Data Metabolism: An Overview

Updated 17 July 2026

Data metabolism is a concept where raw data are iteratively transformed into actionable insights through dynamic curation, feedback, and rebalancing.
It is applied in both vision-language model training and metabolomics to iteratively refine data quality, enhance model accuracy, and derive biological meaning.
The approach integrates methodologies such as constraint-based modeling, dynamical systems, and network-level analyses to bridge raw measurements with computationally interpretable outputs.

Data metabolism denotes a family of practices in which data are treated as active substrates of inference, curation, modeling, and feedback rather than as fixed observations. In the recent vision-language literature, the term is used explicitly for a data-centric, closed-loop development paradigm in which datasets are continuously constructed, diagnosed, revised, and rebalanced during model development. In metabolomics and systems biology, closely related usages describe the conversion of raw measurements into meaningful data objects, the integration of heterogeneous omics into mechanistic models, and the inference of latent metabolic states, fluxes, and phenotypes from those models. Across these usages, the common theme is iterative transformation of measurements into functionally interpretable representations that can guide further analysis or intervention (Zhang et al., 10 Apr 2025, Lakrisenko et al., 2021, Smilde et al., 2019).

1. Terminological scope and major usages

The literature uses data metabolism in multiple, technically distinct senses. In one sense, it is a named methodology for dataset construction and iteration in vision-LLMs. In another, it is a modeling perspective in which metabolomics data are processed through data theory, measurement theory, constraint-based modeling, or dynamical ODE systems to recover metabolic structure and dynamics. Related work at the metabolism–information boundary extends the idea further by coupling metabolic function, inherited information, and viability in protocell simulations (Zhang et al., 10 Apr 2025, Thomsen et al., 2024).

Usage	Core operation	Representative papers
VLM data metabolism	Closed-loop curation, diagnosis, and dataset revision	(Zhang et al., 10 Apr 2025)
Metabolomics data theory	Assign meaning, comparability, and scale to measurements	(Smilde et al., 2019)
Mechanistic metabolic data integration	Infer fluxes, states, and phenotypes from models	(Lakrisenko et al., 2021, Aurich et al., 2016, Mulgrave et al., 2020)

A unifying feature is that data are not regarded as self-interpreting. Raw intensities, counts, captions, image-text pairs, or clinical time series must first be aligned with a structure: a measurement scale, a stoichiometric network, an ODE system, a gene-enzyme rule set, or an iterative failure-analysis loop. This suggests that data metabolism is less a single formalism than a recurrent epistemic pattern: transformation of raw inputs into biologically or computationally actionable states through explicit rules, models, and feedback.

2. Data metabolism as a data-centric schema for vision-LLMs

In the VLM literature, “Data Metabolism” is defined as a data-centric, iterative development paradigm in which the dataset is continuously processed and updated throughout the model lifecycle. The biological analogy is made explicit: metabolism continuously breaks down nutrients and rebuilds the organism, while VLM development continuously constructs high-quality training data, consumes them during training, diagnoses model failures, and uses those failures to revise the dataset for the next round (Zhang et al., 10 Apr 2025).

The framework is organized around two phases. Data Anabolism covers dataset construction from raw web-scale sources, including filtering, deduplication, re-labeling, rewriting, and augmentation. Data Catabolism covers model diagnosis after training, including identification of unexpected behavior, missing knowledge and entity hallucination, task-format inconsistency, and cross-modal inconsistency and reasoning failure. The authors describe these two phases as a closed-loop system in which data curation and model diagnosis feed each other (Zhang et al., 10 Apr 2025).

The standard training pipeline has three stages. Stage 1: Visual Concept Alignment trains only the projector/adaptor, keeps parameters except the adaptor frozen, and does not use AnyRes. Stage 2: Detailed Image Understanding trains on high-resolution images and dense/text-rich data, uses AnyRes with images split into up to 4 sub-images, allows resolution up to 896×896, and makes all parameters trainable. Stage 3: Instruction Tuning focuses on reasoning, interaction, dialogue, and general multimodal tasks, uses AnyRes with up to 9 sub-images, allows resolution up to 1344×1344, and again makes all parameters trainable (Zhang et al., 10 Apr 2025).

The filtering codebook is presented as a six-step ordered pipeline:

Filter duplicated samples across sources
Filter samples with image-text mismatch
Address text-dominant QA pairs
Filter data sources with low quality
Filter low-quality responses with rule-based assessment
Filter low-quality responses with model-based assessment (Zhang et al., 10 Apr 2025)

The reported corpus began with more than 38M samples from 200+ sources and was reduced to about 12M high-quality samples after filtering and rewriting. The authors report removing 43% of the data through duplication cleanup, converting 70% of text-dominant samples to text-only training data, and using an incremental validation procedure in which new data are mixed with 5% to 20% of the original data (Zhang et al., 10 Apr 2025).

The demonstration system, Capybara-VL-7B, uses a SigLIP visual encoder, a Qwen2.5-Instruct-7B LLM, an MLP adaptor with 2×2 pooling, AnyRes processing, and NaViT-style interpolation of ViT positional embeddings. The authors report that the model is best on 15 of 18 metrics below 10B parameters. Their ablation-style table reports an average score progression from 54.2 for Initial to 69.1 for Full, with 67.7 for wo/ QF, and they report about 11.2% improvement on mathematical reasoning from rewriting answers (Zhang et al., 10 Apr 2025).

3. Metabolomics data as numbers with meaning

A different but closely related strand of work argues that metabolomics requires an explicit theory of what counts as data. The central proposition is concise:

data = numbers + meaning (Smilde et al., 2019)

This formulation is used to distinguish raw instrument outputs from measurements that have been assigned biological and analytical meaning. The paper introduces data theory to ask whether comparisons among numbers are meaningful at all, and measurement theory to ask what kind of comparisons are meaningful once the numbers do represent something. Three comparability notions are defined: column-conditional, row-conditional, and matrix-conditional. The paper also emphasizes lack-of-invariance, illustrated by unsynchronized time series in which a value at a nominally identical time point can correspond to different biological states and therefore change in meaning (Smilde et al., 2019).

Measurement theory is framed through an empirical relational system (ERS) and a numerical representation system (NRS). The classical scale types are then summarized by their permissible transformations: nominal by one-to-one transformations, ordinal by monotonic transformations, interval by positive linear transformations $x'=\alpha x+\beta\; (\alpha>0)$ , ratio by similarity transformations $x'=\alpha x\; (\alpha>0)$ , and absolute by the identity transformation $x'=x$ . The key warning is that metabolomics values are not automatically ratio-scaled simply because they are numeric (Smilde et al., 2019).

The paper identifies at least four levels of LC-MS metabolomics data. Level 0 consists of raw numbers. Level 1 consists of aligned, QC-corrected, global-IS-corrected data, which are column-conditional and may be ordinal-scaled or ratio-scaled depending on the calibration region. Level 2 uses group internal-standard correction and can become within-group matrix-conditional. Level 3 consists of calibrated concentrations, which are matrix-conditional and ratio-scaled. Level 4 is “tailor-made data,” in which concentration may no longer be the relevant representation because the biological question concerns activity under a nonlinear response model rather than concentration per se (Smilde et al., 2019).

This theoretical treatment has direct methodological consequences. If data are only ordinal-scaled, methods such as the Mann-Whitney test, nonmetric multidimensional scaling, and optimal scaling methods are appropriate. If data are ratio-scaled, metric methods such as PCA, PLS, and OPLS-DA become justified. The broader implication is that data metabolism in metabolomics begins before model fitting: it begins with specifying what a measurement means, what comparisons it supports, and what operations are legitimate (Smilde et al., 2019).

4. Constraint-based and network-level metabolic data integration

A major systems-biology usage of data metabolism consists of translating metabolomic or transcriptomic observations into network-constrained metabolic states. In this context, metabolism is represented by a stoichiometric matrix and flux vector, and data are used to set bounds, choose reaction activities, or personalize objectives. The standard steady-state constraint is $S v = 0$ , with lower and upper bounds $lb_i \le v_i \le ub_i$ , and an objective such as $\max_v \; c^T v$ used when a flux-balance formulation is adopted (Aurich et al., 2016).

MetaboTools formalizes extracellular metabolomics integration into three stages: prepare metabolomic data and the metabolic model; impose constraints and generate contextualized models; validate and predict phenotypes. The protocol includes metabolite-to-model mapping, transport and exchange curation, medium constraints, detection-limit conversion, uptake/secretion profile definition, semi-quantitative and quantitative constraint integration, compact contextualized subnetwork generation, and optional transcriptomic or proteomic integration. Exchange reactions are interpreted directionally, with negative flux denoting uptake and positive flux denoting secretion. Downstream analyses include single-gene deletion, reaction essentiality, union/intersection models, flux split analysis, steady-state sampling, and perturbation/parameter-plane analysis (Aurich et al., 2016).

Metabolitics extends this logic to personalized disease analysis from biofluid metabolomics measurements. It uses Recon2, with 5,324 metabolites, 7,785 reactions, and 100 non-overlapping pathways, and defines fold change for metabolite $m$ in individual $i$ as

$m_i^{foldChange} = \log m_i - \log H_m^{healthy}.$

The method constructs a personalized linear program with a dynamically built objective, runs flux variability analysis, computes reaction-level and pathway-level diff values relative to healthy controls, applies ANOVA with Benjamini–Hochberg correction, and trains classifiers on pathway-diff features. The reported mean breast-cancer result is about 89.9 ± 4.4% f1-score with logistic regression; Crohn’s disease yields 88% and colorectal cancer 89%. The paper also reports newly highlighted breast-cancer associations: significantly decreased Butanoate metabolism and significantly increased Arginine and Proline Metabolism (2002.04078).

Transcriptomics can be mapped to metabolism without FBA by aggregating expression through gene-enzyme-reaction rules. In “Metabolic enrichment through functional gene rules,” reactions in HMRcore are scored continuously from RNA-seq data using the rule that OR terms are summed and AND terms are reduced by the minimum expression. The proof-of-concept study uses 244 RNA-seq colorectal cancer samples from TCGA, with 229 reactions with rules and 375 metabolic genes associated with those reactions. The analysis reports that MSI tumors are more glycolytic than MSS tumors and that MSI prefers the NAD $^+$ -dependent malate dehydrogenase reaction while MSS prefers the NADP $x'=\alpha x\; (\alpha>0)$ 0-dependent reaction (Maspero et al., 2017).

A further example is the Drosophila hypoxia study, which constrained a stoichiometric model with thorax metabolomics during hypoxia and early recovery. There, the principal age-dependent phenotype appeared not during hypoxia itself but during reoxygenation. The simulations suggested that decreased ATP production during recovery in old flies reflected reduced recovery of mitochondrial respiration pathways and concomitant over-dependence on the acetate production pathway as an energy source (0811.2015).

5. Dynamical models, data assimilation, and multi-scale estimation

Where steady-state approaches are insufficient, data metabolism is carried out through dynamical models. A central review argues that dynamical metabolic models are preferable when the objective is to integrate heterogeneous data, capture transient behavior, and infer latent quantities such as metabolic fluxes, regulation, compartment-specific concentrations, and control points. The canonical form is

$x'=\alpha x\; (\alpha>0)$ 1

with possible expansion to isotopologue systems when stable-isotope tracing data are used (Lakrisenko et al., 2021).

A clinical example is the application of data assimilation to glucose–insulin metabolism in Type 2 diabetes. The model uses seven ODEs based on a previously published longitudinal framework and estimates only two patient-specific parameters: $x'=\alpha x\; (\alpha>0)$ 2 and $x'=\alpha x\; (\alpha>0)$ 3. The workflow combines interior-point deterministic optimization with random-walk Metropolis–Hastings MCMC, using three MCMC chains per patient, 10,000 iterations per chain, and a burn-in of 5,000 iterations. The parameter bounds are $x'=\alpha x\; (\alpha>0)$ 4 and $x'=\alpha x\; (\alpha>0)$ 5, and the proposal step-size coefficient is 0.1. A central finding is that the composite quantity $x'=\alpha x\; (\alpha>0)$ 6 is the most robust to whether insulin measurements are included, and that $x'=\alpha x\; (\alpha>0)$ 7 often jumps upward after gastric or bariatric surgery, capturing improvement in glucose tolerance within one to two OGTTs (Mulgrave et al., 2020).

At the whole-body scale, human metabolism has been modeled as a seven-compartment ODE system comprising Brain (B), Heart and lungs (H), Gut (G), Liver (L), Kidney (K), Muscle tissue (MP), and Adipose tissue (AP). The model tracks 16 metabolites and 2 hormones, insulin and glucagon, and uses stoichiometric reaction networks with Michaelis–Menten kinetics. It is used to simulate 72 hours of fasting after a single meal, a 72-hour regular feeding protocol with 4 meals per day, and 13 days of every-other-day feeding producing 33-hour fasting windows. The simulations show glycogen depletion, rising free fatty acids and glycerol, increasing ketone bodies, adipose triglyceride loss, and muscle protein breakdown under intermittent fasting (Bendsen et al., 2023).

Reproducible implementation is itself a form of data metabolism when historically influential metabolic equations are turned into inspectable, testable software. MetaboliSim provides an open Python implementation of the Mader model with a dynamic five-variable ODE system and a steady-state model for MLSS estimation. The dynamic state vector is

$x'=\alpha x\; (\alpha>0)$ 8

integrated with classical RK4 using frozen metabolic coefficients, with an adaptive RKF45 option at tolerance $x'=\alpha x\; (\alpha>0)$ 9. The implementation reports Newton convergence for PCr recovery in 4–6 iterations, step-size halving from 0.1 s to 0.05 s changing 5-minute blood lactate by less than 0.01 mmol/L, and agreement between dynamic and steady-state solutions below MLSS (Dunst et al., 6 Jun 2026).

At the multicellular scale, a data-enhanced agent-based model integrates intracellular ATP-based metabolism with spheroid mechanics. Each cell’s ATP level is compared against $x'=x$ 0 and $x'=x$ 1 to determine whether the cell is apoptotic, quiescent, or proliferative. Calibration against spheroid-area trajectories from A549 microfluidic experiments used a Gaussian-process surrogate with

$x'=x$ 2

and Sobol analysis identified the ATP consumption rate $x'=x$ 3 as the dominant parameter in the explored regime. Larger spheroids were associated with higher $x'=x$ 4 and lower $x'=x$ 5, linking growth heterogeneity to energetic efficiency and proliferation dynamics (Garcia-Gomez et al., 17 Dec 2025).

6. Visualization infrastructures and the metabolism–information boundary

Data metabolism also depends on infrastructure that makes metabolic structure explorable and interoperable. ReconMap 2.0 was introduced as a manually curated, biochemically intuitive, genome-scale human metabolic map consistent with Recon 2.04 from the Virtual Metabolic Human (VMH) database. It was drawn in CellDesigner, represents reactions as hyperedges and metabolites as nodes, and is delivered through MINERVA built on the Google Maps API. The platform supports content query, interactive navigation, links from every metabolite and reaction back to the corresponding curated entry in the VMH database, external database integration including ChEMBL, custom omics overlays, and user-submitted suggestions and corrections to curators. Flux distributions from the COBRA Toolbox can be visualized by running a simulation, using optimizeCBmodel, calling buildFluxDistLayout, and loading the resulting simple tab-delimited file into ReconMap, where flux values are translated into custom edge thickness and color. The platform also supports SBGN-compliant export and shares 168 metabolites as standard-identifier links with the Parkinson’s disease map (Noronha et al., 2016).

At the boundary of metabolism and information, protocell simulations make the coupling explicit. One model describes a fatty-acid vesicle carrying a ruthenium tris(bipyridine) complex as a photoactive energy transducer and a DNA-based co-factor containing 8-oxo-guanine that functions both as an electron relay and as an inheritable information carrier. Charge transport is modeled by

$x'=x$ 6

and functional information is quantified as

$x'=x$ 7

Using 160 simulated sequences, the authors report about 1.26 bits for $x'=x$ 8, about 0.89 bits for $x'=x$ 9, about 1.34 bits for joint viability, and 63 of the 160 sequences satisfying both thresholds. This work does not use “data metabolism” as a formal label, but it operationalizes a system in which inherited information actively participates in metabolic energy transduction and viability (Thomsen et al., 2024).

Origin-of-metabolism studies and large-scale chirality analyses extend the informational reading of metabolism to deep time. One study argues that metabolism assembled gradually from geochemical chemistry, identifies 166 core metabolic enzymes traceable to LUCA, and reports that 89 enzymes arose on the bacterial stem lineage and 38 on the archaeal stem lineage, supporting the claim that LUCA’s metabolism was enzymatically incomplete. It also argues that 17 cofactors can be functionally replaced by environmental transition metals and that phosphite can phosphorylate AMP to ADP in water using native metals (Mrnjavac et al., 9 Oct 2025). Another study defines a chiral order parameter

$S v = 0$ 0

uses network expansion and data from 40,023 genomes and metagenomes, and reports a chiral-enriched phase of metabolic organization with distinct scaling laws for individuals and ecosystems. This suggests that metabolism can be analyzed as a structured chemical information system whose large-scale organization carries signatures of evolutionary transitions (Malloy et al., 2 May 2025).

Taken together, these strands show that data metabolism spans at least three technical regimes: iterative dataset engineering for AI systems, formal treatment and model-based integration of metabolomics measurements, and explicit coupling of metabolic organization with information, visualization, and viability. The concept is therefore best understood as a cross-disciplinary schema for turning raw observations into structured, feedback-sensitive metabolic knowledge.