Author Response: AI-driven Automated Discovery Tools Reveal Diverse Behavioral Competencies of Biological Networks

Mayalen Etcheverry, Clément Moulin-Frier, Pierre-Yves Oudeyer, Michael LevinOpenalex Unknown 2024 · DOI · OpenAlex

Abstract

Many applications in biomedicine and synthetic bioengineering depend on the ability to understand, map, predict, and control the complex, context-sensitive behavior of chemical and genetic networks. The emerging field of diverse intelligence has offered frameworks with which to investigate and exploit surprising problem-solving capacities of unconventional agents. However, for systems that are not conventional animals used in behavior science, there are few quantitative tools that facilitate exploration of their competencies, especially when their complexity makes it infeasible to use unguided exploration. Here, we formalize and investigate a view of gene regulatory networks (GRNs) as agents navigating a problem space. We develop automated tools to efficiently map the repertoire of robust goal states that GRNs can reach despite perturbations. These tools rely on two main contributions: (1) adapting curiosity-driven exploration algorithms from the AI community to automatically discover the range of reachable goal states of GRNs and (2) proposing a battery of empirical tests inspired by implementation-agnostic behaviorist approaches to assess their navigation competencies. Our data reveal that models inferred from real biological data can reach a surprisingly wide spectrum of steady states, while showcasing various competencies that living agents often exhibit in physiological network dynamics without requiring structural changes. We investigate the applicability of these “behavioral catalogs” for comparing evolved competencies and designing drug interventions or synthetic networks.

Introduction

Understanding and controlling the behavior of chemical and genetic networks is a central goal for biomedicine and synthetic biology. Traditionally, these networks have been viewed as "simple machines" that require physical rewiring—such as gene therapy or editing—to alter their behavior. However, gene regulatory networks (GRNs) are incredibly complex, non-linear, and context-sensitive. This complexity presents two major challenges: the vast space of possible behaviors makes random screening inefficient, and there is a lack of quantitative frameworks to measure the "competencies" of these networks—their ability to reach and maintain specific goals despite environmental noise.

This research shifts the perspective, conceptualizing GRNs as agents navigating a problem space rather than static circuits. By treating these networks as goal-seeking entities, we can apply tools from artificial intelligence and behavior science to map their latent capabilities.

Methodology

The study introduces a two-part automated framework for discovery and assessment:

Curiosity-Driven Exploration (IMGEP)

The authors adapted Intrinsically-Motivated Goal Exploration Processes (IMGEP), algorithms originally developed for robotic skill acquisition. Unlike random sampling, this AI-driven approach selects a "target goal" within the behavior space and learns which interventions (initial conditions) are required to reach it. The algorithm prioritizes novelty, focusing computational resources on discovering previously unseen behaviors.

Behaviorist Robustness Battery

To evaluate the "intelligence" or robustness of discovered behaviors, the networks were subjected to tests inspired by animal behavior science:

  • Noise: Constant fluctuations in gene expression levels.
  • Pushes: Sudden, temporary shifts in the state of specific nodes.
  • Walls: Force fields that prevent the network from entering certain regions of the state space.

The entire pipeline was implemented in JAX, enabling high-performance, parallelized simulation of ordinary differential equation (ODE) models.

Datasets and Benchmarking

The study utilized 30 curated ODE models from the BioModels database, representing various pathways such as the cell cycle, signaling, and metabolism. These were expanded into 432 unique systems by varying the observed nodes to define different behavior spaces. For comparison, a dataset of randomly generated networks was created to establish a baseline for non-evolved dynamics.

Key Insights

  • Latent Versatility: Biological networks can reach a much wider spectrum of steady states than random screening suggests. Many of these "latent" states exist mathematically within the models but are rarely observed without targeted interventions.
  • Basal Cognition: The findings provide evidence for "intelligence" at the sub-organismal level. GRNs can navigate complex transcriptional landscapes to reach goal states even when pushed off-course, mimicking animal homing behavior.
  • Non-Genetic Interventions: The creation of "behavioral catalogs" enables the identification of stimuli-based interventions. In a cancer-related pathway (RKIP-ERK), the tool identified a sequence of chemical "nudges" that could reset a disease state to a healthy state without genetic editing.
  • Evolutionary Signatures: Biological networks proved significantly more versatile and robust than random networks of equivalent size, suggesting that navigational competencies are products of evolutionary pressure.

Results

The curiosity-driven IMGEP significantly outperformed random search in efficiency. In an oscillator design task, IMGEP discovered 1,167 valid configurations, whereas random search found only 42 within the same computational budget.

The framework successfully identified "robust goals"—steady states that the system consistently reaches despite high noise or physical "walls" in the state space. Furthermore, using the RKIP-ERK model, the authors demonstrated a stepwise drug intervention that robustly resets cells from a tumor-associated state to a healthy one, even under stochastic perturbations.

Discussion and Limitations

While these automated tools offer a new path for exploring biological network behavior, several limitations remain. The study analyzes ODE models in isolation, whereas real organisms feature networks embedded in tissues with complex feedback. Currently, behavior spaces are hand-defined by researchers; future iterations could use unsupervised learning to allow the AI to discover interesting behavior dimensions autonomously. Additionally, scaling to genome-wide models will require more efficient exploration strategies. Finally, these in silico discoveries require in vitro validation to confirm that predicted latent states are biologically achievable in "wet lab" settings.

Summary updated

Abstract

Many applications in biomedicine and synthetic bioengineering depend on the ability to understand, map, predict, and control the complex, context-sensitive behavior of chemical and genetic networks. The emerging field of diverse intelligence has offered frameworks with which to investigate and exploit surprising problem-solving capacities of unconventional agents. However, for systems that are not conventional animals used in behavior science, there are few quantitative tools that facilitate exploration of their competencies, especially when their complexity makes it infeasible to use unguided exploration. Here, we formalize and investigate a view of gene regulatory networks as agents navigating a problem space. We develop automated tools to efficiently map the repertoire of robust goal states that GRNs can reach despite perturbations. These tools rely on two main contributions that we make in this paper: (1) Using curiosity-driven exploration algorithms, originating from the AI community to explore the range of behavioral abilities of a given system, that we adapt and leverage to automatically discover the range of reachable goal states of GRNs and (2) Proposing a battery of empirical tests inspired by implementation-agnostic behaviorist approaches to assess their navigation competencies. Our data reveal that models inferred from real biological data can reach a surprisingly wide spectrum of steady states, while showcasing various competencies that living agents often exhibit, in physiological network dynamics and that do not require structural changes of network properties or connectivity. Furthermore, we investigate the applicability of the discovered “behavioral catalogs” for comparing the evolved competencies across classes of evolved biological networks, as well as for the design of drug interventions in biomedical contexts or for the design of synthetic gene networks in bioengineering. Altogether, these automated tools and the resulting emphasis on behavior-shaping and exploitation of innate competencies open the path to better interrogation platforms for exploring the complex behavior of biological networks in an efficient and cost-effective manner. To read the interactive version of this paper, please visit .
BibTeX
@inproceedings{etcheverry2024author,
  author = {Mayalen Etcheverry and Clément Moulin-Frier and Pierre-Yves Oudeyer and Michael Levin},
  title = {{Author Response: AI-driven Automated Discovery Tools Reveal Diverse Behavioral Competencies of Biological Networks}},
  booktitle = {Openalex Unknown},
  year = {2024},
  doi = {10.7554/elife.92683.1.sa3},
}

Loading related content...

Figure