Taxonomy of non-mechanistic 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 problem | Modeled abstraction | Modeling method | What is inferred | Why it is non-mechanistic |
|---|---|---|---|---|
| Ligand changes cellular response | Input-output transfer function | Hill curve, spline, Gaussian process | EC50, response ceiling, response steepness | Does not specify receptor, kinase, transcription-factor, or feedback machinery |
| Protein disappears after synthesis is blocked | Effective abundance-decay process | Exponential or biexponential decay | Apparent half-life, decay rates | Collapses degradation pathways, compartments, binding states, and synthesis history |
| Cells increase in number | Effective population growth rate | Exponential, logistic, Gompertz model | Growth rate, carrying capacity-like parameter | Does not represent the cell-cycle control network producing division |
| Drug combinations alter viability | Drug-response surface | Bliss, Loewe, ZIP, response-surface regression | Synergy or antagonism | Represents interaction at the phenotype level rather than molecular interaction |
| Sequence variants have different activities | Sequence-to-fitness function | Linear model, Gaussian process, random forest, neural network | Predicted activity of unseen variants | Can predict function without representing structure or molecular chemistry |
| Enhancers produce different expression levels | Sequence-to-expression function | Regression, CNN, transformer | Regulatory activity from sequence | Leaves transcription-factor binding, nucleosome dynamics, looping, and transcription kinetics implicit |
| Thousands of genes covary across cells | Low-dimensional cell state | PCA, factor analysis, variational latent-variable model | Latent axes such as activation or differentiation state | State dimensions summarize many causal molecular processes |
| Cells form recognizable transcriptional groups | Discrete cell-state classes | Mixture model, graph clustering, classifier | Cell type or state assignment | Classification can work without identifying what generates each state |
| Cell states seem ordered through differentiation | Position along a developmental trajectory | Principal curve, diffusion pseudotime, graph trajectory | Relative progression through a state transition | Ordering cells does not specify what molecular network drives progression |
| Perturbing genes changes a phenotype | Gene-to-phenotype effect | Regression, generalized linear model, hierarchical model | Effect size of each perturbation | Establishes intervention-response relationships without filling in intermediate molecular steps |
| Pairs of gene perturbations behave unexpectedly | Genetic interaction function | Interaction regression, epistasis score | Synthetic lethality, suppression, enhancement | Characterizes causal relations at the gene level without specifying molecular mediation |
| Many molecular features correlate with age | Effective biological-age coordinate | Penalized regression, random forest, neural network | Predicted chronological or biological age | Age predictor need not encode any causal theory of aging |
| Cell morphology changes under perturbation | Morphological phenotype space | Embedding model, classifier, nearest-neighbor model | Phenotypic similarity or class | Images are mapped to phenotype without reconstructing molecular causes |
| Expression profiles predict treatment response | Molecular-state-to-outcome function | Logistic regression, random forest, neural network | Drug response, survival, subtype | Correlational or predictive mapping can succeed with causal pathway structure unresolved |
| Cells fluctuate between recurrent phenotypic states | Effective state-transition process | Markov model, hidden Markov model | Transition probabilities and dwell times | Specifies state dynamics without specifying the molecular circuitry generating transitions |
| A signaling output oscillates | Effective oscillator | Sinusoidal fit, autoregressive model, state-space model | Period, phase, damping, coherence | Describes temporal behavior without representing the feedback loop producing it |