NOBLE -- Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models

Characterizing the cellular properties of neurons is fundamental to understanding their function in the brain. In this quest, the generation of bio-realistic models is central towards integrating multimodal cellular data sets and establishing causal relationships. However, current modeling approaches remain constrained by the limited availability and intrinsic variability of experimental neuronal data. The deterministic formalism of bio-realistic models currently precludes accounting for the natural variability observed experimentally. While deep learning is becoming increasingly relevant in this space, it fails to capture the full biophysical complexity of neurons, their nonlinear voltage dynamics, and variability. To address these shortcomings, we introduce NOBLE, a neural operator framework that learns a mapping from a continuous frequency-modulated embedding of interpretable neuron features to the somatic voltage response induced by current injection. Trained on synthetic data generated from bio-realistic neuron models, NOBLE predicts distributions of neural dynamics accounting for the intrinsic experimental variability. Unlike conventional bio-realistic neuron models, interpolating within the embedding space offers models whose dynamics are consistent with experimentally observed responses. NOBLE enables the efficient generation of synthetic neurons that closely resemble experimental data and exhibit trial-to-trial variability, offering a $4200\times$ speedup over the numerical solver. NOBLE is the first scaled-up deep learning framework that validates its generalization with real experimental data. To this end, NOBLE captures fundamental neural properties in a unique and emergent manner that opens the door to a better understanding of cellular composition and computations, neuromorphic architectures, large-scale brain circuits, and general neuroAI applications.

💡 Research Summary

The paper introduces NOBLE (Neural Operator with Biologically‑informed Latent Embeddings), a deep‑learning framework designed to capture both the complex nonlinear voltage dynamics of neurons and the intrinsic experimental variability that deterministic biophysical models cannot represent. Traditional bio‑realistic neuron models are built from multi‑compartment cable equations, calibrated via computationally intensive evolutionary optimization to match a set of electrophysiological features. While these Hall‑of‑Fame (HoF) model ensembles can approximate variability by aggregating many deterministic instances, each model remains fixed, and generating new models or exploring parameter space is prohibitively expensive (hundreds of thousands of CPU core‑hours per cell).

NOBLE addresses these limitations by learning a single neural operator that maps from a continuous latent space of interpretable neuron characteristics to the somatic voltage response induced by a current injection. The latent space is defined by two physiologically meaningful parameters extracted from each HoF model: the firing threshold current (I_thr) and the local slope (s_thr) of the frequency‑current (F‑I) curve at that threshold. These parameters are encoded using a NeRF‑style positional encoding—stacked sine and cosine functions of increasing frequencies—producing a time‑varying embedding that captures the influence of each feature across the entire stimulus duration. The current stimulus itself (square‑pulse amplitudes sampled from a skewed normal distribution covering the experimentally relevant range) is concatenated with this embedding and fed into a Fourier Neural Operator (FNO).

The FNO operates in the frequency domain, leveraging fast Fourier transforms on equidistant temporal grids, which aligns naturally with the uniformly sampled voltage traces from both experimental recordings and synthetic PDE simulations. This architecture enables the model to learn global, high‑frequency components of voltage dynamics (spike onset, width, latency) while remaining agnostic to the specific temporal resolution of the training data. Indeed, the authors demonstrate that training on data subsampled by a factor of three (from 0.02 ms to 0.06 ms timesteps) preserves all electrophysiological features of interest, yet reduces the sequence length from ~25 k to ~8.5 k points, dramatically cutting computational load.

Training data consist of synthetic voltage traces generated from 50 HoF models of parvalbumin‑positive (PV + ALB) interneurons and a separate set of VIP interneurons. Each model is simulated under a wide range of current amplitudes, emphasizing the peri‑threshold region where the transition from subthreshold to spiking behavior is highly nonlinear. NOBLE learns to predict the full voltage time series for any point in the (I_thr, s_thr) latent space, effectively interpolating between known HoF models and extrapolating to unseen configurations. Evaluation on 10 held‑out HoF models (out‑of‑distribution) shows that NOBLE’s predictions of voltage waveforms and 16 derived electrophysiological features (spike count, amplitude, width, latency, etc.) fall within the variability observed across real human cortical recordings.

A key contribution is the ability to generate novel, biologically plausible neuron models by sampling or smoothly interpolating within the latent space. Unlike direct interpolation of the underlying PDE parameters—which often leads to numerical instability due to the highly nonlinear nature of the cable equations—latent‑space interpolation yields stable voltage responses that respect the learned distribution of I_thr and s_thr. The authors illustrate this by constructing heat maps and surface plots that relate latent coordinates to specific electrophysiological outcomes, providing an interpretable tool for probing how changes in intrinsic excitability affect observable behavior.

Performance-wise, NOBLE achieves a 4,200‑fold speedup over the original numerical solver while maintaining high fidelity. The framework also supports feature‑specific fine‑tuning: by adding a small auxiliary loss on a targeted electrophysiological metric, the model can improve accuracy on that metric without degrading overall dynamics.

In summary, NOBLE unifies three desiderata for neuron modeling: (1) faithful reproduction of nonlinear voltage dynamics, (2) explicit representation of experimental variability via a biologically grounded latent space, and (3) orders‑of‑magnitude acceleration enabling large‑scale circuit simulations. Its validation on real human cortical data, together with demonstrations on multiple interneuron types, suggests broad applicability to neurophysiology, neuromorphic hardware design, and neuro‑AI research, where rapid generation of diverse yet realistic neuronal responses is essential.