John Holland, Echo and agent-based models in biology

The sad passing of evolutionary computation and genetic algorithm pioneer John H. Holland earlier this week prompted me to think more about how his approach to complex adaptive systems research fits into the biological and biomedical landscape of research of today.   Although we never spoke directly, he visited the Santa Fe Institute when I was in residence there some years ago, and always struck me as an engaging and lively person.  Holland is chiefly remembered for his role in developing genetic algorithms, but I first discovered his work when I was working on the Swarm agent-based modeling tool which was inspired, at least in part, as I understand it, by an earlier framework, Echo, that Holland and other colleagues such as Terry Jones had developed in the early 1990s. Echo was developed to explore the evolution of fitness where an individual’s fitness depended on context.  Through Echo, Holland hoped to explore how the creation (and destruction) of ecological niches could drive ecological population dynamics.

Snapshot of Netlogo implementation of an Echo model

In Echo, each agent could move about the world, and collect resources gained from either the environment or other agents.   Holland’s book Hidden Order: How Adaptation Builds Complexity outlined a grand conception: that Echo could be generalized to encompass systems outside biology such as the emergence of markets.  There was no single “Echo” model, only a family of possible models that could be constructed using the same set of conceptual building blocks.  Echo’s concept of providing a basic set of primitives for modelers was therefore highly influential in the original conception of the Swarm toolkit.   Swarm was also conceived as a set of building blocks to implement a variety of models united by an underlying structure, and not, as sometimes assumed, a single model with parameters that a modeler would “tweak” for a specific question.

Holland was a founding member of the BACH (Arthur Burks, Bob Axelrod and Michael Cohen, John Holland) group at the University of Michigan in the 1980s, which was responsible for helping to train a generation of students many of whom would go on to lay the foundations for complex adaptive systems, network theory and agent-based modeling that continues to thrive to this day.  While agent-based modeling, in particular, has been staple in many scientific disciplines for several decades now, it has a relatively recent, but growing presence in the biomedical research area, including the modelling of tumour cells, the immune system and bacterial growth.  Just to name a few, Gary An and Yoran Vodovotz, have actively integrated agent-based modeling into their research program in the systems biology of inflammation (detailed in their book Translational Systems Biology) and Tony Hunt and Glen Ropella together with others have built detailed agent-based models of the liver.  I have used agent-based modeling approaches in my own research into the evolution of cellular networks and artificial chemistries.  My Amber Biology colleague, Gordon Webster, used an evolutionary computation approach directly inspired by Holland to assist in the reconstruction of phase information from X-ray crystallography data.  Many of the approaches we and others, are pursing in this area owe much to Holland and these early pioneers.

We are now firmly in the Big Data era of biology, and prevailing mainstream approaches seem to favour computational methodologies that work mostly on principles of brute force and scale.  In this era, some of these more abstract approaches pioneered by Holland and others, might seem a little quaint.  Although these large-scale, genome-wide approaches are important, there continues to be a place for examining biological systems at multiple scales and levels of abstraction.  This is because even when we have collected all the data, aligned and data-mined the genomes, transcriptomes and proteomes, visualized the network hairball, we will still need to understand the why, not just the how: why this kind of network, and not that kind of network.   Agent-based modeling, both in Holland’s more abstract pure form, and the more recent system-specific incarnations in biomedicine can allow us to run these kinds of “what-if” experiments, ultimately helping us to answer some of these deeper questions in biology.

Holland’s last monograph from 2012, Signals and Boundaries: Building Blocks for Complex Adaptive Systems shows that he was continuing to tackle these deeper questions.  It’s a  sprawling and expansive book that sets its sights on big game: exploring the underlying dynamics of the co-evolution of signals and semi-permeable boundaries.  Applied to cell biology, say, signals could be proteins or antigen fragments, and boundaries, could be compartments within the cell, or the cell itself.  Holland provides a mathematical and theoretical framework for thinking about these issues.   For example, external extracellular signals can cause changes in internal signalling networks, these networks can, in turn, change boundaries (or break them completely: think diseases that cause lysis).   Multi-agent approaches inspired by Holland’s formalisms can examine these questions.

Signals and Boundaries is also refreshingly free of a dogmatic adherence to a particular set of methodologies, and draws as much upon new work in network theory and evolution, as much as building on his earlier work.  Holland’s specific approach may not be the right one for every kind of biological question (the book considers many non-biological systems, glosses over a lot of biological details) but it seems to me that the kinds of abstractions and the questions he was asking are the right ones.   There are also important practical implications for the treatment of disease: when faced with any complex system (such as a cell) that we wish to change via signalling intervention (such as a drug), we will eventually have to grapple with the subtle dynamics that Holland describes.

Vale Professor Holland.

One comment

  1. “prevailing mainstream approaches seem to favour computational methodologies that work mostly on principles of brute force and scale”: I find such approaches unsatisfying. They may yield practically useful results, but I suspect they’ll often prove brittle, and in any event, they don’t offer the conceptual illumination that for me is the chief pleasure of scientific research.

    I never met Holland, but he was a formative influence on me, in that “Hidden order” cracked open the door through which I eventually passed from physics into biology. Genetic algorithms were my first love among topics even remotely biological, and I still find them fascinating.


Comments are closed.