Sub-second dopaminergic reinforcement orchestrates juvenile social play and is disrupted in Shank3 deficiency. Chen et al. (Enriched version)
Description
This dataset accompanies a study testing whether nucleus accumbens dopamine (DA) acts as an action-contingent reinforcement signal shaping juvenile social play in a sex-divergent manner, and whether Shank3 deficiency disrupts this signal. We collected multi-modal data from wild‑type (WT) and Shank3+/- (PND 35-85) during same-sex dyadic interactions. A nine-camera volumetric system (120 fps) with instance segmentation (Mask RCNN/YOLOv8) and DANNCE tracked 14 body keypoints. The Social-Seq pipeline derived 32 kinematic/interaction features, segmented interactions into 800 ms clips, and used Seq2seq RNN with perspective invariant active learning to generate 36 behavioral syllables (e.g., sniffing, chasing, pouncing, pinning, rearing). Simultaneously, fiber photometry recorded NAc DA dynamics via GRAB-DA3m. In Shank3+/- males, a real-time closed loop system delivered optogenetic stimulation to VTA-NAc projections specifically upon proactive play initiation over 8 training days, followed by 10 stimulation-free days. Key findings: In WT males, proactive play (pouncing, pinning) evoked DA surges, while forced submission suppressed DA. In WT females, DA increased during evasion and rearing but not contact-heavy play. Shank3 mutants showed blunted DA during sniffing/chasing and, critically, a sign-inverted DA response during pouncing , while solitary rearing produced exaggerated DA only in mutant males . A multi-agent reinforcement learning model parameterized with empirical DA amplitudes reproduced mutant phenotypes (preserved sniffing, reduced play, disrupted transitions). Closed-loop DA during play increased targeted play duration persistently and reduced nonsocial leaving, establishing causal sufficiency. Data include raw 3D keypoints coordinates, behavioral syllable labels, synchronized Z-scored DA ΔF/F traces, optogenetic timestamps, and summary metrics (frequencies, durations, transition matrices). Researchers can use these to replicate the ethogram, perform neural‑behavior alignment, parameterize computational models, or evaluate intervention outcomes. Key considerations: sex must be treated as a covariate; behavioral labels have quantitative operational definitions; the Shank3 model is specific to Phelan-McDermid syndrome; and optogenetic results demonstrate sufficiency, not exclusivity, of DA. This multi-scale resource supports studies of developmental social reward and autism-related motivation deficits.
Files
Steps to reproduce
Animals: WT SD and Shank3+/- rats, 5-12 wks, group-housed (12h light/dark, ad libitum) Experiment: （1）Rat fur dyeing : Isoflurane anesthesia, non-toxic blue dye; (2) Photometry recording: AAV-hSyn-DA3m into NAc (AP 1.25, ML 1.85, DV 6.0), 400 nl; fiber 200 μm above. Recording at PND 45-55. (3) Optogenetics: AAV-DIO-ChR2-mcherry and rTH-Cre into VTA (AP 5.5, ML ±1.0, DV 7.85); bilateral fibers above NAc; 470 nm, 40 Hz, 4 mW, PND 45-52. Behavioral tests: Social interaction: 3-day habituation (15 min/day) in cylindrical arena (55 cm dia), 4-6 h isolation, 15 min recording at PND 35-45, 55-65, 75-85. sCPP: two-compartment with different bedding; 10-min pre-test, 5-day conditioning (15 min social vs isolation), 10-min post-test; biased to less-preferred side. Liquid consumption: water-restricted (85-90% BW); 150 μL water/1mM quinine/30% sucrose, randomized, with photometry. Closed-loop optogenetics: baseline 3d, training 8d (behavior-triggered), post 9d. Real-time detection: YOLOv8n-seg + DANNCE + BiLSTM-FCN (36 classes); laser 470 nm, 4 mW, 40 Hz, 0.5-s pulses. Dopamine Recording: Cylindrical arena (55 cm diameter, 50 cm height), red light (20 lux), corncob/soiled bedding. 9 synchronized cameras (1MP, 30 fps, 3840×2400) via OBS, ~3.5 GB/15 min. Calibration: checkerboard (intrinsic) + metal ball tracking (extrinsic) with YOLOv8; code: github.com/chenxinfeng4/multiview_ball_calib. Behavioral analysis: Social-seq pipeline (https://lilab-cibr.github.io/Social_Seq): MaskRCNN; DANNCE pose estimation; jitter correction with SmoothNet. Feature engineering: 32 metrics (kinematics, social distances, sniffing overlaps, occlusions). Clustering: Seq2seq autoencoder → 128D → PCA to 12 PCs (90% variance); K-means (K=20/100) → 42 clusters; UMAP; biLSTM for classification. Manual annotation for reciprocal pairs. Photometry data process: 3 channels 405/480/565 nm LEDs (10 Hz). Signal: median filter, bleach correction, 405 nm regression for motion. Statistics: Behavioral proportions: normalized fold change. Transitions: probability matrices; Kruskal-Wallis + Dunn's with FDR. Dimensionality: PCA/LDA; SVM with LOOCV. DA analysis: align behavior sequence, dompamine signal and 24 motion features. Merge 36 to 10 behavior categories. Response: post-onset peak ΔF/F post-onset minus baseline mean. Correlation analysis: DA-kinematics, DA-duration, DA-antecedent. MARL: 4×4 grid, fully observable (32D one-hot). Action: movement + social intent (none/sniff/pounce/escape) based on Manhattan distance. Rewards calibrated from empirical DA. PPO agents (RLlib), 2×256 tanh layers, branched softmax. Training 120k steps, ε=0.3, entropy=0.01, γ=0.95, lr=3e-4. Evaluation over 30 rollouts. Detailed information: https://lilab-cibr.github.io/Social_Seq/en/figure_reproduce/ !!! Note: Update Fig4S4-Fig6S6_data_update to Social_Seq_DATA to fit pandas2.0 scripts in the figure_reproduce link.
Institutions
- Chinese Institute for Brain ResearchBeijing, Beijing