Long-term memory performance optimization via Neural network-based curve fitting in Drosophila

Abstract

Long-term memory (LTM) formation typically requires extensive training or highly salient experiences, limiting learning efficiency. Operant conditioning is generally thought to produce stronger memory than classical conditioning because of its active learning component. Unexpectedly, however, laser-based social conditioning in Drosophila melanogaster revealed that while classical paradigms yielded lower short-term memory (STM) scores but higher LTM retention, operant paradigms exhibited higher STM scores followed by rapid LTM decay. To resolve this discrepancy, we employed the AI Complex Systems Response (AI-CSR) framework, which reconstructs high-dimensional learning landscapes from sparse sampling and predicts globally optimal training conditions. Following AI-CSR optimization, operant conditioning produced a twofold increase in LTM scores, yielding the strongest 24-hour social memory performance reported in flies to date and revealing the expected superiority of active learning, which had conversely shown poorer performance under standard training protocols. In contrast, AI-CSR did not further enhance classical conditioning performance but reduced training time by 50%. Single-cell RNA sequencing revealed expanded neuronal recruitment marked by the activation and inhibition of various gene combinations. Together, these findings link circuit-level reorganization with molecular programs underlying efficient long-term memory and demonstrate how AI-guided optimization can uncover latent learning capacity in biological systems.