Erin C. McKiernan is a researcher in experimental and computational neuroscience, and professor of mathematics at Instituto Tecnológico y de Estudios Superiores de Monterrey in Xochitepec, Morelos, Mexico. Her research focuses on how the activity of single neurons and small networks contributes to behavior, using electrophysiology, genetic techniques, and computer modeling. She blogs about neuroscience, open science, and open access publishing. Follow on Twitter
Fig.1: Can the complexity of a human brain be captured by computers? (Source: Saad Faruque on Flickr)
The human brain is beautifully complex. And frustrating. Our understanding of it is fragmented, and hindered because, except rarely, we can’t get inside and look around while it is still functioning. Computers offer promise; if we could build an artificial brain that behaves likes a real one, maybe we could pick it apart to see how it works. This is the concept behind the Human Brain Project, recently awarded 1 billion euros by the European Commission as one of two Future and Emerging Technologies Flagships Initiatives. HBP’s co-director, Dr. Henry Markram, says they can realistically simulate a human brain within 10 years. But many neuroscientists argue we don’t know enough about the brain to model it. Where are the gaps in our knowledge and why might they present problems?
Constructing the building blocks
An artificial brain must start with good building blocks. Real neurons are diverse. They extend short and long processes, which branch out in different patterns. These shapes are not just for show; neurons process signals differently depending on their structure. One of HBP’s goals is to model neurons as 3-dimensionally detailed cells.
Neurons can be visualized under microscopes and reconstructed using computer programs. But the structure we can recreate in models is an approximation of the real one. It is difficult to mathematically represent all branches, or their turns and tapers. A greater challenge is ascribing functional properties. In many neurons, it is not clear exactly where in the structure signals from other cells are received, or how signals arriving in different branches are combined.
Neurons are also not static. Their extensions reach out or retract due to development, learning, and injury. Knowledge of how this happens is incomplete. How will modelers decide what rules to implement?
Equipping the building blocks
Fig.2: Neurons have diverse structures. (Source: Ramon y Cajal, ca. 1905)
The building blocks of an artificial brain must also respond and send signals. Real neurons are populated with many different proteins. For example, channel proteins allow charged molecules, ions, to cross the membrane and generate electrical activity. While many types of ion channel have been characterized, many remain to be studied. In most neurons, the complete population of proteins present, or how they all contribute to signaling, is not known. The number and type of proteins in a neuron also changes under a variety of conditions and the mechanisms are poorly understood. Which proteins should be put into different model neurons? And when should protein expression be turned on, off, up, or down?
Finally, where should proteins be placed? Some channels are present only in cell bodies, while others are found in the extensions. The distribution of channels affects the way neurons receive and sends signals, but often isn’t known. How will morphology be coupled with function?
Building small networks
If challenges in modeling single neurons are overcome, the next step is to connect them. Communication between neurons occurs at specialized contacts. We know a lot about the composition of these contacts and the general rules of signal transmission. But important details are missing. Who is connected to whom? Where in single neurons are contacts located? What are the strengths of the connections, and how do strengths change under different conditions? For most neurons, we have limited information, such as potential partners and estimates of connection strengths. Testing all possible pairs of neurons, even within a small region of brain, is not feasible. How will we connect a network of neurons and be sure that the partners, locations, and strengths are correct? Even if the model is built such that connections and strengths can evolve, what will be the rules of evolution? These details will have profound effects on the model’s output.
Connecting across multiple levels
Fig.3: How should a model brain be connected? (Source: Li et al. 2009, PLoS Comp. Biol. 5(5):e1000395)
Many networks must be connected to form a complete brain. Although we know in general terms which brain regions talk to others, we are ignorant as to many of the details of this communication. Which neurons in which regions are connected? What are the feedback loops by which signals travels from one region to another and back again? It is also not clear how information across multiple levels of organization (molecular, cellular) and processed over multiple time scales (seconds, minutes) is integrated. How should the model be bound together?
What will a model brain do?
If all these challenges are surmounted and a human brain simulated, what might the model do? Will it reproduce behaviors that so impress us about real brains? It’s possible. Beyond a certain level of complexity, a model can do many impressive things. But the focus should be on whether we will understand how and why behaviors emerge. We want mechanisms.
The advantage of a model is that the pieces comprising it are known, and if it is sufficiently simple, pieces can be removed to examine their role. But with the level of complexity required for the proposed model, removing one component at a time would not only be extremely cumbersome, it is questionable whether it would increase our understanding. Complex behaviors, if they arise, will likely result from the interaction of many model components. Could we test all the potential contributing interactions?
HBP researchers are right: we cannot continue to study tiny pieces of the brain in isolation and hope to understand how it works. Their goal to integrate information from experiments and computer modeling is a good one. But we must build upon a solid foundation of knowledge. Markram and associated researchers have spent
the last two decades characterizing cells and mapping connections within cortex. The foundation is growing. Yet, large gaps remain. Will they be fatal to the project? Or, will HBP, as proposed, help us fill them? Neuroscientists will have to wait and see.
1. “Will we ever…simulate a human brain?” by Ed Yong. BBC Future, 8 February 2013.
2. “Computer modelling: Brain in a box” by M. Mitchell Waldrop. Nature, 22 February 2012.
The author thanks Marco Herrera Valdez for feedback on earlier drafts, and Ed Yong, John Hewitt, Zen Faulkes, Philippe
Desjardins-Proulx, and Nathan Insel for valuable discussions.
Images: Fig. 1: From Saad Faruque on Flickr (license CC-BY-SA); Fig. 2: From Cajal on Tumblr (or via various sources; image in public domain); Image credit: Santiago Ramòn y Cajal, ca. 1905; Fig. 3: Modified from doi:10.1371/journal.pcbi.1000395 (original Fig. 3, license CC-BY); Image credit: Li et al. (2009), PLoS Comp. Biol., 5(5): e1000395.