One of the challenges facing the field of bioinformatics is the integration of large, multidimensional data sets. To understand biological processes at a systems level, data from multiple physiological components must be analyzed to uncover the dynamics governing their interactions. Currently, there are a multitude of databases covering mRNA and protein expression. To understand how changes in these domains reflect biological processes requires complex dynamic models to collate the many sets of data gathered to describe them. Models of related phenomena can then be compared to discover similarities and differences in their underlying dynamics and to understand changes over time and in response to perturbations. The development of such models should benefit from analytical tools for identifying and classifying expression motifs in bioinformatics data sets. However, such classification problems are fundamentally difficult, taxing even the most advanced techniques currently available.
The brain has evolved over millions of years into a highly efficient signal-processing and integration machine for classifying complex patterns. We propose that the same techniques employed by the brain to tackle particularly intractable problems, such as breaking a signal into discrete objects (signal segmentation), separating the signal from background noise (filtering), and the storage and recognition of complex patterns, can be utilized for the analysis of bioinformatics data sets. Advances in computing power and in our understanding of the brain have given us the ability to formulate usable computer models of neural functioning.
To form a representation of our sensory world, neural systems must interpret multiple, overlapping patterns of activity. For example, detectors of multiple features in the visual system, such as edges and corners, are activated at each point in our visual space. Likewise, each odor activates many olfactory neurons, and sounds trigger activity in a range of tone-sensitive neurons. Similar problems confront bio-informaticians. A disease process produces changes in many different genes/proteins. Multiple, overlapping patterns of gene expression coexist, reflecting the many processes in which each participates. A change in a single gene or protein rarely signifies a pathological state. A neuronal network-based bioinformatics tool could be used to segment the different patterns and to facilitate their classification as either physiological or pathophysiological.
Just as our brains create an internal representation of the sensory world, a neurally based computation tool would form a representation of multiple databases by processing the data and encoding local features as changes in the firing rates of simulated neurons. This tool then would group them into cell assemblies through synchronous oscillations, build up hierarchies of data assemblies, and combine them to form novel associations.
The question of how cell assemblies are formed addresses the core question of how the brain works. How does the brain form a sensory representation of the environment, store this representation, and make associations among its elements? Each sensory modality is initially processed separately in different areas of the cortex, and within each modality the stimuli are broken down further. For example, the visual world is processed into spatial and temporal components (the ventral and dorsal streams).1 Somehow, these distributed representations of the sensory world must be reassembled to form coherent percepts, such as whole objects. Signal processing is a particularly intractable problem. The brain must be able to segment complex sensory stimuli and perform ill-posed, inverse mappings into appropriate categories. For example, our brains must be able to determine, from a flat projection on the retina, the identity and location in three-dimensional space of an object under widely different background and lighting conditions.
Our brains have a remarkable capacity for breaking down and reconstructing the world into coherent shapes that is unmatched by conventional computational techniques. Recent experiments have shed some light on the mechanisms responsible for this. One of these discoveries is that not only is information contained in the average spiking rate (the rate code), but it is also contained in the timing of the spikes (the temporal code).2-9 The rate code (number of spikes in a given time period) indicates how strongly a neuron is activated. The temporal code (the relative timing of the spikes) is a property of a population of cells that can code for global properties of a stimulus,10 for example whether the inputs to those cells are part of the same object or different objects.1113 In the visual system, cells that respond to contiguous stimuli will have a firing rate that synchronously oscillates,1113 while those that respond to different stimuli fire independently. Here we suggest that algorithms inspired by the brain can be used as tools for computer-based analysis of bioinfor-matics data sets and to test different theories about how the brain processes information. These two goals are complementary; better models of neural signal processing will produce better tools for computer-based data processing, while actual computer-based tools can help indicate where the current models can be improved.
This neuromimetic approach has many intrinsic advantages. Neural systems are naturally parallel, which produces advantages in processing speed and robustness to noise. Averaging over a population of neurons that are synchronously coupled allows for context-dependent noise filtering, as only those features that contribute to a given object need to be processed. Temporal coding can also be used to segment signals based on certain criteria, such as contiguity, smoothness, color, etc. The brain represents information in a hierarchical, distributed fashion, analogous to the distributed activation of multiple biomarkers. Because the brain has the ability to learn, it is not necessary to determine how best to represent the data ahead of time. The resulting model can be an emergent property of the data and the neural signal-processing system.
Synchronous oscillations have been found in many brain regions governing most aspects of neural functioning. These include the retina (Figure 2.1),6,14,15 motor cortex,16 hippocampus,17 somatosensory cortex,18,19 visual cortex, 4,1013,2030 and olfactory systems.9,31-37
Temporal coding is likely to be an important component of information processing in all neural systems. This chapter focuses on the role of temporal coding in three model systems that have been particularly well described, both experimentally and computationally — vision, olfaction, and hippocampal memory systems — focusing especially on those aspects potentially most useful for bioinformatics. Our goal is to suggest how lessons learned from these models can be used to build tools for processing bioinformatics data.
Was this article helpful?