Inferring high-dimensional pathways of trait acquisition in evolution and disease

Sam F. Greenbury, Mauricio Barahona, Iain G. JohnstonbioRxiv 2018 · DOI

TL;DR

This paper introduces HyperTraPS, a novel statistical platform for inferring complex, high-dimensional pathways of trait acquisition in biological systems. It excels at distinguishing multiple competing pathways and identifying the most parsimonious explanations for observed data, offering a powerful tool for understanding evolution and disease progression.

What Problem Does This Paper Address?

Understanding how discrete traits (like symptoms, genetic alterations, or mutations) are acquired or lost over time is crucial in fields ranging from evolutionary biology to precision medicine. Existing methods often struggle with the complexity of systems involving many interacting traits, suffer from computational limitations that scale poorly with the number of traits, or are restricted to specific data types (e.g., only phylogenetic data). This leaves a gap in our ability to comprehensively model and predict dynamic processes in high-dimensional biological systems.

Methods

The core of the paper is the HyperTraPS (hypercubic transition path sampling) algorithm. It models trait acquisition as a process of random walkers traversing a hypercubic graph, where each node represents a unique combination of trait presence/absence. The algorithm uses a Bayesian framework to infer the transition probabilities between states, effectively learning the most likely pathways of progression.

Key methodological innovations include:

  • Hypercubic Transition Path Sampling: This approach represents all possible trait combinations as nodes on a hypercube, allowing for efficient inference of complex, interacting pathways.
  • Auxiliary Pseudo-Marginal MCMC (APM MCMC): This technique is employed to improve the efficiency and convergence of the Bayesian inference, enabling the analysis of systems with dozens of sparsely sampled traits.
  • Hierarchical Parameterization and Regularization: To manage the combinatorial explosion of parameters, the authors propose a hierarchy of models (zero, first, and second order) and a greedy backward selection process to identify the most parsimonious model structure that fits the data without overfitting. This helps in distinguishing between independent trait acquisition and more complex, interacting dynamics.
  • Visualization Tools: Novel methods are introduced to visualize the inferred high-dimensional pathways, making complex dynamics intuitively understandable.

Datasets

The authors evaluated their method on:

  • Synthetic Datasets: Three synthetic datasets (DCS1, DCS2, DCS3) were created to specifically test the platform's ability to infer single, multiple competing, and differentially weighted competing pathways.
  • Ovarian Cancer Progression Data: A dataset of copy number variations in ovarian cancer progression from 87 patients was used.
  • Drug-Resistant Tuberculosis Data: A large-scale dataset of genetic polymorphisms in drug-resistant tuberculosis isolates was analyzed, incorporating phylogenetic linkage.

Key Insights and Ideas

  • Uncovering Hidden Pathways: HyperTraPS can reveal previously undetected dynamic pathways, especially in complex systems with multiple interacting traits.
  • Parsimony is Key: The regularization framework effectively balances model complexity with data fit, leading to more interpretable and parsimonious models that adhere to Occam's razor.
  • Distinguishing Competing Dynamics: The method excels at disentangling multiple, potentially conflicting, progression routes within the same dataset, a feat that can be challenging for other approaches.
  • Predictive Power: The inferred models demonstrate strong predictive capabilities, allowing for predictions of future trait acquisition in individual patients or evolutionary lineages.
  • Bridging Fields: HyperTraPS unifies approaches from disease progression and evolutionary inference, offering a generalizable framework applicable to diverse biological questions.

Key Results

  • Synthetic Data: HyperTraPS successfully inferred single pathways (DCS1), multiple competing pathways (DCS2), and dominant/secondary pathways (DCS3), outperforming a DAG-based method (Capri) in distinguishing complex dynamics.
  • Ovarian Cancer: The analysis revealed previously unreported dynamic features and constrained pathways, highlighting the influence of early acquired features on subsequent progression.
  • Tuberculosis: The method identified multiple evolutionary pathways towards drug resistance, suggesting significant flexibility in resistance evolution and providing quantitative support for existing hypotheses.
  • Model Selection: The regularization process effectively identified optimal model structures, with second-order models often being necessary for datasets with interacting traits.
  • Validation: Model validation using training/testing splits demonstrated the predictive power of the inferred models.

Limitations and Future Work

  • Incomplete Data: Handling incomplete data in observed transitions remains a challenge, particularly for uncertain start points of observed transitions. Further development in this area is needed.
  • Model Selection Criteria: While AIC was used for regularization, exploring other Bayesian model selection methods could be beneficial.
  • Prior Information: The current framework assumes uninformative priors. Incorporating prior biological knowledge could further enhance inference.
  • Non-Markovian Dynamics: The current model assumes Markovian dynamics. Exploring non-Markovian extensions could capture more complex temporal dependencies.
  • Explicit Timescales: The current method infers ordering but not explicit timescales, which could be a valuable extension.

Summary updated

Abstract

Abstract The explosion of data throughout the sciences provides unprecedented opportunities to learn about the dynamics of evolution and disease progression. Here, we describe a highly generalisable statistical platform to infer the dynamic pathways by which many, potentially interacting, discrete traits are acquired or lost over time in biological processes. The platform uses HyperTraPS (hypercubic transition path sampling) to learn progression pathways from cross-sectional, longitudinal, or phylogenetically-linked data with unprecedented efficiency, readily distinguishing multiple competing pathways, and identifying the most parsimonious mechanisms underlying given observations. Its Bayesian structure quantifies uncertainty in pathway structure and allows interpretable predictions of behaviours, such as which symptom a patient will acquire next. We exploit the model’s topology to provide visualisation tools for intuitive assessment of multiple, variable pathways. We apply the method to ovarian cancer progression and the evolution of multidrug resistance in tuberculosis, demonstrating its power to reveal previously undetected dynamic pathways.
BibTeX
@misc{greenbury2018inferring,
  author = {Sam F. Greenbury and Mauricio Barahona and Iain G. Johnston},
  title = {{Inferring high-dimensional pathways of trait acquisition in evolution and disease}},
  year = {2018},
  doi = {10.1101/409656},
  howpublished = {bioRxiv},
}

Loading related content...

Figure