Itharajula, M. | Search bioRxiv

bioRxiv2026

Research Process Graph: LLM-Driven Extraction and Hierarchical Organization of Research Logic

Plant biology now publishes thousands of experimental research articles each year, but their core research logic, namely what questions are being asked, with what methods, and what is being found, remains locked inside free text and invisible to systematic analysis. Here we present a structured, 20-year atlas of The Plant Cell in which every paper is converted into a typed, directed Research Process Graph (RPG) of Question (Q), Method (M) and Finding (F) nodes connected by Q[->]M and M[->]F edges. A benchmarked large language model pipeline applied to 2,633 Plant Cell research articles published 2005-2026 recovered >110,000 Q/M/F nodes and >126,000 directed Q[->]M[->]F chains with>98% precision. A second LLM pass generalises each node into a paper-independent canonical form and assigns it to one of 10 top-level (L1) and [~]90 sub-level (L2) categories for each node type, producing the first comprehensive map of plant-biology research logic at the resolution of individual research questions. The atlas reveals that Plant Cell papers fall into seven canonical paper recipes with characteristic Q[->]M[->]F sub-structures, that peripheral experimental techniques have largely turned over while a stable methodological core persisted, and that the strongest correlate of per-PI citation impact is methodological breadth, not productivity or topical breadth. We release the atlas as a public, browsable database with five complementary interfaces: paper views, an LLM-powered research assistant, expert profiles, a taxonomy browser, and a method explorer. The database, available at https://rpg.connectome.tools/, turns the literature into a queryable community resource.

plant biology↗

bioRxiv2026

From Parametric Guessing to Graph-Grounded Answers: Building Reliable ChatGPT-like tools for Plant Science

Large language models (LLMs) are increasingly used by plant biologists to summarize literature, generate hypotheses, and interpret experimental results. However, LLMs are unreliable sources of exhaustive, source-attributed facts, a critical limitation for the list-style queries that pervade plant biology (e.g., "list all transcription factors regulating secondary cell wall (SCW) biosynthesis in Arabidopsis"). Here, we query ChatGPT, Claude, and Gemini with such queries and demonstrate that none return complete gene lists with reliable citations. We trace these failures to how LLMs store knowledge: as statistical patterns distributed across billions of internal parameters, with no mechanism to guarantee completeness, provenance, or reproducibility. We also review fine-tuning mitigation strategies, including multi-task instruction tuning, parameter-efficient methods, and context engineering, that alleviate but do not resolve these limitations. We then discuss retrieval-augmented generation (RAG), which feeds relevant documents to the LLM at query time, and argue that while it improves source attribution, it remains impractical when answers require synthesizing information scattered across hundreds of papers. As an alternative, we advocate graph retrieval-augmented generation (GraphRAG), in which the LLM serves as a reasoning and language interface over a structured, provenance-linked knowledge graph (KG) that returns complete result sets reproducibly. We outline a practical GraphRAG architecture and survey existing plant KG resources. Finally, we discuss open challenges, including entity disambiguation, relation normalization and evidence grading, and propose a roadmap for building open, continuously updated plant KGs that can turn "read 1,000 papers" into a single reproducible query.

bioinformatics↗