A conceptual framework for astrocyte function

Abstract

The participation of astrocytes in brain computation was hypothesized in 1992, coinciding with the discovery that these cells display a form of intracellular Ca²⁺ signaling sensitive to neuroactive molecules. This finding fostered conceptual leaps crystalized around the idea that astrocytes, once thought to be passive, participate actively in brain signaling and outputs. A multitude of disparate roles of astrocytes has since emerged, but their meaningful integration has been muddied by the lack of consensus and models of how we conceive the functional position of these cells in brain circuitry. In this Perspective, we propose an intuitive, data-driven and transferable conceptual framework we coin 'contextual guidance'. It describes astrocytes as 'contextual gates' that shape neural circuitry in an adaptive, state-dependent fashion. This paradigm provides fresh perspectives on principles of astrocyte signaling and its relevance to brain function, which could spur new experimental avenues, including in computational space.

Fig. 1 |. Astrocyte networks are cued by contextual signals to guide neural circuits.

In contextual guidance, neuronal networks operate rapid point-to-point transfer and integration of information, while astrocytic networks transduce slower modulatory information over large spatial scales. In this proposed framework, astrocytes act as the primary receiver of slow contextual signals, such as neuromodulators, sensory inputs and neurohormones, that carry information about the status of the animal with respect to the outside world (left). Owing to their interactions with blood vessels, they are also the first responders to circulating signals, such as corticosterone, glucose, oxygen and CO₂, that carry information about the animal’s bodily status (right). In response, astrocytes signal onto neuronal elements to alter circuit connectivity and/or activity (middle). The net result is a context-specific reconfiguration of the neural network’s topology and functional output. The large spheres represent neurons (in the neuron network), along with their synaptic connections. The small spheres represent astrocytes (in the astrocyte network), with the lines illustrating the gap-junction coupling of astrocytes into functional networks. The dashed arrow from the neuron network represents the influence of local neuron network activity on the astrocyte contextualizing response.

Fig. 2 |. Astrocytes reconfigure circuits in a context-dependent fashion.

a, Stepwise illustration of the contextual guidance proposal, showing an individual astrocyte and embedded neuronal circuit over time. The astrocyte responds to the emergence of a context (shown as Ca²⁺ transients). This leads to the mobilization of circuit-modifying contextualizer(s), which act on select elements (here presynaptic strength as in ref. and interneuron firing as in ref. 48). These modifications reconfigure the network, illustrated as a change in the path of least resistance and resulting output. b, Putative scenarios in which (i) a context 1 yields circuit adaptations a and b in circuits x and y, respectively; (ii) two contexts 1 and 2 yield the same adaptation b in circuits y and z; and (iii) adaptation b is achieved in circuit z in response to contexts 2 or 3.

Fig. 3 |. In contextual guidance, astrocytes are a gate for modulatory inputs.

In current concepts of astrocyte function (left), which are based on ref. , the contribution of astrocytes to neuronal networks is constrained in a feedback loop: synaptic or neuronal signaling is the driver of astrocyte activity, which outputs back onto synapse or neuron activity. Modulatory inputs are assumed to act directly on synapses or neurons. Contextual guidance (right) comprises the same fundamental parts. However, it makes astrocytes a gateway for modulatory signals at large, which are not synaptic by default, de facto describing astrocytes as circuit effectors in a feedforward mechanism. This hierarchical architecture reconciles the slowness of astrocytes and modulatory signaling with the fast timescale of neuronal and synaptic activity. The modulatory contextual inputs that drive astrocyte activity may also comprise, or be influenced by, direct signaling from local neuronal networks.

Fig. 4 |. Astrocytes provide context adaptation and stability to recurrent neural networks: proof of principle.

a, In a conventional recurrent artificial neural network (RNN), synaptic weights are given by a matrix W (left). In a ‘contextual guidance RNN’ (center), synaptic weights are given by the sum W + Γi, where Γi reflects astrocyte-mediated modulation of synaptic strength over two ensembles of connections (red and blue), enacting tiled astrocytic domains (dashed areas). Both RNNs are tasked with producing a specific output (right): bursting activity (harmonic oscillation) or tonic firing, as in ref. . b, RNNs must change their output at the occurrence of a contextual signal (arrow), with error measured as a deviation from the expected output frequency. In this abstraction, where the context is known to the RNNs, a conventional RNN ‘re-learns’ W at each context switch, resulting in high error transients. The ‘contextual guidance RNN’ can change its regime rapidly. In effect, a single W is learned and the network adapts to the context by rapidly modulating synaptic weights through astrocytes (Γi). c, Two LSTM RNNs were built as in a, and trained on an abstraction of the Wisconsin card-sorting task, in which agents receive a set of cards, each with different numbers, shapes and colors, and are tasked with picking one card to match a hidden rule (for example, select the card with triangles). In this task, the rule (that is, context), can be learned only by trial and error, and changes during the task. The RNNs are trained to maximize future rewards by determining and exploiting each new rule. We used 64 neurons and 16 astrocytes, and numerically optimized all free parameters using episodes of 80 trials. Solid traces show the mean error probability, and shading shows the s.e.m. The ‘contextual guidance’ LSTM RNN outperforms a conventional network by adapting to the new context and performing optimally until the next context switch. d, Dimensionality reduction across trials in c shows that the astrocytic modifications of synaptic weights (Γi) are highly context dependent, reminiscent of the contextualizing rules described in the text. Simulations were run on MATLAB (b) and Python (c and d).