Can we train living neuronal networks by tuning their connections, the way we train artificial neural networks?

Reading time: about 7–8 minutes

Artificial neural networks learn by updating connection weights. But what if we try to do something similar with actual neurons? Could we pick two locations in a living network, strengthen the influence of one on the other, and then show that the change persists after stimulation has stopped?

That last question is harder than it sounds. Making a neuron fire is not the same as changing how a network is wired.

In a 2026 paper in Advanced Science, Sreedhar S. Kumar and colleagues combined high-density microelectrode arrays (HD-MEAs), spike-timing-dependent plasticity (STDP), and a new readout called Conditional Activity Metrics (CAM). Their goal was to make neuronal circuit shaping more targeted, measurable, and scalable.

The framework has three essential steps: read the network, stimulate candidate connections, and verify what changed.

1. What is a biological “weight”?

In a conventional neural network, a weight w represents how strongly one unit influences another. A biological excitatory synapse plays a roughly analogous role, but its efficacy cannot be set by simply assigning a number.

Suppose neuron A connects to neuron B. When A generates an action potential (AP), the signal travels along its axon and triggers neurotransmitter release. This can depolarize B, producing an excitatory postsynaptic potential (EPSP).

An EPSP is not itself an action potential. It is one contribution toward B reaching its firing threshold. If A→B becomes more effective, the same input from A may produce a larger EPSP, making B more likely to fire.

This capacity for activity-dependent change is called synaptic plasticity. It is a major component of biological learning, but unlike an artificial model’s weight matrix, it involves receptors, release probability, excitability, and many other cellular processes.

Synaptic efficacy schematic

Figure 1. Original explanatory schematic (English labels): a stronger excitatory pathway can produce a larger postsynaptic effect from the same presynaptic input. This is not a directly measured weight map; compare the study’s conceptual Figure 1a–c.

2. Can spike timing train a connection?

A classic rule of plasticity is spike-timing-dependent plasticity, or STDP. The order of two spikes matters.

For a typical excitatory STDP relationship:

  • Pre → Post: the presynaptic neuron fires shortly before the postsynaptic neuron. This timing often favors long-term potentiation (LTP).
  • Post → Pre: the postsynaptic neuron fires first. This timing often favors long-term depression (LTD).

For example, imagine A firing at 0 ms and B firing at +5 ms. Repeating this pairing may strengthen A→B in suitable biological conditions. The reverse order may favor weakening.

This is not a universal rule: cell type, network activity, and stimulation parameters can reverse or eliminate the expected effect. Still, STDP suggests a testable engineering strategy: impose a carefully chosen stimulation sequence instead of waiting for that sequence to occur by chance.

STDP timing schematic

Figure 2. Original schematic of a classical excitatory STDP rule. We define Δt = tPost − tPre. The arrows show tendencies, not guaranteed outcomes.

Implementing the idea with an HD-MEA

An HD-MEA is a silicon chip covered in tiny electrodes. Neurons or slices placed on the chip generate extracellular electrical signals that the electrodes can record; selected electrodes can also deliver programmed pulses.

The main device in this study provides 26,400 addressable electrodes, spaced 17.5 μm apart, with up to 1,024 simultaneous recording channels (typically sampled at 20 kHz).

The experiment follows a practical sequence:

  1. Record baseline activity. Collect spontaneous spike trains and find where the network is active.
  2. Select candidate pairs. Look for electrodes whose spike trains show consistent temporal relationships. These are putative functional connections, not necessarily proven single synapses.
  3. Deliver paired stimulation. Apply one biphasic pulse at a candidate presynaptic electrode, followed after a controlled delay by a burst of 5–10 pulses at the postsynaptic site. Repeating the sequence attempts to induce plasticity.
  4. Record again. After stimulation, ask whether the relationship between the two sites has changed.

A crucial distinction: accurately timed electrical stimulation commands do not guarantee equally precise neuronal spikes. The authors therefore also evaluated response reliability and stimulation artifacts.

HD-MEA experimental workflow

Figure 3. Original workflow schematic, adapted conceptually from the paper’s Figure 1g. It separates observing a candidate relationship, delivering stimulation, and evaluating the response.

3. Did the synapse really become stronger?

Stimulating A and observing more spikes at B would not, by itself, prove that A→B had strengthened. B might simply have become more excitable, or the whole network might have entered a more active state.

The authors therefore used two complementary kinds of evidence.

Direct electrophysiological evidence: patch clamp

A patch-clamp electrode can record the membrane voltage of a particular postsynaptic neuron. In a selected rat cortical culture, the researchers identified a stimulation site that consistently elicited an EPSP in the patched neuron.

They then delivered 100 STDP pairings. Afterward, when they stimulated the candidate presynaptic site again, the mean EPSP amplitude was approximately 20% greater than before, with a statistically significant difference.

That is stronger evidence than an increase in spike counts alone: the postsynaptic membrane response to a comparable test input became larger.

Experimental EPSP traces and statistics from original Figure 2g–i

Figure 4. Original experimental Figure 2g–i from Kumar et al. (2026). Panel g tracks normalized EPSP amplitudes before and after pairing; h compares the mean EPSP waveforms (about 20% enhancement in this example); i shows the amplitude distributions and statistical test (p < 0.001). These are results from selected recordings—not a 20% improvement for every connection. View the figure at the publisher.

Patch clamp is convincing, but it is difficult to apply to hundreds of cells at once. That motivates a scalable readout from extracellular spike trains.

4. CAM: measuring change without directly reading synaptic weights

The authors introduce Conditional Activity Metrics (CAM). Rather than asking how often B fires overall, CAM asks:

Given that A just fired, what happens next at B?

Imagine observing B for 50 ms after each spike from A. In a simplified example:

Measure Before After
First spike from B 20 ms after A 8 ms after A
Spikes from B within 50 ms 2 4

These numbers are illustrative, not measurements from the paper.

Two aspects have changed:

  • First-spike latency, t₁: B tends to respond earlier after A’s spike.
  • Conditional spike density, ρ: B produces more spikes in the designated post-spike window.

CAM combines these into a two-dimensional metric, CAM = (t₁, ρ). In the actual analysis, it aggregates many presynaptic spike events; the windows for measuring latency and density need not be identical.

For a strengthening-like change, we might observe Δt₁ < 0** and **Δρ > 0. The opposite pattern can be consistent with weakening.

CAM spike timing and density schematic

Figure 5. Original schematic using hypothetical spike times: CAM summarizes when B first responds and how densely B fires after A. It is not a direct measurement of one synaptic weight.

Why control pairs matter

What if every neuron starts firing more frequently after stimulation? A change in CAM for one pair would then be much less persuasive.

The authors compare each target pair’s change in CAM (ΔCAM) with a distribution of changes from matched control pairs in the same network. A robust Mahalanobis distance measures how far the target change lies from normal network variation, accounting for the joint variability of latency and density.

In plain terms: a change only counts as compelling evidence of targeted circuit shaping when it stands out from the controls.

Can CAM track known weights?

To evaluate the metric under controlled conditions, the researchers used a simulated 2,000-neuron spiking network and explicitly modified selected synaptic weights.

The reported correlations were strong:

  • Weight change vs. density change: r = +0.93.
  • Weight change vs. latency change: r = −0.91.

Those numbers come from simulation, where the modified weight is known. They do not imply that the real weight of any biological synapse can be uniquely reconstructed from CAM.

The experimental CAM values were measured mainly at the electrode / multi-unit-activity level, where several neurons may contribute to one channel. They therefore reflect changes in local functional relationships, not exact weights of isolated synapses.

5. What worked in living tissue?

The authors tested the framework in three preparations:

  • Dissociated rat cortical neuron cultures, which provide long-term access to developing networks.
  • Acute mouse cortical slices, which retain some native local circuitry.
  • Mouse stem-cell-derived cerebral organoid slices, which offer a more complex tissue-like environment.

In the 2D cultures, 87 of 227 tested electrode pairs (38.3%) showed statistically significant CAM changes. In the slice and organoid preparations combined, 20 of 52 pairs (38.5%) were significant.

Across all experiments, that is 107 of 279 pairs—roughly 40%.

The team also demonstrated parallel stimulation, using as many as eight stimulation sites. In a subset followed for longer periods, the induced effects persisted for at least 90 minutes.

Experimental detection counts

Figure 6. Original data-summary visualization using counts reported in the paper. “Significant” refers to detected CAM changes—not successful editing of a known single-synapse weight to a prescribed value.

This distinction matters. Some pairs did not change significantly, and some changed in a direction that did not match the simple STDP prediction. The result is an encouraging demonstration of scalability, not a fully predictable biological weight-editing machine.

6. Is this already a trainable biological computer?

Not yet. But the work assembles three capabilities a future system would need:

Read: observe the network and identify candidate functional connections.

Write: deliver precisely timed stimulation intended to reshape those connections.

Verify: measure whether the functional relationship changes after the intervention, and compare the result with controls.

A genuinely task-driven biohybrid computer would also need an input encoding, an output decoder, a task objective, and a feedback strategy that adjusts future stimulation according to performance.

This paper does not demonstrate autonomous task learning in an organoid. It addresses a narrower—and very important—engineering problem: making circuit changes measurable and, to a degree, targetable.

Takeaway

The central advance is not a new law of neuroplasticity. It is an integrated HD-MEA + STDP + CAM workflow that moves beyond merely observing spikes or provoking transient activity.

We are still far from assigning arbitrary weights to living neurons as we do in software. But learning to read, perturb, and verify functional connections is a meaningful step toward more systematic neural-circuit training.

Paper: Sreedhar S. Kumar et al., Engineering Neuronal Network Connectivity Through Precise and Scalable Electrical Modulation, Advanced Science (2026). DOI: 10.1002/advs.75473.

Figures: Original diagrams (1, 2, 3, 5, 6) created for this article; experimental Figure 4 reproduced from Kumar et al., Figure 2g–i. Original paper licensed under CC BY; image credits remain with the authors.