💻 bioinformatics

Zero-Shot Metabolite Prediction from Gene Expression via Physics-Informed Graph Neural Networks

The paper introduces GAZE, a physics-informed graph neural network that integrates gene expression, enzyme functional embeddings, and metabolite descriptors to achieve accurate zero-shot prediction of metabolite concentrations across diverse biological contexts without requiring metabolite-specific training.

Original authors: Novella Rausell, C., Rabelink, T., Mahfouz, A.

Published 2026-07-02

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Novella Rausell, C., Rabelink, T., Mahfouz, A.

Original paper licensed under CC BY 4.0 (https://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 your body as a massive, bustling factory. Inside this factory, genes are the managers who send out instructions, enzymes are the workers who follow those instructions, and metabolites are the final products or raw materials moving through the assembly line.

The big challenge scientists face is: If we know what the managers (genes) are saying, can we predict exactly what the factory is producing (metabolites)?

Previous attempts to solve this were like using a static, paper-based map. They tried to guess the output based on fixed rules, but they often missed the dynamic nature of the factory or ignored the specific chemical "physics" of how things actually move. This meant they couldn't guess the output for new products they hadn't seen before.

Enter GAZE (Graph Attention for Zero-shot metabolite Estimation). Think of GAZE as a super-smart, AI-powered factory supervisor that learns the logic of the entire system rather than just memorizing specific products.

Here is how it works, using simple analogies:

The Unified Map: GAZE builds a giant, interactive map of the factory. It connects 5,414 different "stations" (nodes) with 16,307 "roads" (edges). On this map, it doesn't just look at the gene managers; it also understands the "job descriptions" of the enzyme workers and the unique "chemical fingerprints" of the products.
The "Zero-Shot" Superpower: Usually, to predict a new product, you need to study that specific product for months. GAZE is different. It uses a special tool called a "Metabolite-Conditioned Reader." Imagine this reader as a translator who can look at a product's blueprint (its chemical structure, or SMILES) and instantly understand how it fits into the factory's workflow, even if GAZE has never seen that specific product before. It doesn't need to be retrained for every new item; it just applies the rules it already learned.
The Physics Check: GAZE isn't just guessing; it's "physics-informed." This means it respects the laws of chemistry. It's like a supervisor who knows that you can't build a car without wheels, ensuring the predictions make sense in the real world.

How well did it do?

The researchers tested GAZE in three ways:

The Standard Test: They gave it a massive dataset from a "Cancer Atlas" (involving thousands of genes and hundreds of cell lines). GAZE predicted the output with high accuracy, getting an R² score of 0.816. In plain English, this means its guesses were very close to the actual results.
The "New Product" Test: They hid 50 metabolites from the AI and asked it to guess them without any prior training on those specific items. A standard AI struggled here, often guessing poorly. GAZE, however, cut the error rate in half compared to the standard AI. Remarkably, for 30% of these completely new, unseen products, GAZE actually made a positive, accurate prediction.
The Real-World Test: They took GAZE to a completely different dataset (kidney cancer tissue samples) without changing a single setting. It successfully predicted metabolite levels in this new environment, proving it can generalize its knowledge.

The Bottom Line:
GAZE is a new type of AI that understands the "language" of biology so well that it can predict chemical outputs for things it has never seen before, outperforming all other current methods. It bridges the gap between gene instructions and chemical results by treating the cell like a connected, logical network rather than a list of isolated facts.

Technical Summary: Zero-Shot Metabolite Prediction from Gene Expression via Physics-Informed Graph Neural Networks

Problem Statement
Predicting metabolite concentrations directly from gene expression profiles is a critical challenge for bridging the gap between regulatory programs and metabolic phenotypes. Existing approaches typically rely on static enzyme-metabolite mappings and often lack data-driven learning mechanisms or explicit biochemical constraints. These limitations hinder the ability of current models to generalize effectively, particularly when predicting concentrations for metabolites not explicitly seen during training (zero-shot scenarios).

Methodology: GAZE
The authors introduce GAZE (Graph Attention for Zero-shot metabolite Estimation), a physics-informed graph neural network designed to overcome generalization barriers. The architecture integrates three distinct data modalities within a unified metabolic graph comprising 5,414 nodes and 16,307 edges:

Enzyme Expression: Transcriptional data representing enzyme levels.
Functional Embeddings: Enzyme Commission (EC) functional embeddings.
Metabolite Descriptors: ChemBERTa-derived embeddings based on SMILES strings.

A core component of GAZE is the Metabolite-Conditioned Reader. This module utilizes the SMILES embedding of a specific metabolite to query learned pathway representations. Crucially, this design enables zero-shot prediction by ensuring the model contains no metabolite-specific parameters; instead, it learns a generalizable mapping from gene expression to metabolic states conditioned on the chemical structure of the target metabolite. The "physics-informed" aspect refers to the integration of biochemical constraints into the learning process to guide the model's predictions.

Key Contributions

Unified Graph Framework: Construction of a comprehensive metabolic graph integrating gene, enzyme, and metabolite data.
Zero-Shot Capability: Development of a parameter-efficient architecture that predicts unseen metabolites without retraining or fine-tuning on metabolite-specific data.
Physics-Informed Learning: The incorporation of biochemical constraints to improve prediction robustness compared to standard graph neural network baselines.

Results
GAZE was evaluated across three distinct scenarios using the Cancer Atlas of Metabolic Profiles (18,044 genes, 180 metabolites, 867 cell lines) and an independent external cohort:

Standard Cross-Validation: On the Cancer Atlas, GAZE achieved an $R^2$ of 0.816.
Leave-One-Metabolite-Out (LOMO) Zero-Shot: Across 50 held-out metabolites, the physics-informed variant significantly outperformed a standard GNN baseline. It reduced the median $R^2$ deficit by half (-0.34 vs. -0.72 for the baseline). Notably, 30% of the unseen metabolites achieved positive $R^2$ values.
External Validation: In an independent clear cell renal cell carcinoma tissue cohort (220 samples), GAZE achieved a median Spearman rho of 0.330 across 214 metabolites without any fine-tuning.

In all evaluation settings, GAZE demonstrated superior performance compared to existing tools, including scCellFie, MEBOCOST, and UnitedMet.

Significance
The paper positions GAZE as a significant advancement in linking gene regulation to metabolic outcomes. By successfully addressing the generalization gap in zero-shot prediction, the work demonstrates that integrating biochemical constraints and chemical structure descriptors within a graph neural network framework allows for robust metabolite estimation from transcriptomic data. This capability is instrumental for studying metabolic phenotypes in contexts where metabolite-specific training data is unavailable or limited.

Technical Summary: Zero-Shot Metabolite Prediction from Gene Expression via Physics-Informed Graph Neural Networks

More like this