Our broad goals are to elucidate the functional architecture of local circuits in the mammalian brain and to identify the constraints that have shaped their evolution. A century ago Golgi and Cajal argued about the fundamental nature of these circuits (Nobel Lectures, 1906): are they built from generic neurons that connect imprecisely in an “associative net” (Crick 1979), or from specific cell types that connect precisely like a VLSI circuit (Bullock 1970)? This dispute has now been settled in the retina where connections are sufficiently compact (spanning ~ 0.1 mm) that they have been reconstructed quantitatively by transmission EM of serial, ultrathin sections. Here is what we know.
(i) every neuron does belong to a “natural” type. Types are defined by their clusterings in multiparametric space. There are ~ 60 types, and many are conserved across species from mouse to man.
(ii) every neuron does connect precisely – on several levels. Each neuron contacts a specific set of other types (= 10); each neuron diverges to a specific number of cells; each type converges in specific numbers onto its postsynaptic targets, and each connection employs a specific number of synapses (chemical or electrical) with a precision of ~ 5%. Each synapse expresses a specific complement of pre- and postsynaptic proteins that are regulated down to the level of isoforms and splice-variants.
(iii) structural patterns can be “read” like a VLSI circuit diagram. Thus the wiring of rods and cones suggested separate circuits for daylight, twilight, and starlight, and these were confirmed by subsequent recordings.
Key aspects of the circuitry remain to be elucidated – principally the synaptic interactions between ~30 types of inhibitory interneuron (amacrine), 10 types of excitatory interneuron (bipolar) and 15 types of output neuron (ganglion cell). Also, although the main transmitters are identified (glutamate, GABA, glycine, acetylcholine, dopamine), numerous neuromodulators are present, and many types of G protein-coupled receptors whose contributions to signaling await serious investigation.
Meanwhile, we have begun to investigate how the informational properties of various circuits relate to their functional architecture. Comparisons are facilitated across circuits and conditions when information rates are measured in units of bits s-1. Comparing information rates for various types of ganglion cell, here is what we have learned.
(i) ganglion cell dendritic arbors create gaussian synaptic weightings to maximize their S/N. Neighboring arbors overlap – with cell spacing narrow enough (2sigma) for a flat sensitivity surface but wide enough for independence -- which allows each array to transmit the most information.
(ii) OFF dendritic arbors are narrower and more densely branched than ON arbors. Thus the OFF cell gets the same number of synapses as the ON, but over a narrower visual angle, and the OFF array samples more densely. This predicts that dark regions in natural scenes contain more information -- and a corollary principle: “equal bits for equal synapses”. Model arrays tested on natural images confirm this.
(iii) ganglion cell types with the most synaptic input and the thickest axons send 2-fold more bits/s than types with the least input and thinnest axons. But, because mitochondria always occupy 2% of axoplasm, the metabolic cost of the largest axon is nearly 30-fold greater than the smallest axon. Thus, where higher bit rates are essential, the brain pays dearly – as we also do for broadband signaling.
|