GAZE

GAZE is a physics-informed graph neural network that predicts metabolite concentrations directly from gene expression. Transcriptomic profiling is cheap and ubiquitous, while quantitative metabolomics remains costly and low-throughput, so most expression datasets have no paired metabolite measurements. GAZE closes that gap by inferring metabolite levels from the expression of the enzymes that produce and consume them, casting the problem as message passing over a curated metabolic network rather than a black-box regression.

The model's central innovation is that a single trained checkpoint can score metabolites it never saw during training. Instead of learning a dedicated set of weights for each target metabolite, GAZE uses a Metabolite-Conditioned Reader that queries learned pathway representations with a chemical embedding of the metabolite itself, so it carries no metabolite-specific parameters. This is what lets it generalize zero-shot to unseen metabolites and to independent cohorts without any fine-tuning — the behavior that distinguishes a reusable, pretrained model from a physics-informed network that must be refit for every new dataset.

GAZE was developed by Claudio Novella Rausell, Ton Rabelink, and Ahmed Mahfouz at Leiden University Medical Center and released as a bioRxiv preprint in 2026. The work has not yet been peer-reviewed.

#Key Features

• Physics-informed metabolic graph: Predictions are made over a unified network of 5,414 nodes and 16,307 edges linking genes, enzymes, reactions, and metabolites, so the model respects known metabolic connectivity rather than learning it from scratch.
• Multimodal node representations: Each node combines enzyme expression, functional embeddings derived from EC (Enzyme Commission) numbers, and ChemBERTa embeddings of metabolite SMILES strings.
• Zero-shot Metabolite-Conditioned Reader: The readout module conditions on a metabolite's chemical embedding instead of learned per-metabolite weights, carrying no metabolite-specific parameters and enabling prediction of metabolites absent from training.
• Demonstrated zero-shot generalization: In leave-one-metabolite-out evaluation across 50 held-out metabolites, GAZE halves the median R2 deficit relative to a standard GNN baseline, with 30% of unseen metabolites reaching positive predictive accuracy.
• External validation without fine-tuning: Applied unchanged to an independent clear-cell renal cell carcinoma (ccRCC) tissue cohort of 220 samples and 214 metabolites, it reaches a median Spearman correlation of 0.330.

#Technical Details

GAZE is trained on the Cancer Atlas of Metabolic Profiles, which pairs 18,044 genes with 180 measured metabolites across 867 cancer cell lines. Gene expression is mapped onto the enzymes of a reaction network, and messages propagate over the 5,414-node, 16,307-edge metabolic graph so that a metabolite's predicted concentration is informed by the expression of the enzymes surrounding it. Node features fuse three modalities: enzyme expression levels, EC-number functional embeddings, and ChemBERTa embeddings of metabolite SMILES. The Metabolite-Conditioned Reader uses each metabolite's chemical embedding to query the learned pathway representation, which is why one fixed checkpoint can be evaluated on metabolites and samples never seen in training. On standard held-out evaluation the model reaches R2 = 0.816, and it is benchmarked against the metabolic-inference methods scCellFie, MEBOCOST, and UnitedMet.

#Applications

GAZE lets researchers estimate metabolite concentrations for the large number of transcriptomic datasets that lack paired metabolomics — for example, in cancer cell-line panels and solid-tumor tissue where bulk RNA sequencing is routine but metabolite quantification is not. Because the reader generalizes to unseen metabolites, it can also produce estimates for compounds that were never directly measured in the training data, extending metabolic readouts beyond the assays available in any single study. This benefits cancer metabolism researchers and systems biologists who want a cheaper, expression-derived proxy for experimental metabolomics.

#Impact

GAZE demonstrates that metabolite inference can follow the pretrain-once, apply-broadly pattern: a single trained model transfers zero-shot to unseen metabolites and to an independent tumor cohort without fine-tuning, rather than requiring a fresh fit per dataset. That behavior, driven by conditioning on metabolite chemistry instead of learned per-metabolite weights, is the model's main contribution to metabolic modeling. Its significance is tempered by maturity and openness caveats researchers should weigh: performance on unseen metabolites is modest (30% of held-out metabolites reach positive R2; external median Spearman correlation of 0.330), the training data is cancer-focused, no public code or pretrained-weights repository and no model or data card have been identified, and the work remains a preprint that has not undergone peer review.