Closed-loop isolation of deep-sea extremophiles through in situ microenvironments sensing and preserving

AbstractThis repository contains the comprehensive supplementary dataset and source code supporting the findings of the research article "Closed-loop isolation of deep-sea extremophiles through in situ microenvironments sensing and preserving”, published in Nature Sensor. The provided resources enable the reproduction of the proposed multimodal robotic framework, which integrates environmental perception, admittance-controlled manipulation, and deep learning-based visual servoing. The repository is structured into three primary components: (1) a high-resolution Visual Perception Dataset for training segmentation models, (2) integrated Experimental Data Logs verifying system performance and biological outcomes, and (3) the complete modular Source Code implementation of Algorithms 1–5 described in the manuscript. 1. Visual Dataset for Deep Learning (YOLOv11-seg)A comprehensive collection of approximately ~5,000 high-resolution images capturing microbial plate streaking morphologies on various agar substrates. This dataset was constructed to train, validate, and test the vision-guided robotic system (Algorithm 5) for robust colony instance segmentation. Content: Raw optical captures of microbial colonies, covering challenging scenarios such as dense clusters, overlapping boundaries, and micro-scale growth. Resolution: High-fidelity captures resized to 1280 x 1280 pixels to align with the training hyperparameter (imgsz = 1280). Substrates: Includes five distinct agar types: R2A, ISP2, Postgate, TCBS, GSM-1, and Marine Agar 2216. Format: .jpg (compressed in .zip archive). 2. Supplementary Data Table 1 (Integrated Experimental Logs)An aggregated Microsoft Excel (.xlsx) file organized into ten worksheets, providing the ground truth source data for robotic validation, environmental sensing, and downstream biological analysis: (Sheets 1-2) Robotic System Validation: Haptic feedback logs recording contact forces during the automated plate streaking process, alongside statistical evaluations of kinematic positioning accuracy and overall system stability; (Sheets 3–4) In Situ Environmental Profiling: Real-time physicochemical signals (methane, dissolved oxygen, carbon dioxide, and salinity) acquired during the sampling transect, and hydrostatic pressure dynamics of individual water samplers during the retrieval phase; (Sheet 5) Incubation System Metrics: Operational data for the high-fidelity preservation device, quantifying temporal variations in temperature and pressure, including their respective retention efficiencies and thermal stability percentages; (Sheets 6–7) Microbial Community Structure: Taxonomic datasets presenting the relative abundance of archaea and bacteria at the phylum level, and the normalized relative abundance (Z-scores) of the top 30 microbial genera throughout the incubation period; (Sheets 8–9) Genomic Analysis: Detailed metadata for the study's genome cohort and a functional gene inventory specifically annotated for the deep-sea isolates HP_SR9 and HP_SR10; (Sheet 10) High-pressure Isolate Inventory: Inventory of prokaryotic isolates obtained under high‑pressure in situ conditions using the DISCUSS system. 3. Source Code Algorithms The complete Python implementation of the robotic control and perception framework, modularized into four independent parts corresponding to the distinct subsystems of the platform: Part 1: Multimodal Sensing & Triggering (Module_1)Implements Algorithm 1 (Fusion based on Adaptive Kalman Filter & Dynamic Compensator) and Algorithm 2 (Regime-Aware Probabilistic Sampling Trigger). It handles real-time data ingestion, noise suppression, and autonomous decision-making for sampling initiation. Part 2: Streaking Control System (Module_2)Implements Algorithm 3 (Vision-Guided Compliance Architecture). This module manages the topological trajectory generation (Zigzag/Sector patterns) and the concentration-adaptive admittance controller (Mass-Spring-Damper model) for safe physical interaction with agar surfaces. Part 3: Picking & Perception (Module_3)Implements Algorithm 4 (Vision-Guided Picking). Features visual calibration logic, colony pickability classification (Dual-core consensus), and PID-driven vertical compliance control for precise single-colony isolation. Part 4: Model Training (Module_4)The training pipeline for Algorithm 5 (Deep Learning-Based Instance Segmentation). Contains the configuration and scripts to train the YOLOv11-seg model using the Visual Dataset provided in Section 1. Citation If you utilize the high-resolution microbial colony dataset, the comprehensive in situ data logs, or the robotic algorithms provided in this repository for your research, please cite the following publication: J.-C. Feng. Closed-loop isolation of deep-sea extremophiles through in situ microenvironments sensing and preserving. Nat. Sens., 2025. J.-C. Feng. Closed-loop isolation of deep-sea extremophiles through in situ microenvironments sensing and preserving [Data set]. Zenodo. https://doi.org/10.5281/zenodo.18442830.