GRAFT: Biological Graph and Hypergraph Benchmarks for Linked Gene Expression and Phenotypic Trait Prediction in Arabidopsis thaliana

# GRAFT: Biological Graph and Hypergraph Benchmarks for Linked Gene Expression and Phenotypic Trait Prediction in Arabidopsis thaliana
Source: [https://arxiv.org/html/2606.27413](https://arxiv.org/html/2606.27413)
Manuel Serna\-Aguilera1,6,Vanshika Jindal2,Fiona L\. Goggin2,Jiamei Li2, Aranyak Goswami3,Alexander Bucksch4,Suxing Liu5,Khoa Luu1,6, 1Department of Electrical Engineering and Computer Science, University of Arkansas, AR 2Department of Entomology and Plant Pathology, University of Arkansas, AR 3Department of Animal Science, University of Arkansas, AR 4School of Plant Sciences, University of Arizona, Tucson, AZ 5Georgia State University, GA 6CVIU Lab, University of Arkansas, AR \{mserna, vjindal, fgoggin, jxl080, garanyak, khoaluu\}@uark\.edu bucksch@arizona\.edusliu58@gsu\.edu

###### Abstract

Understanding which genes control which traits in an organism remains one of the central challenges in biology\. Despite significant advances in data collection technology, our ability to map genes to traits is still limited\. This genome\-to\-phenome \(G2P\) challenge spans several problem domains, including plant breeding, and requires methods capable of reasoning over high\-dimensional, heterogeneous, and biologically structured data\. Current datasets and data repositories, however, are not well\-equipped for this task\. Current studies do not link gene expression and trait data, and most focus on very specific traits, limiting the breadth of possible correlations\. To address this gap, we present the novelGene\-GraphRegression forArabidopsisFunctionalTraits \(GRAFT\) dataset, a curated multi\-modal dataset linking gene expression profiles with phenotypic trait measurements inArabidopsis thaliana, a model organism in plant biology\. GRAFT supports tasks such as phenotype prediction and interpretable graph learning\. In addition, we benchmark conventional regression and explanatory baselines, including a biologically\-informed hypergraph baseline, to validate gene\-trait associations\. To the best of our knowledge, this is the first dataset to provide multimodal gene information and heterogeneous trait or phenotype data for the sameArabidopsis thalianaspecimens\. With GRAFT111All benchmark resources will be made publicly available upon acceptance\., we aim to foster research to accurately understand the relationship between genotypes and phenotypes using gene information, higher\-order gene pairings, and trait data from multiple sources\.

## 1Introduction

Of the thousands of genes in an individual’s genome and the hundreds of traits it displays, from their height to their health, which genes control which traits? The answers to this question are essential to nearly all applied life sciences, from crop improvement to animal breeding and medical drug development\. Unfortunately, our ability to provide answers remains quite limited due to data analytics constraints\. Decoding the relationship between an organism’s genetic makeup,i\.e\., its genome, and its traits,i\.e\., its phenome–the total of its different phenotypes, requires identifying complex patterns that interlink multiple high\-dimensional and heterogeneous datasets\. Unfortunately, there is a lack of benchmarking datasets to facilitate the development of computational tools, particularly in plant science\. This hinders plant breeders’ efforts to meet increasing global food demands and combat emerging pests, diseases, and droughts\.

Limitations of Prior Work\.Research to crack the “genome\-to\-phenome” \(G2P\) challenge increasingly relies on high\-dimensional and heterogeneous data to capture as many different aspects as possible of an individual specimen’s genome and phenome\. Plant phenomic data typically has fewer features but is highly heterogeneous, ranging from images of shape and size to manual observations of development to spectrophotometric measurements of processes like photosynthesis\. Combining data from more than one “omics” approach \(i\.e\., multi\-omics\) is more effective at linking genes to specific phenotypes \(observable traits\) than any single omics approach alone[M\. Minervini, A\. Fischbach, H\. Scharr, and S\. A\. Tsaftaris \(2016\)](https://arxiv.org/html/2606.27413#bib.bib94);[59](https://arxiv.org/html/2606.27413#bib.bib95);[68](https://arxiv.org/html/2606.27413#bib.bib96);[D\. Ward, P\. Moghadam, and N\. Hudson \(2019\)](https://arxiv.org/html/2606.27413#bib.bib97);[71](https://arxiv.org/html/2606.27413#bib.bib98);[5](https://arxiv.org/html/2606.27413#bib.bib99);[U\. Seren, D\. Grimm, J\. Fitz, D\. Weigel, M\. Nordborg, K\. Borgwardt, and A\. Korte \(2017\)](https://arxiv.org/html/2606.27413#bib.bib100);[58](https://arxiv.org/html/2606.27413#bib.bib101)\. Despite this, different types of plant omics data are siloed in separate repositories, as shown in Table[1](https://arxiv.org/html/2606.27413#S2.T1), and there is a lack of comprehensive datasets that combine genomic and phenomic profiles from the same individuals to enable correlational analyses\. Due to this lack of benchmarking data, the machine learning community has not kept up with the challenges of analyzing multi\-omics data\. Although there is much discussion in the biological literature about AI models that could infer correlations between samples in omics dataCavillet al\.\([2015](https://arxiv.org/html/2606.27413#bib.bib81)\); Cembrowska\-Lechet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib82)\); Demidchiket al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib83)\); Yanget al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib84)\); Gaoet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib85)\); Depuydtet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib86)\); Yan and Wang \([2023](https://arxiv.org/html/2606.27413#bib.bib87)\); Zhanget al\.\([2022](https://arxiv.org/html/2606.27413#bib.bib88)\); Ali and Mohammed \([2023](https://arxiv.org/html/2606.27413#bib.bib89)\); Baiet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib90)\), no such models seem to exist\. Therefore, there is a critical need for biologically informed models that can map gene expression to heterogeneous phenotypes while enabling investigations into their explainability\.

Problem Motivation\.To address knowledge gaps and limitations in both fields, we present theGene\-GraphRegression forArabidopsisFunctionalTraits \(GRAFT\), a dataset containing genomics \(gene measurements\) and phenomics \(traits\) datalinked to the same specimens, specifically of the foundational or “model” speciesArabidopsis thaliana\. To go beyond gene\-to\-trait regression, GRAFT maps each gene to thousands of biological functions\. Compared to other datasets, as shown in Table[1](https://arxiv.org/html/2606.27413#S2.T1), GRAFT provides linked measurements not only for gene expression but also for many other traits typically examined in entire studies\. To build biologically informed baselines, we use biologically\-informed graphs and hypergraphs to representnn\-order relationships among genes\. We first benchmark regression models, and together with our team of biology experts, we explore SHapley Additive exPlanations \(SHAP\)Lundberg and Lee \([2017](https://arxiv.org/html/2606.27413#bib.bib22)\), graph explanationsYinget al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib24)\); Luoet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib25)\); Zhanget al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib28)\), in collaboration with biology experts, to construct a pipeline that aims to assist biologists in narrowing down top genes out of hundreds or thousands\. Such a framework is beneficial in cutting labor costs and helping plant breeding efforts to zero\-in on a small subset of important genes\.

Contributions of this Work\.We summarize the contributions of this work as follows\.\(i\)We contribute the novel GRAFT dataset, a collection oflinkedgene expression and trait data\.\(ii\)GRAFT additionally provides gene and biological function annotations that connect all genes, which we can interpret as hypergraphs\.\(iii\)We provide insight into the explainability of the baseline regression methods, both quantitatively with our proposed biological explanation recall \(BER\) metric, and expert discussion\.\(iv\)We will publicly release both our data and explanatory framework to foster further work into tackling the G2P challenge\.

## 2Background and Related Work

Table 1:Comparison of publicly\-available datasets or data repositories, which lack diverse, multi\-omics components compared to our dataset—GRAFT\. GRAFT includes gene\-level and heterogeneous phenotype\-level information, while other common datasets and repositories do not\. A blue check✓indicates the dataset contains the corresponding measurement type, while✗indicates otherwise\.![Refer to caption](https://arxiv.org/html/2606.27413v1/figures/our-phenotype-data3.png)Figure 1:The problem that non\-heterogeneous phenomics data poses to downstream tasks\.\(a\)ModelM1M\_\{1\}is trained on homogeneous image\-derived phenotypes on genes\{v1,…,vN\}\\\{v\_\{1\},\\ldots,v\_\{N\}\\\}, and as a result cannot reason across other phenotypes\. This is the inherent limitation of current datasets and benchmarks\.M2M\_\{2\}in case\(b\)suffers from similar limitations asM1M\_\{1\}, so too doesM3M\_\{3\}in\(c\)\.\(d\)With the breadth of phenotypes provided by GRAFT,M4M\_\{4\}learns to correlate genes to heterogeneous traits of the specimens\.Best viewed with zoom and in color\.### 2\.1The Model PlantArabidopsis thaliana

Arabidopsis thaliana, commonly known as the thale cress, is a model plant, much like rats and mice are “model species” that can help us learn about human health\. The thale cress is widely used to study important crops such as corn and riceWoodward and Bartel \([2018](https://arxiv.org/html/2606.27413#bib.bib71)\)\. The thale cress’ relatively small genome and short lifecycle allowed for substantial research progress in the past several decades, making this species a natural choice for this work\. In biology research, it is common to study mutant variants, often termed genotypic “lines” where at least one gene is mutated, leading to an alteration of some biological function, thereby motivating multi\-omics studiesCavillet al\.\([2015](https://arxiv.org/html/2606.27413#bib.bib81)\); Cembrowska\-Lechet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib82)\)\.

### 2\.2Multi\-Omics Analysis and the State of Omics Datasets

“Omics” or “multi\-omics” data refers to measurements from different biological systems in an organismCavillet al\.\([2015](https://arxiv.org/html/2606.27413#bib.bib81)\); Cembrowska\-Lechet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib82)\)\. Transcriptomicsmeasures gene expression patterns, and RNA sequencing \(RNA\-seq\) technology allows scientists to quantify almost all gene expressionsMansooret al\.\([2025](https://arxiv.org/html/2606.27413#bib.bib74)\)\. Genomics, is the study of the structure, function, and evolution of genomes within a given organism or community of organismsHeaveyet al\.\([2022](https://arxiv.org/html/2606.27413#bib.bib80)\)\. Phenomicsmeasures observable traits \(examples in Section[3](https://arxiv.org/html/2606.27413#S3)\)\. Omics data provides researchers with insight into which genes are activated during developmental stages or in response to specific environmental stimuli\. Modern research is interested in multi\-omics data because they provide a fuller picture of genes and traits, as shown in Fig\.[1](https://arxiv.org/html/2606.27413#S2.F1)\. Cembrowska et al\.Cembrowska\-Lechet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib82)\)and many surveysDemidchiket al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib83)\); Yanget al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib84)\); Gaoet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib85)\); Depuydtet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib86)\); Yan and Wang \([2023](https://arxiv.org/html/2606.27413#bib.bib87)\); Zhanget al\.\([2022](https://arxiv.org/html/2606.27413#bib.bib88)\); Ali and Mohammed \([2023](https://arxiv.org/html/2606.27413#bib.bib89)\); Baiet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib90)\)provide further insight into omics research\.

While high\-throughput omics technology has improved considerably, most existing datasets fail to providelinkedmulti\-omics data, forcing studies to operate on limited dataChenget al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib76)\); Floodet al\.\([2016](https://arxiv.org/html/2606.27413#bib.bib73)\); Ubbens and Stavness \([2017](https://arxiv.org/html/2606.27413#bib.bib75)\)\. Major repositories \(Table[1](https://arxiv.org/html/2606.27413#S2.T1)\) often contain only phenotype or only expression data, but rarely both\. Some datasets only provide image\-derived data[M\. Minervini, A\. Fischbach, H\. Scharr, and S\. A\. Tsaftaris \(2016\)](https://arxiv.org/html/2606.27413#bib.bib94);[59](https://arxiv.org/html/2606.27413#bib.bib95);[68](https://arxiv.org/html/2606.27413#bib.bib96);[D\. Ward, P\. Moghadam, and N\. Hudson \(2019\)](https://arxiv.org/html/2606.27413#bib.bib97);[71](https://arxiv.org/html/2606.27413#bib.bib98), or phenotype measurements[5](https://arxiv.org/html/2606.27413#bib.bib99);[U\. Seren, D\. Grimm, J\. Fitz, D\. Weigel, M\. Nordborg, K\. Borgwardt, and A\. Korte \(2017\)](https://arxiv.org/html/2606.27413#bib.bib100), or photosynthetic data[58](https://arxiv.org/html/2606.27413#bib.bib101), or gene\-level data[30](https://arxiv.org/html/2606.27413#bib.bib102);[64](https://arxiv.org/html/2606.27413#bib.bib103);[R\. Edgar, M\. Domrachev, and A\. E\. Lash \(2002\)](https://arxiv.org/html/2606.27413#bib.bib105);[63](https://arxiv.org/html/2606.27413#bib.bib104);[69](https://arxiv.org/html/2606.27413#bib.bib106);[E\. Huala, A\. W\. Dickerman, M\. Garcia\-Hernandez, D\. Weems, L\. Reiser, F\. LaFond, D\. Hanley, D\. Kiphart, M\. Zhuang, W\. Huang, L\. A\. Mueller, D\. Bhattacharyya, D\. Bhaya, B\. W\. Sobral, W\. Beavis, D\. W\. Meinke, C\. D\. Town, C\. Somerville, and S\. Y\. Rhee \(2001\)](https://arxiv.org/html/2606.27413#bib.bib107);[T\. Coxe, D\. J\. Burks, U\. Singh, R\. Mittler, and R\. K\. Azad \(2024\)](https://arxiv.org/html/2606.27413#bib.bib108)\. None of these datasets links multi\-omics data\. Our GRAFT dataset, in contrast, addresses these gaps by offering complete gene expression profiles paired with rich phenotypic traits, enabling direct genotype\-to\-phenotype modeling\. This positions our dataset as a valuable benchmark for developing scalable, integrative machine learning methods\. Details on the datasets from Table[1](https://arxiv.org/html/2606.27413#S2.T1)are provided in the Appendix\.

### 2\.3Graph and Hypergraph Modeling

Graph Convolutional Networks \(GCNs\)Bronsteinet al\.\([2017](https://arxiv.org/html/2606.27413#bib.bib3)\)take simple graphs, represented by node features and an edge list connecting pairs of nodes as its inputs, and produce embeddings of the interactions between nodes,e\.g\., via some message passing aggregation strategy, graph convolutionKipf and Welling \([2017](https://arxiv.org/html/2606.27413#bib.bib5)\); Defferrardet al\.\([2017](https://arxiv.org/html/2606.27413#bib.bib4)\); Morriset al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib6)\), graph attentionVeličkovićet al\.\([2018](https://arxiv.org/html/2606.27413#bib.bib7)\)or transformer operatorShiet al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib8)\)\. Such networks also encode other kinds of data, such as imagesHanet al\.\([2022a](https://arxiv.org/html/2606.27413#bib.bib11)\)or scene graphsNguyenet al\.\([2024b](https://arxiv.org/html/2606.27413#bib.bib51),[a](https://arxiv.org/html/2606.27413#bib.bib50)\)\. Recently, large language models have seen increasing use with knowledge graphs to provide grounded responses on large\-scale databasesChoiet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib13)\); Sunet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib14)\); Chenet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib15)\), or even encoding subgraphs for such modelsPerozziet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib16)\)\. Graphs are also common in biology research for learning gene representations with graph structureLiuet al\.\([2023a](https://arxiv.org/html/2606.27413#bib.bib12)\), a similar but distinct topic to this work\.

Hypergraph Neural Networks \(HGNNs\) utilize hypergraphsBerge \([1989](https://arxiv.org/html/2606.27413#bib.bib34)\); Bolla \([1993](https://arxiv.org/html/2606.27413#bib.bib46)\); Zienet al\.\([1999](https://arxiv.org/html/2606.27413#bib.bib36)\); Rodríguez \([2003](https://arxiv.org/html/2606.27413#bib.bib45)\); Zhouet al\.\([2006](https://arxiv.org/html/2606.27413#bib.bib43)\); Yadati \([2020](https://arxiv.org/html/2606.27413#bib.bib38)\)and encodenn\-ary relations, in which any set ofnnnodes is connected by a hyperedge, rather than binary relations\. In the past decade or so, deep learning has become integrated into hypergraph tasks, motivating the formulation of learnable hypergraph convolutionYadatiet al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib32)\); Fenget al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib30)\)or attentionBaiet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib31)\), and a variety of other investigationsHuang and Yang \([2021](https://arxiv.org/html/2606.27413#bib.bib55)\); Alistarhet al\.\([2015](https://arxiv.org/html/2606.27413#bib.bib47)\); Li and Milenkovic \([2017](https://arxiv.org/html/2606.27413#bib.bib48)\); Kimet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib54)\); Wanget al\.\([2023b](https://arxiv.org/html/2606.27413#bib.bib53),[a](https://arxiv.org/html/2606.27413#bib.bib52)\)\. For more practical tasks, hypergraphs have seen applications in scene generationNguyenet al\.\([2025](https://arxiv.org/html/2606.27413#bib.bib49)\), vision\-natural language scenariosKhanet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib42)\); Kimet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib41)\), federated learningFan and Shuai \([2024](https://arxiv.org/html/2606.27413#bib.bib40)\), recursive hyperedge structureYadati \([2020](https://arxiv.org/html/2606.27413#bib.bib38)\), hypergraph matchingZhenget al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib39)\), contrastive learningWeiet al\.\([2022](https://arxiv.org/html/2606.27413#bib.bib44)\), recommendation systemsHanet al\.\([2022b](https://arxiv.org/html/2606.27413#bib.bib59)\), tabular dataChenet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib37)\), the long tail problemHanet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib61)\); Liuet al\.\([2023b](https://arxiv.org/html/2606.27413#bib.bib60)\), and integration with large language modelsFenget al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib56)\); Chuet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib57)\); Huanget al\.\([2025](https://arxiv.org/html/2606.27413#bib.bib58)\); Luoet al\.\([2025](https://arxiv.org/html/2606.27413#bib.bib62)\)\. Some prior workHwanget al\.\([2008](https://arxiv.org/html/2606.27413#bib.bib33)\)has investigated hypergraphs to connect gene expression and protein interactions; this is distinct from our work, which focuses on gene\-level and phenotype connections\. Thus, this work provides a strong foundation for applying graphs and hypergraphs to the G2P challenge, as we will detail in Section[3](https://arxiv.org/html/2606.27413#S3)\.

### 2\.4Explanatory Methods and Explanatory Gene Discovery

Explanatory methods seek to explain why a machine learning model makes particular predictions for a given input\. They may explain the model prediction—why the model makes a certain prediction, or it may explain a “phenomenon” with respect to the nature of the dataAmaraet al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib26)\)\. Well\-known model\-agnostic methods are LIMERibeiroet al\.\([2016](https://arxiv.org/html/2606.27413#bib.bib23)\)and SHAPLundberg and Lee \([2017](https://arxiv.org/html/2606.27413#bib.bib22)\), which explain which inputs most contribute to the output\. GNNExplainerYinget al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib24)\)PGExplainerLuoet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib25)\)offer efficient graph explanations, while RegExplainerZhanget al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib28)\)specifically addresses the graph regression task by incorporating a graph information bottleneckWuet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib27)\)\.

In the biology literature, several studies have examined correlations between genes and “latent factor” traits, but these studies are designed for different cases than ours\. Approaches such as MOFA\+Argelaguetet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib91)\)and TotalVIGayosoet al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib92)\)were designed forsingle\-cellmulti\-omics data integration, and focus on identifying correlationswithin individual cells\(e\.g\., single\-cell RNAseq\) and cell\-level phenotypeswithin the same specimen\. Our work, in contrast, focuses on specimen\-scale traits \(e\.g\., plant morphology, photosynthesis, development\) derived from bulk gene expression data and other whole\-plant measurements\. We focus on how gene expression patterns across an entire organism influencemacroscopictraits, rather than traits of a single cell\. Similarly, DeepCCAAndrewet al\.\([2013](https://arxiv.org/html/2606.27413#bib.bib93)\), while a powerful statistical tool for finding latent correlations between two datasets, is a general\-purpose method that does not inherently incorporate the biologically\-informed graph structures \(e\.g\., incorporating gene\-gene correlations into a graph\) that are central to our framework’s interpretability and its specific aim of understanding genotype\-to\-phenotype relationships within a biological context\.

## 3The Proposed GRAFT Dataset

We now discuss our proposed dataset—Gene\-GraphRegression forArabidopsisFunctionalTraits \(GRAFT\)\. GRAFT contains gene data from multiple modalities, image\-derived phenotype data, manually collected data, and spectrometry\-based data\. Section[3\.1](https://arxiv.org/html/2606.27413#S3.SS1)discusses data collection, while Section[3\.2](https://arxiv.org/html/2606.27413#S3.SS2)discusses our annotations for genes\. Finally, Section[3\.3](https://arxiv.org/html/2606.27413#S3.SS3)discusses biological functions and their encoding into hyperedges\.

### 3\.1Data Collection

This study utilizes four genetically distinct lines ofArabidopsis thaliana\(thale cress\) that differ in gene expression and therefore in traits\. These differences arise from targeted mutations in specific genes, making the lines well\-suited for studying how genomic variation translates into observable phenotypic differences\. We first collected image\-based traits, manually\-measured traits, and photosynthetic measurements using a spectrometer \(Fig\.[1](https://arxiv.org/html/2606.27413#S2.F1)\)\. Following trait data collection, a subset of the same physical plants was destructively sampled for transcriptomic profiling via RNA sequencing \(RNAseq\), yieldingdirectly paired \(or linked\) genomic and phenotypic observations on the same individual specimens—a property that distinguishes this dataset from the overwhelming majority of published plant genomics resources \(see Table[1](https://arxiv.org/html/2606.27413#S2.T1)\)\. Further details on these genetic lines and important biological background are given in the Appendix\.

#### 3\.1\.1Transcriptomics: Gene Expressions

The GRAFT dataset provides whole\-genome gene expression profiles converingG=34,123G=34,123transcripts currently annotated in theArabidopsis thalianareference genome\. For the purposes of this work, we can treat these are our genes\. This represents a near\-complete transcriptomic coverage of a model plant genome in a benchmarking dataset\. Gene expression values are derived from RNAseq analysis of 24 individual plant specimens \(6 per line\)\. We provide the gene measurements in Fragments Per Kilobase of transcript per Million mapped reads \(FPKM\) form\.

#### 3\.1\.2Phenotypes: Observable Traits

In total, phenotype data were collected for 77 individual thale cress plants drawn from four genetically distinct lines\. Across these 77 plants, 41 phenotypic parameters were measured, spanning morphological, developmental, and physiological modalities\. We focus on five parameters that collectively represent each measurement modality present in GRAFT and exhibit meaningful variation across the four lines\. For the 24 plants that also underwent RNAseq profiling, all five parameters are directly paired with whole\-genome expression measurements on the same individual, forming the core supervised learning pairs used for model training and evaluation\. We do note one specimen must be dropped due to NaN values in its trait data, giving us 23 complete samples in experiments\. Full details of the measurement protocols and per\-line specimen counts are provided in the Appendix\.

The five benchmark traits are:\(1\) Leaf Area: a manual measure of area covered by leaves in unit pixels \(based on image analysis\);\(2\) Heightof the inflorescence or flower stalk: an indicator of how far into reproductive development a plant is \(measured manually\);\(3\) FvP/FmP: a measure of how efficiently the plant can channel light energy into photosynthesis \(collected with a spectrometer\);\(4\) qL: a measure of the chemical state of an important compound called plastoquinone in photosynthesis \(also collected with a spectrometer\); and\(5\) Leaf temperature differential: the difference in temperature between a leaf and its surroundings \(also collected with a spectrometer\)\.

### 3\.2Multimodal Gene Annotations

Beyond raw expression values, GRAFT provides a rich annotation layer for every corresponding gene in theArabidopsis thalianagenome, enabling downstream interpretation of model outputs and explanations\. Annotations were compiled from the National Center for Biotechnology Information \(NCBI222[https://www\.ncbi\.nlm\.nih\.gov/datasets/gene/taxon/3702/](https://www.ncbi.nlm.nih.gov/datasets/gene/taxon/3702/)\) and include: \(i\) gene unique identifies \(UIDs\); \(ii\) common gene symbols \(e\.g\., EX1, FAD7\); and \(iii\) short descriptions\. These annotations are organized in a tabular structure, allowing retrieval of functional descriptions for any gene or gene set of interest—for example, the set of high\-importance genes identified by a model explanation method such as SHAP or a graph explanation\. As illustrated in Fig\.[2](https://arxiv.org/html/2606.27413#S3.F2), we provide a diverse set of annotations per gene \(not all used in this work\)\.

This annotation layer is designed to close the interpretability gap that arises when machine learning models identify statistically important genes: rather than returning anonymous identifiers, users can immediately retrieve biological context for each gene, supporting faster hypothesis generation and experimental follow\-up\. For the GRAFT benchmark specifically, gene annotations support the Biological Explanation Recall \(BER\) evaluation metric \(Section[4\.5](https://arxiv.org/html/2606.27413#S4.SS5)\), which measures whether a model’s explanatory gene set recovers known biological functions relevant to predicted traits\.

![Refer to caption](https://arxiv.org/html/2606.27413v1/figures/hypergraph-data3.png)Figure 2:The gene\-level features are provided by GRAFT\. Genes have identifiers and text descriptions indicating what functions they influence\. We also have the unique nucleotide sequence for each gene, the blueprint encoded as a string of characters\. Finally, we add our gene expression data for all genes across different plant specimens\.Best viewed with zoom and in color\.![Refer to caption](https://arxiv.org/html/2606.27413v1/figures/hyperedge-figure-8.png)Figure 3:A visualization of translating the biological functions of the genes of the thale cress to higher\-order pairings\.\(a\)Thousands of functions exist for many systems in lifeforms, each containing genes that regulate those functions\.\(b\)The sets of genes can form hyperedges\.\(c\)Hyperedges are connected by overlapping genes\.Best viewed with zoom and in color\.
### 3\.3Gene Ontology Structure and Hyperedge Encoding

To capture the functional organization of the genome beyond individual gene annotations, GRAFT incorporates biological function information from the Gene Ontology \(GO\) resource333[https://geneontology\.org/](https://geneontology.org/)\. “GO” terms provide descriptions of the molecular functions, biological processes, and cellular components that gene products are known or predicted to participate in\. Each GO term carries a unique identifier, a standardized text description, and a set of gene associations indicating which genes are known to contribute to that function\. For example, the GO termGO:0047484is described as“regulation of response to osmotic stress,”and is associated with a known and established subset of genes\.

The dataset includes annotations linking genes in our expression matrix to over 6,000 GO terms containing at least one gene\. GO terms span a range of biological functions, both broad and fine\-grained\. Because each GO term defines asetof genes that collectively participate in a shared function, GO terms map directly onto hyperedges in a hypergraph representation of the genome: each hyperedge connects the genes annotated to a given term, encoding biological co\-functionality as graph structure as shown in Fig\.[3](https://arxiv.org/html/2606.27413#S3.F3)\. This hypergraph construction provides the structural prior used by our Hypergraph Neural Network \(HGNN\) baselines and enables biologically grounded explanations\.

Taken together, these three annotation resources—per\-gene identifiers and text annotations, GO term descriptions, and GO\-to\-gene hyperedges—form a self\-contained biological knowledge base bundled with GRAFT\. This design ensures that both the predictive and explanatory outputs of any model trained on GRAFT can be interpreted immediately in biological terms, without requiring users to query external databases separately\.

## 4Methods

### 4\.1Preliminaries

Let us define a specimen𝒮i\\mathcal\{S\}\_\{i\}’s linked data is defined as\(𝐱i,𝐲i\)\(\\mathbf\{x\}\_\{i\},\\mathbf\{y\}\_\{i\}\), where𝐱i∈ℝG×1\\mathbf\{x\}\_\{i\}\\in\\mathbb\{R\}^\{G\\times 1\}holds all theG∈ℤ\+G\\in\\mathbb\{Z\}^\{\+\}gene expressions for𝒮i\\mathcal\{S\}\_\{i\}, and𝐲i∈ℝT\\mathbf\{y\}\_\{i\}\\in\\mathbb\{R\}^\{T\}denotes allT∈ℤ\+T\\in\\mathbb\{Z\}^\{\+\}trait measurements of interest\. There areNN\-many specimens\. We learn a mappingf:ℝG→ℝTf\\colon\\mathbb\{R\}^\{G\}\\to\\mathbb\{R\}^\{T\}under a multi\-output regression objective\. BecauseG≫NG\\gg N, we apply a two\-stagegene filterto prevent data leakage\.

##### Gene Filtering\.

Within each cross\-validation fold, using only training\-set statistics: \(i\)Variance filter: genes with expression variance belowτv=0\.01\\tau\_\{v\}=0\.01are discarded; \(ii\)Spearman filter: the top\-kkgenes ranked by maximum absolute Spearman correlation with any trait are retained, withk=1024k=1024\. The filtered expression vector𝐳∈ℝk\\mathbf\{z\}\\in\\mathbb\{R\}^\{k\}is the input to all downstream modeling\. Further results with varyingkkare provided in the Appendix\.

### 4\.2Graph\-Based Regression

For graphs, we treat each gene as a nodev∈𝒱v\\in\\mathcal\{V\}with initial featurehv\(0\)=𝐳vh\_\{v\}^\{\(0\)\}=\\mathbf\{z\}\_\{v\}\(its scalar expression value\)\. Edges are given by the WGCNARezaieet al\.\([2023](https://arxiv.org/html/2606.27413#bib.bib119)\)Topological Overlap Matrix \(TOM\), which encodes co\-expression similarity derived from the same expression data used for prediction\. Our key insight is that a biologically\-derived adjacency, rather than a learned or arbitrary one, ensures that structural inductive biases align withknown co\-regulation patterns, making subsequent explanationsbiologically interpretable\. In the graph case, we implement the gene\-to\-trait mapping as𝐲^=f​\(𝐳;𝐀\)\\hat\{\\mathbf\{y\}\}=f\(\\mathbf\{z\};\\mathbf\{A\}\), whereffconsists of a GCN backbone and a prediction head, and𝐀\\mathbf\{A\}is the graph adjacency matrix\.

### 4\.3Hypergraph\-Based Regression

Pairwise edges cannot encode multi\-gene biological pathways\. To implement the HGNN, our key insight is to represent the genome as a hypergraphℋ=\(𝒱,ℰH\)\\mathcal\{H\}=\(\\mathcal\{V\},\\mathcal\{E\}\_\{H\}\)where each hyperedgee∈ℰHe\\in\\mathcal\{E\}\_\{H\}is the set of genes connected to a Gene Ontology \(GO\) term, encoded by incidence matrix𝐁∈\{0,1\}k×m\\mathbf\{B\}\\in\\\{0,1\\\}^\{k\\times m\}, wheremmindexes a GO term\. This construction is a direct consequence of our dataset’s GO annotation layer: every GO term becomes a hyperedge that groups genes by shared function or biological process, providing the HGNN with a pathway\-level structural prior unavailable in purely pairwise graph formulations\. In the hypergraph case, we implement the gene\-to\-trait mapping as𝐲^=f​\(𝐳;ℋ\)\\hat\{\\mathbf\{y\}\}=f\(\\mathbf\{z\};\\mathcal\{H\}\), whereffconsists of an HGNN backbone and a prediction head\.

### 4\.4Explanations

A core goal of this benchmark is to assess not only predictive accuracy but also whether a model’s rationale aligns with known biology\. We obtain two complementary explanation signals\.

SHAP\.For every trainedffwe compute SHAPLundberg and Lee \([2017](https://arxiv.org/html/2606.27413#bib.bib22)\)values \(specifically using GradientSHAP, as DeepSHAP introduced complexities during implementation\), yielding a per\-gene, per\-trait importance scoreϕg,t∈ℝ\\phi\_\{g,t\}\\in\\mathbb\{R\}that satisfies the Shapley efficiency axiom\. Scores are averaged in absolute value across the test samples of each fold and then accumulated across folds in the originalGG\-dimensional gene space, so that the final attributionϕ¯g,t\\bar\{\\phi\}\_\{g,t\}is comparable across models regardless of fold\-varying feature subsets\. The top\-kkgenes byϕ¯g,t\\bar\{\\phi\}\_\{g,t\}constitute theexplanatory gene setfor traittt\.

Graph explanations\.To add another source of explanations, within the same analysis framework, for GCNs, we extract subgraph\-level explanations via standardized graph explainersYinget al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib24)\); Luoet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib25)\); Zhanget al\.\([2024](https://arxiv.org/html/2606.27413#bib.bib28)\)\. Discussion of this class of explanations is discussed in the Appendix\.

### 4\.5Explanation Validation

Biological Explanation Recall \(BER\) Score\.For each traitttand each model, the explanatory gene set is input to GO enrichment analysis \(hypergeometric testKlopfensteinet al\.\([2018](https://arxiv.org/html/2606.27413#bib.bib124)\)\)\. BER@kkis the fraction ofa prioritrait\-relevant GO terms recovered among the significantly enriched terms in Eqn[1](https://arxiv.org/html/2606.27413#S4.E1), where𝒯relevant\\mathcal\{T\}\_\{\\mathrm\{relevant\}\}is manually fixed before any model is run \(exact GO terms are given in the Appendix\)\.

BER​@​k=\|𝒯enrich∩𝒯relevant\|\|𝒯relevant\|,\\mathrm\{BER\}@k\\;=\\;\\frac\{\|\\,\\mathcal\{T\}\_\{\\mathrm\{enrich\}\}\\cap\\mathcal\{T\}\_\{\\mathrm\{relevant\}\}\\,\|\}\{\|\\mathcal\{T\}\_\{\\mathrm\{relevant\}\}\|\},\(1\)BER@kkscores are reported acrossk∈\{50,100,200,500\}k\\in\\\{50,100,200,500\\\},kkdenoting top\-genes from SHAP, enabling comparison of explanation quality independently of predictive accuracy\.

DEG List Overlap\.We compare each model type’s top\-kkSHAP genes against the differentially expressed gene \(DEG\) lists returned by statistical analysis of our expression data across the pairwise line comparisons\. We note that this analysis does not relate tospecific traits, which motivates our own analysis\. This validation is independent of GO annotations, thus, a gene identified by both analyses strengthens the gene identification confidence\.

## 5Benchmarks and Evaluation

### 5\.1Trait Regression

Table 2:Root mean\-square error \(rMSE\) for stratifiedKK\-fold cross\-validation resultsfor the top 1024 genesover three random seed splittings\. Note: leaf temperature differential =LTD\.We begin our benchmarking and evaluation on our dataset by testing several regression backbones with gene expressions as input, and report root mean square error \(rMSE\)\.We begin with a baseline MLP with dense connections, then GCNs: GraphConvMorriset al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib6)\), graph transformer \(TConv\)Shiet al\.\([2021](https://arxiv.org/html/2606.27413#bib.bib8)\), and SAGE \(SAGEConv\)Hamiltonet al\.\([2018](https://arxiv.org/html/2606.27413#bib.bib9)\); and HGNNs: Hypergraph convolution \(HConv\)Baiet al\.\([2020](https://arxiv.org/html/2606.27413#bib.bib31)\), HyperGCNYadatiet al\.\([2019](https://arxiv.org/html/2606.27413#bib.bib32)\), and UniGATHuang and Yang \([2021](https://arxiv.org/html/2606.27413#bib.bib55)\)\. To address the high\-dimensionality and low\-sample setting, we train the networks in three different cross\-validation settings: leave\-one\-out \(LOO\), leave\-one\-class\-out \(LOCO\), and stratifiedKK\-fold \(SKFOLD\)\. We report on SKFOLD in the manuscript, as in biological scenarios, we want models to train on all genetic lines or classes, LOO and LOCO results are not as biologically relevant\. From Table[2](https://arxiv.org/html/2606.27413#S5.T2), HGNNs consistently outperform MLPs and GCNs with their compact, biologically\-informed, and interpretable hypergraph design\. Exact training settings, as well as LOO and LOCO results are provided in the Appendix\.

### 5\.2Explanation Validation

Table 3:Mean BER@kkacross stratified folds and model types\. \(\-\) denotes 0\.0±\\pm0\.0\.Table 4:DEG overlap for top\-k=1024k=1024genes\.Biological Explanation Recall \(BER\)\.For each model type \(MLP, GCN, HGNN\), we report the average BER scores across the five traits\. As shown in Table[3](https://arxiv.org/html/2606.27413#S5.T3), HGNN achieves higher BER across all traits and k values, confirming that GO term hyperedges improve the biological coherence of model explanations\. For photosynthetic traits \(qL, FvP/FmP\), BER increases monotonically with k and plateaus neark=500k=500, consistent with large, densely annotated pathway gene sets; for morphological traits \(Leaf Area, LTD\), BER peaks atk=100k=100to200200and declines atk=500k=500, indicating that explanatory signal is concentrated in a small coherent gene set and diluted by including more, “irrelevant”, genes\. GCN achieves near\-zero BER atk≤200k\\leq 200despite competitive regression accuracy, a pattern consistent with over\-smoothing due to dense TOM adjacency\. Pairwise co\-expression edges encode topological proximity rather than functional membership, producing diffuse SHAP attributions that, biologically, do not align with GO pathways\. The MLPs recover no signal for Leaf Area and LTD at any k\. This reflects the polygenic nature of these traits\. MLPs show a modest signal for the more concentrated photosynthetic traits\. Together, these results demonstrate that predictive accuracy and explanation fidelity are dissociable, and that the structural prior, not the regression loss, determines whether a model’s explanations are biologically interpretable\.

Overlap with Differentially Expressed Genes \(DEGs\)\.This experiment uses the−log10\-\\log\_\{10\}hypergeometricpp\-value of the overlap at our primary gene set withk=1024k=1024for each backbone most relevant to each trait\. Results are given in Table[4](https://arxiv.org/html/2606.27413#S5.T4), where the statistical significance of overlap between each model’s top\-k=1024k=1024explanatory genes and independently derived DEG lists\. The HConv backbone achieves strongly significant overlap across all traits \(p−log10p\-\\log\_\{10\}ranges from 5\.4 to 28\.4\), with the photosynthetic traits \(Inflorescence, FvP/FmP\) showing the highest values, indicating that GO\-term hyperedge message passing concentrates explanatory signal on genes that classical differential expression analysis independently identifies as biologically important\. The UniGAT backbone also achieves significant overlap across most traits, confirming that biologically informed models can drive improvement\. HyperGCN, despite sharing the same input\(𝐳,ℋ\)\(\\mathbf\{z\},\\mathcal\{H\}\), exhibits near\-zero overlap\. The GCN and MLP produce low and largely non\-significant overlap, reinforcing our prior conclusion with BER that pairwise co\-expression edges and dense feature\-only regression do not align model explanations with independently validated differential expression patterns\.

## 6Conclusions

In this work, we introduced GRAFT—the first dataset, to our knowledge, that provideslinkedwhole\-genome expression data and heterogeneous phenotypic trait measurements for the same model plant,Arabidopsis thaliana, specimens\. GRAFT also provides full gene and Gene Ontology annotations, enabling biologically informed graph and hypergraph modeling\. Our benchmarking shows that biologically informed and structured models \(HGNNs in particular\) match or exceed dense MLP regression across diverse traits and cross\-validation protocols, and, crucially, produce explanatory gene sets with higher biological fidelity, as measured by BER and by convergent overlap with independently derived DEG lists\. These results demonstrate that incorporating prior biological knowledge directly into model architecture yields interpretability gains that purely data\-driven approaches cannot replicate at this sample size, traits desired by the biology community\. We hope GRAFT lowers the barrier for the machine learning and biology communities to engage with the G2P challenge, and that the benchmark tasks, evaluation metrics, and public code release provide a foundation for future work on explainable multi\-omics regression in plant science and beyond\.

Limitations\.We should reiterate that, while the linked data spans 24 samples, the small sample size makes it considerably expensive and time\-consuming to obtain linked data for a single specimen\. Regardless, limits absolute regression performance and the statistical power of fold\-level evaluations\. Future work in machine learning and biology must account for these limitations\. GRAFT is specific toArabidopsis thaliana, and while this species is the model plant in plant biology, the biology community’s desire for transferability of trained models and explanatory gene sets to crops or other species remains an open question\. Finally, BER scores are bounded by the completeness of GO annotations, which are uneven across the genome, and trait\-GO term mappings made possible by TAIR are a reflection of current knowledge, which will change over time\.

## References

- A comprehensive review of artificial intelligence approaches in omics data processing: evaluating progress and challenges\.International Journal of Mathematics, Statistics, and Computer Science2,pp\. 114–167\.External Links:[Link](https://ijmscs.org/index.php/ijmscs/article/view/8703),[Document](https://dx.doi.org/10.59543/ijmscs.v2i.8703)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- D\. Alistarh, J\. Iglesias, and M\. Vojnovic \(2015\)Streaming min\-max hypergraph partitioning\.InAdvances in Neural Information Processing Systems,C\. Cortes, N\. Lawrence, D\. Lee, M\. Sugiyama, and R\. Garnett \(Eds\.\),Vol\.28,pp\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2015/file/83f97f4825290be4cb794ec6a234595f-Paper.pdf)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- K\. Amara, R\. Ying, Z\. Zhang, Z\. Han, Y\. Shan, U\. Brandes, S\. Schemm, and C\. Zhang \(2024\)GraphFramEx: towards systematic evaluation of explainability methods for graph neural networks\.External Links:2206\.09677,[Link](https://arxiv.org/abs/2206.09677)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1)\.
- G\. Andrew, R\. Arora, J\. Bilmes, and K\. Livescu \(2013\)Deep canonical correlation analysis\.InProceedings of the 30th International Conference on Machine Learning,S\. Dasgupta and D\. McAllester \(Eds\.\),Proceedings of Machine Learning Research, Vol\.28,Atlanta, Georgia, USA,pp\. 1247–1255\.External Links:[Link](https://proceedings.mlr.press/v28/andrew13.html)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p2.1)\.
- \[5\]AraPheno\.Note:[https://arapheno\.1001genomes\.org/](https://arapheno.1001genomes.org/)Accessed: 2025\-05\-12Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.6.4.1)\.
- R\. Argelaguet, D\. Arnol, D\. Bredikhin, Y\. Deloro, B\. Velten, J\. C\. Marioni, and O\. Stegle \(2020\)MOFA\+: a statistical framework for comprehensive integration of multi\-modal single\-cell data\.Genome Biology21\(1\),pp\. 111\(en\)\.External Links:ISSN 1474\-760X,[Link](https://genomebiology.biomedcentral.com/articles/10.1186/s13059-020-02015-1),[Document](https://dx.doi.org/10.1186/s13059-020-02015-1)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p2.1)\.
- S\. Bai, F\. Zhang, and P\. H\. S\. Torr \(2020\)Hypergraph convolution and hypergraph attention\.External Links:1901\.08150,[Link](https://arxiv.org/abs/1901.08150)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2),[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.28.26.26.6)\.
- W\. Bai, C\. Li, W\. Li, H\. Wang, X\. Han, P\. Wang, and L\. Wang \(2024\)Machine learning assists prediction of genes responsible for plant specialized metabolite biosynthesis by integrating multi‑omics data\.BMC Genomics25\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- C\. Berge \(1989\)Graphs and hypergraphs\.Elsevier\.Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- M\. Bolla \(1993\)Spectra, euclidean representations and clusterings of hypergraphs\.Discrete Mathematics117\(1\),pp\. 19–39\.External Links:ISSN 0012\-365X,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/0012-365X%2893%2990322-K),[Link](https://www.sciencedirect.com/science/article/pii/0012365X9390322K)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- M\. M\. Bronstein, J\. Bruna, Y\. LeCun, A\. Szlam, and P\. Vandergheynst \(2017\)Geometric deep learning: going beyond euclidean data\.IEEE Signal Processing Magazine34\(4\),pp\. 18–42\.External Links:ISSN 1558\-0792,[Link](http://dx.doi.org/10.1109/MSP.2017.2693418),[Document](https://dx.doi.org/10.1109/msp.2017.2693418)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- R\. Cavill, D\. Jennen, J\. Kleinjans, and J\. J\. Briedé \(2015\)Transcriptomic and metabolomic data integration\.Briefings in Bioinformatics17\(5\),pp\. 891–901\.External Links:ISSN 1467\-5463,[Document](https://dx.doi.org/10.1093/bib/bbv090),[Link](https://doi.org/10.1093/bib/bbv090),https://academic\.oup\.com/bib/article\-pdf/17/5/891/6687226/bbv090\.pdfCited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.1](https://arxiv.org/html/2606.27413#S2.SS1.p1.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- D\. Cembrowska\-Lech, A\. Krzemińska, T\. Miller, A\. Nowakowska, C\. Adamski, M\. Radaczyńska, G\. Mikiciuk, and M\. Mikiciuk \(2023\)An integrated multi\-omics and artificial intelligence framework for advance plant phenotyping in horticulture\.Biology12\(10\)\.External Links:[Link](https://www.mdpi.com/2079-7737/12/10/1298),ISSN 2079\-7737,[Document](https://dx.doi.org/10.3390/biology12101298)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.1](https://arxiv.org/html/2606.27413#S2.SS1.p1.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- P\. Chen, S\. Sarkar, L\. Lausen, B\. Srinivasan, S\. Zha, R\. Huang, and G\. Karypis \(2023\)HYTREL: hypergraph\-enhanced tabular data representation learning\.External Links:2307\.08623,[Link](https://arxiv.org/abs/2307.08623)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- R\. Chen, T\. Zhao, A\. Jaiswal, N\. Shah, and Z\. Wang \(2024\)LLaGA: large language and graph assistant\.External Links:2402\.08170,[Link](https://arxiv.org/abs/2402.08170)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- C\. Cheng, Y\. Li, K\. Varala, J\. Bubert, J\. Huang, G\. J\. Kim, J\. Halim, J\. Arp, H\. S\. Shih, G\. Levinson, S\. H\. Park, H\. Y\. Cho, S\. P\. Moose, and G\. M\. Coruzzi \(2021\)Evolutionarily informed machine learning enhances the power of predictive gene\-to\-phenotype relationships\.Nature Communications12\(4567\)\.External Links:[Link](https://www.nature.com/articles/s41467-021-25893-w)Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1)\.
- H\. K\. Choi, S\. Lee, J\. Chu, and H\. J\. Kim \(2023\)NuTrea: neural tree search for context\-guided multi\-hop kgqa\.External Links:2310\.15484,[Link](https://arxiv.org/abs/2310.15484)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- Z\. Chu, Y\. Wang, Q\. Cui, L\. Li, W\. Chen, Z\. Qin, and K\. Ren \(2024\)LLM\-guided multi\-view hypergraph learning for human\-centric explainable recommendation\.External Links:2401\.08217,[Link](https://arxiv.org/abs/2401.08217)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- T\. Coxe, D\. J\. Burks, U\. Singh, R\. Mittler, and R\. K\. Azad \(2024\)Benchmarking rna\-seq aligners at base\-level and junction base\-level resolution using the arabidopsis thaliana genome\.PlantsProceedings of the National Academy of SciencesThe Plant CellBioinformaticsJournal of Machine Learning ResearchScientific Reports13\(5\)\.External Links:[Link](https://www.mdpi.com/2223-7747/13/5/582),ISSN 2223\-7747Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.10.8.1)\.
- M\. Defferrard, X\. Bresson, and P\. Vandergheynst \(2017\)Convolutional neural networks on graphs with fast localized spectral filtering\.External Links:1606\.09375,[Link](https://arxiv.org/abs/1606.09375)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- V\.V\. Demidchik, A\.Y\. Shashko, U\.Y\. Bandarenka, G\.N\. Smolikova, D\.A\. Przhevalskaya, M\.A\. Charnysh, G\.A\. Pozhvanov, A\.V\. Barkosvkyi, I\.I\. Smolich, A\.I\. Sokolik, M\. Yu, and S\.S\. Medvedev \(2020\)Plant phenomics: fundamental bases, software and hardware platforms, and machine learning\.Russian Journal of Plant Physiology67,pp\. 397–412\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- T\. Depuydt, B\. De Rybel, and K\. Vandepoele \(2023\)Charting plant gene functions in the multi\-omics and single\-cell era\.Trends in Plant Science28,pp\. 283–296\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- R\. Edgar, M\. Domrachev, and A\. E\. Lash \(2002\)Gene expression omnibus: ncbi gene expression and hybridization array data repository\.External Links:[Document](https://dx.doi.org/10.1093/nar/30.1.207)Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.8.6.1)\.
- Q\. Fan and L\. Shuai \(2024\)Adaptive hyper\-graph aggregation for modality\-agnostic federated learning\.In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 12312–12321\.External Links:[Document](https://dx.doi.org/10.1109/CVPR52733.2024.01170)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- Y\. Feng, C\. Yang, X\. Hou, S\. Du, S\. Ying, Z\. Wu, and Y\. Gao \(2024\)Beyond graphs: can large language models comprehend hypergraphs?\.External Links:2410\.10083,[Link](https://arxiv.org/abs/2410.10083)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- Y\. Feng, H\. You, Z\. Zhang, R\. Ji, and Y\. Gao \(2019\)Hypergraph neural networks\.External Links:1809\.09401,[Link](https://arxiv.org/abs/1809.09401)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- P\. J\. Flood, W\. Kruijer, S\. K\. Schnabel, R\. van der Schoor, H\. Jalink, J\. F\. H\. Snel, J\. Harbinson, and M\. G\.M\. Aarts \(2016\)Discovery and delivery strategies for engineered live biotherapeutic products\.Plant Methods12\(14\)\.Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1)\.
- F\. Gao, K\. Huang, and Y\. Xing \(2023\)Artificial intelligence in omics\.Genomics, Proteomics & Bioinformatics20\(5\),pp\. 811–813\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- A\. Gayoso, Z\. Steier, R\. Lopez, J\. Regier, K\. L\. Nazor, A\. Streets, and N\. Yosef \(2021\)Joint probabilistic modeling of single\-cell multi\-omic data with totalVI\.Nature Methods18\(3\),pp\. 272–282\(en\)\.External Links:ISSN 1548\-7091, 1548\-7105,[Link](https://www.nature.com/articles/s41592-020-01050-x),[Document](https://dx.doi.org/10.1038/s41592-020-01050-x)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p2.1)\.
- \[30\]Gene expression omnibus\.Note:[https://www\.ncbi\.nlm\.nih\.gov/geo/](https://www.ncbi.nlm.nih.gov/geo/)Accessed: 2025\-05\-12Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.8.6.1)\.
- W\. L\. Hamilton, R\. Ying, and J\. Leskovec \(2018\)Inductive representation learning on large graphs\.External Links:1706\.02216,[Link](https://arxiv.org/abs/1706.02216)Cited by:[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.23.21.21.6)\.
- J\. Han, J\. Liu, and J\. Xu \(2024\)Dual\-branch network with hypergraph feature augmentation and adaptive logits adjustment for long\-tailed visual recognition\.Applied Soft Computing167,pp\. 112400\.External Links:ISSN 1568\-4946,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.asoc.2024.112400),[Link](https://www.sciencedirect.com/science/article/pii/S1568494624011748)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- K\. Han, Y\. Wang, J\. Guo, Y\. Tang, and E\. Wu \(2022a\)Vision gnn: an image is worth graph of nodes\.External Links:2206\.00272,[Link](https://arxiv.org/abs/2206.00272)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- Y\. Han, E\. W\. Huang, W\. Zheng, N\. Rao, Z\. Wang, and K\. Subbian \(2022b\)Search behavior prediction: a hypergraph perspective\.External Links:2211\.13328,[Link](https://arxiv.org/abs/2211.13328)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- M\. K\. Heavey, D\. Durmusoglu, N\. Crook, and A\. C\. Anselmo \(2022\)Discovery and delivery strategies for engineered live biotherapeutic products\.Trends in Biotechnology40\(3\),pp\. 354–369\.External Links:ISSN 0167\-7799,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.tibtech.2021.08.002),[Link](https://www.sciencedirect.com/science/article/pii/S0167779921001761)Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- E\. Huala, A\. W\. Dickerman, M\. Garcia\-Hernandez, D\. Weems, L\. Reiser, F\. LaFond, D\. Hanley, D\. Kiphart, M\. Zhuang, W\. Huang, L\. A\. Mueller, D\. Bhattacharyya, D\. Bhaya, B\. W\. Sobral, W\. Beavis, D\. W\. Meinke, C\. D\. Town, C\. Somerville, and S\. Y\. Rhee \(2001\)The arabidopsis information resource \(tair\): a comprehensive database and web\-based information retrieval, analysis, and visualization system for a model plant\.External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1093/nar/29.1.102)Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.10.8.1)\.
- J\. Huang and J\. Yang \(2021\)UniGNN: a unified framework for graph and hypergraph neural networks\.External Links:2105\.00956,[Link](https://arxiv.org/abs/2105.00956)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2),[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.38.36.36.6)\.
- S\. Huang, H\. Li, Y\. Gu, X\. Hu, Q\. Li, and G\. Xu \(2025\)HyperG: hypergraph\-enhanced llms for structured knowledge\.External Links:2502\.18125,[Link](https://arxiv.org/abs/2502.18125)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- T\. Hwang, Z\. Tian, R\. Kuangy, and J\. Kocher \(2008\)Learning on weighted hypergraphs to integrate protein interactions and gene expressions for cancer outcome prediction\.In2008 Eighth IEEE International Conference on Data Mining,Vol\.,pp\. 293–302\.External Links:[Document](https://dx.doi.org/10.1109/ICDM.2008.37)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- A\. U\. Khan, H\. Kuehne, B\. Wu, K\. Chheu, W\. Bousselham, C\. Gan, N\. Lobo, and M\. Shah \(2023\)Learning situation hyper\-graphs for video question answering\.External Links:2304\.08682,[Link](https://arxiv.org/abs/2304.08682)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- E\. Kim, W\. Y\. Kang, K\. On, Y\. Heo, and B\. Zhang \(2020\)Hypergraph attention networks for multimodal learning\.In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 14569–14578\.External Links:[Document](https://dx.doi.org/10.1109/CVPR42600.2020.01459)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- S\. Kim, S\. Kang, F\. Bu, S\. Y\. Lee, J\. Yoo, and K\. Shin \(2024\)HypeBoy: generative self\-supervised representation learning on hypergraphs\.External Links:2404\.00638,[Link](https://arxiv.org/abs/2404.00638)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- T\. N\. Kipf and M\. Welling \(2017\)Semi\-supervised classification with graph convolutional networks\.External Links:1609\.02907,[Link](https://arxiv.org/abs/1609.02907)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- D\. V\. Klopfenstein, L\. Zhang, B\. S\. Pedersen, F\. Ramírez, A\. Warwick Vesztrocy, A\. Naldi, C\. J\. Mungall, J\. M\. Yunes, O\. Botvinnik, M\. Weigel, W\. Dampier, C\. Dessimoz, P\. Flick, and H\. Tang \(2018\)GOATOOLS: A Python library for Gene Ontology analyses\.8\(1\),pp\. 10872\(en\)\.External Links:ISSN 2045\-2322,[Link](https://www.nature.com/articles/s41598-018-28948-z),[Document](https://dx.doi.org/10.1038/s41598-018-28948-z)Cited by:[§4\.5](https://arxiv.org/html/2606.27413#S4.SS5.p1.3)\.
- P\. Li and O\. Milenkovic \(2017\)Inhomogeneous hypergraph clustering with applications\.External Links:1709\.01249,[Link](https://arxiv.org/abs/1709.01249)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- T\. Liu, Y\. Wang, R\. Ying, and H\. Zhao \(2023a\)MuSe\-gnn: learning unified gene representation from multimodal biological graph data\.External Links:2310\.02275,[Link](https://arxiv.org/abs/2310.02275)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- Y\. Liu, H\. Xuan, B\. Li, M\. Wang, T\. Chen, and H\. Yin \(2023b\)Self\-supervised dynamic hypergraph recommendation based on hyper\-relational knowledge graph\.External Links:2308\.07752,[Link](https://arxiv.org/abs/2308.07752)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- S\. M\. Lundberg and S\. Lee \(2017\)A unified approach to interpreting model predictions\.InAdvances in Neural Information Processing Systems,I\. Guyon, U\. V\. Luxburg, S\. Bengio, H\. Wallach, R\. Fergus, S\. Vishwanathan, and R\. Garnett \(Eds\.\),Vol\.30,pp\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2017/file/8a20a8621978632d76c43dfd28b67767-Paper.pdf)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p3.1),[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1),[§4\.4](https://arxiv.org/html/2606.27413#S4.SS4.p2.7)\.
- D\. Luo, W\. Cheng, D\. Xu, W\. Yu, B\. Zong, H\. Chen, and X\. Zhang \(2020\)Parameterized explainer for graph neural network\.External Links:2011\.04573,[Link](https://arxiv.org/abs/2011.04573)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p3.1),[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1),[§4\.4](https://arxiv.org/html/2606.27413#S4.SS4.p3.1)\.
- H\. Luo, H\. E, G\. Chen, Y\. Zheng, X\. Wu, Y\. Guo, Q\. Lin, Y\. Feng, Z\. Kuang, M\. Song, Y\. Zhu, and L\. A\. Tuan \(2025\)HyperGraphRAG: retrieval\-augmented generation via hypergraph\-structured knowledge representation\.External Links:2503\.21322,[Link](https://arxiv.org/abs/2503.21322)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- S\. Mansoor, E\. M\.B\.M\. Karunathilake, T\. T\. Tuan, and Y\. S\. Chung \(2025\)Genomics, phenomics, and machine learning in transforming plant research: advancements and challenges\.Horticultural Plant Journal11\(2\),pp\. 486–503\.External Links:ISSN 2468\-0141,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.hpj.2023.09.005),[Link](https://www.sciencedirect.com/science/article/pii/S2468014124000098)Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- M\. Minervini, A\. Fischbach, H\. Scharr, and S\. A\. Tsaftaris \(2016\)Finely\-grained annotated datasets for image\-based plant phenotyping\.Pattern Recognition Letters81,pp\. 80–89\.External Links:ISSN 0167\-8655,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.patrec.2015.10.013),[Link](https://www.sciencedirect.com/science/article/pii/S0167865515003645)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.3.1.1)\.
- C\. Morris, M\. Ritzert, M\. Fey, W\. L\. Hamilton, J\. E\. Lenssen, G\. Rattan, and M\. Grohe \(2021\)Weisfeiler and leman go neural: higher\-order graph neural networks\.External Links:1810\.02244,[Link](https://arxiv.org/abs/1810.02244)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1),[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.13.11.11.6)\.
- T\. Nguyen, P\. Nguyen, J\. Cothren, A\. Yilmaz, and K\. Luu \(2025\)HyperGLM: hypergraph for video scene graph generation and anticipation\.External Links:2411\.18042,[Link](https://arxiv.org/abs/2411.18042)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- T\. Nguyen, P\. Nguyen, X\. Li, J\. Cothren, A\. Yilmaz, and K\. Luu \(2024a\)CYCLO: cyclic graph transformer approach to multi\-object relationship modeling in aerial videos\.External Links:2406\.01029,[Link](https://arxiv.org/abs/2406.01029)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- T\. Nguyen, P\. Nguyen, and K\. Luu \(2024b\)HIG: hierarchical interlacement graph approach to scene graph generation in video understanding\.External Links:2312\.03050,[Link](https://arxiv.org/abs/2312.03050)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- B\. Perozzi, B\. Fatemi, D\. Zelle, A\. Tsitsulin, M\. Kazemi, R\. Al\-Rfou, and J\. Halcrow \(2024\)Let your graph do the talking: encoding structured data for llms\.External Links:2402\.05862,[Link](https://arxiv.org/abs/2402.05862)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- \[58\]PhotosynQ\.Note:[https://photosynq\.org/](https://photosynq.org/)Accessed: 2025\-05\-12Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.7.5.1)\.
- \[59\]Plant phenotyping datasets\.Note:[https://www\.plant\-phenotyping\.org/datasets\-home](https://www.plant-phenotyping.org/datasets-home)Accessed: 2025\-05\-12Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.3.1.1)\.
- N\. Rezaie, F\. Reese, and A\. Mortazavi \(2023\)PyWGCNA: a python package for weighted gene co\-expression network analysis\.39\(7\),pp\. btad415\.External Links:ISSN 1367\-4811,[Document](https://dx.doi.org/10.1093/bioinformatics/btad415),[Link](https://doi.org/10.1093/bioinformatics/btad415),https://academic\.oup\.com/bioinformatics/article\-pdf/39/7/btad415/50920596/btad415\.pdfCited by:[§4\.2](https://arxiv.org/html/2606.27413#S4.SS2.p1.5)\.
- M\. T\. Ribeiro, S\. Singh, and C\. Guestrin \(2016\)"Why should i trust you?": explaining the predictions of any classifier\.External Links:1602\.04938,[Link](https://arxiv.org/abs/1602.04938)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1)\.
- J\.A\. Rodríguez \(2003\)On the laplacian spectrum and walk\-regular hypergraphs\.Linear and Multilinear Algebra51\(3\),pp\. 285–297\.External Links:[Document](https://dx.doi.org/10.1080/0308108031000084374),[Link](https://doi.org/10.1080/0308108031000084374),https://doi\.org/10\.1080/0308108031000084374Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- \[63\]Sequence read archive\.Note:[https://www\.ebi\.ac\.uk/ena/browser/home](https://www.ebi.ac.uk/ena/browser/home)Accessed: 2025\-07\-9Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1)\.
- \[64\]Sequence read archive\.Note:[https://www\.ncbi\.nlm\.nih\.gov/sra/](https://www.ncbi.nlm.nih.gov/sra/)Accessed: 2025\-05\-12Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.8.6.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.9.7.1)\.
- U\. Seren, D\. Grimm, J\. Fitz, D\. Weigel, M\. Nordborg, K\. Borgwardt, and A\. Korte \(2017\)AraPheno: a public database for arabidopsis thaliana phenotypes\.External Links:[Document](https://dx.doi.org/10.1093/nar/gkw986)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.6.4.1)\.
- Y\. Shi, Z\. Huang, S\. Feng, H\. Zhong, W\. Wang, and Y\. Sun \(2021\)Masked label prediction: unified message passing model for semi\-supervised classification\.External Links:2009\.03509,[Link](https://arxiv.org/abs/2009.03509)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1),[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.18.16.16.6)\.
- J\. Sun, C\. Xu, L\. Tang, S\. Wang, C\. Lin, Y\. Gong, L\. M\. Ni, H\. Shum, and J\. Guo \(2024\)Think\-on\-graph: deep and responsible reasoning of large language model on knowledge graph\.External Links:2307\.07697,[Link](https://arxiv.org/abs/2307.07697)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- \[68\]Synthetic arabidopsis dataset\.Note:[https://doi\.org/10\.25919/5c36957c0af41](https://doi.org/10.25919/5c36957c0af41)Accessed: 2025\-05\-12External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.25919/5c36957c0af41)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.4.2.1)\.
- \[69\]The arabidopsis information resource\.Note:[https://www\.arabidopsis\.org/](https://www.arabidopsis.org/)Accessed: 2025\-05\-12Cited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.10.8.1)\.
- J\. R\. Ubbens and I\. Stavness \(2017\)Deep plant phenomics: a deep learning platform for complex plant phenotyping tasks\.Frontiers in Plant ScienceVolume 8 \- 2017\.External Links:[Link](https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2017.01190),[Document](https://dx.doi.org/10.3389/fpls.2017.01190),ISSN 1664\-462XCited by:[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1)\.
- \[71\]UNL plant phenotyping datasets\.Note:[https://plantvision\.unl\.edu/unl\-plant\-phenotyping\-datasets/](https://plantvision.unl.edu/unl-plant-phenotyping-datasets/)Accessed: 2025\-05\-12Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.5.3.1)\.
- P\. Veličković, G\. Cucurull, A\. Casanova, A\. Romero, P\. Liò, and Y\. Bengio \(2018\)Graph attention networks\.External Links:1710\.10903,[Link](https://arxiv.org/abs/1710.10903)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p1.1)\.
- P\. Wang, S\. Yang, Y\. Liu, Z\. Wang, and P\. Li \(2023a\)Equivariant hypergraph diffusion neural operators\.External Links:2207\.06680,[Link](https://arxiv.org/abs/2207.06680)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- Y\. Wang, Q\. Gan, X\. Qiu, X\. Huang, and D\. Wipf \(2023b\)From hypergraph energy functions to hypergraph neural networks\.External Links:2306\.09623,[Link](https://arxiv.org/abs/2306.09623)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- D\. Ward, P\. Moghadam, and N\. Hudson \(2019\)Deep leaf segmentation using synthetic data\.External Links:1807\.10931,[Link](https://arxiv.org/abs/1807.10931)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2606.27413#S2.T1.5.4.2.1)\.
- T\. Wei, Y\. You, T\. Chen, Y\. Shen, J\. He, and Z\. Wang \(2022\)Augmentations in hypergraph contrastive learning: fabricated and generative\.External Links:2210\.03801,[Link](https://arxiv.org/abs/2210.03801)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- A\. W\. Woodward and B\. Bartel \(2018\)Biology in bloom: a primer on the arabidopsis thaliana model system\.Genetics208\(4\),pp\. 1337–1349\(en\)\.Cited by:[§2\.1](https://arxiv.org/html/2606.27413#S2.SS1.p1.1)\.
- T\. Wu, H\. Ren, P\. Li, and J\. Leskovec \(2020\)Graph information bottleneck\.External Links:2010\.12811,[Link](https://arxiv.org/abs/2010.12811)Cited by:[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1)\.
- N\. Yadati, M\. Nimishakavi, P\. Yadav, V\. Nitin, A\. Louis, and P\. Talukdar \(2019\)HyperGCN: a new method of training graph convolutional networks on hypergraphs\.External Links:1809\.02589,[Link](https://arxiv.org/abs/1809.02589)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2),[§5\.1](https://arxiv.org/html/2606.27413#S5.SS1.p1.1),[Table 2](https://arxiv.org/html/2606.27413#S5.T2.33.31.31.6)\.
- N\. Yadati \(2020\)Neural message passing for multi\-relational ordered and recursive hypergraphs\.InAdvances in Neural Information Processing Systems,H\. Larochelle, M\. Ranzato, R\. Hadsell, M\.F\. Balcan, and H\. Lin \(Eds\.\),Vol\.33,pp\. 3275–3289\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/217eedd1ba8c592db97d0dbe54c7adfc-Paper.pdf)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- J\. Yan and X\. Wang \(2023\)Machine learning bridges omics sciences and plant breeding\.Trends in Plant Science28,pp\. 199–210\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- Y\. Yang, M\. A\. Saand, L\. Huang, W\. B\. Abdelaal, J\. Zhang, Y\. Wu, J\. Li, M\. H\. Sirohi, and F\. Wang \(2021\)Applications of multi\-omics technologies for crop improvement\.Frontiers in Plant Science12\.External Links:[Link](https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2021.563953),[Document](https://dx.doi.org/10.3389/fpls.2021.563953),ISSN 1664\-462XCited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- R\. Ying, D\. Bourgeois, J\. You, M\. Zitnik, and J\. Leskovec \(2019\)GNNExplainer: generating explanations for graph neural networks\.External Links:1903\.03894,[Link](https://arxiv.org/abs/1903.03894)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p3.1),[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1),[§4\.4](https://arxiv.org/html/2606.27413#S4.SS4.p3.1)\.
- J\. Zhang, Z\. Chen, H\. Mei, L\. Da, D\. Luo, and H\. Wei \(2024\)RegExplainer: generating explanations for graph neural networks in regression tasks\.External Links:2307\.07840,[Link](https://arxiv.org/abs/2307.07840)Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p3.1),[§2\.4](https://arxiv.org/html/2606.27413#S2.SS4.p1.1),[§4\.4](https://arxiv.org/html/2606.27413#S4.SS4.p3.1)\.
- R\. Zhang, C\. Zhang, C\. Yu, J\. Dong, and J\. Hu \(2022\)Integration of multi\-omics technologies for crop improvement: status and prospects\.Frontiers in bioinformatics2\.Cited by:[§1](https://arxiv.org/html/2606.27413#S1.p2.1),[§2\.2](https://arxiv.org/html/2606.27413#S2.SS2.p1.1)\.
- Q\. Zheng, M\. Zhang, and H\. Yan \(2024\)CURSOR: scalable mixed\-order hypergraph matching with cur decomposition\.External Links:2402\.16594,[Link](https://arxiv.org/abs/2402.16594)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- D\. Zhou, J\. Huang, and B\. Schölkopf \(2006\)Learning with hypergraphs: clustering, classification, and embedding\.InAdvances in Neural Information Processing Systems,B\. Schölkopf, J\. Platt, and T\. Hoffman \(Eds\.\),Vol\.19,pp\.\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2006/file/dff8e9c2ac33381546d96deea9922999-Paper.pdf)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.
- J\.Y\. Zien, M\.D\.F\. Schlag, and P\.K\. Chan \(1999\)Multilevel spectral hypergraph partitioning with arbitrary vertex sizes\.IEEE Transactions on Computer\-Aided Design of Integrated Circuits and Systems18\(9\),pp\. 1389–1399\.External Links:[Document](https://dx.doi.org/10.1109/43.784130)Cited by:[§2\.3](https://arxiv.org/html/2606.27413#S2.SS3.p2.2)\.