Unpublished draft

Taxonomy of non-mechanistic and causal models in biology

Mechanistic models in molecular cell biology explain complicated phenomena like nutrient sensing or apoptosis by decomposing them into:

  • the molecules (receptor A, kinase B, cell membrane…)
  • how these molecules act on each other (A dimerizes, B phosphorylates C, D undergoes conformation change…)

A circle-and-arrow chart of these molecules is called the mechanism. Knowing mechanisms is useful for many things, like for example picking targets for a therapeutic intervention.

In contrast, many other models are non-mechanistic. Non-mechanistic models abstract away molecules and interactions because they are unknown, irrelevant, or too complicated to model. Instead, these models introduce higher-level abstractions. These abstractions are also useful to know.

Biological problemModeled abstractionModeling methodWhat is inferredWhy it is non-mechanistic
Ligand changes cellular responseInput-output transfer functionHill curve, spline, Gaussian processEC50, response ceiling, response steepnessDoes not specify receptor, kinase, transcription-factor, or feedback machinery
Protein disappears after synthesis is blockedEffective abundance-decay processExponential or biexponential decayApparent half-life, decay ratesCollapses degradation pathways, compartments, binding states, and synthesis history
Cells increase in numberEffective population growth rateExponential, logistic, Gompertz modelGrowth rate, carrying capacity-like parameterDoes not represent the cell-cycle control network producing division
Drug combinations alter viabilityDrug-response surfaceBliss, Loewe, ZIP, response-surface regressionSynergy or antagonismRepresents interaction at the phenotype level rather than molecular interaction
Sequence variants have different activitiesSequence-to-fitness functionLinear model, Gaussian process, random forest, neural networkPredicted activity of unseen variantsCan predict function without representing structure or molecular chemistry
Enhancers produce different expression levelsSequence-to-expression functionRegression, CNN, transformerRegulatory activity from sequenceLeaves transcription-factor binding, nucleosome dynamics, looping, and transcription kinetics implicit
Thousands of genes covary across cellsLow-dimensional cell statePCA, factor analysis, variational latent-variable modelLatent axes such as activation or differentiation stateState dimensions summarize many causal molecular processes
Cells form recognizable transcriptional groupsDiscrete cell-state classesMixture model, graph clustering, classifierCell type or state assignmentClassification can work without identifying what generates each state
Cell states seem ordered through differentiationPosition along a developmental trajectoryPrincipal curve, diffusion pseudotime, graph trajectoryRelative progression through a state transitionOrdering cells does not specify what molecular network drives progression
Perturbing genes changes a phenotypeGene-to-phenotype effectRegression, generalized linear model, hierarchical modelEffect size of each perturbationEstablishes intervention-response relationships without filling in intermediate molecular steps
Pairs of gene perturbations behave unexpectedlyGenetic interaction functionInteraction regression, epistasis scoreSynthetic lethality, suppression, enhancementCharacterizes causal relations at the gene level without specifying molecular mediation
Many molecular features correlate with ageEffective biological-age coordinatePenalized regression, random forest, neural networkPredicted chronological or biological ageAge predictor need not encode any causal theory of aging
Cell morphology changes under perturbationMorphological phenotype spaceEmbedding model, classifier, nearest-neighbor modelPhenotypic similarity or classImages are mapped to phenotype without reconstructing molecular causes
Expression profiles predict treatment responseMolecular-state-to-outcome functionLogistic regression, random forest, neural networkDrug response, survival, subtypeCorrelational or predictive mapping can succeed with causal pathway structure unresolved
Cells fluctuate between recurrent phenotypic statesEffective state-transition processMarkov model, hidden Markov modelTransition probabilities and dwell timesSpecifies state dynamics without specifying the molecular circuitry generating transitions
A signaling output oscillatesEffective oscillatorSinusoidal fit, autoregressive model, state-space modelPeriod, phase, damping, coherenceDescribes temporal behavior without representing the feedback loop producing it

Causal inference frameworks

When doctors, economists, and computer scientists reason about genetic or cellular effects, their claims often bottom out in probabilities or effect sizes — statements like “this gene causes a 50% higher risk of cancer” or “these alleles add +5 IQ points.” That style is best thought of as predictive causality, distinct from the mechanistic causality sought in wet-lab biology.

Philosophers of science and quantitative biologists distinguish several causal frameworks:

FrameworkMain ideaTypical use in biology
Interventionism (Woodward)X causes Y if manipulating X changes YKnockouts, CRISPR perturbations, drug treatments
Manipulationist causation (Pearl/Woodward tradition)Causation defined through interventions rather than correlationsGenetic perturbation studies
Potential Outcomes Framework (Rubin Causal Model)Compare outcome under treatment vs no treatmentClinical trials, increasingly genomics
Structural Causal Models (Pearl)Directed acyclic graphs and structural equationsEpidemiology, systems biology, genomics
INUS conditions (Mackie)Cause is an insufficient but necessary part of an unnecessary but sufficient complexMultifactorial disease causation
Mechanistic causationExplain causal relations by identifying entities and activities producing an effectDominant style in molecular biology
Probabilistic causation (Suppes)Causes raise probabilities of effectsCancer epidemiology, risk factors
Counterfactual causation (Lewis)If X had not occurred, Y would not have occurredEvolutionary biology, epidemiology
Actual causationWhich particular event caused a particular outcome?Precision medicine, pathology
Causal network inferenceInfer causal graph from observational dataTranscriptomics, signaling networks
Dynamical systems causationCausation as state transitions in dynamical systemsDevelopmental biology, physiology
Information-theoretic causalityCausal influence measured via information flowNeuroscience, systems biology

See also