Figure 3. Acoustic characterization of ultrasonic vocalizations (USVs) elicited by optogenetic activation of preoptic area (POA) neurons.

(A) The variational autoencoder (VAE) takes spectrograms as input (left), maps the spectrograms to low-dimensional latent representations using an encoder (middle), and approximately reconstructs the input spectrogram using a decoder (right). (B) (Left) Dimensionality reduction techniques such as PCA or UMAP can be used to visualize the resulting latent vectors. (Right) Interpolations in latent space correspond to smooth USV syllable changes in spectrogram space exhibiting realistic dimensions of variation. (C) UMAP projections of latent syllable representations of female-directed USVs (red) and optogenetically elicited USVs (blue) from two example mice. (D) Maximum Mean Discrepancy (MMD) was calculated between distributions of latent syllable representations to generate three comparisons: female-directed USVs from the first half of the recording session vs. female-directed USVs from the second half, all female-directed USVs vs. opto-USVs (N = 16 experimental mice), and female-directed USVs from two different recordings sessions in N = 10 control mice (gray points). Note that larger MMD values indicate that distributions are more dissimilar (E) UMAP projections of latent descriptions of female-directed (red) and optogenetically elicited USVs (blue) for all mice (N = 15). (F) UMAP projections from panel E, color-coded by total energy (left) and frequency bandwidth (right). Example spectrograms of opto-USVS and female-directed USVs are plotted below, and the location of each example USV in UMAP space is indicated by the colored dots on the grayscale UMAP projection on the bottom left. See also Figure 3—source data 1.