A Vibe Coding Workflow for AI-Assisted Analysis and Visualization of Genbank Records of Pepper Infecting Viruses

Article information

Plant Pathol J. 2026;42(1):103-108

Publication date (electronic) : 2026 February 1

doi : https://doi.org/10.5423/PPJ.NT.10.2025.0154

Eseul Baek¹, Ho-Jong Ju^1,2, Jiwan Seo^3,, Ju-Yeon Yoon^1,4,5,

¹Department of Plant Protection and Quarantine, Jeonbuk National University, Jeonju 54896, Korea

²Department of Agricultural Biology, Jeonbuk National University, Jeonju 54896, Korea

³School of Computer Science and Engineering, Chung-Ang University, Seoul 06974, Korea

⁴Department of Agricultural Convergence Technology, Jeonbuk National University, Jeonju 54896, Korea

⁵Department of Bioenvironmental Chemistry, Jeonbuk National University, Jeonju 54896, Korea

^*Co-corresponding authors. J. Seo Phone) +82-63-270-4186, FAX) +82-63-270-2531, E-mail) jwseo@ec.cse.cau.ac.kr J.-Y. Yoon Phone) +82-63-270-4188, FAX) +82-63-270-2531, E-mail) jyyoon@jbnu.ac.kr

Handling Editor : Ki-Tae Kim

Received 2025 October 23; Revised 2025 November 27; Accepted 2025 December 13.

Plant viral diseases pose a global threat to plant health and the stability of agricultural systems. A comprehensive understanding of their epidemiological dynamics is essential for protecting agricultural productivity and developing effective management strategies.

Traditional field surveys and individual reports, however, are limited by regional and temporal constraints, making it challenging to capture broader occurrence patterns. The National Center for Biotechnology Information (NCBI) GenBank systematically collects nucleotide sequences submitted by researchers worldwide, along with associated metadata such as collection date, geographic origin, and host species. This makes GenBank a promising resource for spatiotemporal trend analysis (Benson et al., 2012; NCBI Resource Coordinators, 2016). Nevertheless, these data are aggregated in a fragmented manner, and the GenBank interface itself does not provide automated visualization of annual or regional reporting counts. Although PubMed offers a “publication year graph” for visualizing the number of related studies reported by year, this feature reflects only the volume of published articles rather than the number of deposited viral sequences, limiting its utility for direct epidemiological insights. For researchers who wish to collect and analyze these data, the requirement to establish computational pipelines often presents a barrier, as the integration and visualization processes are technically demanding and time-consuming. To overcome these challenges, a vibe coding approach was introduced. Vibe coding refers to a programming paradigm in which research intentions, expressed in natural language (the “vibe,” i.e., the conceptual intent), are translated into executable software modules through interactive dialogue with a generative AI coding agent (Meske et al., 2025). This large language model (LLM)-based interactive development allows non-programmers to rapidly perform complex tasks such as data integration and visualization. Building on the potential of NCBI data, the present study applied the vibe coding approach to visualize global research reporting trends on pepper-infecting viruses and explored its extension into predictive modeling, thereby contributing to agricultural biosecurity and policy decision-making. As a case study, peppers (Capsicum spp.) were analyzed by collecting sequence and metadata records of viruses such as pepper mild mottle virus (PMMoV) and tomato spotted wilt virus (TSWV) deposited in GenBank. These choices are substantiated by epidemiological surveys in Korea indicating that PMMoV and TSWV are among the principal viruses affecting pepper crops (Kwon et al., 2018). Moreover, PMMoV has been recognized as a globally significant threat to Capsicum production, causing severe yield losses (Kumari et al., 2023). An initial pipeline for visualizing spatiotemporal distributions and genetic variation patterns of reported sequences, alongside the construction of a simple predictive model. This framework demonstrates how public GenBank data can be transformed into structured epidemiological insights, and it holds potential for broader applications across diverse viruses and host plants.

This study combined LLMs with a vibe coding approach to develop a system that automatically collects and visualizes genetic and metadata information of pepper-infecting viruses deposited in NCBI, given user-defined host species and temporal ranges. The proposed method is significant in that it enables researchers without programming expertise to perform complex biological data analyses. LLMs are large-scale deep learning models trained on vast textual corpora in a self-supervised manner, capable of understanding context and generating appropriate outputs. Their architecture is based on the Transformer framework (Vaswani et al., 2017), and recent developments have rapidly expanded from traditional natural language processing tasks (Brown et al., 2020; Devlin et al., 2019; Liu and Lapata, 2019) to multimodal and agent-oriented applications spanning images, speech, and human-computer interaction. In addition to general-purpose models like GPT and Gemini, there has been a rise in domain-specific coding assistants that focus on code completion and debugging. Notably, the vibe coding paradigm, first introduced by Andrej Karpathy in 2025, allows users to issue natural language queries that are interpreted and executed as code. This innovation enhances productivity and lowers the barriers to programming.

In this study, natural language queries were used to request code generation, iteratively refining scripts through follow-up queries based on execution results. To implement this workflow, we utilized widely accessible AI agents, specifically OpenAI’s ChatGPT and Google’s Gemini (free tier). This interactive process allowed even those with limited programming experience to complete advanced data analysis workflows effectively. Reliable virus metadata were retrieved through the NCBI Entrez Programming Utilities (E-utilities) API, using the BioPython Entrez library as an interface. A pipeline was built to automatically retrieve sequence information of pepper-infecting viruses reported during the past decade in major East and Southeast Asian producing countries such as Korea, China, Japan, Thailand, and Indonesia. While manual searches through the GenBank web interface are possible, API-based approaches enable automated large-scale data collection and periodic updates. When combined with server-side scheduling, this framework allows for near real-time data acquisition.

Queries to the AI assistant were structured in two modes. The first involved simple prompts (e.g., “Generate a program to collect pepper virus data from NCBI”), which produced basic for-loop scripts. The second mode involved structured prompts specifying objectives, expected outputs, and required parameters (Fig. 1). In these cases, by providing structured prompts, the assistant produced function-based, modular code with adjustable parameters, enabling efficient retrieval of metadata such as nucleotide sequences, collection year, and geographic origin. Errors occurred in both approaches, particularly in retrieving geographic information: although NCBI provides this in the geo_loc_name field, the AI agent initially attempted to access a non-existent country field. In addition, to collect nucleotide sequences encoding specific regions of the coat protein (CP), additional conditions were required: for example, “nucleocapsid protein” in the case of TSWV and “17 kDa protein” for PMMoV. While this issue was resolved through iterative queries, users lacking prior knowledge of the database structure may encounter delays.

Fig. 1

Comparison between simple queries (left) and structured queries when interacting with the AI assistant (right). Simple queries generated basic scripts with limited flexibility, whereas structured queries produced function-based code with adjustable parameters and richer metadata outputs. NCBI, National Center for Biotechnology Information; CSV, comma-separated values.

These findings highlight that domain knowledge remains indispensable when analyzing specialized datasets such as those in bioinformatics. At the same time, they suggest that domain experts without advanced programming skills can nonetheless construct competitive analytical workflows by iteratively interacting with AI coding agents. Providing a structured description of the program’s goals and required conditions further enhances the quality of the generated code. Collected data were organized into Pandas DataFrame structures and exported in CSV and Excel formats, facilitating integration into databases or visualization tools.

Exploratory data analysis (EDA) was performed prior to data visualization to understand of the datasets quality and distributional properties. EDA typically involves inspecting data distributions, detecting missing values and outliers, and summarizing key statistical properties. The same two types of prompts used for pipeline development were applied for EDA. Since LLM-based analyses are context-sensitive, outputs may vary across repeated prompts; however, results in this study were generally consistent. Both approaches produced outputs including Pandas DataFrame summaries, missing value counts, and frequency plots stratified by country, year, and virus species. This consistency likely reflects the standardized nature of the EDA process, which leads to similar outputs even across different contexts.

The dataset comprised approximately 1,700 records, with China contributing the highest number of reports, followed by Korea, with concentrations in 2021 and 2023 (Fig. 2). Such imbalances may introduce risks of reduced generalization performance in predictive modeling. While techniques such as resampling or boosting could mitigate these biases, the present study focused on exploratory visualization rather than model training. Accordingly, no resampling was applied, although data subsets with extremely low sample counts were excluded from certain visualizations to improve interpretability. The collected viral sequences and metadata were visualized in two major ways.

Fig. 2

Temporal and geographic visualization of major pepper viruses, based on metadata records retrieved from NCBI GenBank, with data summarized by country and year. Approximately 1,700 entries reported between 2015 and 2025 are included, highlighting differences in reporting frequency across countries with reported cases in East and Southeast Asia. NCBI, National Center for Biotechnology Information.

First, virus reporting frequencies from 2015 to 2025 were visualized by country and year (Fig. 2). To facilitate interpretation, the top ten most frequently reported viruses were separated from others, and the five most frequently reporting countries identified during EDA were plotted in bar graph format. As in earlier analyses, the process involved iterative collaboration with the AI assistant. Data were searched in real time using the NCBI API, with a final search deadline of December 2025. To ensure consistency and reproducibility, only records containing complete CP sequences were included, and partial genomes or unannotated submissions were excluded.

In Korea, reports of new virus strains discovered in Capsicum spp. have been consistently submitted over the past decade, with the only exception being in 2019. This suggests ongoing surveillance and continued GenBank entry. TSWV was first reported in 2017, peaked in 2018, and continued to be reported in 2022 and 2023. In 2024, Korea reported only one PMMoV strain, one cucumber mosaic virus (CMV) strain, and two additional minor records, with no additional reports in 2024.

China exhibited a broader viral spectrum, with PMMoV and CMV isolates being the most frequently reported in 2015, followed by a variety of pepper-infecting viruses. In 2022, TSWV accounted for the largest proportion in China, a pattern similar to that observed in South Korea in the same year.

Indonesia has consistently reported Pepper yellow leaf curl Indonesia virus (PepYLCIV) over several years, with 108 new isolates recorded in 2025, the highest annual number of reports among the analyzed countries. Notably, this virus was also detected in Thailand in 2017, 2020, and 2025, suggesting a wider geographical distribution. Thailand also reported multiple numbers isolates of Pepper yellow leaf curl virus (PepYLCV), including pepper yellow leaf curl Thailand virus (PepYLCThV).

Second, gene-level analyses were performed to identify hotspots and conserved regions within the CP nucleotide sequence of the same virus (Fig. 3). In these alignments, conserved sites were represented by the predominance of a single nucleotide (A, T, G, or C), whereas variable sites were displayed as mixed distributions across all four nucleotide rows, with low counts shown in violet. Regions of highest variability were shown in the accompanying variation rate plots, allowing easy identification of positions with elevated variation levels. When examining the ten top-ranked nucleotide(nt) positions by variation rate, the major variable sites in TSWV were located at 342, 315, 658, 732, 186, 762, 591, 376, 763, and 738 nt, whereas in PMMoV they were located at 57, 81, 99, 168, 117, 213, 276, 474, 165, and 357 nt (Supplementary Tables 1, 2). These visualizations, generated from multiple sequence alignments performed using MAFFT (Multiple Alignment using Fast Fourier Transform) are valuable for understanding evolutionary features and epidemiological dynamics, and have practical applications in primer design, mutation monitoring, and diagnostic development. Genetic variation analyses were based on organism and coding sequence (CDS) feature annotations from NCBI. Sequence alignments were conducted, mutation rates were calculated, and frequency distributions at specific positions were visualized. This process was iteratively refined with the AI assistant, which provided suggestions and improvements regarding data handling, mutation calculations, and visualization formats. To improve usability, the system was also designed to return consistent results regardless of whether scientific or common virus names were used in the search.

Fig. 3

Alignment-based visualization of nucleotide sequences encoding the coat protein of tomato spotted wilt virus (TSWV; alignment of 363 sequences, upper panel) and pepper mild mottle virus (PMMoV; alignment of 130 nucleotide sequences, lower panel). Conserved positions are represented by the predominance of a single nucleotide (A, T, G, or C), with the dominant base indicated in yellow. Variable positions are displayed as mixtures of nucleotides across the four rows (A, T, G, C), with counts approaching zero shown in violet. Regions of highest variability are shown in the accompanying variation rate plots, where the ten top-ranked sites are marked with black arrows. For each dataset, multiple sequence alignment was performed to compare positional variation across the coat protein coding region, in both panels, respectively Searches were configured to return consistent results regardless of whether scientific or common virus names were used.

In conclusion, we established a foundational system for efficient analysis and visualization of large-scale viral datasets collected from NCBI through an LLM-based vibe coding approach. Collaboration with an AI assistant demonstrated that even abstract queries could yield usable outputs, while more precise, domain-informed prompts enabled the rapid generation of refined analyses. By instructing the AI coding assistant to incorporate data validation procedures in the prompts, high-quality datasets were obtained without requiring advanced programming expertise. Furthermore, the analytical framework automatically integrates newly deposited sequence data from NCBI, eliminating the need for repeated manual updates. Using these data, alignment-based visualizations of viral sequences were generated, and virus reporting frequencies from 2015 to 2025 were mapped by country and year. These outputs enabled spatiotemporal analysis of outbreak patterns. The analytical framework developed in this study was demonstrated using pepper-infecting viruses as a representative case. However, the system is designed to accommodate other major horticultural crops and their associated viruses reported in East and Southeast Asia, allowing users to explore genomic variation and reporting trends across years and geographic regions. The overall web interface is shown in Supplementary Fig. 1.

As increasing volumes of biological information become digitized and more accessible, this approach is expected to reduce barriers of time and expertise in constructing analytical and predictive models in plant pathology. In line with this, recent work has noted that as LLMs may improve their capacity for systematic review tasks, such as identifying relevant publications and extracting specific metadata, they will further enable the integration of GenBank data with published studies and streamlining both discovery and curation (Tao et al., 2025). These findings also underscore the importance of active researcher engagement in reviewing and refining AI-generated outputs to ensure scientific rigor. Overall, LLM-based vibe coding represents a practical and adaptable tool that can enhance research productivity across data collection, analysis, and visualization, ultimately contributing to greater efficiency and scalability in plant virus research.