Skip to content

Systems Biology, Data Science, Digital Health and Personalised Medicine

Networks and motifs, flux balance analysis, pathway and disease modelling, omics data science, digital health and personalised medicine for CSIR NET LS-BT.

  • 20 min read
  • Updated
On this page · 15 sections
  1. Prerequisites
  2. The Hook: The Part List Is Not the Machine
  3. What Systems Biology Is
  4. Biological Networks
  5. Quantitative Models
  6. Data Science in Biology
  7. Pathway and Disease Modelling
  8. Digital Health and Personalised Medicine
  9. Memory Tricks
  10. Don’t Confuse
  11. Common Mistakes
  12. Practice
  13. Quick Revision Table
  14. Key Takeaways
  15. What’s Next

Prerequisites

The Hook: The Part List Is Not the Machine

How CSIR NET asks it. New topic in the revised LS-BT syllabus, with no past CSIR papers. Expect Part B on network terms (hub, degree, scale-free, motif), model types (ODE vs Boolean vs constraint-based), what FBA assumes, and what pharmacogenomics means. Expect Part C on reading a small network, a steady-state or Michaelis–Menten calculation, sensitivity and specificity of a biomarker test, or interpreting a machine-learning result.


What Systems Biology Is

Systems biology studies biological systems as integrated networks of interacting components, combining high-throughput measurement with computational models, and it works in cycles:

graph LR
    A["Measure
omics, imaging,
perturbation screens"] --> B["Build model
network, ODEs,
Boolean, FBA"] B --> C["Simulate and
generate predictions"] C --> D["Test by new
experiment
(perturbation)"] D --> A
ContrastReductionist approachSystems approach
QuestionWhat does this gene do?How do these components together produce this behaviour?
DataOne or few variables, deepMany variables, genome-wide
ModelVerbal or localExplicit, quantitative, predictive

Emergent properties cannot be read off single parts: robustness to perturbation, bistability (two stable states, giving a switch, as in cell-fate commitment), oscillation (cell cycle, circadian clock, as with negative feedback plus delay), ultrasensitivity, adaptation and homeostasis.

Top-down work starts from genome-wide data and infers network structure; bottom-up work builds a detailed kinetic model of a well-characterised module. Synthetic biology is the engineering counterpart: build the circuit (a toggle switch, a repressilator) to test whether the design principle holds.


Biological Networks

NetworkNodesEdges
Protein–protein interaction (interactome)ProteinsPhysical interactions (undirected); from yeast two-hybrid, affinity purification–mass spectrometry; STRING, BioGRID, IntAct
Gene regulatoryTranscription factors and target genesDirected activation or repression; from ChIP-seq, perturbation and expression data
MetabolicMetabolites (and reactions)Enzymatic conversions, directed and stoichiometric; KEGG, MetaCyc, genome-scale reconstructions
SignallingReceptors, kinases, second messengersDirected activation and modification; Reactome
Co-expressionGenesCorrelation of expression across samples (WGCNA modules); correlation is not causation
Disease / phenotypeDiseases, genesShared genes or pathways (“diseasome”); drug–target networks

Graph vocabulary

TermMeaning
Degree (k)Number of edges at a node; in directed networks, in-degree and out-degree
HubA high-degree node; in many biological networks hubs are enriched for essential genes
Degree distributionMany biological networks are approximately scale-free (\(P(k) \sim k^{-\gamma}\)): most nodes have few links, a few have very many. This makes them robust to random node loss but vulnerable to targeted attack on hubs
Path length, diameterSteps between nodes; small-world networks have short average paths with high clustering
Clustering coefficientHow often a node’s neighbours are also neighbours of each other
Betweenness centralityHow many shortest paths pass through a node (“bottleneck”)
Module / communityA densely connected group, often a functional complex or pathway

Network motifs

Motifs are small wiring patterns that occur far more often than in randomised networks with the same degree distribution, each with a characteristic dynamic behaviour:

MotifWiringBehaviour
Negative autoregulationX represses XSpeeds the response, reduces noise in the steady state
Positive autoregulationX activates XSlows the response; with cooperativity can give bistability
Coherent feed-forward loopX → Y, X → Z, Y → Z, same signsSign-sensitive delay: filters brief input pulses
Incoherent feed-forward loopOne path activates, the other repressesPulse generator and speed-up; can give fold-change detection
Single-input moduleOne regulator, many targetsCoordinated, temporally ordered expression
Dense overlapping regulonMany regulators, many targetsCombinatorial control
Negative feedback with delayLoop with lagOscillation

Quantitative Models

FrameworkRepresentsNeedsGivesTypical use
ODE (deterministic kinetic)Concentrations changing in timeRate laws and parametersTime courses, steady states, bifurcationsSignalling and metabolic modules
Stochastic (Gillespie / chemical master equation)Individual molecular eventsSame, plus molecule countsDistributions, noiseLow copy numbers, gene expression bursts
Boolean / logicalNodes ON or OFF with logic rulesOnly the wiring and logicAttractors (steady states, cycles)Large networks with unknown parameters
Constraint-based (FBA)Steady-state reaction fluxesStoichiometry, exchange bounds, an objectiveFlux distributions, growth predictionsGenome-scale metabolism
Petri nets, rule-based, agent-basedDiscrete events, rules, individual cellsVariesQualitative or multiscale behaviourCombinatorial complexes, tissue models
Spatial (PDE)Concentration in space and timeDiffusion constants, geometryGradients, patternsMorphogen gradients

Rate laws and standards

Mass action: \(v = k[\mathrm{A}][\mathrm{B}]\). Enzymes: Michaelis–Menten, \(v = V_{max}[S]/(K_m + [S])\). Cooperative or ultrasensitive steps: the Hill function \(v = V_{max}[S]^n/(K^n + [S]^n)\), where \(n > 1\) sharpens the response, and large \(n\) with feedback can produce a switch.

Models are exchanged in SBML (Systems Biology Markup Language) and deposited in BioModels; CellDesigner, COPASI, COBRA and Tellurium are common tools, and SBGN is the standard graphical notation. Reproducibility conventions ask for the equations, the parameters, their provenance and the initial conditions.

Flux balance analysis

FBA predicts metabolic fluxes at steady state from a genome-scale metabolic reconstruction (gene–protein–reaction associations built from an annotated genome, then curated).

$$\mathbf{S}\,\mathbf{v} = 0, \qquad v_{\min} \le v_i \le v_{\max}, \qquad \text{maximise } \mathbf{c}^{\mathsf T}\mathbf{v}$$

with S the stoichiometric matrix (metabolites × reactions), v the flux vector and the objective usually biomass production. What it assumes and does not need: steady state (no metabolite accumulation) and an optimality assumption, but no kinetic parameters or concentrations. It therefore predicts fluxes and growth, not time courses. Standard uses: predicting growth on different carbon sources, in silico gene knockouts (essentiality), and choosing engineering targets for product yield (see Scale-up, Plant Design and Metabolic Engineering). Extensions: FVA (flux variability), pFBA (parsimonious), dynamic FBA, and methods that integrate expression data. The solution is usually not unique: many flux distributions attain the same optimum.


Data Science in Biology

The data and the problems

Omics layers: genomics, epigenomics, transcriptomics (bulk and single-cell), proteomics, metabolomics, microbiomics, plus imaging and clinical records. Shared statistical difficulties:

  • High dimensionality, small n (\(p \gg n\)): thousands of features, tens of samples, so overfitting is the default outcome.
  • Multiple testing: 20,000 tests at α = 0.05 give about 1,000 false positives by chance. Control the family-wise error rate (Bonferroni) or, more usually in omics, the false discovery rate (Benjamini–Hochberg) and report q-values.
  • Batch effects and confounding: process samples in randomised batches, record covariates, correct explicitly (for example with established batch-correction methods); never compare cases run in one batch with controls run in another.
  • Normalisation before comparison, and missing values in proteomics and metabolomics.
  • Reproducibility: version the data and code, record software versions, follow FAIR principles (findable, accessible, interoperable, reusable) and deposit in public archives.

Common analysis steps

TaskTypical methods
Dimensionality reduction and visualisationPCA (linear, variance-maximising), t-SNE and UMAP (non-linear, for visualisation only; distances between clusters are not quantitative)
ClusteringHierarchical, k-means, graph-based (Leiden/Louvain in single-cell work)
Differential analysisModerated tests for microarrays, count models (negative binomial) for RNA-seq, with FDR control
Functional interpretationGO and pathway enrichment (over-representation tests), gene-set enrichment on ranked lists
Network inferenceCorrelation and mutual information, Bayesian networks, regression-based methods; causality needs perturbation or time-course data
Prediction (classification, regression)Regularised regression (LASSO, ridge), random forests, gradient boosting, support vector machines, neural networks
Multi-omics integrationJoint factor models, similarity network fusion, multi-view learning

Machine learning, honestly used

  • Supervised (labels known: diagnosis, survival) vs unsupervised (structure discovery: clusters, factors) vs reinforcement learning.
  • Split the data: train, validation and held-out test sets; use cross-validation; an independent external cohort is the real evidence. Feature selection must happen inside the cross-validation loop, or performance is inflated.
  • Bias–variance trade-off: overfitting (great training, poor test performance) vs underfitting.
  • Class imbalance: with 1% positives, a model that always says “negative” is 99% accurate, which is why accuracy alone is a bad metric. Report sensitivity, specificity, precision, recall, F1, ROC-AUC and precision–recall AUC.
  • Interpretability matters clinically: feature-importance methods, and the reminder that a predictive feature is not a causal one.
  • Deep learning has done most for structured problems with large training sets: variant effect prediction, regulatory-sequence models, image-based pathology, and protein structure prediction (AlphaFold).

Pathway and Disease Modelling

  • Pathway databases to models: curated maps (KEGG, Reactome, WikiPathways) give the topology; adding rate laws turns a map into a simulable model. Pathway enrichment tells you which processes a gene list touches; it does not tell you flux or direction.
  • Disease as network perturbation: disease genes cluster in network neighbourhoods (“disease modules”), and comorbidity often reflects shared modules. A drug acts at a point in the network, so its effects, including side effects and resistance, are network properties. Network pharmacology studies drug–target–disease networks instead of one drug–one target.
  • Cancer systems biology: driver versus passenger mutations, pathway-level recurrence (two tumours may hit different genes in the same pathway), clonal evolution and resistance, and signalling models that explain why blocking one node triggers compensatory feedback.
  • Infectious disease and immunology: host–pathogen interaction networks, within-host dynamics, and population-level epidemiological models (compartment models such as SIR, where the basic reproduction number \(R_0\) determines whether an outbreak grows, and the herd-immunity threshold is \(1 - 1/R_0\)).
  • Physiological and multiscale models: whole-cell models, organ-level models (for example cardiac electrophysiology), and physiologically based pharmacokinetic (PBPK) models used to predict dosing. Digital twin proposals extend this to an individual patient’s model, which remains largely a research goal.
  • Model validation is the whole point: a model is judged on predictions it was not fitted to, usually a perturbation the data did not contain. Fitting many parameters to one time course proves little; sensitivity analysis shows which parameters matter and identifiability analysis shows which can be determined at all.

Digital Health and Personalised Medicine

Digital health

Electronic health records, telemedicine, mobile health apps, wearables and sensors (continuous heart rate, activity, glucose), clinical decision support and clinical-grade AI tools. Real promise, real caveats: data quality (records are written for billing and care, not research), interoperability standards (HL7 FHIR, and coding systems such as ICD, SNOMED CT and LOINC), privacy and consent, algorithmic bias when training data under-represent a population, regulatory approval of software as a medical device, and the digital divide in access. De-identification, controlled-access repositories and federated analysis (models travel to the data, the data do not travel) are the usual technical answers.

Personalised (precision) medicine

Tailoring prevention, diagnosis and treatment to the individual’s genotype, phenotype, environment and lifestyle, in place of one-size-fits-all.

ComponentWhat it meansExample type
PharmacogenomicsGenotype predicts drug response or toxicityVariation in drug-metabolising enzymes makes someone a poor or an ultra-rapid metaboliser, changing the effective dose; HLA alleles predict hypersensitivity reactions; a companion diagnostic test decides eligibility for a targeted therapy
Molecular stratificationSub-classifying a disease by molecular profileExpression or mutation subtypes that carry different prognosis and treatment
Targeted therapyDrug matched to the tumour’s or patient’s molecular lesionChosen on the basis of a tested marker rather than tissue of origin alone
Risk predictionPolygenic risk scores summing many small-effect variantsUseful on average, weaker for individuals, and poorly transferable across ancestries when derived from one population
Liquid biopsy and monitoringCirculating tumour DNA and other analytes tracked over timeEarly detection of relapse or resistance
Clinical interpretationVariant classification and curated evidenceACMG/AMP criteria, ClinVar, pharmacogenomic guidelines

Supporting infrastructure: large biobanks and population genome programmes, standards for variant reporting, and the clinical genomics pipeline from variant calling to a reported result.

Ethical, legal and social issues (ELSI) are examinable: informed consent for broad future use, incidental (secondary) findings, the right not to know, genetic discrimination in insurance or employment, return of results to families, data sovereignty, equity of access, and the under-representation of non-European ancestries in reference data, which limits how well these tools work for most of the world.


Memory Tricks

Don’t Confuse

ThisvsThat
Hub: high degreevsBottleneck: high betweenness (can be low degree)
Scale-free: hub-dominated degree distributionvsSmall-world: short paths, high clustering
Network motif: over-represented wiring patternvsModule: densely connected functional group
ODE model: parameters, time coursesvsBoolean model: logic only, attractors
FBA: steady-state fluxes, optimality, no kineticsvsKinetic model: dynamics, needs parameters
DeterministicvsStochastic: matters at low molecule numbers
Co-expression (correlation)vsRegulation (causation): needs perturbation or time data
PCA: linear, distances meaningfulvst-SNE / UMAP: visualisation, inter-cluster distances not quantitative
FWER (Bonferroni)vsFDR (Benjamini–Hochberg): expected proportion of false discoveries
Sensitivity and specificity: prevalence-independentvsPPV / NPV: depend on prevalence
Pharmacogenomics: genotype → drug responsevsPolygenic risk score: many variants → disease risk
Training accuracyvsExternal validation: the only convincing evidence

Common Mistakes


Practice

All questions below are original practice questions written in CSIR NET style. They are not previous-year questions; this unit is new in the revised syllabus.

Level 1: Foundation

  1. What does \(\mathbf{S}\mathbf{v} = 0\) express in flux balance analysis?
  2. Which network motif, with cooperativity, can produce bistability?
  3. Which two quantities of a diagnostic test are independent of disease prevalence?
Answers
  1. Steady state: for every internal metabolite, production equals consumption.
  2. Positive autoregulation (positive feedback).
  3. Sensitivity and specificity (PPV and NPV are not).

Level 2: Part B Style

Q1 (practice). In a scale-free protein interaction network, removing a few of the highest-degree nodes:

  • (a) has little effect, because the network is robust to all perturbations
  • (b) fragments the network, because connectivity depends disproportionately on hubs
  • (c) increases the average path length only marginally, as in a random network
  • (d) changes the degree distribution to a Poisson distribution
Answer and explanation

(b). Scale-free networks are error-tolerant but attack-vulnerable: random removal usually hits low-degree nodes and barely matters, whereas targeted removal of hubs breaks the network into components. In a random (Poisson) network, by contrast, targeted and random removal differ far less.

Q2 (practice). Which statement about flux balance analysis is correct?

  • (a) It requires \(K_m\) and \(V_{max}\) values for every enzyme.
  • (b) It predicts time-dependent metabolite concentrations.
  • (c) It needs a stoichiometric matrix, flux bounds and an objective function, but no kinetic parameters.
  • (d) It always yields a unique optimal flux for each reaction.
Answer and explanation

(c). FBA is constraint-based: stoichiometry plus bounds plus an objective, solved by linear programming. Kinetic parameters (a) and time courses (b) belong to ODE models, and alternative optimal solutions (d) are common.

Q3 (practice). An RNA-seq experiment tests 20,000 genes at α = 0.05 without correction. The expected number of genes called significant by chance alone is:

  • (a) 5
  • (b) 50
  • (c) 1,000
  • (d) 20,000
Answer and explanation

(c). \(0.05 \times 20{,}000 = 1{,}000\). This is why FDR control (Benjamini–Hochberg q-values) or FWER control (Bonferroni, here α = 0.05/20,000 = 2.5 × 10⁻⁶) is mandatory.

Q4 (practice). A patient carries a variant making them a poor metaboliser of a drug that is inactivated by the enzyme concerned. Compared with a normal metaboliser given the same dose, this patient is expected to have:

  • (a) lower plasma drug levels and reduced effect
  • (b) higher plasma drug levels and a greater risk of dose-related toxicity
  • (c) unchanged plasma levels, since metabolism does not affect exposure
  • (d) higher levels only if the drug is a prodrug
Answer and explanation

(b). Slower inactivation means slower clearance, so exposure rises and dose-related adverse effects become more likely; dose reduction or an alternative drug is the usual response. Option (d) inverts the case: for a prodrug needing that enzyme for activation, a poor metaboliser would instead get too little active drug, which is a favourite twist in questions.

Level 3: Part C Style

Q1 (practice). A screening test for a condition with prevalence 2% has sensitivity 90% and specificity 90%. It is applied to 50,000 people.

(i) Construct the 2 × 2 table. (ii) Calculate PPV and NPV. (iii) The specificity is improved to 99% with sensitivity unchanged: recompute the PPV. (iv) What does this say about screening rare conditions?

Answer and worked solution
  1. Diseased = 0.02 × 50,000 = 1,000; healthy = 49,000. TP = 900, FN = 100; FP = 0.10 × 49,000 = 4,900, TN = 44,100.
  2. PPV = 900/(900 + 4,900) = 900/5,800 = 15.5%. NPV = 44,100/(44,100 + 100) = 99.77%.
  3. With specificity 99%: FP = 0.01 × 49,000 = 490, so PPV = 900/(900 + 490) = 900/1,390 = 64.7%.
  4. When the condition is rare, the false positives are drawn from the huge healthy majority, so specificity dominates the PPV: raising specificity from 90% to 99% lifted PPV from 16% to 65%, while the same change in sensitivity would hardly move it. Hence two-stage screening (a sensitive first test, a highly specific confirmatory test) and the caution about screening low-prevalence conditions.

Q2 (practice). A gene X represses its own transcription. Transcription follows \(v_{syn} = \dfrac{\beta}{1 + (x/K)^n}\) and the protein is degraded or diluted at rate \(\gamma x\), with \(\beta = 10\) nM min⁻¹, \(K = 5\) nM, \(n = 1\) and \(\gamma = 0.1\) min⁻¹.

(i) Write the steady-state condition and find the steady-state concentration. (ii) What happens to the steady state if \(\beta\) doubles? (iii) Qualitatively, what changes if the autoregulation were positive instead, with \(n = 2\)?

Answer and worked solution
  1. Steady state: \(\dfrac{\beta}{1 + x/K} = \gamma x\). Substituting, \(\dfrac{10}{1 + x/5} = 0.1x\), so \(50 = 0.1x(5 + x)\), giving \(x^2 + 5x - 500 = 0\) and \(x = \dfrac{-5 + \sqrt{25 + 2000}}{2} = \dfrac{-5 + 45}{2} = \mathbf{20}\) nM (the negative root is unphysical).
  2. With \(\beta = 20\): \(100 = 0.1x(5 + x)\) gives \(x^2 + 5x - 1000 = 0\), so \(x = (-5 + \sqrt{4025})/2 = (-5 + 63.44)/2 \approx \mathbf{29.2}\) nM. Doubling the synthesis rate raised the steady state by only about 1.5-fold, not 2-fold: that buffering of output against parameter change is exactly what negative autoregulation provides (along with a faster approach to steady state and lower noise).
  3. With positive autoregulation and \(n = 2\), the production curve is sigmoidal and can intersect the degradation line at three points: two stable states separated by an unstable one. The system becomes bistable — a switch with hysteresis, able to hold a cell-fate decision after the inducing signal is gone.

Q3 (practice). A group profiles tumours from 300 patients (transcriptome, exome, methylome) and trains a classifier that predicts response to a therapy with 94% accuracy on their data. Consider:

(A) The reported accuracy is weak evidence unless it comes from data held out from all training and feature-selection steps, and ideally from an independent cohort. (B) If only 6% of patients are responders, 94% accuracy could be achieved by predicting “non-responder” for everyone, so sensitivity, specificity and precision–recall must be reported. (C) Features with high importance in the model are thereby shown to be causal drivers of response. (D) Samples should be checked for batch effects and confounding (for example responders processed in a different run), because a classifier can learn the batch instead of the biology.

Which statements are correct?

  • (a) A, B and D only
  • (b) A and C only
  • (c) B, C and D only
  • (d) A, B, C and D
Answer and worked solution

(a) A, B and D.

  • (A) Correct: with \(p \gg n\), in-sample performance is inflated; feature selection must sit inside the cross-validation loop, and external validation is the standard of evidence.
  • (B) Correct: this is the class-imbalance trap, which is why accuracy alone is uninformative.
  • (C) Wrong: predictive importance is association, not causation. Causal claims need perturbation experiments or a credible causal design.
  • (D) Correct: technical confounding is one of the commonest reasons omics classifiers fail to replicate.

Quick Revision Table

ItemFact
Systems biologyNetworks + models + high-throughput data, in measure–model–predict–test cycles
Emergent propertiesRobustness, bistability, oscillation, ultrasensitivity, adaptation
Network typesPPI (STRING, BioGRID, IntAct), gene regulatory, metabolic (KEGG, MetaCyc), signalling (Reactome), co-expression (WGCNA), disease
Graph termsDegree, hub, betweenness (bottleneck), clustering coefficient, module, path length
Scale-free\(P(k) \sim k^{-\gamma}\); robust to random loss, fragile to hub attack; small-world = short paths + clustering
MotifsNegative autoregulation (fast, low noise); positive autoregulation (bistable); coherent FFL (delay/filter); incoherent FFL (pulse); SIM; DOR; delayed negative feedback (oscillation)
Model typesODE, stochastic (Gillespie), Boolean, constraint-based (FBA), rule/agent-based, PDE (spatial)
Rate lawsMass action; Michaelis–Menten \(v = V_{max}[S]/(K_m+[S])\); Hill \(n > 1\) ultrasensitive
StandardsSBML, BioModels, SBGN; COPASI, COBRA, CellDesigner
FBA\(\mathbf{Sv} = 0\), flux bounds, maximise objective (biomass); no kinetics, no time course; non-unique optima; FVA, pFBA, dynamic FBA
Data science issues\(p \gg n\); multiple testing (Bonferroni FWER vs BH FDR/q-values); batch effects; normalisation; FAIR data
MethodsPCA vs t-SNE/UMAP; clustering; differential expression; GO and pathway enrichment; network inference; LASSO, random forest, boosting, neural networks
ML disciplineTrain/validation/test split, cross-validation with feature selection inside, external cohort; bias–variance; class imbalance; ROC-AUC and PR-AUC
Disease modellingDisease modules, network pharmacology, driver vs passenger, pathway-level recurrence, SIR models (\(R_0\), herd immunity \(1 - 1/R_0\)), PBPK, digital twins
Digital healthEHR, telemedicine, wearables, decision support; FHIR/HL7, ICD, SNOMED CT, LOINC; privacy, bias, regulation, federated analysis
Personalised medicinePharmacogenomics (metaboliser status, HLA risk alleles, companion diagnostics), molecular stratification, targeted therapy, polygenic risk scores, liquid biopsy, ACMG/AMP interpretation
Test arithmeticSensitivity and specificity prevalence-independent; PPV depends on prevalence; specificity dominates PPV for rare conditions
ELSIConsent, incidental findings, right not to know, discrimination, data sovereignty, equity, ancestry bias in reference data

Key Takeaways

What’s Next

Advertisement
Advertisement