Neural Network Reads Soma Activity from Dendritic Recordings with Sparse Supervision

The soma, the cell body of a neuron, generates the electrical output that determines how the cell communicates, yet in many experimental setups it cannot be measured directly. Researchers often have access only to dendritic recordings, which capture a filtered and attenuated version of the somatic signal. Reconstructing the somatic membrane potential from these distal measurements is a difficult inverse problem, and it becomes even harder when the soma and dendrites are weakly coupled, because much of the somatic detail is lost in transmission.

A new study published in Advances in Continuous and Discrete Models on 27 October 2026 by Abdeltif Oujbara, Benjamin Ambrosio and M. A. Aziz-Alaoui tackles this reconstruction task with a physics-informed neural network, or PINN, constrained by a two-compartment Hodgkin–Huxley model of a neuron. The work appears as an open access research article, shared early as a citable version of record in progress.

The approach trains the network on dense dendritic voltage recordings together with the known injected current. Crucially, the method also uses a small number of somatic voltage samples, amounting to at most 5% of the time points, as sparse supervision. With this limited anchoring, the network reconstructs the full somatic trajectory while simultaneously adjusting a selected set of somatic maximal conductances, the parameters that govern the neuron’s ion channels.

The researchers tested the method on synthetic data generated under four stimulation protocols, allowing them to quantify precisely how reconstruction quality depends on the amount of somatic supervision available. The results draw a clear boundary between failure and success. Without any somatic samples, the present formulation returns only a smooth trajectory in which action potentials are entirely absent and the subthreshold voltage level is biased, with a root-mean-square error of 10 to 14 millivolts.

The picture changes dramatically with even minimal supervision. Just 1% of the somatic time points is enough for the method to recover every spike in the somatic trajectory. Increasing the supervision to 5% of the time points brings the root-mean-square error down to about 2 millivolts, spikes included, and yields relative errors below 0.1% on the reconstructed sodium and delayed-rectifier conductances, the key parameters shaping the neuron’s firing behavior.

To benchmark their approach, the authors compared the PINN with an unscented Kalman filter constrained by the same two-compartment model and assimilating exactly the same observations. They also assessed how robust the reconstructions are to measurement noise and to random initialization of the network, two factors that often determine whether inverse methods are practical outside idealized conditions.

The authors are explicit about the limits of what their results show. The reconstructions reported rely on sparse somatic anchoring in addition to the dendritic recordings, so the method does not dispense with somatic information altogether. Moreover, the findings come from a synthetic, weakly supervised two-compartment setting, and the study is framed as delimiting what can be inferred under these conditions rather than demonstrating performance on experimental data.

Within that scope, the study identifies the amount of somatic information required by the present formulation, offering a quantitative guide for anyone attempting to recover somatic dynamics from distal measurements. By pairing a physics-based Hodgkin–Huxley constraint with a data-driven network and quantifying the supervision threshold, the work clarifies where the boundary lies between reconstructing a neuron’s output and losing it entirely in the filtered dendritic signal.

Subject of Research: Reconstructing somatic membrane potential from dendritic recordings using a physics-informed neural network in a weakly coupled two-compartment neuron model

Article Title: Physics-constrained inference of somatic dynamics from dendritic recordings with sparse somatic supervision in weakly coupled two-compartment neuron model

Article References: Oujbara, A., Ambrosio, B., & Aziz-Alaoui, M. A. (2026). Physics-constrained inference of somatic dynamics from dendritic recordings with sparse somatic supervision in weakly coupled two-compartment neuron model. Advances in Continuous and Discrete Models. https://doi.org/10.1186/s13662-026-04133-9

Image Credits: AI Generated

DOI: 10.1186/s13662-026-04133-9

Keywords: physics-informed neural network, Hodgkin–Huxley model, two-compartment neuron, somatic membrane potential, dendritic recordings, inverse problem, sparse supervision, unscented Kalman filter, maximal conductances, computational neuroscience, Physics-constrained, inference