
- TRACKASTRA: Transformer-Based Cell Tracking for Live-Cell Microscopy
- Cell Behavior Video Classification Challenge, a benchmark for computer vision methods in live-cell imaging
- Intermediate cell states in epithelial-to-mesenchymal transition
- Systems analysis reveals a transcriptional reversal of the mesenchymal phenotype induced by SNAIL-inhibitor GN-25
- Impact of Nischarin on EMT regulators in breast cancer cell lines
- A Spiking Heterogeneous Harmonic Resonate-and-Fire State Space Model for Time Series