Systems Biology and Multiscale Modeling

Published

July 7, 2026

Systems biology is the control layer of AI for biology. A single-cell embedding, pathway score, or tissue image becomes actionable only when it connects to a mechanism that survives perturbation. The central question is not whether a model summarizes biological state. The question is whether it identifies a regulatory, metabolic, spatial, or physiological relationship strongly enough to guide the next experiment.

Learning Objectives

Use this chapter to:

  • Explain how networks, pathways, feedback, metabolism, tissues, and multiscale models control biological state.
  • A network edge, pathway score, flux model, and tissue simulator are different claims with different evidence requirements.

Prerequisites: Single-Cell Foundation Models for the representation layer; Perturbation Prediction and Virtual Cells for intervention response; Evaluation Principles for Life Sciences AI for split design and baseline discipline.

Summary: Explain how networks, pathways, feedback, metabolism, tissues, and multiscale models control biological state. The field has mature pieces in metabolic modeling and perturbational genomics, while general multiscale prediction is still limited by identifiability and validation.

Key point: A network edge, pathway score, flux model, and tissue simulator are different claims with different evidence requirements. Open question: whether network and multiscale models can make prospective predictions that survive perturbation.

Bottom line: Systems biology is the connective tissue between molecular data, cells, tissues, virtual organisms, therapeutics, and experimental design.

Field Guide

What is this field trying to solve? Explain how networks, pathways, feedback, metabolism, tissues, and multiscale models control biological state.

What is the core idea? A network edge, pathway score, flux model, and tissue simulator are different claims with different evidence requirements.

What is the current state of the field? The field has mature pieces in metabolic modeling and perturbational genomics, while general multiscale prediction is still limited by identifiability and validation.

What do we know, and what remains open? Known reference points include SCENIC, BEELINE, CellOracle, Perturb-seq, NicheNet, SBML, BioModels, flux balance analysis, BiGG Models, PhysiCell, Morpheus, and whole-cell models. What remains open is whether network and multiscale models can make prospective predictions that survive perturbation.

Why does this matter? Systems biology is the connective tissue between molecular data, cells, tissues, virtual organisms, therapeutics, and experimental design.

Introduction

Biology is not organized as a set of independent features. Genes regulate other genes, proteins assemble into pathways, metabolites constrain growth, cells communicate with neighbors, tissues respond to mechanics and inflammation, and organisms adapt over time. Systems biology provides the language for those dependencies. It asks which components interact, which relationships change after intervention, which parameters are identifiable, and which predictions remain stable across context.

AI enters this field from two directions. The first direction is data-rich representation: single-cell models, spatial models, image models, genome models, protein models, and literature graphs. These systems organize observations at scale. The second direction is mechanistic modeling: differential equations, Boolean networks, agent-based tissue models, constraint-based metabolic models, and curated pathway simulators. These models encode assumptions about how a biological system operates.

The most useful current work sits between those traditions. Machine learning improves representation, denoising, parameter fitting, surrogate modeling, and experiment prioritization. Mechanistic modeling supplies constraints, conservation laws, causal structure, and falsifiable predictions. The combination is useful because neither tradition solves the whole problem alone. Representation learning without mechanism can mistake batch structure for biology. Mechanistic simulation without measurement can remain underdetermined.

Gene regulatory network inference illustrates the problem. Single-cell RNA-seq and chromatin-accessibility data can nominate transcription factors and target genes. SCENIC connected co-expression, motif enrichment, and regulon activity to infer regulatory programs from single-cell data (Aibar et al., 2017). Benchmark work showed that GRN inference methods vary substantially across settings and require careful evaluation rather than one default method (Pratapa et al., 2020). A recent review of GRN inference in the single-cell multi-omics era emphasizes the same point: multi-omic measurements improve the evidence base, but perturbational and experimental assessment remain central (Badia-i-Mompel et al., 2023).

Whole-cell modeling shows the other anchor. Karr and colleagues built a genotype-to-phenotype model of Mycoplasma genitalium by integrating many cellular processes into one executable model (Karr et al., 2012). The result remains important because it proved that whole-cell modeling is possible in a constrained organism, not because it solved whole-cell biology in general. The information burden was the lesson: even a minimal cell required curated mechanisms, parameter estimates, model integration, and phenotype checks.

What Counts as a Systems Biology Claim?

Systems biology claims differ by what they assert about the biological system:

Claim Type Example Evidence Needed
Association Gene A and gene B co-vary across cells Matched data, confounding checks, donor or batch-aware splits
Regulatory hypothesis Transcription factor A regulates target B Motif or chromatin support, expression timing, perturbation evidence
Causal intervention Knockdown of A changes B or pathway C Perturbation design, measured effect size, controls, time points
Dynamical mechanism Feedback loop creates oscillation or state transition Time-course data, fitted parameters, out-of-sample trajectory checks
Metabolic constraint Flux through pathway X limits growth or product formation Stoichiometric model, constraints, flux data or growth validation
Tissue-scale behavior Local cell rules reproduce measured tissue pattern Spatial calibration, independent tissue validation, uncertainty analysis
Multiscale prediction Molecular perturbation changes tissue or organism phenotype Evidence across scales, not only a molecular correlate

The word “causal” should be reserved for designs that support intervention reasoning. A network edge inferred from expression and chromatin data may be biologically plausible. It is not causal until changing the upstream node changes the downstream readout in the expected direction under a credible design. Even then, the claim is scoped to the cell type, perturbation strength, time point, and assay used.

Gene Regulatory Networks and Causal Perturbation Evidence

Gene regulatory network inference is a high-value use case because it turns high-dimensional molecular data into a set of testable regulatory hypotheses. The field has several evidence sources:

  • Co-expression across cells or samples
  • Transcription-factor motif enrichment
  • Chromatin accessibility near target genes
  • Enhancer-promoter contacts or three-dimensional genome data
  • Perturbational readouts from CRISPR, RNA interference, small molecules, or environmental exposures
  • Literature and curated pathway evidence

Each source carries different weight. Co-expression is weak because genes can move together downstream of a third factor. Motif evidence is useful but not sufficient because motif presence does not prove binding or activity. Chromatin accessibility adds context, but accessible DNA can be permissive rather than regulatory. Perturbation evidence carries more weight because it measures what happens after an intervention.

SCENIC is a useful example of structured inference because it does not equate expression correlation with regulation. It adds motif enrichment and regulon activity to improve biological plausibility (Aibar et al., 2017). CellOracle extends the logic toward in silico perturbation by using base GRNs and single-cell data to estimate downstream consequences of transcription-factor perturbation (Kamimoto et al., 2023). These are valuable hypothesis engines. Their outputs still require experimental confirmation.

Perturb-seq and related pooled perturbation methods are therefore central to systems biology evidence. Perturb-seq linked CRISPR perturbations to single-cell transcriptomic readouts, creating a measured genotype-to-expression bridge (Dixit et al., 2016). Genome-scale Perturb-seq expanded the data layer for genotype-phenotype maps (Replogle et al., 2022). These assays do not solve causality in every sense, but they move the claim from observational structure toward intervention evidence.

The practical rule is simple: a GRN edge should carry an evidence label. “Observed co-expression,” “motif-supported,” “chromatin-supported,” “literature-derived,” “perturbation-supported,” and “validated in independent context” are different claims. Collapsing them into one edge weight creates false precision.

Pathway, Metabolic, and Mechanistic Models

Pathway analysis is often the first systems layer added to AI biology. It is also one of the easiest to over-read. Pathway enrichment asks whether a list of genes overlaps a curated gene set. It is useful for triage, but it does not encode pathway dynamics, directionality, feedback, cell-type context, or flux. A pathway score is not a pathway model.

Mechanistic pathway models make stronger commitments. They specify nodes, interactions, parameters, timing, and sometimes equations. Standards such as SBML support exchange and reuse of computational models across tools (Keating et al., 2020). Model repositories such as BioModels provide curated computational models for reuse and inspection (Le Novere et al., 2006). These standards matter because a model that cannot be inspected, reproduced, or exchanged is hard to validate.

Metabolic modeling is the most mature systems biology example for many programs. Flux balance analysis uses a stoichiometric model plus constraints to estimate feasible metabolic states (Orth et al., 2010). Genome-scale reconstructions such as Recon 2 and BiGG Models provide reusable metabolic network infrastructure (Thiele et al., 2013; King et al., 2016). These models are not generic AI systems. They are constraint-based biological tools, and their value comes from mass balance, reaction curation, objective functions, measured constraints, and validation against growth, secretion, isotope tracing, or other metabolic readouts.

AI adds value around these mechanistic layers in several ways:

  • Parameter estimation from noisy high-dimensional data
  • Surrogate models that approximate expensive simulations
  • Active learning to choose the next perturbation or measurement
  • Model selection across competing mechanisms
  • Integration of transcriptomic, proteomic, metabolomic, and imaging data
  • Uncertainty estimation over unmeasured parameters and unobserved states

The danger is that a learned surrogate may be mistaken for the mechanism it approximates. A surrogate trained on a narrow simulation grid inherits the assumptions, blind spots, and invalid regions of the simulator. A hybrid model should therefore report both the mechanistic assumptions and the learned component’s domain of validity.

Whole-Cell, Virtual-Tissue, and Multiscale Models

Whole-cell modeling and virtual-tissue modeling make different commitments. A whole-cell model tries to represent cellular processes inside one cell or cell type. A virtual-tissue model represents many cells interacting in space, often with mechanical, chemical, and signaling rules. A multiscale model links molecular, cellular, tissue, and sometimes organism-level behavior.

The Karr M. genitalium model remains the canonical whole-cell reference because it integrated many cellular processes in a minimal organism and produced genotype-to-phenotype predictions for selected outcomes (Karr et al., 2012). Its relevance to AI is not that modern neural models have replaced it. Its relevance is that executable biology requires scope control, curated mechanisms, and measured phenotypes.

Virtual tissues often use agent-based or hybrid discrete-continuum simulation. PhysiCell provides an open-source platform for multicellular systems biology, including cell agents, diffusion, mechanics, and phenotype rules (Ghaffarizadeh et al., 2018). Morpheus similarly supports modeling and simulation of multicellular systems (Starruss et al., 2014). These tools are useful when the biological question is spatial: tumor growth, tissue patterning, wound healing, immune infiltration, morphogenesis, or organoid architecture.

Multiscale modeling becomes difficult because errors compound across scales. A gene-expression prediction may be adequate for a transcriptional endpoint, weak for a protein endpoint, weaker for cell behavior, and inadequate for a tissue or organism phenotype. Each scale transition adds unmeasured variables, latency, feedback, and context dependence. The model must therefore state what is actually being forecast: a transcript, a pathway activity score, a flux, a cell-state proportion, a spatial pattern, a functional assay, or a phenotype.

Current Models, Datasets, and Benchmarks

No single benchmark covers systems biology. The field requires task-specific evidence. GRN inference, perturbation response, metabolic modeling, and tissue simulation use different ground truth.

Resource or Method Main Use Evidence Role
SCENIC Regulon inference from single-cell data Structured GRN hypothesis generation (Aibar et al., 2017)
BEELINE GRN inference benchmark suite Method comparison across synthetic and experimental data (Pratapa et al., 2020)
CellOracle GRN-informed in silico transcription-factor perturbation Hypothesis generation for regulatory perturbations (Kamimoto et al., 2023)
NicheNet Ligand-target inference across interacting cells Cell-cell communication hypotheses (Browaeys et al., 2020)
Perturb-seq CRISPR perturbations with single-cell RNA readout Direct perturbational evidence (Dixit et al., 2016)
Genome-scale Perturb-seq Larger genotype-phenotype maps Training and validation layer for regulatory response (Replogle et al., 2022)
SBML Computational model exchange standard Reproducibility and reuse (Keating et al., 2020)
BioModels Curated computational model repository Reuse and model inspection (Le Novere et al., 2006)
Flux balance analysis Constraint-based metabolic modeling Feasible metabolic-state reasoning (Orth et al., 2010)
BiGG Models Genome-scale metabolic model repository Reusable metabolic reconstructions (King et al., 2016)
Whole-cell model of M. genitalium Executable minimal-cell model Whole-cell modeling precedent (Karr et al., 2012)
PhysiCell Agent-based multicellular simulation Virtual-tissue modeling (Ghaffarizadeh et al., 2018)
Morpheus Multicellular systems simulation Virtual-tissue and developmental simulation (Starruss et al., 2014)

The benchmark question should match the scientific claim. A GRN benchmark tests edge recovery or regulatory program structure. A perturbation benchmark tests response under held-out genes, combinations, cell types, donors, or laboratories. A metabolic benchmark tests growth, secretion, flux, isotope tracing, or product formation. A virtual-tissue benchmark tests spatial pattern, cell population dynamics, mechanics, or independent imaging. A general accuracy number across these tasks is not meaningful.

Evidence Standards for Causal and Multiscale Claims

Systems biology evidence should be graded by claim type:

Claim Minimum Evidence Stronger Evidence
Regulatory association Co-expression, chromatin, motif, or literature support Matched multi-omics across donors and contexts
Regulatory causality Perturbation of upstream node changes downstream readout Orthogonal perturbations, rescue, time-course, independent replication
Pathway mechanism Known pathway members move together Directional effect, dose-response, time ordering, functional readout
Metabolic constraint Stoichiometric feasibility Fluxomics, isotope tracing, growth or product validation
Tissue behavior Spatial model reproduces calibration pattern Independent tissue, perturbation, and quantitative imaging validation
Cross-scale prediction Molecular feature correlates with higher-scale outcome Prospective intervention showing the higher-scale outcome changes

The phrase “causal model” should not be used for a model that only predicts labels from observational data. Causal evidence begins with a defined intervention. It strengthens with temporal ordering, negative controls, positive controls, dose-response, independent replication, and a falsifying experiment. It strengthens further when the effect direction matches a mechanistic prior and when rescue experiments reverse the phenotype.

For multiscale claims, the main risk is scale inflation. A model may explain a molecular readout and then be discussed as if it explains a phenotype. A serious multiscale claim should name every bridge: molecular measurement to pathway activity, pathway activity to cell behavior, cell behavior to tissue state, and tissue state to organism phenotype. Missing bridges should be reported, not hidden.

Failure Modes

Observational Edges Treated as Causal Edges

The common failure is a network graph where all edges look equally authoritative. Co-expression, motif support, chromatin accessibility, and perturbation evidence have different meanings. A visually clean network can hide weak evidence.

Pathway Enrichment Treated as Mechanism

Enrichment results are often written as if the pathway was measured directly. In reality, a gene list overlapped a gene set. That finding may be useful, but it does not prove direction, timing, feedback, cell type, or function.

Batch, Donor, and Cell-State Confounding

Single-cell and spatial datasets often carry donor, laboratory, platform, dissociation, and batch structure. A network inferred from these data may recover technical or sampling structure rather than biology unless the analysis design addresses those variables.

Parameter Non-Identifiability

Many mechanistic models have more plausible parameter combinations than the data can distinguish. A model may fit the observed curve well while multiple mechanisms remain possible. Parameter uncertainty is therefore part of the result, not a technical appendix.

Surrogate Models Outside Their Valid Region

Machine-learning surrogates are useful when they approximate expensive simulations. They are unsafe when applied outside the simulation grid, training domain, or biological regime used for calibration.

Cross-Scale Overreach

The largest error is moving from molecular prediction to organism phenotype without measurement at the intermediate scales. A transcriptomic response does not automatically imply a protein response, pathway flux, tissue effect, or clinical endpoint.

Practical Validation Checklist

Before treating a systems biology output as actionable, a researcher or program leader should document:

  • Biological scale: gene, regulatory element, pathway, metabolite, cell, tissue, organ, organism, or ecosystem
  • Claim type: association, regulatory hypothesis, perturbation response, dynamical mechanism, metabolic constraint, tissue pattern, or multiscale prediction
  • Evidence label: observational, motif-supported, chromatin-supported, literature-derived, perturbation-supported, mechanistic, or externally validated
  • Intervention: genetic perturbation, compound, environmental exposure, cell-cell signal, mechanical change, or no intervention
  • Readout: RNA, protein, chromatin, imaging, viability, flux, secretion, tissue architecture, behavior, or phenotype
  • Timing: baseline, early response, late response, steady state, developmental stage, or disease stage
  • Controls: negative controls, positive controls, rescue where feasible, batch controls, donor splits, and technical replicates
  • Baseline: simple statistical baseline, mechanistic baseline, random network, known pathway, or domain standard
  • Uncertainty: parameter uncertainty, model uncertainty, measurement error, and sensitivity to preprocessing
  • External check: independent dataset, cell type, tissue, organism, laboratory, or prospective experiment
  • Falsifying result: the specific measurement that would weaken or reject the claim

This checklist is not administrative overhead. It is the difference between a useful systems biology model and a convincing diagram.

What is demonstrated?

Demonstrated capability includes structured GRN inference as a hypothesis-generation layer. SCENIC, BEELINE, CellOracle, NicheNet, and related methods provide evidence-weighted ways to nominate regulatory programs, ligand-target relationships, and candidate perturbations. The demonstrated claim is bounded: these methods organize evidence and prioritize experiments. They do not turn observational data into causal truth.

Demonstrated capability also includes perturbational genomics as a stronger evidence layer. Perturb-seq and genome-scale Perturb-seq measure how genetic interventions change transcriptomic state. These data support causal-leaning claims within the tested cell type, perturbation design, time point, and readout. They do not automatically transfer to all tissues, doses, cell states, or organism phenotypes.

Demonstrated capability includes pathway and metabolic models in domains where curation, constraints, and validation exist. Flux balance analysis, genome-scale metabolic reconstructions, SBML, BioModels, and BiGG Models are established infrastructure for mechanistic reasoning. Their strength comes from explicit assumptions and inspectable structure.

Demonstrated capability includes narrow whole-cell and virtual-tissue models. The M. genitalium whole-cell model, PhysiCell, and Morpheus show that executable biological models are feasible when scope is controlled and validation endpoints are clear.

What is theoretical?

Theoretical capability includes general hybrid models that connect foundation-model representations with mechanistic simulators across many biological scales. This direction is plausible because learned representations compress high-dimensional measurements, while mechanistic models impose biological constraints. The open question is whether the combined system remains calibrated after new interventions, tissues, developmental stages, and environments.

Theoretical capability also includes AI-assisted model discovery, where competing mechanisms are proposed, parameterized, and ranked by active learning. The use case is strongest for regulatory networks, signaling pathways, metabolic engineering, and tissue simulation. The evidence threshold is high because many mechanisms fit the same observations.

Theoretical capability includes multiscale pipelines that link genomic variation, regulatory programs, cell states, tissue architecture, and organism phenotype. Current evidence supports pieces of that chain. It does not yet support a general bridge across the full chain.

What is beyond current capability?

Beyond current capabilities includes treating observational regulatory networks as causal maps. Without intervention, time, and orthogonal evidence, most edges remain hypotheses.

Beyond current capabilities includes a general whole-cell simulator for arbitrary mammalian cells across disease, development, and environmental context. Existing whole-cell models are important precedents, not general solutions.

Beyond current capabilities includes reliable organism-level prediction from molecular measurements alone. Development, physiology, immune state, microbiome, behavior, and environment add causal structure that is not captured by a single omics layer.

Beyond current capabilities includes replacing wet-lab perturbation evidence with network inference. A network helps choose experiments. It does not substitute for them.

What would make this more promising?

A systems biology claim moves upward when it survives perturbation, context shift, and independent measurement. The strongest improvements would include:

  • Prospective perturbation experiments chosen before seeing the outcome
  • Held-out cell types, donors, tissues, disease states, or laboratories
  • Time-course measurements that distinguish direct from downstream effects
  • Orthogonal readouts, such as RNA plus protein, chromatin plus RNA, flux plus metabolite, or imaging plus function
  • Rescue experiments showing reversal of the predicted phenotype
  • External replication using a different platform or laboratory
  • Public benchmarks with simple baselines and preregistered held-out tasks
  • Uncertainty estimates that remain calibrated under distribution shift

The claim weakens when the result disappears under a linear baseline, donor-level split, batch-aware analysis, alternative preprocessing pipeline, or independent perturbation assay.

What should researchers, biotech teams, funders, and program leaders do with this?

Use systems biology models as decision aids for experiment selection. Ask the model to reduce the search space, rank perturbations, expose missing evidence, and make assumptions inspectable.

Require every network edge, pathway result, or cross-scale forecast to carry an evidence label. A program should know which parts are observational, which are mechanistic, which are perturbation-supported, and which are externally validated.

Fund validation at the same time as modeling. A systems biology program without perturbation capacity often produces attractive hypotheses and weak evidence. The validation budget should include controls, orthogonal readouts, held-out contexts, and failure analysis.

Keep the decision local. For target discovery, the decision may be which perturbation to run next. For metabolic engineering, it may be which pathway design to test. For tissue modeling, it may be which spatial mechanism to measure. The model output should be tied to a concrete next experiment.

Further Reading and Related Chapters

FAQ

Does a gene regulatory network prove causality?

No. A GRN is a structured hypothesis unless it includes perturbation evidence. Causal language requires an intervention, a measured response, appropriate controls, and validation in the relevant biological context.

What is the difference between pathway enrichment and pathway modeling?

Pathway enrichment tests overlap between a gene list and a curated gene set. Pathway modeling represents interactions, directionality, constraints, timing, or flux. Enrichment is triage. Modeling is a mechanistic claim.

When is a multiscale model credible?

A multiscale model is credible when each scale transition is specified and tested. Molecular to pathway, pathway to cell behavior, cell behavior to tissue pattern, and tissue pattern to phenotype all require evidence.

What is the strongest validation experiment?

The strongest experiment is usually prospective perturbation with an orthogonal readout in a held-out biological context. A rescue experiment adds further weight when feasible.

How should program leaders use systems biology AI today?

Use it to choose experiments, organize evidence, and test mechanisms. Do not use it as a replacement for perturbation evidence or as a shortcut around biological validation.