GENIUS-LLM: An evidence-traceable, gene-centered multi-omics inference framework for plant gene function analysis Authors: Beizhuo Li, Jiaqi You, Jie Zheng, Xianda Zheng, Xianlong Zhang, Maojun Wang, Zeyu Zhang Published: 2026 Dear Editor, Gene function analysis is fundamental to modern agricultural research because it informs targeted molecular breeding, helps decipher complex biological mechanisms, and supports efforts to enhance crop yield and stress tolerance (Liu et al., 2020; Sreenivasulu et al., 2022). However, the integration of rapidly expanding plant multi-omics resources remains challenging. Diverse datasets provide complementary clues, but differences in data format, coverage, and annotation context make integrated analysis labor-intensive. Deep learning methods have improved predictive performance; however, their black-box nature limits interpretability, and their use often requires substantial computational expertise (Wang et al., 2020). Meanwhile, querying existing plant databases usually requires prior knowledge of exact identifiers or trait terms, raising the barrier to actionable biological insight (Zhang et al., 2025). We therefore asked whether a large language model (LLM)-based workflow that organizes multi-omics evidence into traceable reasoning units could generate gene-function hypotheses that are more useful and interpretable than those produced by standard predictive approaches (Wang et al., 2020; Pazos Obregón et al., 2022). To address this question, we developed GENIUS-LLM, a gene-centered multi-omics inference framework. Rather than treating the task as direct label prediction, GENIUS-LLM organizes heterogeneous biological observations into a controlled evidence assembly and reasoning workflow, converting them into an auditable functional report (Figure 1A). Using an architecture inspired by retrieval-augmented generation (Lewis et al., 2020), we organized seven categories of gene-centered information across cotton, rice, and Arabidopsis (Supplemental Table 1). For a queried gene, the system retrieves relevant records, converts them into standardized natural language blocks, and submits them to GPT-4o using a fixed scaffold that constrains the reasoning order and the handling of conflicting evidence (Figure 1B). A central design principle is differential evidence weighting: direct or curated functional cues, such as transcriptome-wide association studies (TWAS) signals and curated GO/KEGG annotations, are assigned higher priority; expression profiles and annotation data serve as validation; and co-expression or sequence similarity provide supporting context (Gusev et al., 2016). This architecture addresses two recurrent problems in practical LLM use: long prompts dilute attention, and loosely organized context causes the model to overemphasize salient but indirect evidence. By fixing retrieval, priority weighting, and reasoning into a reusable scaffold, GENIUS-LLM keeps inference narrow enough to audit while supporting iterative dialogue as users add or update literature. Implementation details and design principles are provided in Supplemental Figure 2 and the supplemental methods.Figure 1Integrated summary of GENIUS-LLM.(A) Conceptual overview of the framework, showing gene-centered multi-omics input, prompt-guided analysis, and feedback-enabled hypothesis generation.(B) Overview of GENIUS-LLM.(C) Benchmark comparison across cotton, rice, and Arabidopsis using precision, recall, AUC, and F1, showing the performance of GENIUS-LLM relative to that of six deep-learning baselines and standalone GPT-4o.(D) Interpretability evaluation of GENIUS-LLM. The bar chart shows the mean scores (1–5) for each interpretability metric across cotton, Arabidopsis, and rice.(E) Performance comparison across different GO levels and combinations.(F) Initial case study inference for Ghir_A09G005620/GhBEE3, illustrating evidence-traceable functional interpretation. The main text further elaborates on the optimized inference results obtained after supplementing information through multiple rounds of dialogue. Integrated summary of GENIUS-LLM. (A) Conceptual overview of the framework, showing gene-centered multi-omics input, prompt-guided analysis, and feedback-enabled hypothesis generation. (B) Overview of GENIUS-LLM. (C) Benchmark comparison across cotton, rice, and Arabidopsis using precision, recall, AUC, and F1, showing the performance of GENIUS-LLM relative to that of six deep-learning baselines and standalone GPT-4o. (D) Interpretability evaluation of GENIUS-LLM. The bar chart shows the mean scores (1–5) for each interpretability metric across cotton, Arabidopsis, and rice. (E) Performance comparison across different GO levels and combinations. (F) Initial case study inference for Ghir_A09G005620/GhBEE3, illustrating evidence-traceable functional interpretation. The main text further elaborates on the optimized inference results obtained after supplementing information through multiple rounds of dialogue. Holding GPT-4o constant as the inference engine within this scaffold, we compared GENIUS-LLM with six deep-learning-based association prediction models, including graph-based models (Heterogeneous Graph Attention Network [HAN], Graph Attention Network [GAT], Heterogeneous Graph Transformer [HGT], and Relational Graph Convolutional Network [RGCN]), non-graph neural network methods (Convolutional Neural Network [CNN] and Multilayer Perceptron [MLP]), and standalone GPT-4o run without the GENIUS-LLM evidence-assembly workflow. Across cotton, rice, and Arabidopsis, GENIUS-LLM achieved the strongest overall performance, ranking first in 11 of the 12 species-by-metric comparisons summarized in Figure 1C. Its advantage was most evident in precision, area under the receiver operating characteristic curve (AUC), and F1-score (F1), and the framework consistently outperformed the best deep-learning baselines in balanced overall performance, with F1 scores reaching 0.6885 in cotton, 0.6444 in rice, and 0.7150 in Arabidopsis. Standalone GPT-4o captured some of the functional signal but was generally weaker and less consistent. These results suggest that GENIUS-LLM’s advantage stems from its ability to organize heterogeneous evidence into a priority-aware reasoning workflow. To systematically assess interpretability, we conducted an independent expert-style assessment of 150 randomly sampled functional reports (50 per species) using a predefined 1–5 scoring rubric. The framework consistently achieved high scores across five dimensions—evidence traceability, reasoning linkage, multi-evidence integration, calibration, and overall coherence (Figure 1D)—supporting the conclusion that it produces logically transparent hypotheses rather than opaque predictions. Detailed scoring criteria and full results are provided in Supplemental Note 1. We next examined whether GENIUS-LLM maintained biological specificity and remained robust when key evidence was incomplete. First, we carried out GO hierarchical analyses across GO levels 3, 4, and 5, using both single- and multi-level settings, to test whether the framework could maintain functional specificity at different ontology resolutions. Across the three species, single-level predictions at GO level 5 generally achieved the highest precision, whereas the level-3/4 combination was often the strongest among hierarchical combinations, especially in cotton (Figure 1E). This pattern suggests that GENIUS-LLM can support both fine-grained inference and context-aware reasoning across neighboring GO levels. Second, we performed ablation experiments by removing individual evidence categories to assess the framework’s robustness to missing evidence and to clarify which evidence types contributed most strongly. Removing annotation or expression evidence generally caused larger performance declines, whereas removing co-expression or BLAST-derived context produced smaller changes and occasionally yielded modest gains (Supplemental Table 2). Together, these analyses indicate that the framework behaves in a biologically consistent way: direct and well-grounded evidence contributes most strongly, whereas indirect association signals are useful but should remain subordinate. To better evaluate the practical value of the framework, we selected a recently cotton gene function study that was published after the release of the backbone model used here, thereby reducing the possibility that the prediction could be explained by prior parametric knowledge alone (Sun et al., 2025). The queried gene in our system, Ghir_A09G005620, corresponds to the experimentally characterized GhBEE3, although the two names derive from different TM-1 annotation versions; the version-matching details are provided in Supplemental Table 3. In the first dialogue round, before the newly published study was provided, GENIUS-LLM inferred that this gene functions as a positive regulator of brassinosteroid signaling and early fiber development, in agreement with the subsequently reported experimental findings (Figure 1F). Importantly, this result was traceable to evidence rather than being generated as an opaque prediction: the report linked the cotton gene to Arabidopsis BEE1 through homology, used GO annotations to support transcription-regulatory activity, and interpreted its developmental role from the observed tissue expression pattern. After the newly published study was introduced in a second dialogue round, GENIUS-LLM further refined the inference by sharpening the temporal interpretation toward the fiber initiation stage while preserving the original core conclusion. This case therefore illustrates that the framework can not only generate an initial evidence-traceable functional hypothesis but also update and improve that hypothesis when new evidence is provided; the full multi-round output is shown in Supplemental Figure 1. In summary, we developed GENIUS-LLM as an evidence-traceable, multi-omics inference framework for plant gene function analysis. Its primary role is to synthesize heterogeneous, gene-centered evidence into auditable functional hypotheses that researchers can inspect, challenge, and test experimentally. In this sense, GENIUS-LLM should be viewed as a hypothesis generator rather than a locus-discovery tool or a replacement for statistical mapping analyses and empirical validation (Hu et al., 2025). Several limitations remain. First, there is currently no widely adopted community benchmark that directly matches the report-style, evidence-integration task examined here. Second, many fine-grained omics data types are not yet directly compatible with the current pipeline and will require additional preprocessing or tool-calling support before they can be incorporated systematically. Third, the present evaluation remains restricted to the species, datasets, and experimental settings examined here; therefore, broader validation across additional plant systems and evidence conditions will be needed to establish the framework’s generalizability. Despite these limitations, the central conclusion remains consistent across benchmark comparisons, GO-hierarchy analyses, ablation experiments, and the GhBEE3 case study: in plant functional genomics, progress depends not only on accessing more data or deploying larger models, but also on organizing heterogeneous evidence into conflict-aware reasoning units that remain interpretable across species and can be updated as new evidence emerges. Data and code availability The original and processed data underlying this study are available at https://doi.org/10.6084/m9.figshare.28269782, and the code supporting this study is publicly available in the following GitHub repository: https://github.com/yimengzhiyan/GENIUS-LLM. Funding This study was supported by the 2025 Hubei Provincial Cotton Coordinated Fund-Supported Industrial Development Project for Biological Breeding and Efficient Cultivation Technologies of Machine-Harvested Cotton after Wheat. Additional support was provided by the Major Project of the Science and Technology Department of Xinjiang (2024A02002), the 10.13039/501100012166National Key Research and Development Program of China (2021YFF1000900), the 10.13039/501100012226Fundamental Research Funds for the Central Universities (BC2025101), and the Xinjiang Leading Talents Program – Key Universities Project (XJRC-2025-BTJY-YJ-GX-QNQZ-001). Acknowledgments We thank the high-performance computing platform at the National Key Laboratory of Crop Genetic Improvement, Huazhong Agricultural University, for computational support. No conflict of interest declared. Author contributions Conceptualization, B.L. and Z.Z.; methodology, B.L.; investigation, Z.Z. and M.W.; data curation, B.L.; formal analysis, B.L.; software, B.L. and J.Z.; visualization, B.L., J.Z., and J.Y.; writing – original draft, B.L.; writing – review and editing, B.L., J.Y., Z.Z., M.W., X. Zhang, and X. Zheng; funding acquisition, Z.Z. and M.W.; supervision, Z.Z., M.W., and X. Zhang; project administration, Z.Z.