🧬 biology

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

This paper introduces GRAFT, a novel multi-modal dataset linking gene expression profiles with phenotypic trait measurements in *Arabidopsis thaliana*, and establishes benchmarks for predicting traits and validating gene-trait associations using graph and hypergraph learning methods to address the genome-to-phenome challenge.

Original authors: Manuel Serna-Aguilera, Vanshika Jindal, Fiona L. Goggin, Jiamei Li, Aranyak Goswami, Alexander Bucksch, Suxing Liu, Khoa Luu

Published 2026-06-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Manuel Serna-Aguilera, Vanshika Jindal, Fiona L. Goggin, Jiamei Li, Aranyak Goswami, Alexander Bucksch, Suxing Liu, Khoa Luu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: The "Who Did What?" Mystery

Imagine you have a massive instruction manual for building a complex machine (like a car or a robot). This manual has 34,000 different pages (genes). The machine also has hundreds of different features, like how fast it goes, how much fuel it uses, or how tall it is (traits).

For a long time, scientists have been trying to figure out: Which specific pages in the manual control which specific features of the machine?

The problem is that the data is usually split up.

One library has the instruction manual pages.
Another library has photos of the finished machines.
A third library has measurements of how fast they go.

Because these libraries don't talk to each other, it's like trying to solve a puzzle where half the pieces are in a box in New York and the other half are in a box in London. You can't see the full picture.

The Solution: The GRAFT Dataset

The authors of this paper created a new dataset called GRAFT (Gene-Graph Regression for Arabidopsis Functional Traits). Think of GRAFT as a super-organized detective's case file.

Instead of having separate boxes for instructions and results, GRAFT puts them all in one folder for the exact same plant specimens.

The Subject: They used Arabidopsis thaliana (a small weed often used in labs, kind of like the "lab rat" of the plant world).
The Link: They took a specific plant, measured its physical traits (like how tall it is or how green its leaves are), and then immediately took a sample of that same plant to read its genetic code.
The Result: Now, for every single plant, they know exactly what its genes said and what the plant actually looked like.

How They Solved It: The "Social Network" Analogy

Once they had the data, they needed a way to analyze it. They tried three different approaches, which they compare to different ways of organizing a social network:

The "List" Approach (MLP): Imagine trying to guess a person's personality just by reading a list of their favorite hobbies, one by one, without seeing how they connect. This is what standard computer models usually do. It works okay, but it misses the big picture.
The "Friendship" Approach (Graph/GCN): Imagine looking at who is friends with whom. If Gene A is friends with Gene B, and Gene B is friends with Gene C, maybe they are all part of the same club. This helps a bit, but it only looks at pairs of genes.
The "Club" Approach (Hypergraph/HGNN): This is the paper's big innovation. Imagine a gene isn't just friends with one other gene; it's part of a book club, a sports team, and a choir all at the same time. In biology, groups of genes work together to perform specific tasks (like "making leaves green" or "fighting drought").
- The authors built a Hypergraph. Instead of just connecting two dots (genes), they drew a big circle (a "hyperedge") around a whole group of genes that belong to the same "club" (biological function).
- The Result: The computer model that used this "Club" approach (the Hypergraph) was much better at predicting plant traits and, more importantly, could explain why it made those predictions in a way that made sense to biologists.

The "Why It Matters" Test: The Detective's Report

The authors didn't just want the computer to guess the right answer; they wanted it to give a good excuse for why it guessed that.

They used a special test called Biological Explanation Recall (BER).

The Analogy: Imagine a detective (the computer model) points to a suspect (a gene) and says, "This person did it!"
The Test: The scientists then check the police records (Gene Ontology databases) to see if that suspect actually has a history of committing that specific type of crime.
The Outcome: The "Club" model (Hypergraph) was the best detective. When it pointed to a gene, that gene was almost always part of the correct "crime ring" (biological pathway). The other models (the "List" and "Friendship" approaches) often pointed to suspects who had no connection to the crime, even if they guessed the right answer by luck.

The Bottom Line

The paper claims that by organizing genes into "clubs" (hypergraphs) based on what they actually do, rather than just looking at them individually or in pairs, scientists can:

Better predict what a plant will look like based on its DNA.
Get explanations that biologists can actually trust and use.

A Note on Limits: The authors are honest that this is a small dataset (only 24 plants). It's like solving a mystery with only 24 witnesses. While the method works great on this small group, they admit it will take more work to prove it works on huge, complex crops like corn or wheat. But, they have provided the "case file" (the dataset) and the "detective tools" (the code) for other scientists to keep solving the mystery.

Technical Summary of GRAFT: Biological Graph and Hypergraph Benchmarks

Problem Statement
The central challenge in modern biology is mapping the relationship between an organism's genome (genotype) and its observable characteristics (phenotype), known as the Genome-to-Phenome (G2P) problem. While high-throughput technologies have generated vast amounts of multi-omics data, current datasets and repositories are ill-equipped for G2P analysis. Existing resources typically silo data types: some contain only image-derived phenotypes, others only gene expression profiles, and very few link genomic and phenomic data from the same individual specimens. Furthermore, most existing benchmarks focus on specific, homogeneous traits, failing to capture the heterogeneous nature of biological data (e.g., combining morphological, developmental, and physiological measurements). This lack of integrated, multi-modal benchmarking data hinders the development of computational tools capable of reasoning over high-dimensional, structured biological data to identify which genes control which traits.

Methodology
To address these gaps, the authors introduce GRAFT (Gene-Graph Regression for Arabidopsis Functional Traits), a curated multi-modal dataset and benchmarking framework.

Dataset Construction: GRAFT links whole-genome gene expression profiles with heterogeneous phenotypic trait measurements for the same Arabidopsis thaliana specimens.
- Genomics: RNA sequencing (RNA-seq) data for 34,123 transcripts from 24 individual plants (6 per line across 4 genetically distinct lines).
- Phenomics: 41 phenotypic parameters measured across 77 plants, with 23 specimens having complete paired data. The benchmark focuses on five traits representing different modalities: Leaf Area (image-derived), Inflorescence Height (manual), and three photosynthetic/physiological metrics (FvP/FmP, qL, Leaf Temperature Differential) collected via spectrometry.
- Annotations: The dataset includes rich gene annotations (UIDs, symbols, descriptions) and Gene Ontology (GO) terms. Crucially, GO terms are encoded as hyperedges, where each hyperedge connects a set of genes participating in a specific biological function, forming a hypergraph structure.
Modeling Approach: The paper benchmarks regression models under a multi-output objective ( $f: \mathbb{R}^G \to \mathbb{R}^T$ ) to predict traits from gene expression.
- Preprocessing: A two-stage gene filter (variance and Spearman correlation) reduces dimensionality to the top 1,024 genes to prevent data leakage.
- Baselines: The study compares three classes of models:
  1. MLP: Dense feed-forward networks.
  2. Graph Neural Networks (GCNs): Utilizing a WGCNA-based Topological Overlap Matrix (TOM) to define pairwise gene co-expression edges.
  3. Hypergraph Neural Networks (HGNNs): Utilizing the GO-term annotations to define hyperedges, capturing n-ary relationships among genes involved in shared biological processes.
- Explainability: The framework employs SHAP (GradientSHAP) for feature importance and graph explainers (e.g., GNNExplainer) for subgraph extraction.
Evaluation Metrics:
- Predictive Accuracy: Measured via Root Mean Square Error (rMSE) using stratified K-fold cross-validation.
- Biological Explanation Recall (BER): A novel metric evaluating whether the top- $k$ genes identified by a model's explanation recover known, trait-relevant GO terms (via hypergeometric enrichment testing).
- DEG Overlap: Statistical significance ( $-\log_{10}$ p-value) of the overlap between model-explained genes and independently derived Differentially Expressed Gene (DEG) lists.

Key Results

Predictive Performance: HGNNs consistently outperformed both MLPs and GCNs across all five traits and cross-validation settings. For instance, the HGNN with UniGAT achieved the lowest rMSE for Leaf Area (0.961) and Inflorescence Height (0.999), compared to MLPs (1.058, 1.051) and GCNs (1.197, 1.338).
Explanatory Fidelity:
- BER Scores: HGNNs demonstrated significantly higher Biological Explanation Recall than GCNs and MLPs. For photosynthetic traits, BER increased monotonically with $k$ for HGNNs, whereas GCNs achieved near-zero BER at lower $k$ values, suggesting their explanations were biologically diffuse.
- DEG Overlap: HGNNs (specifically HConv) showed strongly significant overlap with independent DEG lists (p-values ranging from 5.4 to 28.4), whereas GCNs and MLPs showed low or non-significant overlap.
Interpretation: The results indicate that while dense regression (MLP) and pairwise graph structures (GCN) can achieve moderate predictive accuracy, they fail to produce explanations aligned with known biology. The structural prior provided by hypergraphs (GO-term groupings) is critical for generating biologically interpretable gene sets.

Significance and Claims
The authors claim that GRAFT is the first dataset to provide linked multimodal gene information and heterogeneous phenotypic data for the same Arabidopsis thaliana specimens. Its primary significance lies in:

Bridging the Data Gap: It addresses the critical lack of benchmarking resources that combine genomic and phenomic profiles from the same individuals, enabling direct genotype-to-phenotype modeling.
Validating Structural Priors: The work demonstrates that incorporating biological knowledge (via hypergraphs) directly into model architecture yields interpretability gains that purely data-driven approaches cannot replicate at this sample size.
Facilitating G2P Research: By providing a public dataset, benchmark tasks, and an explanatory framework, GRAFT aims to lower the barrier for the machine learning and biology communities to tackle the G2P challenge, specifically aiding in the identification of important genes for plant breeding and trait prediction.

Limitations
The authors acknowledge that the small sample size (24 linked specimens) limits absolute regression performance and statistical power. Additionally, the dataset is specific to Arabidopsis thaliana, and the transferability of these models to crops or other species remains an open question. Finally, the BER metric is bounded by the completeness and current state of Gene Ontology annotations.