SCIENCE@home

Of Toasters and Molecular Ticker Tapes

Konrad Paul Kording

Published 2011 in PLoS Comput. Biol.

ABSTRACT

Experiments in systems neuroscience can be seen as consisting of three steps: (1) selecting the signals we are interested in, (2) probing the system with carefully chosen stimuli, and (3) getting data out of the brain. Here I discuss how emerging techniques in molecular biology are starting to improve these three steps. To estimate its future impact on experimental neuroscience, I will stress the analogy of ongoing progress with that of microprocessor production techniques. These techniques have allowed computers to simplify countless problems; because they are easier to use than mechanical timers, they are even built into toasters. Molecular biology may advance even faster than computer speeds and has made immense progress in understanding and designing molecules. These advancements may in turn produce impressive improvements to each of the three steps, ultimately shifting the bottleneck from obtaining data to interpreting it. This is an ‘‘Editors’ Outlook’’ article for PLoS Computational Biology Moore’s law has characterized progress in microprocessor techniques (see Figure 1, dashed). In very good approximation, computers have doubled in speed every 2 years. At the same time, computing has progressively gotten cheaper, and we can now successfully solve many computing problems that once seemed hard, such as speech recognition. We are also solving entirely new computational problems, such as searching billions of web pages for information on neuroscience or molecular biology. Such exponential growth makes solutions to seemingly insurmountable problems seem trivial given a bit of time. Meanwhile, simple computers have gotten cheaper over time. This decrease in costs was so dramatic that many of today’s toasters contain microprocessors for time-keeping, switching on or off heat, light feedback of heating state, and the handling of key presses. This makes building toasters simpler and ultimately cheaper. When computers were invented, toasters were certainly not an expected application of computing. Importantly, the speed of computers has increased dramatically faster than the number of neurons that can be simultaneously recorded [1] (Figure 1, dotted). Still, the increase of the number of simultaneously recorded neurons has allowed the development of advanced algorithms that take advantage of this growing number. Indeed, the field that analyzes multivariate neural data is large now and can analyze complicated interactions between large numbers of neurons [2–4], electroencephalography (EEG) or magnetoencephalography (MEG) recordings, optical recordings, or functional magnetic resonance imaging (fMRI) voxels [5,6]. However, as the amount of data increases, so does the complexity of the questions. For example, many current studies of neural data analysis ask how neurons interact, but as the number of neurons grows, the number of potential connections grows quadratically as each neuron may interact with each other neuron. This, in turn, leads to models with many free parameters, which requires new statistical methods of fitting these parameters. Molecular biology has seen accelerating progress over the last decades. One readily quantifiable cost in molecular biology is that of sequencing DNA. The development of a host of different methods has allowed the cost of sequencing each base pair to dramatically decrease over time (Figure 1, solid). The rate of improvement is much faster than that of neuron recording techniques or even Moore’s law. This development allowed sequencing the entire human genome at a price of billions of dollars in the year 2003, and sequencing the same genome at higher quality now costs less than $2,000. The current push is to sequence an entire genome for less than $1,000 [7]. This development allows solving many problems of obvious importance, such as the search for gene-related markers of disease [8]. From a computational perspective, a central objective of neuroscience is to understand how neurons convert their inputs into outputs and collectively produce action based on stimuli and internal processes, such as memory and attention. This leads to what I would call the three central steps of experimental approaches in systems neuroscience. (1) Select the signals that are important for a given neuroscience question. As long as we cannot approach understanding the entire brain at the same time, it is highly useful to select what to stimulate and what to measure. (2) Get stimuli into the brain. To understand what neurons do, inputs need to be defined or known. (3) Get data out of the brain. Only large amounts of data allow meaningful statistical inferences. Virtually all experimental approaches to systems neuroscience can be phrased in these terms. It is interesting to ponder a few well-known examples. In a typical single-cell visual cortex experiment [9] that studies how the visual cortex encodes visual stimuli, we would put an electrode into primary visual cortex (1), show various visual stimuli on a monitor in front of the animal (2), and record neural activity from the electrode (3). In a typical fMRI experiment about visual cortex [10], we would choose a contrast that tells us about changes in blood flow in the brain (which is a proxy for average firing rate) (1), stimulate subjects by using a set of fixed visual stimuli (2), and read Citation: Kording KP (2011) Of Toasters and Molecular Ticker Tapes. PLoS Comput Biol 7(12): e1002291. doi:10.1371/journal.pcbi.1002291 Editor: Karl J. Friston, University College London, United Kingdom Published December 29, 2011 Copyright: 2011 Konrad P. Kording. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Funding: Research was funded by the NIH grants 1R01NS063399 and 2P01NS044393. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. * E-mail: kk@northwestern.edu Competing Interests: The author has declared that no competing interests exist. PLoS Computational Biology | www.ploscompbiol.org 1 December 2011 | Volume 7 | Issue 12 | e1002291 out data using resonance signals (3). In a typical slice work experiment [11], we might identify a specific cell type under the microscope (1), activate other cells using glutamate uncaging (2), and then record signals from the intracellular electrode to find out how the other cells affect the recorded cell (3). In all these cases, selection, stimulation, and reading the data (‘‘data out’’) are crucial aspects of the work. Each of these steps has its own criteria for being maximally useful. For the selection step (1), we would like to select all the relevant signals and nothing else. For the stimulation step (2), we Figure 1. Comparison of the scaling laws between neuroscience (exponential fit from [1]), computer science (exponential fit of years 2000–2007 from [29]), and DNA sequencing, see [30]. Remarkably, the steepness of the curves is worth comparing while the offset on the y-axis is arbitrary. doi:10.1371/journal.pcbi.1002291.g001 Figure 2. The three central steps that define experiments in systems neuroscience. The question mark denotes areas where the author expects exciting developments. (A) Methods for selecting where neural signals come from. (B) Methods for stimulating neurons. (C) Methods for reading out the data. See text for detail. doi:10.1371/journal.pcbi.1002291.g002 PLoS Computational Biology | www.ploscompbiol.org 2 December 2011 | Volume 7 | Issue 12 | e1002291 want to be able to get in large amounts of well defined data with high information bandwidth and low noise [12]. Lastly, for the ‘‘data out’’ step (3), we want to get a lot of data out with low noise and high bandwidth. Lastly, given limited budgets, we want the techniques to be cheap. Certainly, as we scale up the analysis towards understanding larger systems, the price per unit of information must be limited. Each of the currently used approaches has limitations along the three steps. For example, fMRI has a huge number of channels, and can select essentially any brain area. Yet, it seems unlikely that it could be used to record only from certain cell types, such as interneurons [13], noise levels are high, and spatiotemporal resolution is low. Single-cell slice physiology is good at selecting neurons based on observable features like brain region and cell morphology, and it is low noise, but data out bandwidth is quite low and there may be biases in the selection of the recorded cells. Similarly, other current techniques are all rather limited on at least one of the three axes. Molecular biology is starting to offer powerful solutions to overcome limitations of the three steps [14], and I want to start by reviewing past progress. (1) The selection step. There are many different levels of selection. We might want to select individuals that have certain diseases for which there are genetic markers. In this case, we can select these individuals through genetic tests [15]. We might want to select neurons but not other cells (Figure 2A, upper), and this is possible through a set of well-established genetic methods that enable or disable gene expression using tissue-specific promoters [16,17]. We might want to select only a subset of all the cells (Figure 2A, middle) [18]. Lastly, we might only want to only select certain neurons that have defined physiological properties (Figure 2A, lower). Interestingly, it is even possible to select each cell individually and assign it a random color that is visible under the microscope [19]. Importantly, genetically selecting neurons enables certain ways of stimulating just those neurons and reading out only from those neurons (see below). Figure 3. Molecular ticker tapes. Neural activity affects an intracellular concentration. DNA polymerase copies a

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-28 of 28 references · Page 1 of 1

CITED BY

Showing 1-26 of 26 citing papers · Page 1 of 1