🧬 biology

Graph Neural Network based Hierarchy-Aware Embeddings of Knowledge Graphs: Applications to Yeast Phenotype Prediction

This paper presents a graph neural network method that incorporates ontology-derived semantic loss to generate hierarchy-aware knowledge graph embeddings, which significantly improve the prediction and biological interpretation of yeast gene deletion effects, including double and triple knockouts.

Original authors: Filip Kronström, Alexander H. Gower, Daniel Brunnsåker, Ievgeniia A. Tiukova, Ross D. King

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Filip Kronström, Alexander H. Gower, Daniel Brunnsåker, Ievgeniia A. Tiukova, Ross D. King

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine you are trying to teach a computer to understand the complex biology of yeast (a tiny fungus used in baking and brewing). You have two huge piles of information:

The Fact Book: A massive list of specific observations, like "If we delete Gene A and Gene B, the yeast grows 20% slower."
The Rule Book: A structured hierarchy of concepts, like "Enzymes are a type of Protein," and "Proteins are a type of Molecule." This is called an ontology.

For a long time, computer models were great at reading the Fact Book but often ignored the Rule Book. They learned the specific facts but missed the big picture logic. This paper introduces a new way to teach the computer to respect both the specific facts and the logical rules simultaneously.

Here is how they did it, using some creative analogies:

1. The "Box" Metaphor: Organizing the Rules

Usually, computers represent ideas as single points in space (like a dot on a map). If you want to say "All Proteins are Molecules," you have to force the "Protein" dot to be very close to the "Molecule" dot.

This paper uses Box Embeddings. Instead of dots, imagine every concept is a box (like a cardboard shipping box).

The "Molecule" box is huge.
The "Protein" box is smaller and sits inside the "Molecule" box.
The "Enzyme" box is even smaller and sits inside the "Protein" box.

This creates a perfect visual hierarchy. If a box is inside another, the computer "knows" that the smaller thing is a type of the larger thing. This is much better for handling the "Rule Book."

2. The "Messenger" Metaphor: The Graph Neural Network (GNN)

To make predictions, the computer needs to look at how genes interact. They used a Graph Neural Network (GNN), which acts like a team of messengers running through a city (the knowledge graph).

Each messenger stands at a gene.
They run to their neighbors, pick up information, and bring it back.
After a few rounds, every messenger knows not just about their own gene, but about the whole neighborhood of genes connected to it.

3. The "Strict Teacher" Metaphor: Semantic Loss

Here is the paper's main innovation. Usually, the messengers (the GNN) learn by trial and error to predict the outcome (like yeast growth). Sometimes, in their rush to get the answer right, they might accidentally break the rules (e.g., putting a "Protein" box outside the "Molecule" box).

The authors added a "Strict Teacher" (called a semantic loss) to the training process.

Every time the messengers update their knowledge, the Strict Teacher checks the boxes.
If a "Protein" box is floating outside the "Molecule" box, the teacher gives them a "penalty" (a math error score).
The computer has to learn to predict the yeast growth while keeping all the boxes neatly nested inside each other.

The Results: What Did They Find?

1. Better Predictions
When they tested this on predicting how yeast grows after deleting two genes (a "double knockout"), the model with the "Strict Teacher" performed significantly better than models without it.

The Analogy: It's like a student who studies both the specific practice problems and the textbook chapters on theory. They get better grades than the student who only memorized the practice problems.
The Score: They achieved a prediction accuracy (R² score) of 0.377, which is a solid improvement over the baseline.

2. Predicting the Unseen
They tested the model on "triple knockouts" (deleting three genes at once), which the model had never seen before. It still worked well, suggesting the model learned the logic of the biology, not just memorized the specific data points.

3. Discovering New Biology (The "Aha!" Moment)
Because the model understands the relationships between genes, the researchers could ask it: "Which traits are interacting to cause this result?"

The model flagged a connection between Inositol (a nutrient) and Salt Stress (NaCl).
The Experiment: The researchers went to a real lab and tested this. They grew yeast with and without inositol under salty conditions.
The Discovery: The experiment confirmed the model's guess! Adding inositol actually helped the yeast survive the salt stress. The model had successfully predicted a hidden biological relationship that hadn't been explicitly written down in the data before.

4. Checking the "Rule Book" for Errors
Finally, they showed that this "box" system could be used to check if the Rule Book itself was wrong. If you add a new rule (a new link in the graph) and it causes the boxes to get messy or the "Strict Teacher" to scream, it might mean the new rule doesn't fit with the rest of the knowledge. This helps scientists spot errors in their databases.

Summary

This paper built a computer brain that doesn't just memorize facts; it understands the structure of knowledge. By forcing the computer to keep its "concept boxes" neatly organized (like a well-stocked warehouse), it became better at predicting real-world biological outcomes and even helped scientists discover a new interaction between nutrients and stress in yeast.

Technical Summary: Hierarchy-Aware Embeddings of Knowledge Graphs via GNNs

Problem Statement
Knowledge Graphs (KGs) are widely used to represent structured domain knowledge, particularly in the life sciences where they are enriched with ontological information. While standard KG embedding (KGE) methods effectively learn from observed relational triples, they often fail to explicitly enforce global semantic constraints derived from underlying ontologies, such as class hierarchies (subClassOf) and disjointness. Standard Graph Neural Networks (GNNs), while powerful for aggregating local neighborhood information, do not inherently respect these hierarchical structures, potentially learning representations that violate known domain logic. The paper addresses the challenge of integrating volumetric representations (specifically box embeddings) with GNNs to create "hierarchy-aware" embeddings that align with ontological constraints while remaining effective for downstream quantitative prediction tasks.

Methodology
The authors propose a framework combining Graph Neural Networks with box-based embeddings, termed Hierarchy-aware GNNs.

Graph Construction: The method treats the KG as a set of TBox-style axioms. Class assertions ( $C(a)$ ) are rewritten as subsumption axioms ( $\{a\} \sqsubseteq C$ ), and role assertions ( $r(a, b)$ ) as existential restrictions ( $\{a\} \sqsubseteq \exists r.\{b\}$ ). This unifies instance-level facts and terminological knowledge into a single computational graph where nodes represent classes and nominals, and edges represent logical relations.
Architecture:
- GNN Backbone: The model utilizes a heterogeneous GNN based on the GraphSAGE architecture. It employs relation-specific message passing (SAGEConv modules) to aggregate information from local neighborhoods across different edge types.
- Box Embeddings: The output of each GNN layer is treated as a latent variable $\theta$ that is transformed into an axis-aligned hyperrectangle (box) using a parameterization that ensures valid box coordinates (lower bound $\le$ upper bound).
- Semantic Loss: To enforce ontological constraints, the model incorporates semantic loss terms during training. Two types of losses are evaluated for subsumption ( $C \sqsubseteq D$ $C ⊑ D$ ):
  - Distance Loss ( $L_{distance}$ ): Penalizes the geometric distance between a subclass box and the boundary of its superclass box.
  - Overlap Loss ( $L_{overlap}$ ): Penalizes based on the volume of intersection relative to the subclass volume.
  - Disjointness constraints ( $C \sqcap D \sqsubseteq \bot$ ) are enforced via corresponding negative losses ( $L^{-}_{distance}$ or $L^{-}_{overlap}$ ).
- Joint Training: The total loss function combines a task-specific loss (e.g., Mean Squared Error for regression) with the weighted semantic losses and regularization terms. This allows the model to learn task-specific features while adhering to the class hierarchy.

Key Contributions

Hierarchy-Aware GNN Framework: A novel architecture that integrates box embeddings into GNNs, enabling the simultaneous learning of graph structure and ontological hierarchies.
Semantic Loss Integration: The application of semantic loss terms (derived from ontology axioms) directly into the training of GNNs for quantitative prediction tasks, rather than just for static ontology embedding.
Yeast Phenotype Prediction: Construction of a comprehensive, ontology-enriched KG for Saccharomyces cerevisiae integrating data from SGD, GO, ChEBI, APO, and BioCyc.
Hypothesis Generation: A method for identifying interacting traits by analyzing edge-pair importance in the trained model, leading to experimentally validated biological hypotheses.
Link Evaluation: A proposed method for ranking potential KG revisions based on the distance changes in the resulting embeddings and the impact on semantic loss.

Results
The method was evaluated on the task of predicting cell growth fitness for double gene knockouts in yeast.

Predictive Performance: The hierarchy-aware GNN achieved a mean $R^2$ score of 0.360 using prior box embeddings. Incorporating the semantic loss (specifically the distance-based loss) further improved performance to 0.377, significantly outperforming baselines including standard GNNs without box embeddings, LightGBMs on sparse phenotype features, and ComplEx embeddings.
Generalization: The model demonstrated generalization capabilities when tested on triple gene knockouts (data not seen in training), achieving an $R^2$ of 0.415 with semantic loss, suggesting the learned representations capture robust biological traits.
Hypothesis Validation: The model identified a potential interaction between inositol utilization and osmotic stress resistance. An automated biological experiment confirmed this finding, showing that inositol supplementation significantly rescued cells from NaCl-induced stress, validating the model's ability to guide discovery.
Embedding Analysis: Visualizations of 2D embeddings for the molecular function domain showed that the GNN transforms prior embeddings to better respect disjointness and subsumption, though some hierarchical violations persisted, likely due to the complexity of the domain and the dimensionality of the embedding space.
Link Evaluation: Initial tests on ranking graph revisions showed that the distance metric could distinguish between true edges and randomly drawn edges for certain relation types, though variance was high.

Significance and Claims
The paper claims that incorporating high-level qualitative knowledge (class hierarchies) via semantic losses improves quantitative prediction performance in biological domains. The authors argue that this approach provides a regularizing effect and semantic grounding for the model's representations.

Crucially, the work demonstrates that models trained with semantic grounding can be used to generate interpretable hypotheses that lead to real-world scientific discoveries. The successful experimental validation of the inositol-NaCl interaction serves as a proof of concept for using computational models to guide biological experimentation. The authors remain modest regarding the link evaluation method, noting it is currently rudimentary and requires further investigation to effectively discriminate between true and random edges in complex graphs. They also acknowledge that the current setup does not utilize sequence information, which could further enhance gene representations, and that the simplification of individuals as boxes (rather than points) is a trade-off for enforcing hierarchical constraints.