Inference of Delayed Biological Regulatory Networks from Time Series Data

Emna Ben Abdallah, Tony Ribeiro, Morgan Magnin, Olivier F. Roux, Katsumi InoueCMSB 2016 · DOI

TL;DR

This paper introduces a novel method, MoT-AN, to infer biological regulatory networks (BRNs) from time-series data. It extends existing automata network models by explicitly incorporating time delays, crucial for accurately representing biological processes, and demonstrates its effectiveness on challenging benchmark datasets.

What Problem Does This Paper Address?

Understanding the intricate web of interactions within biological systems, particularly gene regulatory networks (BRNs), is fundamental to deciphering cellular behavior and disease mechanisms. While many methods exist to infer these networks from gene expression data, a significant challenge remains in accurately capturing the temporal dynamics of these interactions. Biological processes are rarely instantaneous; there are inherent delays between a regulatory signal being triggered and its effect being observed. Existing methods often simplify these dynamics, leading to incomplete or inaccurate models. This paper addresses this gap by proposing a framework that explicitly models these time delays, aiming to create more realistic and predictive BRN models.

Methods

The core of this work is the development of Timed Automata Networks (T-ANs), an extension of traditional Automata Networks (ANs). ANs model biological systems as networks of finite-state machines where the state of one machine can influence the transitions of others. T-ANs enhance this by allowing timed local transitions, where each transition is associated with a specific delay. This delay represents the time it takes for a regulatory event to propagate.

The paper then presents an algorithm, MoT-AN, to learn these T-AN models from time-series data. The algorithm takes as input:

  • Background knowledge: Information about which genes influence which others (the network topology).
  • Time-series data: Discretized gene expression levels over time (represented as chronograms).

MoT-AN works by:

  1. Detecting changes: Identifying points in the time-series data where gene expression levels change.
  2. Generating candidate transitions: For each detected change, it proposes potential timed local transitions that could explain this change, considering the known influences and the expression levels of regulators at previous time points.
  3. Finding minimal subsets: It then identifies minimal sets of these candidate transitions that collectively explain all observed changes in the data.

The algorithm leverages Answer Set Programming (ASP) for its implementation, which allows for the exhaustive enumeration of all possible models.

Datasets

The primary dataset used for evaluation is from the DREAM4 challenge. This challenge provides benchmark datasets for systems of varying sizes (10 and 100 genes) and includes various types of data, such as:

  • Time series data under different perturbations.
  • Steady-state data.
  • Data from gene knockouts and knockdowns.

The paper specifically focuses on using the time-series data from the DREAM4 challenge for learning and prediction.

Key Insights and Ideas

  • The critical role of time delays: The paper strongly emphasizes that accurately modeling time delays is not just an enhancement but a necessity for realistic BRN inference. Biological processes are inherently dynamic and sequential.
  • T-ANs as a powerful formalism: Timed Automata Networks provide a flexible and expressive way to capture both the discrete state changes and the temporal aspects of biological regulation.
  • ASP for exhaustive model generation: Using Answer Set Programming allows the algorithm to explore the entire space of possible timed local transitions, ensuring completeness in finding explanations for the observed data.
  • Discretization strategy matters: The paper highlights that the choice of discretization for gene expression levels (e.g., using 2 to 6 qualitative values) significantly impacts the precision and performance of the learned models, with a trade-off between accuracy and overfitting.

Key Results

  • Accurate prediction: The MoT-AN algorithm achieved competitive results on the DREAM4 challenge, with Mean Square Errors (MSE) comparable to other leading methods, ranging from 0.033 to 0.086 for the 100-gene benchmarks.
  • Scalability: The algorithm demonstrated scalability, handling benchmarks with up to 100 genes. Learning and predicting for 10-gene networks took less than 30 seconds, while 100-gene networks took approximately 3 hours and 20 minutes on a single processor.
  • Impact of parameters: Experiments showed that the run time of the algorithm is exponentially influenced by the indegree of local transitions and the level of discretization chosen.

Limitations and Future Work

  • Assumption of known topology: A significant limitation acknowledged is that the algorithm assumes the network topology (which genes influence which) is known beforehand. The paper suggests that this information can often be obtained from existing biological databases.
  • Handling noisy data: While the input data is assumed to be pre-processed and discretized, the paper notes that robustness against noisy and scarce data is a future direction, particularly for challenges like the DREAM8 breast cancer network inference challenge.
  • Further analysis of T-ANs: The authors suggest exploring the relationship between the learned timed models and their untimed counterparts for deeper biological insights.
  • Asynchronous vs. Synchronous behavior: While the paper discusses both asynchronous (per-automaton) and synchronous (global) behavior, further work could delve deeper into analyzing these aspects.

Summary updated

BibTeX
@inproceedings{abdallah2016inference,
  author = {Emna Ben Abdallah and Tony Ribeiro and Morgan Magnin and Olivier F. Roux and Katsumi Inoue},
  title = {{Inference of Delayed Biological Regulatory Networks from Time Series Data}},
  booktitle = {CMSB},
  year = {2016},
  doi = {10.1007/978-3-319-45177-0\_3},
}
Figure