Disinhibition gates how fruit flies lock a goal heading into a recurrent memory circuit

A network tuned to hold activity stably is hard to switch off once it starts, a trade-off between stability and flexibility that attractor models of working memory share ^[4]. Models usually handle it by adding a gate. How a real circuit would build that gate out of synapses has been unclear, the authors write ^[5].

The fruit fly offers a place to look at the synapses directly. Its fan-shaped body is a highly recurrent part of the navigation centre thought to encode goals. It holds about 30 types of local neuron, and the connectome maps their wiring at synaptic resolution ^[6]. In the behaviour studied here, a brief encounter with odour sends a fly on a persistent run, usually biased upwind ^[10]. The same group had earlier tied hDeltaK neurons to that goal: sparse activation sent flies walking in repeatable directions relative to wind, and silencing made their trajectories less persistent after the odour was gone ^[7].

The new piece is the split. In the model, inhibiting hDeltaK leaves PFG free to follow whatever the compass system reports ^[3]. Lifting that inhibition lets the recurrent loop lock the current reading in place as a heading memory ^[3]. So a turning fly keeps an updated heading in PFG, and when inhibition drops, the direction it faces becomes the goal it holds ^[3]. The recordings match that timing, with inhibition onto hDeltaK rising during turns and suppressed during odour and goal-directed runs ^[15].

The persistence half rests on direct measurement. Whole-cell recordings showed that hDeltaK persistence depends on recurrence ^[12], and that the cells get slow recurrent excitation and fast inhibition from their partners ^[13]. A model built on those timings stayed persistent across a range of synaptic strengths, so the memory did not depend on one precisely set weight ^[14].

I think the switching half is a strong hypothesis, supported by when the inhibition arrives. Inhibition onto hDeltaK changes at the right moments ^[15], and the model reproduces the decoupling seen during turns and rest when inhibition is used as the switch ^[2]. The thing this doesn't tell you is whether imposing or blocking that inhibition is enough to write or erase a goal. The abstract does not report that manipulation or say which neurons supply the inhibition.

The authors set the work against prefrontal cortex. There, persistent activity during working-memory tasks outlasts the cells' membrane time constants by a wide margin, evidence that it arises from network interactions ^[8]. Their conclusion is stated in general terms. "Our work reveals how disinhibition can serve as a gate to rapidly write an ongoing measurement to a recurrent circuit," the authors wrote ^[9].