Data related to Lillvis JL et al., 2023: Nested neural circuits generate distinct acoustic signals during Drosophila courtship

Data related to Lillvis JL et al. 2023, Nested neural circuits generate distinct acoustic signals during Drosophila courtship.This dataset includes the raw and analyzed audio files for all optogenetic experiments and the ground truth song annotations used to train the model used to segment song in SongExplorer. Details of each can be found below.RAW AND ANALYZED AUDIO FILES:This dataset includes courting male D. melanogaster optogenetic silencing experiments (GtACR_MaleFemalePairs.zip), courting male optogenetic activation experiments (CsChrimson_MaleFemalePairs.zip), and isolated male optogenetic activation experiments (CsChrimson_isolatedMales.zip).The file naming scheme is as follows:date_recordingID_chamber_genotypegenotype e.g., SS46542CHRIM = SS46542>CsChrimson = CsChrimson expressed in pIP10 or SS46542GTACR = SS46542>GtACR1 = GtACR expressed in pIP10.WAV files are the raw audio files-ambient, -mel-ipi, -mel-pulse, -mel-sine, -other, -poly .wav files are the probabilities of each characteristic as classified in SongExplorer-predicted-1.0pr.csv are the ethograms of each recording after thresholding the probabilities for each wordLED2.WAV is the timing and relative amplitude of the 525 nm light (GtACR experiments)the intensities used for GtACR_MaleFemalePairs.zip were (in uW/mm2) 3, 5, 8, 10, 18, 24, 30, 36LED1.WAV is the timing and relative amplitude of the 625 nm light (CsChrimson experiments)the intensities used for CsChrimson_MaleFemalePairs.zip were (in uW/mm2) 2, 4, 6, 8, 17, 25, 34, 42the intensities used for CsChrimson_isolatedMales.zip were (in uW/mm2) 0.3, 0.7, 1, 1.5, 2, 2.4, 2.8, 3.2, 3.7, 4.1, 4.5, 4.9, 5.4, 5.8, 6.2, 8.4, 10.5, 12.7, 14.8, 17, 19, 21, 25, 29, 34, 38, 42GROUND TRUTH DATA FOR SONGEXPLORER MODELThis dataset includes audio .wav files and annotated .csv files used to train the classifier that was used to segment all audio recordings.See https://github.com/JaneliaSciComp/SongExplorer for detailed instructions on training a convolutional neural network and analyzing data using SongExplorer. In brief, the songSegmentationModel_groundTruthData.zip files shared here can be inserted into the groundtruth directory used for training. Input to the model was a 204.8 ms interval of the raw microphone oscillogram centered around points randomly chosen from within each annotated interval, which equates to 1024 ticks in time sampled at 5000 Hz. There were four blocks of layers, each with a 1D convolution, a kernel size of 128, eight feature maps, a ReLU non-linearity, and 50% dropout. The last three layers were also strided by two so as to downsample the output resolution by eight to 625 Hz. Notably, the network did not explicitly create a spectrogram or use a Fourier transform. Output were six taps which represented the predicted probability of the labelled classes (mel-pulse, mel-sine, mel-ipi, poly, ambient, other). In total, the model had 23,630 parameters, all of which were trainable. Training used a batch size of 32, a learning rate of 1e-6, and the Adam optimizer. Ten percent of the data were withheld for validation, and the number of training steps, 4 million, was chosen to ensure that the validation accuracy plateaued.The version of SongExplorer used was dated 20 Feb 2022, and corresponds approximately to commit dce7cb87cb8f in the git repository. The exact values of the model parameters that were learned and used are contained in the included PB file (songSegmentationModel.zip). The model used in Lillvis JL et al., 2023 (songSegmentationModel.zip) will not work in the latest version of SongExplorer. However, training a new model using this data (songSegmentationModel_groundTruthData.zip) and similar parameters to those indicated above will generate a model with similar ability to segment D. melanogaster song.本数据集关联Lillvis JL等人2023年发表的《嵌套神经环路在果蝇求偶过程中产生不同声学信号》（Nested neural circuits generate distinct acoustic signals during Drosophila courtship）研究。数据集包含所有光遗传（optogenetic）实验的原始与分析后音频文件，以及用于训练SongExplorer音频分段模型的真实标注歌曲数据。各部分详情如下。 ### 原始与分析后音频文件 本数据集包含求偶雄性黑腹果蝇（Drosophila melanogaster）的光遗传沉默实验数据（GtACR_MaleFemalePairs.zip）、求偶雄性光遗传激活实验数据（CsChrimson_MaleFemalePairs.zip），以及孤立雄性个体光遗传激活实验数据（CsChrimson_isolatedMales.zip）。 文件命名规范如下：日期_录制ID_实验腔_基因型/基因型 示例：SS46542CHRIM = SS46542>CsChrimson，即pIP10脑区表达CsChrimson；或SS46542GTACR = SS46542>GtACR1，即pIP10脑区表达GtACR1。 .wav格式文件为原始音频文件；-ambient、-mel-ipi、-mel-pulse、-mel-sine、-other、-poly.wav文件为SongExplorer分类得到的各特征概率；predicted-1.0pr.csv为对各录制样本的概率进行阈值处理后得到的行为谱图（ethograms）。LED2.WAV记录了525 nm光的时序与相对振幅（对应GtACR实验）；GtACR_MaleFemalePairs.zip所用的光照强度（单位：μW/mm²）为3、5、8、10、18、24、30、36。LED1.WAV记录了625 nm光的时序与相对振幅（对应CsChrimson实验）；CsChrimson_MaleFemalePairs.zip所用的光照强度（单位：μW/mm²）为2、4、6、8、17、25、34、42；CsChrimson_isolatedMales.zip所用的光照强度（单位：μW/mm²）为0.3、0.7、1、1.5、2、2.4、2.8、3.2、3.7、4.1、4.5、4.9、5.4、5.8、6.2、8.4、10.5、12.7、14.8、17、19、21、25、29、34、38、42。 ### SongExplorer模型训练用真实标注数据 本数据集包含用于训练音频分段分类器的.wav音频文件与标注.csv文件，该分类器用于对所有音频录制样本进行分段。有关训练卷积神经网络（convolutional neural network）及使用SongExplorer分析数据的详细说明，请参见https://github.com/JaneliaSciComp/SongExplorer。简言之，本数据集提供的songSegmentationModel_groundTruthData.zip文件可直接放入训练所用的真实标注目录中。 模型输入为以每个标注区间内随机选取的点为中心的204.8 ms原始麦克风示波图（oscillogram）区间，对应5000 Hz采样率下的1024个时间刻度。模型包含四个卷积块，每个块均含一维卷积、卷积核尺寸128、8个特征图、ReLU非线性激活函数及50%的Dropout正则化；后三层均采用步长为2的卷积，将输出分辨率降采样至625 Hz（原分辨率降为1/8）。值得注意的是，该网络未显式生成声谱图或使用傅里叶变换。模型输出共6个分支，分别对应标注类别（mel-pulse、mel-sine、mel-ipi、poly、ambient、other）的预测概率。该模型总计包含23630个可训练参数。训练采用批次大小32、学习率1e-6及Adam优化器（Adam optimizer）；10%的数据被预留作为验证集，训练步数设为400万步以确保验证准确率趋于平稳。 本次实验所用SongExplorer版本日期为2022年2月20日，大致对应Git仓库中提交号dce7cb87cb8f的版本。所学得的模型参数精确值包含在附带的PB文件（songSegmentationModel.zip）中。Lillvis JL等人2023年研究中使用的模型（songSegmentationModel.zip）无法在最新版SongExplorer中运行，但使用本数据集（songSegmentationModel_groundTruthData.zip）及上述类似参数训练得到的新模型，可实现相近的黑腹果蝇音频分段效果。