Inferring temporal dynamics from cross-sectional data using Langevin dynamics

TL;DR

This paper introduces a novel method to infer temporal dynamics from cross-sectional data using Langevin dynamics, enabling the development of predictive models in systems where longitudinal data is unavailable.

Contribution

The paper presents a new approach to derive stochastic differential equations from cross-sectional data assuming local equilibrium, applicable across various domains like biology and social sciences.

Findings

Validated against longitudinal datasets with significant predictive accuracy

Demonstrated improvement when incorporating domain expert knowledge

Addresses a key challenge in modeling from cross-sectional data

Abstract

Cross-sectional studies are widely prevalent since they are more feasible to conduct compared to longitudinal studies. However, cross-sectional data lack the temporal information required to study the evolution of the underlying processes. Nevertheless, this is essential to develop predictive computational models which is the first step towards causal modelling. We propose a method for inferring computational models from cross-sectional data using Langevin dynamics. This method can be applied to any system that can be described as effectively following a free energy landscape, such as protein folding, stem cell differentiation and reprogramming, and social systems involving human interaction and social norms. A crucial assumption in our method is that the data-points are gathered from a system in (local) equilibrium. The result is a set of stochastic differential equations which capture the temporal dynamics, by assuming that groups of data-points are subject to the same free energy landscape and amount of noise. Our method is a 'baseline' method which initiates the development of computational models which can be iteratively enhanced through the inclusion of expert knowledge. We validate the proposed method against two population-based longitudinal datasets and observe significant predictive power in comparison with random choice algorithms. We also show how the predictive power of our 'baseline' model can be enhanced by incorporating domain expert knowledge. Our method addresses an important obstacle for model development in fields dominated by cross-sectional datasets.