University:
Boston University
Course:
CAS BI 108 | Biology 2
Academic year:

2024
Views:

177

Pages:

1
Author:

Jarrett Richard

The Impact of Structural Changes on Learning Capacity in the Fly Mushroom Body Circuit Introduction The Drosophila mushroom body (MB) is involved in olfactory learning and memory. The MB network is comprised of projection neurons (PNs), Kenyon cells (KCs), and mushroom body output neurons (MBONs). ● The synaptic plasticity of the KC to MBON synapses plays a role in the 1 learning process . ● Our question: How do perturbations to the structure and connectivity of the KC to MBON neural circuit affect the MBONs’ ability to learn an odor classification task? Results Discussion/Conclusions We ran two types of ablation experiments (removing 1, 5, 10, 20, and 30 KCs), with ablating 0 KCs as a control: 1. Randomly removing a number of presynaptic KCs from each MBON. 2. Removing presynaptic KCs with the highest synaptic weight values from all MBONs. Figure 3: Average and SD of Error Rate over Time for 4 different MBONs Certain MBONs’ final average error rates (ER) were higher than others, and those MBONs struggled more to learn the task. Figure 6: Random Pruning - Number of KCs Removed vs. the Difference in the Average Final Error Rate after Pruning and without Pruning Figure 4: Number of KC Connections vs. Average Final Error Rate MBON-n1 left has only three presynaptic KCs. We hypothesized that the number of KCs may affect MBON performance. Figure 1: PN to KC to MBON Neural Network Model Approximation Figure 7: Targeted Ablation - Number of KCs Removed vs. the Difference in the Average Final Error Rate after Ablation and without Ablation This neural network model is based 2 on the connectivity data from an electron microscopic reconstruction of the mushroom body of Drosophila melanogaster. ● 40 PNs, 110 KCs, 19 MBONs There was only a minor difference in MBONs’ performance after the targeted and random removal. Further experiments with more complicated tasks are needed to determine the difference in impact of the two types of KC removal. Figure 5: Type of KCs Removed in Targeted Ablation Experiments Methods PN Activity ● For the odors, we generated three 40-dimensional Gaussian distributions (absolute values of coordinates), which represent the projection neuron activity Figure 2: 2D Slice of Distributions of PN activity in Response to Three Odors Training the model 3 ● We used the delta rule to adjust our weights at each step (new input point) during the training. Δweight = input * error * learning rate