Energy limitation as a selective pressure on the evolution of sensory systems

The evolution of the accessory structures (e.g. lenses in the eye),peripheral receptors and regions of the central nervous system that together form sensory systems is often viewed solely in terms of the benefits they provide to an animal, i.e. information about the animal's internal and external environments. Sensory systems differ widely in their complexity and size, from clusters of G-protein coupled receptors for chemosensation in bacteria (Maddock and Shapiro,1993) to the olfactory and gustatory systems of insects and vertebrates (for reviews, see Laurent,2002; Shepherd et al.,2004), and from single receptors innervating insect mechanosensory hairs (e.g. French and Sanders,1981), whose activity is processed locally (e.g. Burrows and Newland, 1993; Burrows and Newland, 1994), to vertebrate somatosensory systems (e.g. Penfield and Boldrey, 1937; Nelson et al., 1980). Extracting germane information from internal and external environments requires sensory receptors (with their accessory structures) to sample these environments and central circuits to analyse and interpret the incoming information. In the case of the very simplest organisms, the entire machinery for sensing the environment and acting upon it is found within the same cell,whereas at their most elaborate, for example the mammalian visual system,peripheral sensory structures may consist of millions of neurons, with even greater numbers of neurons involved in processing the information they obtain within the central nervous system. Irrespective of its size and complexity,however, the more reliable the information a sensory system can extract from the environment, the more accurate the decision making and motor control it facilitates.

Several studies have now shown that there are high energetic costs incurred by neural tissue, including that of sensory systems, both whilst processing information and at rest (Ames et al.,1992; Attwell and Laughlin,2001; Lennie,2003; Niven et al.,2003a; Nawroth et al.,2007; Niven et al.,2007). There are also likely to be considerable energetic costs associated with the development and carriage of nervous systems. Thus, nervous systems are subject to two conflicting selective pressures: the need to minimise energy consumption and also to generate adaptive behaviour under fluctuating environmental conditions. More specifically, in sensory systems there will be a trade-off between the energetic cost of a sensory structure encoding a particular sensory modality and the amount of reliable, germane information obtained.

Selective pressures to reduce energy consumption and improve behavioural performance can affect all levels of organization within an individual, from sub-cellular structures and single cells to brain regions and entire brains. Equally, these selective pressures can affect any life history stage. For example, a large visual system with high acuity may allow more accurate assessment of potential mates facilitating better mate choice, but at an energetic cost that may reduce individual fecundity. This emphasizes that information obtained by sensory systems must affect behaviour in a way that is beneficial to the individual or it will be selected against. Furthermore,these behaviours must ultimately be realized as increased fitness if they are to be selected (e.g. Krebs and Davies,1993).

Thus, understanding how energy acts as a selective pressure on the evolution of sensory systems requires assessment not only of the relationship between energy and performance at the cellular and sub-cellular levels but also at the levels of sense organs, brain regions, entire brains and the entire organism. Although information about a particular sensory modality may be obtained by specific peripheral sense organs and processed, at least initially, by discrete brain regions, it is essential to consider the benefits and the costs, not only in the context of the specific neural circuits involved but also in the context of the whole organism. To determine the impact of energy as a selective pressure on the evolution of the nervous system it is important to know both how and when energy is expended within specific sensory systems or whole nervous systems, and what proportion of energy is consumed by these processes, relative to the energy budget of the whole organism.

Sensory processing consumes a proportion of the total energy consumption of the nervous system and, therefore, is limited both by an animal's total energy budget and the distribution of energy costs throughout the nervous system. An animal's total energy consumption can be measured using its metabolic rate, of which there are several available measures (for reviews, see Hammond and Diamond, 1997; White and Seymour, 2003; Savage et al., 2004; Weibel et al., 2004; Nagy, 2005; Suarez and Darveau, 2005; Weibel and Hoppeler, 2005). An important consideration is which of these measures is most relevant to understanding the limitations on the energy available for sensory processing. Basal metabolic rates (BMRs) have been measured in vertebrates, as have resting metabolic rates (RMRs) in invertebrates (see Lovegrove, 2000; White and Seymour, 2003; Bokma, 2004; Savage et al., 2004; Niven and Scharlemann, 2005; Chown et al., 2007). They are composed of the energy requirements of various housekeeping functions such as protein synthesis, membrane turnover and maintenance of membrane potentials in a range of tissues and organs as well as oxygen transport and, in endotherms,the maintenance of body temperature. The peripheral and central nervous systems represent a significant component of the BMR or RMR in many animals. For example, in humans (Homo sapiens), although the brain is just 2%of the body mass it consumes approximately 20% of the BMR(Clarke and Sokoloff, 1999). Likewise, in blowflies (Calliphora vicina), the retina alone is estimated to consume approximately 8% of the resting metabolic rate(Howard et al., 1987). The high proportional energy consumption of neural tissue suggests that it may have a significant effect upon the overall BMR or RMR of an animal, but this is not supported by the empirical data; although in mammals the scaling exponent of absolute brain size and BMR with body mass is similar(Martin, 1981; Mink et al., 1981), plotting the deviation in brain size versus the deviation in BMR from their predicted scaling relationships with body mass reveals no correlation(McNab and Eisenberg, 1989)(but see Isler and van Schaik,2006a) (Fig. 1). This would be expected if the energy consumption of neural tissue can be traded off against the energy consumption of other `expensive tissues' such as kidney or gut. Indeed, trade-offs between brain and gut have been suggested to play an important role in the evolution of human and primate brain size(Aiello and Wheeler, 1995).

Animals do not necessarily spend large amounts of time at their BMR,however, and the field metabolic rate (FMR) (for a review, see Nagy, 2005), which is a measure of an animal's energy consumption when it is freely moving through its natural environment, is likely to be a more relevant measure to understanding the effects of energy on the evolution of the brain and sensory systems. As far as we are aware, no direct comparison of FMR and brain size is available,although the blood flow to, and oxygen uptake of, the brain has been measured in some mammals, including humans, during exercise. In humans, periods of exercise, when the skeletal muscles consume large amounts of oxygen, cause a reduction in the blood supply to most of the organs that contribute to the BMR such as the kidneys or liver; however, the blood supply to the brain remains relatively constant (Ide and Secher,2000). For example, when cycling there is no change in global blood flow to the brain (Madsen et al.,1993). However, voluntary movements, such as hand movements, do evoke increased local blood flow to cortical areas that receive sensory inputs as well as motor areas involved in execution(Raichle et al., 1976; Orgogozo and Larsen, 1979). This suggests that global blood flow to the brain is relatively constant,despite changes in regional blood flow. The relative constancy of global brain blood flow during behaviour is reinforced by data from Weddell seals(Leptonychotes weddellii), which whilst undertaking simulated dives,substantially reduce the blood flow to the heart, liver and pancreas and partially reduce the blood flow to the lungs and the retina, but not to the cortex or cerebellum (Fig. 2)(Zapol et al., 1979). This makes sense because during hunting sensory systems must continue to monitor the environment and the brain must continue to make decisions and execute motor patterns. The need for a constant supply of oxygen to the brain to maintain function is emphasized by the consequences of an interruption to the blood supply to the mammalian brain: a decrease in blood flow below critical levels can cause a 90% drop in ATP within 5 min(Erecinska and Silver, 2001). Thus, even short interruptions to the supply of oxygenated blood may lead to severe long-term functional consequences including swelling, lysis and ischemic cell death (for a review, see Lipton, 1999). The effects of a reduction in the amount of oxygen received by the brain can be observed in humans under hypoxic conditions, such as at high altitudes, where blood oxygen levels drop and both sensory and motor functions are severely impaired(Hornbein, 2001). To prevent such oxygen depletion the nervous systems of both vertebrates and insects are supplied with oxygen through dense networks of capillaries or tracheoles(Fig. 3), respectively. The neurons of some vertebrates such as the western painted turtle (Chrysemys picta), the common frog (Rana temporaria) and the crucian carp(Cyprinus carpio) have evolved to tolerate hypoxic or anoxic conditions (for reviews, see Lutz,1992; Boutilier,2001; Nilsson,2001; Bickler and Donohoe,2002). However, in both C. picta and R. temporaria, neural function is compromised under these conditions. However, neurons within the crucian carp brain remain active in anoxic conditions and continue to sense their environment and generate behaviour(Nilsson, 2001). Many invertebrates have remarkable tolerance under hypoxic or anoxic conditions,possibly due to their low metabolic rates in comparison to vertebrates (e.g. Niven and Scharlemann, 2005; Chown et al., 2007). However,there is little information about the effects of hypoxic/anoxic conditions upon invertebrate neurons, sensory systems or behaviour. Nevertheless, data from vertebrates clearly suggest that the sustained supply of oxygen to sensory systems and the brain has important implications for their evolution because nervous systems are a constant energy sink, consuming energy irrespective of whether they are at rest or active.

There are numerous physiological processes within neural circuits that contribute to the overall costs of neural tissue: transduction, electrical signal transmission within the neuron and synaptic transmission(Fig. 4). For neurons that possess chemical synapses, these costs include the maintenance of concentration gradients for ions including Na⁺, K⁺,Ca²⁺, H⁺, Cl^– and HCO₃^– and potential differences across the cell membrane, the synthesis of neurotransmitter molecules, the loading of neurotransmitter molecules into vesicles, the docking of neurotransmitter vesicles at the pre-synaptic membrane and the breakdown or reuptake of neurotransmitter from the synaptic cleft. In addition, there are costs associated with the synthesis of macromolecules such as proteins and fatty acids.

The basic currency of energy within cells is the phosphate–phosphate bond of the adenosine triphosphate (ATP) molecule (for a review, see Alexander, 1999). In neurons,the energy released in breaking this bond is harnessed by the 3Na⁺/2K⁺ ATPase, which exudes 3Na⁺ ions and admits 2K⁺ ions for each molecule of ATP that is converted to ADP,for the maintenance of the membrane potential(Skou, 1957; Post et al., 1972; Glynn, 1993). Numerous symporters and antiporters are coupled to the 3Na⁺/2K⁺ATPase, linking the movements of ions such as Ca²⁺,Cl^– and HCO₃^– to the energy consumed by the 3Na⁺/2K⁺ ATPase(Fig. 4). The potential and concentration gradients that drive the movements of ions through ion channels are also dependent upon the 3Na⁺/2K⁺ ATPase. Additionally, the V-ATPase, which transports one H⁺ ion across the membrane for each molecule of ATP converted to ADP, is critically important to neurons (Harvey, 1992; Moriyana et al., 1992). The V-ATPase pumps H⁺ ions into the lumen of synaptic vesicles,creating a concentration gradient and potential difference between the lumen and the surrounding cytoplasm that is used to drive the uptake of neurotransmitter molecules (Fig. 4).

Electrical models of single fly photoreceptors based upon intracellular measurements of their input resistance and membrane potential have been used to estimate the total energy consumption of the 3Na⁺/2K⁺ATPase within single neurons (Laughlin et al., 1998; Niven et al.,2003a; Niven et al.,2007). Comparison of the estimated ATP consumption from Calliphora vicina photoreceptors with estimates of the total costs derived from whole retina oxygen consumption measurements are remarkably similar (Laughlin et al.,1998; Pangrsic et al.,2005; Niven et al.,2007). This suggests that, at least within photoreceptors,movements of ions across the neuronal cell membrane, which are driven by concentration gradients and potential differences maintained by the activity of the 3Na⁺/2K⁺ ATPase, are the major energy cost.

Both electrical models of individual photoreceptors and oxygen measurements from whole retinas suggest that the energy consumption of neural tissue at rest is extremely high, e.g. rabbit (Ames et al., 1992), fly (Laughlin et al., 1998; Niven et al.,2003a; Niven et al.,2007; Pangrsic et al.,2005). This high cost, even in the absence of any signalling, is due to the activity of the 3Na⁺/2K⁺ ATPase, which is necessary for the maintenance of the resting potential(Niven et al., 2007). Comparison of signalling and resting costs in photoreceptors suggests that they are related, the resting costs being approximately 25% of the cost at the highest light levels (Fig. 5). Resting costs and signalling costs are probably related in all neurons (both spiking and non-spiking), although the precise relationship is likely to be different depending on the specific neural type.

Recently, three energy budgets have been produced that attempt to divide the total energy consumption of neural tissue into its constituent components(Attwell and Laughlin, 2001; Lennie, 2003; Nawroth et al., 2007). For example, Attwell and Laughlin (Attwell and Laughlin, 2001) divide the energy consumption of a single action potential in rat grey matter into the voltage-gated currents producing the action potential, pre-synaptic Ca²⁺ entry at the synapse, the recycling of glutamate released into the synaptic cleft and loading of vesicles, vesicle endo- and exocytosis and the activation of post-synaptic receptors (NMDA, non-NMDA and mGluR) (Fig. 6A). The costs calculated for a single component are then multiplied by the total number of these components within the nervous system to give the total energy consumption. These energy budgets emphasize the high costs of maintaining the membrane potential at rest as well as the extremely high costs of action potential conduction. Indeed, within an active olfactory glomerulus the energetic costs were dominated by the demands of action potential transmission within the afferent olfactory neurons and their synaptic outputs, despite postsynaptic dendrites comprising at least half the total glomerular volume (Nawroth et al.,2007) (Fig. 6B). Other functions of neurons such as the loading of vesicles with glutamate and its recycling following release into the synaptic cleft constitute a relatively small proportion of these energy budgets(Attwell and Laughlin, 2001; Nawroth et al., 2007).

These studies emphasize that the movements of ions across the neuronal cell membrane at rest and whilst signalling are a major cost in both spiking and non-spiking neurons. The costs themselves are likely to differ between cell types depending on the input resistance at rest, the precise combinations and densities of ion channels in the membrane, the total membrane area, the number of output synapses and type of neurotransmitter they release.

Whilst oxygen consumption and blood flow measurements, electrical circuit models and energy budgets can explain the mechanistic causes of the high energetic costs associated with single neurons, sensory processing regions or grey matter; they do not in themselves explain why such costs exist. To understand why specific components within the nervous system cost particular amounts of energy we need to understand the function of these components. Two particularly important factors that affect the energetic cost of neural information processing are noise and response speed, which determine the signal-to-noise ratio (SNR) and bandwidth, respectively(Laughlin, 2001; Niven et al., 2007). Noise in sensory systems is both intrinsic and extrinsic (for review, see Faisal et al., 2008). Extrinsic noise is derived from the sensory stimuli themselves, which because they are either quantum-mechanical or molecular in nature do not perfectly convey information about the environment(Hecht et al., 1942; Barlow et al., 1971; Berg and Purcell, 1977; Baylor et al., 1979; Lillywhite and Laughlin, 1979; Aho et al., 1988). Intrinsic noise occurs at all stages of sensory processing, including the transduction of the sensory stimulus into an electrical signal, the transmission of electrical signals within neurons and synaptic transmission of signals between neurons (Barlow, 1956; Katz and Miledi, 1970; Lillywhite and Laughlin, 1979; Aho et al., 1988; Mainen and Sejnowski, 1995; Berry et al., 1997; de Ruyter van Steveninck et al.,1997). One potential way to improve the SNR of single neurons is to increase their numbers of receptor molecules and ion channels(Weckström and Laughlin,1995; Laughlin,1996; Niven et al.,2003b; Vähäsöyrinki et al.,2006; Niven et al.,2007). However, each additional receptor that is activated or ion channel that is opened consumes energy. Likewise, an improved SNR could be conveyed to postsynaptic neurons by releasing greater numbers of synaptic vesicles but each additional vesicle will consume more energy. At the circuit level, noise reduction can be produced by averaging the outputs of sensory receptors or neurons in space, time, or both. In the peripheral nervous system, averaging the signals from neighbouring receptors may eliminate noise to some extent if the noise in these receptors is uncorrelated. However,spatial averaging reduces the resolution with which the sensory receptors sample the environment, requiring an increase in receptor density to restore the resolution. Thus, averaging increases the energetic costs of sensory processing because greater numbers of neurons are required.

Increasing the number of receptors and ion channels in a neuron not only affects its SNR but also its bandwidth. In particular, increasing the density of ion channels will decrease the input resistance and the membrane time constant. However, greater densities of receptors and ion channels allow a greater flow of ions across the neuronal cell membrane during signalling that require the 3Na⁺/2K⁺ ATPase to restore them after signalling. The bandwidth and SNR both contribute to the information rate. Recent comparative studies have explained the links between fly photoreceptor energy consumption and information rates through their bandwidth and SNR(Niven et al., 2007). In both small and large photoreceptors the cost of a unit of information (bit) drops as more is encoded (Fig. 7). This is due to the resting costs, which are divided by the amount of information encoded. Nevertheless, smaller photoreceptors are more efficient than their larger counterparts despite having lower information rates because their energetic costs are substantially lower(Niven et al., 2007).

Given the finite energy budgets of animals, the high energetic demands of sensory systems and the need for animals to generate adaptive behaviour,energy efficient solutions to the challenges faced by sensory systems should be strongly selected for during evolution. Energy efficiency, which leads to an increase in the ratio between the information encoded and the energy expended by sensory systems, can occur due to adaptations in their morphology,physiology, or both. Numerous examples of energy efficient strategies exist at all levels of neural organization, from the combinations of ion channels within single neurons (Niven et al.,2003a) and their size (Niven et al., 2007) to the coding of information within populations of neurons (Levy and Baxter,1996; Baddeley et al.,1997; Vinje and Gallant,2000; Perez-Orive et al.,2002; Hromádka et al.,2008) and computational maps(Chklovskii and Koulakov,2004).

The efficiency with which single neurons code information is critically dependent upon the biophysical properties of their membranes, such as the total surface area or the combinations of ion channels that they express,because the movement of ions across the membrane is the major energetic cost of neurons (Laughlin et al.,1998; Attwell and Laughlin,2001; Niven et al.,2007). Numerous features of neurons such as their conduction velocity, time constants and space constants are dependent upon the combinations and densities of ion channels within the membrane(Hille, 2001).

The most obvious impact of ion channels upon the energy efficiency of single neurons is the generation of action potentials by voltage-gated sodium channels. A high density of voltage-gated sodium channels can support the production of action potentials, which are used to code information digitally. However, some neurons lack voltage-gated sodium channels (or a sufficient density of voltage-gated sodium channels) and code information as graded changes in membrane potential. These non-spiking neurons are found in the peripheral visual, auditory and vestibular systems of vertebrates (e.g. photoreceptors, bipolar cells and hair cells) and throughout the visual and mechanosensory systems of insects and crustaceans (e.g. photoreceptors, motion detector neurons, local non-spiking interneurons)(Tomita, 1965; Ripley et al., 1968; Autrum et al., 1970; Werblin and Dowling, 1969; Hagins et al., 1970; Harris et al., 1970; Kaneko, 1970; Burrows and Siegler, 1976; Yau et al., 1977). These neurons all transmit information to post-synaptic neurons as graded or analogue signals.

Simulations of large monopolar cells in the fly retina suggest that transmission of analogue information (graded potentials) is at least as costly per unit of information (measured in bits) as digital transmission (action potentials) but far more information can be transmitted per second using analogue coding (Laughlin et al.,1998). This suggests that many more spiking neurons are required to transmit a given amount of information per second. In this case, even if the cost per bit of information were the same with both coding schemes,digital coding would incur greater overall energy consumption because when no information is coded the resting potentials of a larger number of spiking neurons must be maintained. Thus, analogue coding is a more efficient solution to transmitting a given amount of information within a limited amount of time. However, graded potentials degrade over long distances and so there is a reduction in the reliability of the analogue signals being transmitted. Digital coding using action potentials also has the benefit of being able to threshold out synaptic noise, which may accumulate in networks due to synaptic transmission (Sarpeshkar,1998; Laughlin et al.,2000). This reduction in reliability means that the relative energy efficiency of analogue versus digital coding drops as the distance information is to be transmitted increases. The greater efficiency of analogue coding over short distances and digital coding over longer distances suggests that the nervous system should use a mixture of the two schemes. Such hybrid coding schemes are indeed observed in neural circuits, which combine graded potentials (including postsynaptic potentials) and action potentials. Indeed, a theoretical prediction of the optimal mixture of analogue and digital coding in electronic devices closely resembles that observed in cortical neurons (Sarpeshkar,1998).

As well as influencing whether neurons code information as graded or action potentials, ion channels can also alter energy efficiency through their density and their activation/inactivation properties. For example, in Drosophila melanogaster photoreceptors, a considerable proportion of Shaker voltage-gated K⁺ channels are activated at the steady-state resting potential, reducing the current-to-voltage gain(Niven et al., 2003b). Photoreceptors exposed to bright lights depolarize, inactivating the Shaker voltage-gated K⁺ channels and increasing the current-to-voltage gain and improving the spread of the signal across the voltage range. The Shaker conductance incurs an energetic cost at rest, but not at more depolarized potentials when it is inactivated(Niven et al., 2003a; Niven et al., 2003b). In Drosophila mutant photoreceptors that lack the Shakerconductance, it is replaced with a leak conductance that restores the current-to-voltage gain at rest but it does not inactivate at more depolarized potentials (Niven et al.,2003b). This leak conductance incurs an energetic cost both at rest and at more depolarized potentials. Thus, photoreceptors experience a twofold cost associated with the loss of the Shaker channel, a drop in their information rate and an increase in their energy expenditure(Niven et al., 2003a; Niven et al., 2003b)(Fig. 8). More generally,altering the precise types of voltage-gated ion channels and their densities within a neuron will alter the relationship between energy consumption and information processing. Voltage-gated ion channels may increase the energy efficiency of information coding in neurons by altering the relationship between resting costs and signalling costs. For example, voltage-gated ion channels that activate at high voltages do not incur an energetic cost at rest. In spiking neurons, increasing the threshold potential would significantly reduce energy consumption because fewer action potentials would be initiated, although this will alter the information transmitted.

The properties of other classes of ion channels and metabotropic receptors may also alter the energy efficiency of information transmission in sensory receptors or at synapses. For example, several different receptor types,including AMPA, Kainate, NMDA and metabotropic glutamate receptors, are found on the post-synaptic membrane of glutamatergic synapses, which are found throughout vertebrate sensory systems. The AMPA receptors have evolved to activate and deactivate on millisecond timescales in response to glutamate,whereas NMDA and metabotropic glutamate receptors operate on longer timescales(Attwell and Gibb, 2005). It has been shown that energy consumption limits neuronal time constants and the millisecond timescale of AMPA receptors(Attwell and Gibb, 2005). It is suggested that the properties of the NMDA and metabotropic glutamate receptors are linked to their role as coincidence detectors and their involvement with synaptic plasticity. Altering the precise combination of AMPA, Kainate, NMDA and metabotropic glutamate receptors will alter the energetic cost of that particular synapse.

Energy efficiency can also be achieved by matching the filter properties of neuronal components to the signals they process(Laughlin, 1994; Laughlin, 2001; Niven et al., 2007). For example, insect photoreceptors have a region of photosensitive membrane, the rhabdom, and photoinsensitive membrane that filters the light-induced current generated by the rhabdom. The filter properties of the photoinsensitive membrane are determined by the combination and density of ion channels expressed by the photoreceptor (Vallet et al., 1992; Laughlin and Weckström, 1993; Laughlin, 1994; Weckström and Laughlin,1995; Laughlin,1996; Juusola et al.,2003; Niven et al.,2003b; Niven et al.,2004; Vähäsöyrinki et al.,2006). An ability to process information at frequencies beyond those necessary for the generation of adaptive behaviour (having excess bandwidth in the photosensitive filter) will be severely penalized by increased energetic costs over the entire signalling range and at rest(Niven et al., 2007). Indeed,the filter properties of the photoinsensitive membrane would be expected to be reduced to the absolute minimum necessary for maintaining function. Similarly,it has been suggested that the timescale of activation and deactivation of AMPA receptors at glutamatergic synapses is matched to the membrane time constants of neurons and, hence, their signalling speed(Attwell and Gibb, 2005).

Strategies to improve the energy efficiency of neural coding are not restricted to single neurons but can also occur within populations of neurons(Levy and Baxter, 1996; Vinje and Gallant, 2000; Balasubramanian et al., 2001; Willmore and Tolhurst, 2001; De Polavieja, 2002; Perez-Orive et al., 2002; Schreiber et al., 2002; Olshausen and Field, 2004; Hromádka et al., 2008). Energy efficiency within neural populations is still constrained by the properties of individual neurons, such as the relationship between the energetic cost of maintaining a neuron at rest and whilst signalling. This relationship is particularly important for sparse coding, an energy efficient coding strategy in which a small proportion of the neurons in a population represent information (for a review, see Olshausen and Field, 2004). The optimum proportion of neurons within a population is dependent upon the relationship between resting costs and signalling costs. High signalling costs and low resting costs will favour extremely sparse representations of information, so called `grandmother neuron' codes in which a single event is associated with the activity of a single neuron(Attneave, 1954; Barlow, 1969), whereas higher fixed costs favour denser neural codes in which a greater number of neurons within a population are active (Levy and Baxter, 1996; Attwell and Laughlin, 2001) (Fig. 9). Studies in both vertebrates and invertebrates have shown that neural codes at higher levels of sensory systems, such as the primary visual cortex of primates, may be sparse (e.g. Vinje and Gallant, 2000; Perez-Orive et al., 2002; Hromádka et al.,2008).

Energy efficiency can also be achieved by the placement of components within nervous systems/sensory systems to minimize energetic costs by `saving wire' (Cherniak, 1994; Cherniak, 1995). For example,the placement of brain regions with high interconnectivity adjacent to one another will reduce the total length of axons. This will reduce the amount of axonal membrane and the distance over which action potentials are transmitted. The placement of neural components at several levels of organization, from visual cortical areas to single neurons, has been shown to closely match the minimized wire length. For example, the interconnectivity of neurons within the cortex appears to minimize wiring volume(Chklovskii, 2004).

Although energy efficient coding strategies and component placements occur in many sensory systems, constraints also exist that prevent energy savings. Noise is a constraint both upon energy efficient coding and the minimization of wiring costs within the nervous system(Balasubramanian et al., 2001; De Polavieja, 2002; Faisal et al., 2005). For example, an optimum energy efficient code that maximizes the information coded by a given number of spikes (the Boltzmann distribution) predicts that neural spike rates should follow an exponential distribution. Populations of neurons encoding natural stimuli adhere to this prediction at high spike rates but not at low spike rates, which are used less often than predicted, because they are less reliable (Baddeley et al.,1997; Balasubramanian et al.,2001; Balasubramanian and Berry, 2002; De Polavieja,2002). Functional constraints, such as timing or the positions of peripheral nerves, may also prevent the nervous system adopting energy efficient strategies (Chen et al.,2006; Kaiser and Hilgetag,2006; Niven et al.,2008). These constraints limit the extent to which selective pressures on the nervous system to reduce energetic costs can affect the coding of information and the placement of components.

A complementary strategy to directly reducing the cost of processing information is to reduce the amount of predictable or redundant information within sensory systems (Atteneave, 1954; Barlow, 1961; Srinivasan et al., 1982). The most obvious source of predictable sensory information is the sensory feedback generated by motor activity. For example, limb movements can produce predictable mechanosensory feedback; however, it is the deviation from the expected sensory feedback that is essential for maintaining limb control and stability (Gossard et al.,1990; Gossard et al.,1991; Wolf and Burrows,1995). An efference copy of the expected movements can be used to selectively gate-out the predictable sensory feedback and allow only the novel sensory feedback to be processed (Sillar and Skorupski, 1986; Bell and Grant, 1989; Gossard et al.,1990; Gossard et al.,1991; Wolf and Burrows,1995; Li et al.,2002; Poulet and Hedwig,2006). By reducing the number of predictable signals being processed, the overall energetic cost is reduced but not the total amount of information. For example, within the visual system, redundant information can arise because adjacent photoreceptors sample neighbouring points on natural scenes that are highly correlated (Atteneave, 1954; Barlow, 1961; Srinivasan et al., 1982). Again, eliminating redundant signals reduces the overall energetic cost but not the total information.

Many animals possess enlarged or elaborated sensory systems for the acquisition of a specific sensory modality. The elaboration of sensory structures is often correlated with behavioural and/or ecological specialization and improved performance in particular behavioural tasks. For example, the insectivores show considerable behavioural and ecological diversity and have sensory specializations in both the peripheral sense organs and cortical regions (for a review, see Catania, 2005). The East African hedgehog (Atelerix albiventris) possesses large eyes and ears as well as whiskers and microvibrissae that suggest they are generalists, not relying wholly on one specific sensory modality. This is reflected in the organization of their cortex, which contains large prominent auditory,somatosensory and visual areas (Fig. 10A). In contrast moles, and in particular the star-nosed mole(Condylura cristata), are specialized for a subterranean lifestyle with reduced eyes and a highly modified nose that contains 22 tactile appendages and a relatively enlarged cortical somatosensory representation(Fig. 10B). Ant species with more visual behaviour, likewise, show an enlargement of the optic lobes and mushroom body lip – brain regions associated with visual processing and possibly visual memory, respectively(Gronenberg and Hölldobler,1999).

There is often an implicit assumption that relatively larger structures within sensory systems are associated with an increase in information processing and, consequently, improved behavioural performance. Relatively larger sensory structures may have greater numbers of neurons, neurons with a greater volume, or both. It also seems likely that relatively larger sensory structures will incur a greater absolute energetic cost. However, for most sensory systems, these relationships remain assumptions without empirical support. These relationships between information processing and energy consumption have been assessed in the fly retina(Niven et al., 2007). Comparison of different sized photoreceptors showed that the largest had substantially higher information rates than the smallest but also incurred substantially higher energetic costs, both at rest and whilst signalling(Fig. 7). Thus, information processing in the largest photoreceptors was less energy efficient than in their smaller counterparts. The largest photoreceptors were from flies with the greatest number of photoreceptors suggesting that, in this case, the largest retina had both the highest total energy consumption and processed the greatest quantity of information.

Although similar relationships can be envisaged for other peripheral sensory structures, the relationship between information processing, energy consumption and the size of higher centres remains unclear. The efficacy of energy-efficient coding schemes can change with size (e.g. graded versus action potentials, sparse coding), making direct comparisons difficult. Thus, direct quantification of the energetic costs, performance and size of a particular sensory system is essential for understanding the cost–benefit trade-offs that have influenced its evolution. This is particularly important in comparisons of phylogenetically distant species,among which it is not reasonable to assume that a specific volume of neural tissue consumes similar amounts of energy. For example, elasmobranchs have a larger relative brain size when compared to teleost fish of the same body mass(Nilsson et al., 2000)(Fig. 11A). Early studies assumed that elasmobranch brains consumed considerably more energy than those of teleosts with similar body mass. However, measurements from elasmobranch and teleost neural tissue have shown that they have very different specific rates of energy consumption but that overall their brains consume similar amounts of energy (Fig. 11B,C).

If the trade-off between the performance of sensory systems and their energetic costs has influenced their evolution, then under conditions in which the need to maintain performance in a particular sensory system is reduced,the sensory system itself should be reduced. Many animals living in environments in which information from a particular sensory modality is no longer useful, and is therefore under reduced selective pressure, do indeed show marked reductions in the size of structures associated with that modality. For example, the peripheral and central components of the visual systems of animals that live in caves or subterranean environments are often reduced or absent. Populations of the Mexican cave fish (Astyanax mexicanus) isolated in caves without light have undergone convergent eye loss at least three times within the last 1 million years, whereas populations that have lived continuously on the surface have retained their eyes(Fig. 12A)(Strecker et al., 2004; Wilkens, 2007; Borowsky, 2008; Niven, 2008a). Likewise, fish that live permanently in caves have relatively smaller visual processing regions in their brain compared to those that are found both in caves and on the surface (Fig. 12B)(Poulson and White, 1969). Blind mole rats (Spalax ehrenbergi) also show a marked reduction in eye size, the remaining structures being maintained subcutaneously for circadian rhythm generation (David-Gray et al., 1998). Like the star-nosed mole, blind mole rats have a reduced thalmocortical visual system and an expanded somatosensory representation associated with a subterranean lifestyle. In some cases,reductions in the volume of sensory processing regions can occur within individuals following the transition to a new environment. For example, virgin female ants (Messor pergandei or Pogonomyrmex rugosus) make mating flights before removing their wings and excavating a subterranean nest to found a new colony. Brain regions specifically associated with visual processing such as the medulla are reduced in volume in these mature mated ant queens relative to virgin female ants(Julian and Gronenberg,2002).

The reductions of specific sensory structures and their associated brain regions also occur in animals living on islands (see Niven, 2005; Niven, 2007; Niven, 2008b). For example, a fossil bovid, Myotragus balearicus, found on two Mediterranean islands, has a reduced orbit and endocast volume relative to bovids found on the mainland (Kohler and Moya-Sola,2004) suggesting a reduction in visual processing due to a reduced need for vigilance in the absence of mainland predators on these islands(Kohler and Moya-Sola, 2004; Niven, 2005). Fruit flies(Drosophila melanogaster) bred under laboratory conditions do not require vision to locate food or mates reducing the selection pressure on the visual system (Fig. 13). These flies show an overall reduction in compound eye size and facet size that is related to their time in captivity (Tan et al., 2005).

These reductions or losses of sensory structures in numerous independent lineages suggest that there is an advantage to the loss of sensory structures in environments in which they cannot provide information, implying that they represent a cost. An alternative explanation may be that in the absence of selection pressure to retain sensory structures they are lost through disuse,not because of their cost, as proposed by Darwin(Darwin, 1859):

As we discussed above, however, neural tissue is energetically expensive both to use and to maintain (Ames et al.,1992; Nilsson,1996; Attwell and Laughlin,2001; Lennie,2003; Niven et al.,2003a; Nawroth et al.,2007; Niven et al.,2007). Thus, in the case of subterranean vertebrates and invertebrates, maintaining the retina and central visual processing regions would incur considerable energetic costs, even in the dark. This suggests that the reduction of peripheral and central visual structures is indeed due, at least in part, to their high energetic costs. The high energetic costs of maintaining and using neural structures should influence the evolution of both central and peripheral structures irrespective of the particular sensory modality. Indeed, these high energetic demands should favour the reduction of peripheral and central structures associated with a particular sensory modality to a functional minimum (Niven et al., 2007). However, the strength of the selective pressure for reduced energetic costs of sensory systems will depend critically upon the precise environmental circumstances in which a specific animal finds itself(Niven, 2005; Niven, 2007; Niven, 2008a; Niven, 2008b). For example,animals living in caves or on islands are often extremely energy limited,increasing the need for energy saving by reducing sensory structures to a functional minimum.

The limited energy budgets of animals coupled with the high energetic costs of the brain have led to the suggestion that the additional energy invested in the development, maintenance and use accompanying an expansion in brain size is traded off against a reduction in the size of another energetically expensive tissue. Aiello and Wheeler proposed that during primate evolution the expansion of the brain relative to body mass was accompanied by a relative reduction in gut size: the expensive-tissue hypothesis(Aiello and Wheeler, 1995). A similar correlation has also been found in teleost fish(Kaufman et al., 2003). However, more recent tests of this theory in birds and bats have failed to find strong support for a trade-off between brain size and gut size(Jones and MacLarnon, 2004; Isler and van Schaik, 2006b). For example, in birds pectoral muscle mass is negatively correlated with brain size, suggesting a trade-off, whilst reproductive costs are positively correlated (Isler and van Schaik,2006b). One possibility suggested is that species with larger brains are better able to provision their offspring(Isler and van Schaik,2006b).

A further implication of the expensive-tissue hypothesis is that a reduction in brain size relative to body could accompany increased energy usage by another tissue/organ. For example, some ant queens use their energy stores to produce the initial workers within their colony. The provisioning of eggs requires large amounts of energy and queens use additional energy gained from the breakdown of flight muscles and possibly from visual processing structures for reproduction(Hölldobler and Wilson,1990; Julian and Gronenberg,2002).

Trade-offs may also occur between sensory systems, the increase in the volume of peripheral sense organs or central sensory processing regions in one modality being accompanied by a reduction in another sensory modality. For example, as mentioned above, both the star-nosed mole and the blind mole rat have reduced a thalmocortical visual system and an expanded somatosensory representation (Cooper et al.,1993; Catania,2005). Numerous examples of enhanced somatosensory systems in animals with reduced visual systems also occur in both vertebrates and invertebrates that inhabit cave systems. These trade-offs between different sensory modalities may be important because they may not affect the total energetic cost of sensory processing within the brain substantially and,therefore, do not necessarily affect energetic demands. However, the expansion or reduction of peripheral or central sensory processing structures may be limited by the extent to which regions within the brain can evolve independently (mosaic evolution) (Finlay and Darlington, 1995; Barton and Harvey, 2000; Striedter,2005). Within the mammalian cortex, however, substantial developmental plasticity can occur with sensory processing regions being co-opted for different sensory modalities depending on experience, including trauma (for a review, see Krubitzer and Kaas, 2006). Such plasticity between different sensory modalities within the cortex may be particularly important because it facilitates rapid adaptation to novel environmental circumstances without substantially affecting the total energetic cost of sensory processing within the brain.

Energy consumption affects all aspects of animal life from cellular metabolism and muscle contraction to growth and foraging(Alexander, 1999). Yet despite early studies on energy metabolism in neural tissue (e.g. Kety, 1957), the impact of energy consumption upon the evolution of nervous systems has only recently begun to be generally appreciated(Laughlin, 2001). Recent studies have made substantial advances in relating the energy consumption of neural tissue to neural function. Together these studies show that there are high energetic costs associated with the nervous system both at rest and whilst neurons are signalling (Laughlin et al., 1998; Attwell and Laughlin, 2001; Niven et al.,2007). Crucially for the evolution of the nervous system, and in particular sensory systems, these costs are incurred even during activity. Thus, animals pay an energetic cost associated with nervous system irrespective of the demands of other tissues such as skeletal muscle.

Evidence from fly photoreceptors suggests that the energetic costs incurred by neurons at rest are linked to their energetic costs whilst signalling by basic biophysical relationships (Niven et al., 2007). Thus neural function, and therefore the production of adaptive behaviour, is linked to neural energy consumption. Excess signal processing capacity in sensory systems is severely penalized by increased energetic costs producing a Law of Diminishing Returns. The precise relationship between energy consumption and signalling is likely to depend on the specific neuronal type; these relationships can be adjusted by the specific combinations of ion channels, and possibly synaptic inputs, within the neuronal cell membrane. Numerous strategies for reducing the costs incurred by the sensory systems have been found in both insect and vertebrate sensory systems (e.g. Vinje and Gallant,2000; Perez-Orive et al.,2002; Niven et al.,2003a; Niven et al.,2003b; Hromádka et al.,2008). These strategies aim to reduce the energetic costs within sensory systems by filtering out predictable inputs to sensory systems,reducing the amount of redundant information that is encoded and representing this information more efficiently.

Energy limitations appear to have affected the evolution of sensory systems, causing trade-offs between sensory systems encoding different modalities. The effects of energy limitation appear to be especially obvious in animals living in on islands or in caves, which tend to be energy-limited environments (Kohler and Moya-Sola,2004; Niven,2007; Borowsky,2008; Niven,2008a; Niven,2008b). For these animals, reductions or complete losses of visual structures are relatively common and appear to confirm the penalty for excess capacity found at the level of single neurons.

The acquisition of sensory information for many modalities, including vision, requires muscular movements that are largely ignored by most analyses of energy consumption within sensory systems. For example, eye and/or head movements are essential in both mammals and insects for obtaining certain types of visual information, such as parallax. Active senses such echolocation in bats and electrosensation in fish also depend on motor activity to generate the initial signal. The energetic costs associated with this muscle activity may be substantial and will further increase our estimates of the costs of sensory systems.

We would like to thank John Douglass, Biswa Sengupta and Bill Wcislo for comments. This study was supported by the Royal Society (J.E.N.), the BBSRC(S.B.L.) and the Frank Levinson Family Foundation to the STRI Laboratory of Behavior and Evolutionary Neurobiology (J.E.N.).