Deep learning unlocks label-free cell analysis through multiplexed impedance measurement

Cells reveal themselves through electricity. When a basophil—a white blood cell that triggers allergic reactions—passes through a narrow channel lined with electrodes, it distorts the electrical field around it in ways that encode its state: whether it has been activated by an allergen, how its membrane has stiffened or softened, what has shifted inside its nucleus. For decades, researchers have measured these electrical signatures one cell at a time, in single channels, trading sensitivity for speed. Now a team has built a system that measures four channels in parallel, reading the electrical properties of cells without any fluorescent dyes, and using artificial intelligence to untangle the signals when multiple cells arrive at once.

The challenge is straightforward in principle but vicious in practice. When you run four parallel detection channels through a single pair of electrodes—a trick called code-multiplexing, borrowed from telecommunications—the signals overlap and interfere with each other. A cell in channel one looks like noise to a decoder tuned to channel two. Prior attempts to separate these signals relied on linear mathematics: assume each signal is a simple scaled copy of a template, subtract it, move to the next one. This works reasonably well in clean conditions. But in a microfluidic channel, cells squeeze and deform. Flow rates fluctuate. Signals stretch and compress in ways that violate the linear assumption. The old methods lost accuracy fast.

The researchers built a neural network that unfolds the classical signal-separation algorithm—successive interference cancellation—into a learnable architecture. Rather than treating each overlapping signal as a rigid template, the network predicts the intensity and duration of each electrical "bit" separately, bit by bit, allowing for the nonlinear distortions that actually occur in flowing cells. It also predicts which channel each signal came from. All three tasks—intensity, duration, and channel identity—are learned by a single shared encoder, so information flows between them. The network was trained on thirty million synthetic overlap scenarios, generated by layering real single-cell measurements on top of each other at random time offsets, then tested on data from donors it had never seen. On singlets, it achieved 99.85% accuracy in identifying the source channel. On doublets and triplets—two or three cells arriving nearly simultaneously—accuracy dropped to 96.93% and 88.83%, but still far outperformed the linear baseline, which made errors at rates of 1.6, 17.5, and 26.6 percentage points higher.

The real power emerges when you measure at multiple frequencies. Cells have different electrical properties at different frequencies—their membranes respond differently, their cytoplasm behaves differently. By measuring at 101.3 kilohertz, 811 kilohertz, 1.13 megahertz, and 4.33 megahertz, you can build a frequency-dependent electrical profile of each cell. The researchers used the network's prediction at 811 kilohertz as a template, then used least-squares fitting to map that template onto the measurements at the other frequencies. This works because the underlying physics—the Maxwell mixture model—predicts that signals at different frequencies should have the same shape but different amplitudes. The reconstructed impedance distributions at each frequency correlated with ground truth at r ≈ 0.89, meaning the population-level electrical profiles were accurate even when individual cells overlapped.

To prove the method works on real biology, the team measured basophil activation in blood from six donors—four presumed nonallergic, two with documented food allergies. They split each sample into four aliquots, stimulated them with different doses of anti-IgE antibody or food allergens, and ran all four through the microfluidic channels in parallel. As basophils activate in response to allergen, their membranes change: they upregulate activation markers like CD63 and CD203c, and their electrical properties shift. The researchers measured something called "opacity"—the ratio of impedance at high frequency to impedance at low frequency—and found it correlated strongly with activation as measured by standard fluorescence flow cytometry. At 811 kilohertz, the correlation was r = −0.92 to −0.93. Basophils from the allergic donors showed activation only to the allergen they were actually allergic to, not to irrelevant food proteins. The nonallergic donor's basophils showed dose-dependent changes with anti-IgE but no activation to allergens. The electrical measurements tracked the fluorescence measurements perfectly.

What matters here is not just the accuracy but the throughput. By running four channels in parallel, the system measures four different stimulation conditions in the same twenty minutes it would take to measure one. The electrical measurement needs no labels, no staining, no antibodies—just cells flowing through a channel. The deep-learning decoder handles the signal separation automatically, learning from data rather than relying on hand-tuned algorithms. The method is sensitive enough to catch rare cells—basophils are only one percent of circulating white blood cells—and specific enough to distinguish activation states. For clinical testing of allergic responses, or for screening how cells respond to drugs, or for any application where you need to measure cell function at scale without labels, this is a significant step forward.