A computer model of the fly’s entire central nervous system learned to remember a cue with its connections fixed: the memory trace is stored within cells, and the model can retrieve it only in the state in which it learned

A computer model of the fly’s entire central nervous system learned to remember a cue with its connections fixed: the memory trace is stored within cells, and the model can retrieve it only in the state in which it learned

On October 29, Han Wang, Chunjie Tian and colleagues at the Affiliated Hospital of Guangdong Medical University posted a preprint, a paper that has not yet undergone peer review, on bioRxiv. Using the connectome, a map of the connections between all neurons in the central nervous system of an adult male fruit fly, they built a model with 167 106 cells. They trained it to push boxes onto targets in the puzzle game Sokoban.

Previous connectome-based models, the authors write, either operated without learning, as in the virtual fly that found odor sources, or learned within a single sensory pathway. Here, the cell parameters come from the fly brain model developed by Philip Shiu and colleagues, with adaptation added: after a spike, a cell responds less strongly to incoming signals for a fraction of a second. According to the authors, memory that persists without sustained neuronal activity is usually attributed to synapses, while adaptation has mainly been studied as a supporting mechanism in computation. The network of 25,6 million connections remained fixed; training adjusted the gain and firing threshold of each cell type, along with a separate layer that selected moves based on cell activity.

In the memory task, the model sees a board with one of two targets marked for 40 ms, followed by a blank delay and then the board with no targets marked. It must then push a box toward the previously marked target. The delay duration (0–640 ms) and the number of board frames shown before the decision (0–8) produced 35 combinations; the model was trained on nine, with delays of 0, 40 and 160 ms. Spiking activity that distinguishes the cues subsides within 22–42 ms after the cue disappears. The cue leaves a trace in adaptation: of all the variables describing a cell’s state in the model, adaptation alone retains a difference between cues. When the next board frame arrives, cells with different levels of adaptation respond with different spikes, allowing the model to read the trace. If, at the end of the delay, the states of all cells are replaced with those from a trial with the other cue, the model moves toward the other target on the trained combinations.

The model uses the cue only on the combinations it was trained on; with two cues, these amount to 18 trials. On the remaining combinations, the trace in the cells is at least as strong as on the trained combinations, but it no longer guides the model to push the box toward the target: after a 320 ms delay, cell activity at the time of the decision is largely opposite to the activity following the trained delay of 160 ms. Access to the trace depends on the state in which the model learned: the time elapsed since the cue, which the authors call a lock, and the excitability of the cells.

Of 24 impairments reported in flies, four could plausibly be represented in the model; one concerns visual memory: loss of the potassium channel Kv4. In flies, this loss is caused by amyloid-beta Aβ42, a protein fragment associated with Alzheimer’s disease, as shown in a 2015 study. Without Kv4, cells become more excitable; in the model, this is represented by lowering the firing threshold of every cell. Introducing the loss into the trained model, at a severity calibrated against fly data, reduces the number of successfully completed trials from 18 to 7. A model trained with the loss present from the outset completes all 18 trials at both half and twice that severity. Restoring normal excitability in such a model reduces its performance to an average of just 8,3 out of 18.

In mice modeling early Alzheimer’s disease, light activation of engram cells, which store a memory, restored recall when ordinary cues failed: retrieval was impaired, while the stored memory remained intact. The authors conclude: “Memory failure in this model need not mean that the stored memory has been lost: it may mean that the key no longer fits the lock.”