When Everyone Can Build a Model: Mechanistic Inference in the Age of AI
Authors: Marvin van Aalst, Adèle H. Ribeiro, Anna Matuszyńska
Organizations: Computational Life Science, Faculty of Mathematics and Natural Sciences, RWTH Aachen University, Germany · Causal and Explainable Artificial Intelligence in Life Sciences, Faculty of Mathematics and Natural Sciences, RWTH Aachen University, Germany · Center for Computational Life Sciences, RWTH Aachen University, Germany
Modern deep learning and AI-assisted code generation are lowering the technical barrier to building mathematical models of biological systems that reproduce complex observations. This lets researchers with domain expertise build semi-mechanistic, hybrid models constrained by prior knowledge and augmented with learned structure without extensive training in modelling. However, easier construction does not make interpretation easier: predictive success of such so-called grey-box models does not establish that an inferred mechanism corresponds to the one that generated the data, and the learned component can absorb errors in the mechanistic part. Many biological modelling problems are underdetermined, so several, possibly infinitely many, models may reproduce the same data equally well. This ambiguity enters at two modelling choices: which variables can affect others, and which response function describes the underlying mechanism. For the first choice, causal inference provides a formal framework stating which causal relationships and intervention effects are identifiable from the data and available knowledge. For the second, ensemble analyses quantify the uncertainty and the range of behaviours still allowed by models that fit equally well. Together, they turn a single good fit into an explicit account of what is supported, what remains open, and which experiment would decide.
Figures & tables
Figure 1: Prediction or causation? Different questions, different ends. Starting from observational data sampled from an unknown structural causal model (SCM) M , the question decides the route. A prediction question asks for P(y∣x) , short for P(Y∣X=x) , and is answered by fitting a black-box model or, if domain knowledge is used, a grey-box model. Observational data alone do not determine how almost any M responds to interventions, whatever the model class (Causal Hierarchy Theorem, Bareinboim et al., 2022 ), so the fitted black-box model M′ generally differs from M under intervention. The fitted grey-box model M′′ is guaranteed to match the true effect only if its domain knowledge encodes correct causal constraints under which that effect is identifiable. The symbol ≡? between the two lower ends marks this comparison. A causal question asks for P(y∣do(x)) , short for P(Y∣do(X=x)) . Causal modelling or causal discovery yields causal constraints on the observed variables, encoded, for example, in a partial ancestral graph that represents the causal structures compatible with the data and domain knowledge, while accounting for latent confounding. If the effect is identifiable from these constraints, an identification formula expresses it in terms of the available data (here, adjustment for B ). If not, further domain knowledge or experiments can refine the constraints.
Figure 2: Feature choice decides whether the learned term of a grey-box model carries the true causal effect. All data are simulated from the structural causal model in panel 1, and the coupled mechanistic model is an illustrative Windkessel model. (1) Drug dose X affects blood pressure Y , while dashed circles mark unobserved variables. A is an instrument, which affects Y only through X ; B is an observed cause of X that shares an unobserved cause with Y ; and C is a collider of unobserved causes of X and Y . Feature sets that block the backdoor path through B , a non-causal route from X to Y , and exclude the collider are valid; all others are invalid. (2) Expected change in blood pressure as a function of dose, from linear regressions on different feature sets S . The valid set {X,B} recovers the true effect (black), whereas the confounded set {X} and the collider set {X,A,B,C} give positive slopes. The in-sample fit ( R2 ) is highest for the collider set, so goodness of fit does not reveal which set is valid. (3) The regression on S yields a learned term r(X∣S) , which scales the vascular resistance R of the Windkessel model and forms the grey-box model. (4) After a step increase in dose, the grey-box model with the valid set follows the true trajectory. The model with the collider set predicts a rise in pressure instead of a fall.
Modern Machine Learning (ML) and Artificial Intelligence (AI) models, especially large language models (LLMs), are increasingly used to generate scientific hypotheses and mechanistic explanations from observational data. This position paper argues that in the high-dimensional proxy regimes where modern ML excels, mechanistic learning is generically underdetermined: many incompatible mechanisms induce essentially the same observational relationships on the support of the data, so predictive success and coherent explanations are insufficient evidence of mechanism discovery. This underdetermination becomes uniquely hazardous with large language models (LLMs), which tend to collapse large equivalence classes of explanations into a single fluent narrative. This paper proposes concrete standards for ``mechanistic ML,'' and argues these norms are necessary if LLM-centered workflows are to support science rather than merely simulate it.
Tyler H. McCormick
Departments of Statistics and Sociology, University of Washington, Seattle, WA, USA.
A primary goal of science is to learn mechanistic world models from limited experimental data, both to explain observations and to predict novel interventions. We introduce the Model Discovery Agent (MDA), which combines LLM proposals for M-open model discovery, experiment design based on Value of Information, and approximate Bayesian inference over model structures, parameters, and stochastic latent trajectories. We apply MDA to learn symbolic reaction rate laws for ChemBench \citep{kabra2026autoscilab}, partially observed ODE models for GlucoseBench \citep{xie2018simglucose,kovatchev2009insilico}, and partially observed SDE models for a new stochastic single-neuron simulator we create. In the appendix, we also show results on various other domains from BoxingGym \citep{gandhi2025boxinggym}. We show that MDA has improved sample efficiency compared to various baseline methods, and the learned models are good predictors but also provide interpretable abstractions of each domain.
Kevin Murphy
Dept. Computer Science Univ. British Columbia, Canada.
Recent advances in foundation models have transformed AI for Science, enabling remarkably accurate predictive performance across domains ranging from protein folding to weather forecasting. Yet prediction alone does not constitute scientific discovery. Scientific understanding depends on uncovering the reusable explanatory mechanisms that generate observations, whereas contemporary machine learning remains fundamentally organised around predictive mappings rather than explanatory structure. In this paper, we argue that scientific discovery is fundamentally a problem of knowledge organisation. To this end, we introduce Mechanistic World Models, a new design paradigm that places reusable mechanisms at the centre of representation, computation and learning. Drawing on insights from the philosophy of science, we derive the computational capabilities required for discovery, identify the design principles and inductive pressures that encourage explanatory knowledge to emerge, and formalise the anatomy of a mechanism-centric world model. Finally, we show how diverse research directions including mechanistic interpretability, causal representation learning, equation discovery and modular architectures capture complementary ingredients of this paradigm while lacking a unified framework. We propose Mechanistic World Models as a conceptual foundation and computational blueprint for moving AI beyond predictive forecasting towards autonomous scientific discovery.
Ingmar Posner, Anson Lei, Bernhard Schölkopf
Applied AI Lab., University of Oxford, United Kingdom. · MPI for Intelligent Systems & ELLIS Institute, Tübingen, Germany.