A Cerebellar Neuroprosthetic System: Computational Architecture and in vivo Test

Abstract

Emulating the input-output functions performed by a brain structure opens the possibility for developing neuroprosthetic systems that replace damaged neuronal circuits. Here, we demonstrate the feasibility of this approach by replacing the cerebellar circuit responsible for the acquisition and extinction of motor memories. Specifically, we show that a rat can undergo acquisition, retention, and extinction of the eye-blink reflex even though the biological circuit responsible for this task has been chemically inactivated via anesthesia. This is achieved by first developing a computational model of the cerebellar microcircuit involved in the acquisition of conditioned reflexes and training it with synthetic data generated based on physiological recordings. Secondly, the cerebellar model is interfaced with the brain of an anesthetized rat, connecting the model's inputs and outputs to afferent and efferent cerebellar structures. As a result, we show that the anesthetized rat, equipped with our neuroprosthetic system, can be classically conditioned to the acquisition of an eye-blink response. However, non-stationarities in the recorded biological signals limit the performance of the cerebellar model. Thus, we introduce an updated cerebellar model and validate it with physiological recordings showing that learning becomes stable and reliable. The resulting system represents an important step toward replacing lost functions of the central nervous system via neuroprosthetics, obtained by integrating a synthetic circuit with the afferent and efferent pathways of a damaged brain region. These results also embody an early example of science-based medicine, where on the one hand the neuroprosthetic system directly validates a theory of cerebellar learning that informed the design of the system, and on the other one it takes a step toward the development of neuro-prostheses that could recover lost learning functions in animals and, in the longer term, humans.

Keywords: association learning; cerebellum; classical conditioning; inferior olive; memory; neuroprosthetics; nucleo-olivary pathway; timing.

Figure 1

Biological microcircuit and synthetic counterpart. Recording (PN and IO) and stimulation sites are shown. After amplification and filtering of the signals recorded in the afferent structures, discrete events retrieved from multi unit activity are isolated by the event detection stages of the system, such that they are fed to their counterparts in the synthetic cerebellum (PN and IO). In the intact circuit, the repeated coincidence of CS and US signals within the cerebellar cortex induces plasticity causing the cerebellum to respond to the CS with a CR. In our model, once such a CR is acquired, it is relayed via the synthetic DN to the facial nucleus (FN) of the rat as an electrical stimulation that causes the animal to trigger the behavioral CR, i.e., the eye-blink. In addition, within the model, the CR triggered by the DN inhibits the IO, preventing a US-derived signal from reaching the cerebellum once a protective action has already been issued. Since anesthesia prevents acquisition in the rodent’s cerebellum, behavioral CRs expressed in the set up studied here are controlled by the synthetic circuit.

Figure 2

Intrinsic latencies of the eye-blink conditioning preparation. (A) ISI, inter-stimulus interval; ω_CS, latency between the peripheral CS stimulation and the detection of its associated neuronal response in the PN; t_CR, internal response timing learned by the model between the CS detection and the CR triggering; ω_CR, latency between the neuronal triggering of the CR and the effective eyelid closure, Λ_noi, delay between the CR trigger and the onset of the negative feedback loop inhibition; ω_US, latency between the US-trigger and the detection of its associated neural response in the IO. (B) Same latencies as in (A) for the minimum learnable ISI.

Figure 3

Functional model of the cerebellum. The processes in the top row (white boxes) map PN activity into action; in the case of eye-blink conditioning, tone detections into eye-blinks. Such mapping is controlled by the memory parameter w. The shaded processes adapt the mapping, namely, they are involved in the adjustment of w. The numbers identify specific processes. The latencies affecting each of the recording and stimulating channels as well as the parameters used in each process (see main text for an explanation).

Figure 4

Raster plots of the inputs and outputs of the model with and without stability constraint. (A) Model with stability constraint. PN detections (green), IO detections (black), and CR triggers (blue, well-timed thick, and late thin). CS (yellow area) and US (pink area) periods. The horizontal dashed red line separates acquisition and extinction phases. Vertical blue line marks the limit for well-timed CRs. (B) Model without stability constraint. Data plotted as in(A).

Figure 5

Behavior of the model with simulated data. (A) Behavioral performance. Percentage of CRs per block of trials of the model fitted with stability constraint (solid line) and without (dashed line). The vertical dotted line separates acquisition and extinction training. (B) Trajectory of w in the model fitted with stability constraint (dashed line) and without (solid line). The horizontal green dotted line marks the level above which the model does not trigger any CRs. Blocks of 10 trials.

Figure 6

Results with and without delayed NOI. (A) Raster plot with the output of the model with the delay of the NOI set to 0 s. (B) Trajectory of w in the model with a delay of 100 ms in the NOI (solid line) and with no delay (dashed line). The horizontal green dotted line marks the level above which the model does not trigger any CRs. Blocks of 10 trials.

Figure 7

Results with non-delayed NOI inhibition in different conditions. Raster plots with the output of the model with the delay of the NOI set to 0 s. (B) the model constraint to acquire CRs twice as fast or with (A). A ratio of TDs in the IO lowered to 50%. (C) Traces of the synaptic efficacy w for the simulation in Figure 6 (black) compared to the simulations in (A) (dotted red) and (B) (dotted blue).

Figure 8

Effect of the delayed plasticity trace on the behavior. (A) Model with plasticity trace starting Λ_noi seconds after each PN detection. (B) Model with plasticity trace starting right after each PN detection. Data plotted as inFigure 4.

Figure 9

Event detection performance for the recording sites. ROC curves for the IO (A) and the PN (B) event detections.

Figure 10

Performance of the experiment predicted by the training data. (A) Trajectory of the memory parameter after 2500 simulations plotted in blocks of 10 trials. The simulated experiment contained 120 trials of acquisition and 180 trials of extinction. Distribution of the block-by-block values of w (grayscale) with mean (blue) and output of a sample simulation (red) are shown. We indicate the levels of the weight that result in late (upper green line) and well-timed CRs (lower green line). The transition from acquisition to extinction training is marked by a vertical line. (B) Predicted behavioral performance after 2500 simulations. Percentage of well-timed CRs. Distribution of the block-by-block performance (grayscale) with mean (blue) and result (red) of a sample simulation [same as in(A)].

Figure 11

Event detections and triggers during the online experiment. Raster plot with the PN detections (blue dots; well-timed PNs are thicker), IO detections (black), and CR triggers (blue dots; well-timed triggers are thicker). The CS (yellow area) and US (pink area) periods are indicated. Blue line separates well-timed from late CRs. The horizontal dashed red line separates acquisition and extinction phases.

Figure 12

Quantitative results. (A) Events detected in the PN. Histogram of PN detections relative to the CS-trigger: TDs (black bars) and FAs (gray bars); in this case all FAs are late CS-detections. CS period (yellow area) and US period (pink area). (B) Events detected in the IO. Detections in the IO sorted relative to the US-trigger. Data plotted as in (A). (C) Behavioral performance of the bio-hybrid. Percentage of well-timed CRs during acquisition and extinction (solid line) are shown. CRs that were not triggered at least 20 ms ahead of the US-trigger appear as late CRs (dashed line). Each block contains 10 trials. (D) Timing of CRs. Histogram of the CRs: well-timed (black bars) and late ones (gray bars). CS period (yellow area) and US period (pink area) are indicated. The information is extracted from trials 118 to 190.

Figure 13

Weight trajectory during the experiment. The dashed vertical line separates the acquisition and extinction phases.

Figure 14

Fluctuations in the spontaneous IO rate. Mean IO rate in each trial of the experiment. The horizontal dotted line marks the 1.14 Hz level of activity. The vertical dashed line marks the transition from acquisition to extinction trials.

Figure 15

Observed performance vs. performance during simulated unpaired acquisition. (A) Acquisition during paired CS–US training versus simulated unpaired CS–US. Trajectory of the weight during the acquisition phase of the experiment (black line) plotted against results of 20,000 simulations of unpaired training. Distribution of the simulation results (grayscale), median (blue dotted line), and the 0.05 bottom of the distribution (red line) is shown. Blocks of 10 trials. (B) Behavioral performance during acquisition against performance in the simulations. Percentage of CRs during acquisition in the experiment (black line) plotted against the percentage obtained in the simulations. Distribution of the simulation performances (grayscale), median (blue dotted), and the upper 0.1 of the distribution (red line).

Figure 16

Performance with the adaptive calibration method. (A) Trajectory of the memory parameter w plotted in blocks of 10 trials. Distribution of the performance of the simulations of unpaired CS–US stimulation (grayscale), with mean (dotted blue) and lower 10% (dotted red), for a total of 2500 simulations of 36 blocks. Trajectory of the simulated classical conditioning experiment (solid black), with 18 blocks of acquisition and 18 blocks of extinction. The transition from acquisition to extinction training is marked by a vertical line. (B) Behavioral performance of the same simulations. Percentage of well-timed CRs per block is shown. Results plotted with the same convention as in (A).