henhaobio
Sign in

Molecular PlantPlant BiologyPeer-reviewed journal article

PlantScience.ai: An LLM-Powered Virtual Scientist for Plant Science

Haopeng YuiD, Shasha Zhou, Mingyu Huang, Ling Ding, Yuxuan Chen, Yinru Wang, Yingyu Ren, Nuo Cheng, Xinya Wang, Jie Liang, Huakun ZhangiD, Yiliang DingiD, Ke Li

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

The accelerating growth of plant science literature presents a major challenge for researchers seeking to extract accurate, up-to-date knowledge from an increasingly fragmented and domain-specific corpus. General-purpose large language models (LLMs), while powerful, often misinterpret plant science terminology and lack mechanisms for source traceability. We created PlantScience.ai, a virtual plant biology scientist powered by our automated scientific knowledge graph construction pipeline (AutoSKG). PlantScience.ai exhibits expert-level reasoning in plant biology and maintains scholarly rigour in its citations. Through continuous learning, it integrates the latest research, ensuring that its knowledge base remains current and scientifically robust. Apart from providing the answers to the scientific questions, PlantScience.ai can interact with human scientists, follow instructions, and retrieve information with citation awareness, grounding each response in primary sources to ensure accuracy and verifiability. PlantScience.ai marks a pivotal advance toward a collaborative scientific paradigm in which virtual and human plant scientists work synergistically to accelerate discovery while preserving the unique value of human insight. PlantScience.ai is available at https://plantscience.ai.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (28 Oct 2025) · CC BY 4.0

PlantScience.ai couples an LLM-built, continuously updated plant science knowledge graph (4.0 million entities, 10.8 million edges) with graph-based retrieval for citation-traceable plant biology answers.

General-purpose LLMs invent citations and misread plant-specific terms and abbreviations (for example ABA, or FT and CO), and existing plant LLM agents such as PlantGTP and SeedLLM depend on static, species-focused databases. PlantScience.ai keeps its knowledge current across broad plant science, links every answer to its sources, and stores chat history only on the user's device.

Key results

  1. ShownThe automated AutoSKG pipeline grew the plant science knowledge graph (PSKG) steadily from June to October 2025. Part of the growth came from merging PlantConnectome, a published knowledge graph built from the literature, into PSKG.

    June 2025: ~1.2 million entities, 4.0 million edges; September 2025: 1.8 million entities, 6.5 million edges; October 2025, after integrating PlantConnectome (built from ~71,000 articles): 4.0 million entities, 10.8 million edges; ~1,000 knowledge categories, expanded dynamically. · Main; Construction of Plant Science Knowledge Graph

  2. SuggestedDomain experts rated PlantScience.ai above three general-purpose LLMs on factual accuracy, relevance and comprehensiveness when answering plant science questions.

    60 questions (molecular biology, genetics, physiology, pathology) curated by JIC/TSL senior researchers; compared with Claude 3.5 Sonnet, GPT-4o and Llama 3.1 8B under default settings; anonymised, randomly ordered answers scored 1-10, then normalised 0-1 within each question and evaluator; t-tests (P < 0.05 to P < 0.001). The text gives no mean scores. · Main; Benchmark Design and Evaluation; Figure legends

  3. ClaimedGraph-based retrieval builds a knowledge chain from the query's entities through intermediate nodes and typed relations. This lets each answer be traced to entity descriptions and original sources, and shown as an interactive subgraph.

    GPT-4o finds entities in the query; voyage-3-large embeddings rank nodes by cosine similarity; seed nodes plus immediate neighbours, then semantically similar nodes, then k-hop expansion with Dijkstra shortest paths. No retrieval ablation or citation-accuracy metric is reported. · Main; Graph-based RAG; Figure legends

  4. ClaimedAutoSKG pulls entities and relations from text by prompting an LLM over overlapping windows, adds critique rounds meant to reduce missed or hallucinated entities, and keeps hashed file IDs so every item can be traced to its source.

    GPT-4o-mini reads 1,500-token windows with 200-token overlap, guided by five expert-written examples; follow-up checks for coverage and hallucination; synonyms merged; GPT-4o assigns node categories. Extraction precision and recall are not quantified. · AutoSKG pipeline

  5. ClaimedThe knowledge base is refreshed two ways: user queries trigger web searches for new papers, and users can upload their own findings; AutoSKG processes both. The system was refined over more than a year of expert testing.

    New material is processed in weekly batch updates; uploads go through plantscience.ai/upload; about 150 scientists from JIC and TSL took part in internal testing. The trial's results are not quantified. · Main; Construction of Plant Science Knowledge Graph

Methods, data and reproducibility

AutoSKG prompts GPT-4o-mini to extract entities and relations over overlapping text windows, with critique rounds, then summarises source passages into descriptions and incrementally updates a plant science knowledge graph stored in Neo4j. To answer a query, GPT-4o finds its entities, voyage-3-large embedding search and k-hop/Dijkstra path expansion retrieve a knowledge chain, and an LLM writes the answer from it.

  • Plant Science Knowledge Graph (PSKG), built by AutoSKG from tens of thousands of full-text, peer-reviewed plant science articles · Arabidopsis, wheat, rice, maize, and a broadening range of plant taxa · n = 4.0 million entities and 10.8 million edges (Oct 2025)
  • Seed datasets contributed by John Innes Centre and Sainsbury Laboratory researchers and curated by experts before AutoSKG processing · Arabidopsis and crops (wheat, rice, maize)
  • PlantConnectome, a published LLM-derived plant knowledge graph, merged and reconciled into PSKG · plants · n = approximately 71,000 research articles
  • Expert-validated benchmark questions on plant molecular biology, genetics, physiology and pathology (Table S1) · plants · n = 60 questions
Methods & tools
AutoSKG (automated scientific knowledge graph pipeline), GPT-4o-mini (entity and relation extraction), GPT-4o (query entity recognition, node categorisation, benchmark baseline), voyage-3-large embeddings (Voyage AI), Neo4j graph database, Dijkstra's algorithm (k-hop shortest-path expansion), Optical character recognition (OCR), FastAPI, Vue.js 3 with Pinia, Axios and ECharts, PlantConnectome, Claude 3.5 Sonnet (benchmark baseline), Llama 3.1 8B (benchmark baseline)
Software & environment
Uses GPT-4o-mini (extraction), GPT-4o (query entity recognition, node categorisation), voyage-3-large embeddings, Neo4j, FastAPI and Vue.js 3; no library versions or containers are given.
Reproducibility gaps
Only the AutoSKG code is released. The text offers no download for the PSKG graph, the retrieval or server code, or the benchmark answers and scores. The pipeline depends on commercial APIs (GPT-4o, GPT-4o-mini, voyage-3-large).
Limitations (2)
  • The 60-question benchmark text gives no mean scores, effect sizes, evaluator count, rater identity or inter-rater agreement. Questions were curated by JIC/TSL researchers, whose institutes also helped test the system. Four criteria are mentioned but only three defined. · noted in this summary
  • No metrics are given for how accurately AutoSKG extracts entities and relations, or for what graph-based retrieval adds over plain retrieval. Web-searched and user-uploaded content enters PSKG automatically, and no validation step is described. · noted in this summary

Read this if you want a design for a literature-grounded, continuously updated knowledge-graph retrieval assistant in plant biology, or plan to build private knowledge graphs with the open AutoSKG code.

Citations
7
FWCI
14.36
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISItharajula '26Lim '25Xiao '24Burda '25Wu '26D’Souza '25
76% semantic matchBioinformatics
Manoj Itharajula, Shan Chun Lim, Marek Mutwil·bioRxiv·Apr 6, 2026