The Biophysical Limits of Structural-Functional Connectivity Correlation in Neuroimaging Psychiatry

Abstract

Correlation between structural connectivity (SC) from diffusion MRI and functional connectivity (FC) from fMRI is widely used in neuroimaging psychiatry as a measure of structure-function coupling (SFC) and is often taken as evidence of tight coupling between anatomy and neural communication. This Perspective examines conceptual and biophysical limitations of interpreting SFC as evidence of neural coupling and mechanistic dysfunction in psychiatric disorders. SFC quantifies statistical dependence between network summaries but does not capture biophysical causation. Local circuit properties such as excitation-inhibition balance, synaptic gain, and neuronal responsiveness can alter FC without detectable changes in macroscale white matter structure. Global, static correlations further ignore temporal constraints such as conduction delays and frequency-dependent dynamics. FC is additionally filtered through neurovascular coupling, often altered in psychiatric conditions, weakening the interpretability of SFC as a proxy for neural coupling. Biophysically constrained generative models offer an alternative framework in which SC provides a fixed anatomical scaffold for simulating neural dynamics and explaining FC differences via explicit circuit parameters. SFC remains useful as a descriptive network measure, but its interpretation as a marker of biophysical coupling is unwarranted. Generative modeling approaches are needed to support mechanistic inference and translational relevance.

Recent research in neuroimaging psychiatry frequently relies on the correlation between structural connectivity (SC) derived from diffusion MRI and functional connectivity (FC) derived from resting state fMRI to characterize network alterations in mental disorders. This measure, commonly referred to as structure-function coupling (SFC), has been widely adopted to study network-level alterations across psychiatric and neurological disorders [1, 2]. Although this approach is sometimes framed within computational psychiatry, many such studies do not employ explicit generative or algorithmic models. Instead, SFC is used as a summary statistic to support inferences about structure-function relationships and is often implicitly treated as a mechanistic marker of “coupling”. While SFC provides a useful descriptive index of network alignment, interpretation of a simple pairwise correlation coefficient as evidence for biophysical coupling remains problematic and risks overstating mechanistic inference [3]. In this perspective, I argue that such inferences should instead be grounded in biophysically constrained generative models that treat SC as an anatomical scaffold for dynamic simulations.

A clear distinction is required between three related but nonequivalent concepts. First, anatomical constraint refers to the fact that white matter architecture limits possible communication pathways between brain regions. Second, statistical dependence reflects covariance between empirical SC and FC matrices. Third, biophysical causation concerns the neurophysiological mechanisms through which structure shapes dynamic neural interactions. SFC therefore quantifies only statistical dependence; no direct inference about biophysical causation follows from this statistic alone. Yet in neuroimaging psychiatry, disease- related changes in SFC are frequently discussed as if they indexed altered neural coupling, blurring this crucial distinction [1–3 ].

Several studies have interpreted reduced SFC as a direct index of impaired communication efficiency or dysconnectivity in psychiatric populations. For example, in schizophrenia, reduced SFC has been linked to disrupted integration across large-scale networks and proposed as a multimodal biomarker for dysconnectivity [1]. In major depressive disorder, regional SFC alterations have been related to impaired top-down regulation and evaluated as candidate markers for diagnosis or symptom prediction [4]. However, these interpretations rely on the assumption that correlation strength reflects underlying biophysical coupling, an assumption that remains insufficiently validated.

SFC is more appropriately understood as a statistical descriptor of alignment between structural constraints and functional co-activation patterns. Empirically, this measure varies systematically across the cortical hierarchy, with stronger correspondence in unimodal regions and weaker correspondence in transmodal association cortex [2]. It is also shaped by developmental stage and cognitive state, suggesting that it reflects a composite of anatomical constraints, network dynamics, and context-dependent modulation rather than a direct measure of biophysical coupling.

The functional organization of the brain emerges from nonlinear and state dependent dynamics operating at the local circuit level. Parameters such as excitation-inhibition balance, synaptic gain, and neuronal responsiveness strongly influence functional interactions. These parameters are frequently implicated in psychiatric pathophysiology, including schizophrenia and autism spectrum disorder [5]. Substantial alterations in these local dynamics can produce marked changes in FC without detectable changes in macroscale white matter structure. Consequently, an intact or only modestly altered SFC profile does not rule out profound local circuit dysfunction, and a static correlation between SC and FC is poorly suited for testing hypotheses centered on such mechanisms.

Temporal and dynamical constraints further limit the interpretability of SFC. Structural connections shape signal propagation speed and transmission delays, which critically determine large-scale synchrony and oscillatory patterns. Global, static FC estimates collapse these temporal features into time-averaged associations, obscuring delay-dependent and frequency-specific effects. Static SFC relies on time-averaged FC and therefore obscures temporal variability in network interactions. Evidence from dynamic functional connectivity studies indicates that functional coupling fluctuates over time [6], even in the absence of structural changes. This suggests that static SFC measures capture only a constrained summary of a temporally evolving system and may underestimate the contribution of dynamic processes that are not directly tied to fixed anatomical pathways. Methodological developments in related neuroimaging domains illustrate this limitation. In naturalistic and hyperscanning studies, static inter-subject correlation measures have increasingly been supplemented by dynamic and directional analyses of information flow [7, 8]. The broader methodological lesson is clear; biologically plausible inference requires models that respect temporal dynamics rather than reliance on static correlation alone.

Additional complexity arises from the indirect nature of the BOLD signal. Functional connectivity derived from fMRI is filtered through neurovascular coupling, a process known to be altered in many psychiatric and neurological conditions [9], including in unmedicated patients. Changes in neurovascular coupling can modify observed FC independently of underlying neural communication, further weakening the link between SFC and true neural coupling. In such cases, altered SFC may reflect vascular rather than neural mechanisms, complicating mechanistic interpretation in clinical cohorts.

To support mechanistic inference in neuroimaging psychiatry, a shift toward biophysically constrained generative modeling is required. In these frameworks, the structural connectome is treated as a fixed anatomical backbone rather than a correlative partner. Neural mass or mean field models simulate dynamic activity constrained by this structure, and model parameters are adjusted to reproduce empirical FC [10, 11]. For instance, in schizophrenia, models can be parameterized to simulate alterations in excitation- inhibition balance or synaptic gain while keeping SC constant, and model validity is assessed by the extent to which simulated functional patterns reproduce empirical observations [10, 11]. Importantly, changes in simulated FC are often achieved by altering local inhibitory gain, synaptic scaling, or neuromodulatory tone while keeping SC constant. Disease-related differences are therefore expressed in explicit biophysical parameters rather than abstract correlation coefficients, providing testable and clinically interpretable hypotheses about circuit dysfunction. Within this framework, SFC can still serve as a descriptive summary, but mechanistic claims are grounded in the model parameters rather than the correlation itself.

SFC retains value as a descriptive network metric and as a compact summary of how empirical FC aligns with the underlying connectome. However, interpretation of SFC as a proxy for biophysical coupling is not warranted in the absence of explicit dynamical models. Progress in neuroimaging psychiatry will depend on approaches that explicitly model how structure gives rise to function through state-dependent, dynamic neural processes. Greater emphasis on biophysically grounded generative models will strengthen causal inference, sharpen hypotheses about circuit-level pathology, and improve the translational relevance of connectivity findings.