2 Methods

2.1 Architecture and design

The microbiONT package is developed using Python and employs the Streamlit (https://github.com/streamlit/streamlit) framework for its graphical user interface (GUI). The platform is designed with a local-execution architecture, ensuring that all computational tasks, including sequencing data analysis and AI inference, are conducted exclusively on the host machine. This design prioritizes data privacy and operational capability in offline environments. The system utilizes Python’s subprocess module to manage the execution of underlying bioinformatics binaries, offering a unified interface that integrates command-line tools into a graphical environment. The general architecture of microbiONT is illustrated in Fig. 1.

Figure 1: The general architecture of microbiONT. Overview of the integrated analysis pipeline and user interface components. (A) Automated “One-Click” workflows for routine analysis; (B) An interactive AI Assistant for natural language guidance and code generation; (C) Modular customization options for adjusting specific analysis parameters; and (D) A selection of 16S or 18S rRNA gene reference databases. The central flowchart illustrates the standardized processing steps from basecalling to format conversion utilizing established bioinformatics tools.

2.2 Standardized analysis workflow

The primary analytical function is accessible through a standardized procedure that has been tailored for routine 16S/18S amplicon sequencing. This automated pipeline executes a predetermined sequence of processing steps (Fig. 1): (i) Basecalling: Raw Pod5 files are processed using Dorado v1.1.1 (https://github.com/nanoporetech/dorado), utilizing the SUP (Super-accurate) model to enhance sequence accuracy. The resulting BAM files are converted to FASTQ format using SAMtools v1.22.0 (Li et al. 2009). (ii) Filtering: Reads are subjected to quality score (Q-score) and length filtering via NanoFilt v2.8.0 (De Coster et al. 2018) to eliminate low-quality and chimeric sequences. (iii) Demultiplexing: Adapter removal and sample demultiplexing are performed using Porechop v0.2.4 (https://github.com/rrwick/Porechop) to segregate barcoded samples. (iv) Quality control: Sequencing statistics and quality metrics are visualized using NanoPlot v1.46.2 (De Coster and Rademakers 2023) to facilitate performance assessment. (v) Taxonomy: Taxonomic assignment is conducted using Emu 3.5.1 (Curry et al. 2022), an efficient classifier optimized for 16S full-length reads. To accommodate diverse research needs, microbiONT provides a multiple selection of pre-configured reference databases derived from various authoritative sources, covering both 16S and 18S rRNA genes. For 16S analysis, the platform includes both the default Emu database (accessed May 2023) and the NCBI 16S gene database (accessed May 2025). Furthermore, microbiONT incorporates two 18S rRNA gene databases targeting invertebrates and broader microbial groups (spanning Chromista, Plantae, and Protozoa), which were sourced from the MetaZooGene Barcode Atlas and Database (MZGdb, accessed January 2023) (Bucklin et al. 2021).

While microbiONT is equipped with an 18S rRNA gene database to facilitate eukaryotic profiling, we emphasize that the 18S analytical pipeline has not been rigorously benchmarked for quantitative accuracy in this study. The primary limitation stems from the extreme scarcity of standardized eukaryotic mock communities and publicly available datasets with established ground truths. Furthermore, quantitative 18S analysis faces intrinsic limitations: extreme intra- and inter-species gene copy number variations among microeukaryotes (Gong and Marchetti 2019), and the relative scarcity of curated eukaryotic reference databases compared to prokaryotic ones (Guillou et al. 2013, Bucklin et al. 2021). Consequently, we recommend that users currently utilize the 18S module primarily as an exploratory tool for qualitative community assessment, and interpret the resulting relative abundance metrics with appropriate caution.

This allows users to select the most appropriate reference dataset based on their specific biological context. (vi) Format conversion: Specialized Python scripts automatically convert taxonomy results into TSV abundance/taxonomy tables and Newick phylogenetic trees compatible with MicrobiomeAnalyst (Lu et al. 2023), as well as appropriately formatted count tables and representative FASTA sequences for functional prediction tools including PICRUSt2 (Douglas et al. 2020) and FAPROTAX (Louca et al. 2016).

2.3 User interface and pipeline flexibility

To complement the standardized workflow, microbiONT offers extensive operational flexibility through a user-friendly Streamlit interface. The GUI is structurally organized with a sidebar for module selection and a comprehensive main panel for execution monitoring, where real-time processing logs provide immediate feedback on workflow status while simultaneously hosting the interactive AI chat interface (Supplementary Fig. S1, available at supplementary material Bioinformatics Advances online). In addition to the standardized One-Click workflow (Supplementary Fig. S2A, available at supplementary material Bioinformatics Advances online), users can customize parameters for each processing step (i–vi) directly through the interface, allowing each module to be operated independently (Supplementary Fig. S2B, available at supplementary material Bioinformatics Advances online). Furthermore, the platform features a Custom Pipeline mode for assembling tailored workflows (Supplementary Fig. S2C and D, available at supplementary material Bioinformatics Advances online). For advanced requirements, an integrated command-line interface allows users to directly modify and execute raw commands (Supplementary Fig. S2E, available at supplementary material Bioinformatics Advances online). Additionally, users can submit analysis requests using natural language; the AI assistant generates responses and executable commands, assisting users with parameter explanations and troubleshooting alongside their active analysis tasks (Supplementary Fig. S3, available at supplementary material Bioinformatics Advances online).

2.4 AI integration and system deployment

microbiONT incorporates a local Large Language model (LLM), Llama 3.1 (https://ai.meta.com/blog/meta-llama-3/), through the Ollama framework (https://github.com/ollama/ollama). To optimize performance for specific analytical tasks, the model is configured using a custom Modelfile tailored to microbiONT architecture. This module interprets natural language queries to offer technical explanations of analysis parameters and bioinformatics guidance, functioning as an interactive reference system.

To evaluate the reliability of the AI assistant, we conducted a rigorous benchmark using 30 queries across six core workflow categories (e.g. basecalling, quality control, taxonomy). Following standard LLM evaluation methodologies, accuracy was measured using the pass@k metric (Chen, et al., 2021), assessing both the strict correctness (accurate parameter selection) and completeness (executable syntax) of the generated commands. Validation results indicate that the model achieves an initial accuracy (pass@1) of 71.3% (Supplementary Table S3, available at supplementary material Bioinformatics Advances online). Notably, the use of iterative prompting (pass@5), increases accuracy to over 96%, demonstrating the system’s robust capability to generate correct responses and executable commands even from simplified or variable user prompts (See Supplementary File 1 and Supplementary Table S3, available at supplementary material Bioinformatics Advances online). To further facilitate practical application, we provide a curated set of verified, pass@1 prompt templates derived from this benchmark (Supplementary File 2, available at supplementary material Bioinformatics Advances online). These templates have been optimized for common analytical tasks, allowing users to directly copy or minimally adapt them (e.g. modifying input filenames) for immediate and highly accurate AI execution.

To address challenges related to software compatibility and environment configuration, microbiONT employs a user-friendly, shell script-based installation mechanism designed for simplicity. The deployment process is streamlined into two steps: users execute the install.sh script to automatically establish the required environment and dependencies, and subsequently launch the platform via the runtime script (./run.sh), which initiates the web-based interface.

2.5 Criticality and mitigation of AI-generated errors

Although the customized Modelfile significantly improves prompt adherence, the inherently generative nature of LLMs leaves room for functional errors. While the overall success rate remains high, we have detailed the potential error patterns that may arise. To assess the consequences of these errors and justify the system’s user interface design, the observed AI failures were categorized by their potential impact on data integrity (Supplementary Table S4, available at supplementary material Bioinformatics Advances online). This evaluation distinguishes between “loud” and “silent” failures. Loud errors, such as the generation of non-existent tool parameters (Parameter Hallucinations), represent “safe failures”; the terminal immediately rejects the invalid commands, thereby precluding data corruption. Conversely, a more substantial risk arises from silent errors, specifically defined as Logical Misinterpretations. In such cases, the AI produces syntactically valid but methodologically flawed commands. If executed without manual verification, these commands proceed without system warnings, potentially causing undetected over- or under-correction of reads and subsequent taxonomic misclassifications. To address the risk of silent analytical failures, microbiONT incorporates an AI-assisted, human-in-the-loop framework. By requiring explicit user validation of the generated bash commands prior to execution, the platform ensures that researchers retain continuous technical oversight, preventing logical inaccuracies from propagating into downstream biological interpretations.