Inferring neural codes from natural behavior in fruit fly social communication

The cost of experimental neural recordings strongly limits the stimuli or conditions one can present. Ideally, we would therefore like to know how neural responses observed in specific experiments generalize beyond them. E.g., how does the brain’s response to isolated sounds presented in the lab compare to its response to those sounds within natural language in complex social settings? We addressed this problem in fly courtship–a complex, acoustically mediated social behavior. Starting with recorded responses from diverse female neurons to isolated elements of male courtship song, we found these were equally well fit by multiple encoding models. It was unclear, however, if certain models better explained responses to natural song, whose elements unfold in complex sequences. As recording neural responses during courtship or to stimuli adequately spanning natural song space was infeasible, we pursued a novel approach of using the fit encoding models to generate artificial time-varying population responses to songs extracted from a separate, pure-behavior dataset of natural courtship rituals. Reading out the artificial activity to predict future female locomotion extracted from the same rituals let us score each encoding model relative to the behavior data. This revealed one model’s responses–where song inputs were gated by multiplicative adaptation (MA)–predicted locomotion far better than others’, suggesting this model better reflects song coding in natural courtship. Iterating our procedure revealed the MA code relied on slowly integrating, fast-adapting song responses across a multi-timescale neural population, which we found should exhibit a testable signature of accumulator-like natural song responses, and should efficiently transform song into neural activity. This represents a new approach for using behavioral data to gain mechanistic insight into neural coding in natural settings, with which we are presently guiding novel neuroimaging experiments, and sheds new light on how neural population dynamics process temporally complex communication signals.