Modeling functional networks of the primate cortex With: Maria Ercsey-Ravasz Department of Physics & Interdisciplinary Center for Network Science and Applications (iCeNSA) University of Notre Dame Henry Kennedy, Kenneth Knoblauch, Nikola Markov, Marie-Alice Gariel, etc. Stem Cell and Brain Research Institute, INSERM U846 LYON Markov N, Misery P, Falchier A, Lamy C, Vezoli J, Quilodran R, Giroud P, Gariel MA, Ercsey-Ravasz M, Pilaz LJ, Huissoud C, Barone P, Dehay C, Toroczkai Z, Van Essen DC, Kennedy H, Knoblauch K. Area-specific visual cortical connectivity profiles revealed by quantitative brain-wide analysis. Cereb Cortex, in press. (2010) Markov N, Ercsey-Ravasz MM, Gariel MA, Falchier A, Quilodran R, Vezoli J, Lamy C, Misery P, Giroud P, Clavagnie S, Sallet J, Barone P, Dehay C, Toroczkai Z, Kennedy H, Knoblauch K. Heterogeneity of connection strengths and a distance rule specify cortical networks. subm. (2010) How does the brain work? Understanding the brain would greatly benefit: Medicine Computing, science, industry Artificial Intelligence (AI) started out with huge optimism: 1965: H.A. Simon: “Machines will be capable, within twenty years, of doing any work a man can do.” 1970: M. Minsky: “In from three to eight years we will have a machine with the general intelligence of an average human being” 1980’s: Progress stalled 2005: Jeff Hawkins: “Neural network research ignores essential properties of the human cortex” AI tries to get around real brain research, led by the belief that computing paradigms are not so many, and that they could be “guessed”. Nature already figured Well, the brain is not this: out efficient computing: - 0.09g (green lizard)… 1400g (human)… 7800g (sperm whale). Get clues from brain experiments?! So what’s taking so long? It is this “goo”: Networks of the real world Networks are physical systems that can be represented as graphs with processes on them. Where are networks? • Infrastructures: transportation nw-s (airports, highways, roads, rail, water) energy transport nw-s (electric power, petroleum, natural gas) • Communications: telephone, microwave backbone, internet, email, www, etc. • Biology: brain, gene regulatory network, protein-protein interactions, metabolism, cell-signaling, the food web, etc. • Social systems: acquaintance (friendship) nw-s, collaboration networks, epidemic networks • Geology: river networks Real-world networks perform functions and are characterized by heterogeneous structures This typically, already shows up at the level of connectivity (degree) distributions: Binomial Random graph (Erdős-Rényi graph) Poisson Capacity achieving degree distribution of Tornado code. The decay exponent -2.02. [ M. Luby, et.al. ,Proc. 29th ACM Symp. Theor. Comp. pg. 150 (1997). ] networks in biology the metabolic pathway Chemicals Bio-Chemical reactions Cellular Networks: The Bio-Map GENOME Protein-gene interactions PROTEOME Protein-Protein interactions METABOLISM Bio-chemical reactions Citrate Cycle social networks person relationship R. S. Burt (2000) “Network structure of social capital” Networks can be dynamic and evolving. Visualization by Aaron Koblin, http://www.aaronkoblin.com/info.html Human built networks are purposely functional, and they seem to be heterogeneous, as it was seen in the case of the Internet and WWW. More examples: Electronic circuits and devices: Software (subroutine is a node, a call is a link): N=104 element logic circuits (iscas89, itc99). R. F. I Cancho, C. Janssen and R.V. Sole. Phys. Rev. E 64, 046119 (2001). S. Valverde and R.V. Sole. Europhys. Lett. 72, 858 (2005). Biological networks are nature-built, via evolutionary processes, and they are clearly functional. Archae a Bacteria Bacteria Eukaryote s H. Jeong et.al. Nature 407, 651 (2000). Eukaryote s The cerebral cortex The neuron (brainmaps.org) An adult primate brain contains about 100 billion (1011) neurons, with branches that connect at more than 100 trillion (1014) points. The network of ~1011 neurons and ~1014 synapses is too large to be genetically encoded synapse by synapse. Has randomness, modified during learning: Information encoding emerges at larger scales. Statistical description should provide reproducible measures. The best method is functional decomposition: a TOP-DOWN approach 10 Functional areas of the cortex Segmentation based on histological criteria. Cytoarchitectonic details. Macaque cortex: 83 areas 11 Experiments: Retrograde Tracing Kennedy team (Lyon): (70 man-years work!) Retrograde tracing uncovers inter-areal connectivity. Tracers: Fast Blue (FsB), Diamino Yellow (Dy) A neuron is a directed element “Target” NO secondary labeling (of pre-synaptic neurons)! Measurements: - number of labeled neurons in source areas - projection lengths through the white matter An injection reveals the FULL set of INCOMING projections. Macaque cortex: 26 areas injected: V1, V2, V4, TEO, DP; PBr, STPr, STPi, STPc, 7A, 7B, 2, 5, F1, F2, F5, F7, 24c, ProM, 46d, 9/46d, 9/46V, 8A, 8B, 45B ? ? “Sources” “Target” 12 Injections sites in the macaque cortex Low-power fluorescence photomicrographs obj X10 filter Injection in V1, horiz section plane Injection in V2, coronal section plane Injection in V4, coronal section plane 13 40 μm histological sections 14 Fraction of Labeled Neurons (FLN) Assume that an injection was performed in area X. Define: FLN(X Y ) NY Ne - the “probability” that an extrinsic labeled neuron is projecting into X from Y. The FLN values are highly characteristic across animals: FLN • Varies over 6 orders of magnitude! • This heterogeneity is a signature of functional organization. Distribution by reach ~79% within 2mm (local) ~16% to neighboring areas ~4% long-range ~1% sub-cortical Inter-areal network The FLN distribution Close to a lognormal (f=FLN) if 2 with Strong heterogeneity in weight distribution. log10(FLN) 16 Why Lognormal? I. Projection lengths: Exponential Distance Rule (EDR): p(d) exp(d) 0.088mm1 1 11.36mm This can be interpreted as the probability of a single neuronal projection having length d. this is an average physical property across the whole cortex: We assume that FLN( j i) f ij p(dij ) exp( dij ) ln f ij dij ln Thus, on average, distances are in linear relationship with the corresponding ln(FLN) values. 17 II. Distribution of inter-areal distances (Geometry) This is the fraction of area pairs i,j separated by a given distance d. Expectation: unimodal distribution. Fitted by a (truncated) Gaussian: Substituting ( f ) d 1 ln f 1 ln from the exponential distance rule: 1 1 2 1 2 exp 2 1 ln f 1 ln exp 2 ln f 2 2 2 2 ln lognormal 18 Others have also found lognormals in the brain: Distance rule locally: N(d) N0 exp(sd) s: “Song” decay rate (mm-1) Kennedy measurements, intrinsic labeling, distance distribution. 22% loss at d1 = 50μm = 0.05mm 78% are not lost at this distance N(d1)/N0 0.78 exp(sd 1) s 4.96mm 1 4.35mm1 d2 = 100μm = 0.1mm 56% are not lost at 44% loss at this distance s 5.79mm 1 Network properties • directed • weighted • spatially embedded ? - 26 injected nodes - all 83 nodes projecting into the set of 26 - found total of E=1232 directed edges/projections - there are M=430 edges between the 26 nodes The graph amongst the 26 nodes forms which is an edge-completed sub-graph of Density of G26x26 : 2626 ? G26x26 G26x26 G83x83 G83x83 430 430 0.66.. 26 25 650 In-degree distrib. of the 26 nodes within G83x83 It is very dense! 66% At such high densities, the density of the graph is the most determining factor for its “binary” properties: Necessarily it will have: -short average path length ( = 1.34 for G26x26) -small diameter ( = 2) -high clustering coefficients (= 0.87). SOUNDS LIKE SMALL-WORLD… But it really isn’t… too dense. Neither is Worst case: 1232/83(83-1) = 0.181 (18%). Extrapolating the same mean: 57% G83x83 20 Two Random Graph models: Constant distance rule (CDR): p(d) = constant. Exponential distance rule (EDR): p(d) = λexp(λd). Using the set of distances D measured between the areas: - choose a connection length d from D according to the distance rule p(d) - pick uniformly at random an area pair (i,j) whose separation is d (same bin) - insert a connection in the graph directed from j to i - multiple connections from j to i are allowed, thus generating its weight - the process stops once all the 430 binary connections are reached 21 Degree and 3-motif distributions EDR clearly outperforms CDR A non-local, binary property: Average (binary) shortest path length as function of density, while sequentially removing the weakest links. A non-local weighted property: Communication capacity: ln pk i j ln f k i ln f i j . j k pk i j f k i f i j , positive, additive along paths i Link “resistance”:wij ln f ij ln 1/ f ij 0 rij = lowest resistance path weight (sum of the w-s along paths). Global “conductance”: Eg 1 1 N(N 1) i j rij V Latora, M Marchiori, Europhys J B (2003) Local “conductance”: El 1 1 N i ki (ki 1) j k{i} 1 rjk / i I Vragovic, E Louis, A Diaz-Guilera, PRE (2005) High bandwidth backbone: (Kamada-Kawai algorithm used for layout) Unweighted network Weighted network visual prefrontal parietal temporal sensori-motor 80 strongest connections just before un-reachability 24 Conclusions – so far • Massive involvement in local circuitry. • Link strengths are consistent across individuals • Range of strengths varies over 6 orders of magnitude • The interareal cortical network has high density. • Binary features are not very informative. • Weights, distances and geometry are important • Exponential distance rule: strong economy of lengths. • An EDR based simple model captures binary properties • Fluctuations around the EDR encodes bio-specificity: e.g., new projections of PERI onto amygdala support recent theories of inference based on memory prediction (Hawkins et.al. 2009). • There is direct access between almost any two areas, likely ensuring efficient high-level integration • Massive processing of information happens locally. • To block out cacophony, information is filtered/encoded using hierarchies of weights (6 orders) • Weight/geometry - based hierarchical layering in cortical architecture and processing. 25 Plasticity of the cortex: Brainport: http://www.pbs.org/kcet/wiredscience/video/286-mixed_feelings.html Rewiring ferret’s brain at birth: L. Melchner, S.L. Pallas & M.Sur. Nature, 404, 871 (2000) 26