Learning virulence-transmission relationships using causal inference

The relationship between traits that influence pathogen virulence and transmission is part of the central canon of the evolution and ecology of infectious disease. However, identifying directional and mechanistic relationships among traits remains a key challenge in various subfields of biology, as models often assume static, fixed links between characteristics. Here, we introduce learning evolutionary trait relationships (LETR), a data-driven framework that applies Granger-causality principles to determine which traits drive others and how these relationships change over time. LETR integrates causal discovery with generative mapping and transfer-operator analysis to link short-term predictability with long-term trait distributions. Using a synthetic myxomatosis virus-host data set, we show that LETR reliably recovers known directional influences, such as virulence driving transmission. Applying the framework to global pandemic (SARS-CoV-2) data, we find that past virulence improves future transmission prediction, while the reverse effect is weak. Invariant-density estimates reveal a long-term trend toward low virulence and transmission, with bimodality in virulence suggesting ecological influences or host heterogeneity. In summary, this study provides a blueprint for learning the relationship between how harmful a pathogen is and how well it spreads, which is highly idiosyncratic and context-dependent. This finding undermines simplistic models and encourages the development of new theory for the constraints underlying pathogen evolution. Further, by uniting causal inference with dynamical modeling, the LETR framework offers a general approach for uncovering mechanistic trait linkages in complex biological systems of various kinds.

💡 Research Summary

**
The paper introduces a novel data‑driven framework called Learning Evolutionary Trait Relationships (LETR) that combines Granger‑style predictive causality with discrete dynamical systems to uncover directional, mechanistic links between pathogen traits such as virulence and transmission. Traditional approaches to the virulence‑transmission trade‑off rely on static correlations and often fail to capture context‑dependent dynamics. LETR addresses this gap by first identifying Granger‑causal drivers from multivariate time‑series using information‑theoretic measures—geometric information flow (GeoC), transfer entropy, and causation entropy—capable of detecting nonlinear dependencies. When a candidate variable improves one‑step forecasts of a focal trait (e.g., virulence µ), it is retained as an upstream driver.

In the second stage, LETR fits a conditional generative update map f(µₙ, yₙ) that predicts the next generation of the focal trait given its current value and any identified drivers. The map can be a simple analytical function (e.g., logistic map) or a regularized supervised learner (e.g., LASSO, random forest, neural network). Once the map is learned, the framework approximates the Frobenius‑Perron transfer operator associated with f, enabling the computation of invariant densities π that describe the long‑run distribution of the trait across the population. This dual focus on short‑term predictability and long‑term equilibrium bridges empirical time‑series analysis with theoretical evolutionary dynamics.

The authors validate LETR on a synthetic myxomatosis dataset that encodes a known causal direction—virulence drives transmission. LETR recovers this direction with high accuracy, outperforming standard linear Granger tests. They then apply the framework to global SARS‑CoV‑2 data, using case‑fatality rates as a proxy for virulence (µ) and reproduction numbers or incidence as transmission metrics (β). Results show that past virulence significantly improves forecasts of future transmission, whereas the reverse effect is weak or absent. This asymmetry suggests that, during the COVID‑19 pandemic, evolutionary pressures favored reductions in virulence without compromising transmissibility, consistent with theoretical expectations of pathogen adaptation to human hosts.

Invariant‑density analysis reveals a long‑term trend toward low values of both virulence and transmission, indicating a possible movement toward a stable, less harmful equilibrium. Notably, the virulence density exhibits bimodality, implying the coexistence of two distinct evolutionary attractors—potentially reflecting geographic, demographic, or intervention‑driven heterogeneity. The transmission density, by contrast, shows a single peak, suggesting a more uniform selective landscape for spread.

Methodologically, LETR’s strengths lie in (1) its ability to detect directional causality directly from time‑series without imposing linearity, (2) the integration of causal discovery with generative modeling, and (3) the use of transfer‑operator theory to link short‑term dynamics with long‑term statistical steady states. Limitations include sensitivity to short or noisy time‑series (which can destabilize GeoC and entropy estimates), the need for sufficient data to reliably estimate high‑dimensional conditional densities, and the current focus on pairwise trait interactions rather than full multivariate networks.

Future work proposed by the authors includes extending LETR to Bayesian structural learning to quantify uncertainty in causal graphs, incorporating reinforcement‑learning based policy simulations to evaluate intervention strategies (e.g., vaccination, social distancing), and applying the framework to other pathogens (influenza, malaria) and ecological contexts (host‑parasite coevolution, conservation biology).

In summary, LETR provides a rigorous, flexible pipeline for uncovering mechanistic trait linkages in complex biological systems. By demonstrating that static trade‑off models are insufficient for describing virulence‑transmission dynamics, the study encourages the development of dynamic, context‑aware theories of pathogen evolution and offers a generalizable tool for evolutionary biologists, ecologists, and epidemiologists.