Connectivity biases generate a learning hierarchy in the Drosophila mushroom body

Abstract

Learning and memory centers must balance maximizing coding capacity with prioritizing biologically relevant information. Expansion layers, a circuit motif common to many learning and memory centers, including the insect mushroom body, transform dense sensory representations into sparse, distributed ones, and theoretical models propose that random connectivity within these layers maximizes coding capacity by generating highly discriminable responses. Yet this solution creates a fundamental problem: purely random connectivity treats all stimuli equally, without prioritizing survival-relevant cues over neutral ones. Here, we show that the Drosophila melanogaster mushroom body resolves this capacity-selectivity trade-off through systematic biases in projection neuron-Kenyon cell connectivity. Although connectivity is random at the single-cell level, some projection neuron types connect up to 15-fold more frequently than others. These biases translate directly into function: Kenyon cell responses scales with projection neuron connectivity, and the breadth of odor-evoked responses predicts learning performance. Odors activating more than 20% of Kenyon cells drive robust associative memories, whereas those activating fewer than 10% are poorly learned. VL1 projection neurons are a notable exception: despite their weak connectivity, they elicit broad Kenyon cell activity but fail to support learning, revealing a circuit-level gate on learning. These results show that the mushroom body embeds a learning hierarchy in its connectivity architecture, prioritizing ethologically relevant odors while preserving coding capacity for diverse associations.