Neuroinformatics: sharing, organizing and accessing data and models Arnd Roth Wolfson Institute for Biomedical Research University College London

The optogenetics revolution Fuhrmann et al., 2015

The optogenetics revolution Fuhrmann et al., 2015

The connectomics revolution Helmstaedter et al., 2013

The connectomics revolution Helmstaedter et al., 2013

Connectomics data mining Jonas & Körding, 2015

Connectomics data mining Jonas & Körding, 2015

Deep artificial neural networks Mnih et al., 2015

Neuroinformatics: sharing, organizing and accessing experimental data Allen Institute Janelia Research Campus Open Connectome Project Cell Image Library Human Brain Project INCF

Single neuron and network simulators NEURON GENESIS MOOSE PSICS NEST

Meta-simulators: simulator- independent model description PyNN neuroConstruct NeuroML NineML

12 neuroConstruct

13 neuroConstruct Software tool (written in Java) developed in Angus Silver’s Laboratory of Synaptic Transmission and Information Processing Facilitates development of 3D network models of biologically realistic cells through graphical interface Allows anatomical positioning of cells and complex connectivity of axons/dendrites Automatically generates scripts for running simulations in NEURON/GENESIS/MOOSE/PSICS/PyNN & more Support for import, export & conversion of NeuroML

14 neuroConstruct – latest developments neuroConstruct can generate code for Parallel NEURON - Most widespread platform for large scale detailed neuronal simulations - Near linear speedup of simulations up to hundreds of cores Python scripting interface - Python becoming language of choice for neuroinformatics applications - Gives access to all functionality “behind the GUI” Open Source Brain - Platform for sharing & collaboratively developing models in computational neuroscience - Many neuroConstruct projects from multiple brain regions available

3D version of Traub et al 2005 Thalamocortical column model Parallel simulation durations scale approx. linearly up to 200 processors & 10,000 cells Example using Python interface & Parallel NEURON

16 Wider interoperability framework

Towards multiscale simulation: from molecules to circuits MCell CellBlender STEPS TrakEM2 TREES toolbox

Public databases of neural models ModelDB NeuroMorpho.org BigNeuron OpenSourceBrain Human Brain Project

19http:// How to make computational neuroscience a more accepted scientific approach? Reproducibility: easy to rerun and validate simulation result reported in a scientific paper. Accessibility: available to theoretical and experimental neuroscientists in an understandable format Portability: cross-simulator validation and exchange of models and components enabling reuse Transparency: exposure of internal properties and automated validation

20http:// Neuroinformatics infrastructure NeuroML A simulator-independent language for describing and exchanging detailed neuronal and network models LEMS Compact and flexible model description language that underlies NeuroML 2 The Open Source Brain Initiative Accessible repository of standardized models and infrastructure for collaborative, open source model development

21 The Open Source Brain repository

22 Current model development life-cycle

23http:// Current model development life-cycle

24http:// OSB collaborative development scenario

OSB iterative development through critical evaluation Validate Experiment Model

26

A Whole Community Approach Must bring experimental and theoretical & computational neuroscience closer. While the latter seek minimal models, the former want hard earned experimental facts not to be ignored. As the functional principles of neuronal networks in the brain remain elusive, and the interactions are often highly non-linear, ignoring biological facts without thought to errors can easily result in misleading conclusions, and erroneous theories of brain function. Adhoc simplification is a matter of taste

Level of detail: A rift in neuroscience 1.Simplify the details – minimal model for hypothesis-driven science – Adhoc simplification – Minimal for which question? vs 2.Consider all known – data-driven is data-ready – Hypothesis-free integration of facts – Algorithms fill in gaps from sparse data – Fewer free parameters! – Avoid wasting time hand tuning parameters for a given model “island”

“We find that the major obstacle that hinders our understanding the brain is the fragmentation of brain research and the data it produces. Our most urgent need is thus a concerted international effort that can integrate this data in a unified picture of the brain as a single multi- level system...” The HBP-PS Consortium 2012:8