Computational neuroscience frequently operates in two separate, parallel worlds. On one side sit highly successful statistical and phenomenological models—such as Generalized Linear Models (GLMs), latent variable recurrent neural networks (RNNs), and deep networks. These models excel at describing neural trajectories, decoding behavior, and fitting experimental data. On the other side sit biophysical models, which describe the physical hardware of the brain: ion channels, synaptic receptors, membrane conductances, and cellular morphology.
The disconnect between these two domains creates a descriptive-mechanistic gap. Statistical models can capture what a neural population is doing with remarkable accuracy, but they often fail to explain how the underlying biological circuitry implements that behavior. Grounding abstract mathematical descriptions in physical, biophysical mechanisms is essential if we want our models to explain, rather than just describe, the brain.
The Illusion of Fit: Why Description is Not Explanation
It is tempting to assume that if a model fits neural data perfectly, it must capture something fundamental about the underlying biology. However, statistical models are often flexible enough to mimic biological phenomena without sharing any of their physical constraints.
For example, Weber et al. (2017) demonstrated that a standard Poisson GLM can replicate a vast dynamical repertoire of complex biophysical neurons, including bursting, adaptation, and bistability. Yet, the abstract filters within these statistical models lack any direct mapping to physical components like ion channels or pumps. The model predicts spikes perfectly, but its parameters cannot tell us why the cell bursts or which specific potassium current drives its adaptation.
This gap spans every scale of neuroscience:
- Network Dynamics: Maximum entropy models from statistical mechanics can perfectly capture global network synchrony across hundreds of sensory neurons. However, as noted in Searching for Collective Behavior in a Large Network of Sensory Neurons (2014), these models rely on abstract mathematical constraints that fail to identify the specific biological circuits or inhibitory pathways that actually enforce the synchrony.
- Deep Learning and Psychiatry: In clinical applications, Durstewitz et al. (2019) point out that deep learning approaches prioritize statistical correlation over biological mechanism. While these networks can predict disease trajectories from complex data, they lack physiological validity, making it difficult to extract actionable therapeutic targets.
- High-Level Cognitive Frameworks: High-level statistical frameworks can describe probabilistic relationships in neural data—such as treating the visual cortex as a generative model drawing samples from a posterior distribution (Denfield et al. (2023)). Yet, they leave open the question of how actual biophysical circuits perform these calculations on physical hardware.
The Parameter-to-Mechanism Mapping Problem
The mathematical root of the descriptive-mechanistic gap is degeneracy: multiple distinct biological mechanisms can produce the exact same statistical signature. When a model relies on abstract parameters to fit experimental curves, it treats the underlying biology as a black box, making it impossible to distinguish between competing physiological explanations.
Consider neural gain control and suppression. Statistical methods like spike-triggered covariance can identify the functional axes along which a neuron suppresses its sensitivity (Schwartz et al. (2001)). However, a mathematical divisor in a normalization model cannot distinguish whether that suppression is caused by synaptic inhibition, metabolic shunting, or pre-synaptic vesicle depletion (Senn (2002)).
Similarly, mathematical learning rules frequently utilize abstract "integrator" variables to track spike history over time. As Pfister et al. (2006) show, while these variables elegantly bridge timing-dependent and rate-based plasticity mathematically, they leave a gap between high-level detectors and their physical counterparts, such as intracellular calcium signaling or NMDA receptor states.
Without an explicit mapping between parameters and mechanisms, abstract models remain isolated from physical reality. Abstract phase-coupling terms in network models cannot be mapped directly to physical, long-range synaptic connections (Greenwood et al. (2016)), and phenomenological noise abstractions cannot tell us how individual ion channel fluctuations interact with vesicle docking sites (Moezzi et al. (2014)).
Strategies for Biophysical Grounding
To bridge this gap, researchers are developing frameworks that force statistical models to respect biological constraints, or conversely, simplify biophysical models so they can be fitted directly to data.
1. Enforcing Biological Constraints on Latent Spaces
Instead of allowing latent variable models or RNNs to explore arbitrary mathematical spaces, researchers can bake biological laws directly into their architectures. Jha et al. (2024) introduced Cell-Type Dynamical Systems (CTDS), a framework that explicitly separates latent variables into excitatory and inhibitory populations, thereby enforcing Dale's Law. By constraining the statistical model to match physical neural architectures, the resulting latent trajectories become interpretable and can successfully predict the outcomes of cell-specific perturbations like optogenetics.
2. Analytical Order Reduction
Rather than guessing abstract models, we can take highly detailed, high-order biophysical models (like the Hodgkin-Huxley equations) and mathematically derive simplified "meta-models." Kepler et al. (1990) pioneered a framework that collapses the dimensionality of complex gating variables into "Equivalent Potentials." Because this reduction is derived analytically rather than through heuristic curve-fitting, the resulting low-order parameters maintain a direct, unbroken mathematical link to the underlying biological gating variables.
3. Bottom-Up Emergence Verification
Another powerful strategy is to demonstrate how high-level, abstract computations naturally emerge from simple, realistic cellular components. Keil (2011) demonstrated this with the $\psi$-model, showing that complex multiplicative operations (such as tracking looming objects) do not require abstract mathematical formulas but emerge naturally from shunting inhibition and membrane power laws.
Similarly, Slee et al. (2005) linked abstract high-pass filtering and sub-millisecond differentiation directly to specific ion channel dynamics in the auditory brainstem. In the hippocampus, while phenomenological models can characterize how ordered population sequences emerge from independent single-cell rules (Chadwick et al. (2014)), researchers must ultimately map these dynamics onto specific biological architectures—such as the distinct sequencing and sharpening roles of the dentate gyrus and medial entorhinal cortex (Ahmadi et al. (2025))—to achieve true biophysical grounding.
Why Grounding Matters
Closing the descriptive-mechanistic gap is not merely an exercise in mathematical rigor; it changes how models are used in practice. If a continuous-time model using Neural Ordinary Differential Equations (NODEs) identifies a geometric feature like a "sink" or local minimum during an epileptic seizure (Jia et al. (2026)), that mathematical abstraction must be mapped to a specific biological neural population to guide clinical interventions or neuromodulation. By grounding statistical descriptions in physical mechanisms, computational neuroscience moves from merely cataloging the brain's trajectories to understanding the biological rules that generate them.
Key Foundational Work
Order Reduction for Dynamical Systems Describing the Behavior of Complex Neurons (1990) — contributed 1 concept elements to this direction.
Effects of Firing Synchrony on Signal Propagation in Layered Networks (1989) — contributed 1 concept elements to this direction.