Inferential Mechanics Part 1: Causal Mechanistic Theories of Machine Learning in Chemical Biology with Implications

Machine learning techniques are now routinely encountered in research laboratories across the globe. Impressive progress has been made through ML and AI techniques with regards to large data set processing. This progress has increased the ability of the experimenter to digest data and make novel predictions regarding phenomena of interest. However, machine learning predictors generated from data sets taken from the natural sciences are often treated as black boxes which are used broadly and generally without detailed consideration of the causal structure of the data set of interest. Work has been attempted to bring causality into discussions of machine learning models of natural phenomena; however, a firm and unified theoretical treatment is lacking. This series of three papers explores the union of chemical theory, biological theory, probability theory and causality that will correct current causal flaws of machine learning in the natural sciences. This paper, Part 1 of the series, provides the formal framework of the foundational causal structure of phenomena in chemical biology and is extended to machine learning through the novel concept of focus, defined here as the ability of a machine learning algorithm to narrow down to a hidden underpinning mechanism in large data sets. Initial proof of these principles on a family of Akt inhibitors is also provided. The second paper containing Part 2 will provide a formal exploration of chemical similarity, and Part 3 will present extensive experimental evidence of how hidden causal structures weaken all machine learning in chemical biology. This series serves to establish for chemical biology a new kind of mathematical framework for modeling mechanisms in Nature without the need for the tools of reductionism: inferential mechanics.

💡 Research Summary

**
The manuscript “Inferential Mechanics: Causal Mechanistic Theories of Machine Learning in Chemical Biology” presents a comprehensive theoretical and empirical investigation into why machine‑learning (ML) models in chemical biology often fail to generalize despite being trained on high‑quality data. The authors argue that the root cause lies in the neglect of underlying causal structures that govern the relationship between chemical structure and biological activity. They introduce a formal framework based on Judea Pearl’s causal calculus, defining a causal model C as a triple (U, V, F) of unobserved background variables, observed variables, and functional relationships. This model is represented as a Directed Acyclic Graph (DAG) that captures the flow of causation from experimental conditions to observed outcomes.

A central distinction is made between Total Effect (TE) and Direct Effect (DE). While TE measures the overall impact of an intervention on an outcome, chemical‑biological research is typically interested in DE—the effect of a structural change on activity when all other variables (including hidden confounders) are held constant. The authors demonstrate how Back‑door and Front‑door adjustment formulas can be applied to identify DE even when key variables are unobserved, thereby allowing causal inference from observational datasets.

The novel concept of “focus” is introduced as a meta‑learning capability that enables an ML algorithm to discover hidden mechanistic substructures within a large, apparently homogeneous dataset. Traditional feature engineering reduces molecules to fingerprint bit vectors, which may discard crucial mechanistic information such as binding site identity (M) and binding dynamics. Focus, by contrast, partitions the data according to inferred mechanistic categories (e.g., distinct binding pockets) and trains separate models on each partition. This approach effectively isolates the Direct Effect of structure on activity within each mechanistic context, mitigating the dilution of signal caused by mixing disparate mechanisms.

To illustrate the theory, the authors analyze a set of Akt kinase inhibitors. Two series (s1 and s2) inhibit the same protein but bind to different pockets (m1 and m2). When the full dataset is used for training, the aryl‑methyl (Ar‑Me) fingerprint bit appears statistically unrelated to activity, and the model ignores it. However, after applying focus to separate the data by binding pocket, the Ar‑Me bit emerges as a decisive predictor for the series that binds at m1. This reversal exemplifies Simpson’s paradox: a trend that disappears or reverses when data are aggregated but becomes clear when stratified by a hidden variable. The authors also reinterpret the superior performance of the “Causal‑Chemprop” model on seed‑like compounds as an implicit focus on a shared hidden mechanism rather than a purely structural similarity effect.

The paper concludes that integrating causal calculus with the focus paradigm yields an “inferential mechanics” framework that (1) makes hidden causal structures explicit, (2) forces ML models to learn true mechanistic relationships rather than spurious correlations, and (3) provides a systematic diagnostic for causal errors that cannot be fixed by conventional best‑practice guidelines alone. The authors outline a three‑part series: Part 2 will formalize chemical similarity within this causal context, and Part 3 will present extensive experimental validation showing how hidden causal structures degrade ML performance across diverse chemical‑biological datasets. Overall, the work offers a rigorous, mathematically grounded pathway to more reliable, mechanistically interpretable machine‑learning models in chemical biology.