A Dual-Head Transformer-State-Space Architecture for Neurocircuit Mechanism Decomposition from fMRI

Precision psychiatry aspires to elucidate brain-based biomarkers of psychopathology to bolster disease risk assessment and treatment development. To this end, functional magnetic resonance imaging (fMRI) has helped triangulate brain circuits whose functional features are correlated with or even predictive of forms of psychopathology. Yet, fMRI biomarkers to date remain largely descriptive identifiers of where, rather than how, neurobiology is aberrant, limiting their utility for guiding treatment. We present a method for decomposing fMRI-based functional connectivity (FC) into constituent biomechanisms - output drive, input responsivity, modulator gating - with clearer alignment to differentiable therapeutic interventions. Neurocircuit mechanism decomposition (NMD) integrates (i) a graph-constrained, lag-aware transformer to estimate directed, pathway-specific routing distributions and drive signals, with (ii) a measurement-aware state-space model (SSM) that models hemodynamic convolution and recovers intrinsic latent dynamics. This dual-head architecture yields interpretable circuit parameters that may provide a more direct bridge from fMRI to treatment strategy selection. We instantiate the model in an anatomically and electrophysiologically well-defined circuit: the cortico-basal ganglia-thalamo-cortical loop.

💡 Research Summary

The paper introduces a novel framework—Neurocircuit Mechanism Decomposition (NMD)—that transforms conventional functional connectivity (FC) derived from fMRI into mechanistic descriptors of brain circuitry. The authors argue that static FC conflates multiple underlying processes (upstream drive, downstream sensitivity, and modulatory gating) and thus provides limited guidance for treatment selection. To address this, they propose a dual‑head architecture that jointly learns (i) a graph‑constrained, lag‑aware transformer to infer directed, pathway‑specific routing probabilities and temporal delays, and (ii) a measurement‑aware state‑space model (SSM) that deconvolves the hemodynamic response and recovers latent neural dynamics.

Both heads share a lightweight temporal encoder that converts each ROI’s BOLD time series into token embeddings. The transformer head respects an anatomical adjacency mask, factorizes attention into spatial and temporal components, and produces a routing tensor π(i,ℓ→j) that quantifies (a) the strength of influence from source i to target j, (b) the most predictive lag, and (c) the precision of the timing distribution. Using the deconvolved neural states x̂(t) from the SSM, the transformer computes a drive signal û̂_j(t) for each target as a weighted sum of lagged sources.

The SSM head models latent neural activity with a linear discrete‑time state equation x(t+1)=A·x(t)+B·û(t)+w(t) and an observation equation y(t)=h∗(C·x(t))+v(t) that explicitly incorporates region‑specific HRF convolution. Matrix A captures intrinsic dynamics, while B encodes baseline input sensitivity (how strongly each pathway responds to its endogenous drive). The observation model includes physiologically constrained HRF parameters and noise covariances. Crucially, gradients are stopped at the interface between the two heads: the SSM first deconvolves BOLD into x̂(t), the transformer then operates on x̂(t) to infer routing, but the transformer’s loss does not back‑propagate into the SSM. This asymmetric coupling prevents parameter trade‑offs and preserves identifiability of four distinct quantities: (1) transformer‑derived drive magnitude, (2) transformer‑derived timing (lag profile), (3) SSM‑derived baseline sensitivity, and (4) SSM‑derived state‑dependent gating.

State‑dependent modulation is modeled by extracting a time‑varying gain g(t) from a designated modulatory subsystem (e.g., midbrain SN/VTA) and applying it multiplicatively to B inside the SSM, yielding an effective sensitivity B_eff(t)=B·(I+β·G(t)). This captures dynamic amplification or suppression of routing that may reflect arousal, neuromodulation, or symptom‑related switches.

Training optimizes a composite loss: the primary term is the negative log‑likelihood of the observed BOLD under the SSM (computed via differentiable Kalman filtering/smoothing). Auxiliary transformer losses include multi‑step forecasting (2–4 TR ahead) to emphasize predictive causality, and optional reconstruction of the final frame. Regularization comprises L1 sparsity on routing weights, smooth unimodal priors on lag embeddings, physiological priors on HRF shape, and spectral stability constraints on A (eigenvalues within the unit circle). Hyperparameters are deliberately fixed to favor interpretability: a single attention block, four attention heads (two spatial, two temporal), hidden dimension 128, dropout 0.1, and low‑rank factorization of B (rank‑3).

The authors demonstrate the method on the cortico‑basal‑ganglia‑thalamo‑cortical loop. The resulting output metrics include: (i) cortical drive profiles (weights of cortical ROIs contributing to striatal activity), (ii) striatal receptivity (scalar sensitivity from B), (iii) directed pathway influence (summed routing probabilities), (iv) lag timing statistics (peak, centroid, concentration), (v) time‑resolved drive signals û̂(t), and (vi) dynamic gating g(t). Compared with traditional FC, these metrics disentangle drive, timing, and modulation, offering a more direct bridge to therapeutic hypotheses such as targeted TMS, DBS, or pharmacological interventions.

Limitations are acknowledged: the model introduces many parameters and requires substantial data for stable estimation; the reliance on hard anatomical masks may miss unconventional pathways; and the linear SSM may not capture all nonlinear neural dynamics. Future work is suggested to integrate multimodal data (EEG/MEG/PET), adopt Bayesian uncertainty quantification, and validate mechanistic features against clinical treatment outcomes. Overall, the paper presents a sophisticated, biologically grounded approach that moves fMRI analysis from descriptive connectivity toward actionable mechanistic insight.