Design for a Genetic System Capable of Hebbian Learning Chrisantha Fernando Systems Biology Centre Birmingham University January 2006 Chrisantha Fernando Systems Biology Centre Birmingham University January 2006

Hebbian Learning  Donald Hebb “ Let us assume that the persistence or repetition of a reverberatory activity (or "trace") tends to induce lasting cellular changes that add to its stability. When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased. ”  Long term potentiation in neurons subsequently demonstrated, along with LTD.  Hebbian Learning can implement  Associative Learning, e.g. LTP in Hippocampus.  Auto-associative Memory, e.g. Cerebellar motor memories.  Self-Organized Map Formation, e.g. ocular dominance columns.  Donald Hebb “ Let us assume that the persistence or repetition of a reverberatory activity (or "trace") tends to induce lasting cellular changes that add to its stability. When an axon of cell A is near enough to excite a cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased. ”  Long term potentiation in neurons subsequently demonstrated, along with LTD.  Hebbian Learning can implement  Associative Learning, e.g. LTP in Hippocampus.  Auto-associative Memory, e.g. Cerebellar motor memories.  Self-Organized Map Formation, e.g. ocular dominance columns.

Examples of Hebbian Learning Associative learning: Classical Conditioning UCS = Collision CS = Proximity

Hebbian Learning in Rabbits

Self-Organized Maps

How is Hebbian Learning Implemented in Brains? Pre Post +/-

What about Learning in Single Cells?  No conclusive evidence that unicellular organisms can do associative learning on their own.  Bacteria can habituate and be sensitized but they have not been shown to undertake associative learning.  No conclusive evidence that unicellular organisms can do associative learning on their own.  Bacteria can habituate and be sensitized but they have not been shown to undertake associative learning.

Vibration Shock

Slower Replicating Eukaryotes in Complex Environments Might Benefit from Lifetime Learning.  Fine tuning of obstacle avoidance/motion.  Learn that chemical A (or a spatio-temporal pattern of A) is associated with food or harm rather than chemical B or C or D…  Achieve robust development in the face of lifetime noise due to external perturbations and internal (mutational) perturbations.  Fine tuning of obstacle avoidance/motion.  Learn that chemical A (or a spatio-temporal pattern of A) is associated with food or harm rather than chemical B or C or D…  Achieve robust development in the face of lifetime noise due to external perturbations and internal (mutational) perturbations.

Intra-cellular Molecular Networks  Dennis Bray (Nature 2003). Neural networks have a remarkable ability to learn different patterns of inputs by changing the strengths of their connections (10). They are widely used in a variety of tasks of machine recognition. From the standpoint of a living cell, the closest approximation to a neural network is probably found in the pathways of intracellular signals (11). Multiple receptors on the outside of a cell receive sets of stimuli from the environment and relay these through cascades of coupled molecular events to one or a number of target molecules (associated with DNA, for example, or the cytoskeleton). Because of the directed and highly interconnected nature of these reactions, the ensemble as a whole should perform many of the functions commonly seen in neural networks. Thus, in aggregate, the signaling pathways of a cell are capable of recognizing sets of inputs and responding appropriately, with their connection "strengths" having been selected during evolution.1011  Eric Winfree and Hopfield view gene transcription networks as CTRNNs, pto.  Dennis Bray (Nature 2003). Neural networks have a remarkable ability to learn different patterns of inputs by changing the strengths of their connections (10). They are widely used in a variety of tasks of machine recognition. From the standpoint of a living cell, the closest approximation to a neural network is probably found in the pathways of intracellular signals (11). Multiple receptors on the outside of a cell receive sets of stimuli from the environment and relay these through cascades of coupled molecular events to one or a number of target molecules (associated with DNA, for example, or the cytoskeleton). Because of the directed and highly interconnected nature of these reactions, the ensemble as a whole should perform many of the functions commonly seen in neural networks. Thus, in aggregate, the signaling pathways of a cell are capable of recognizing sets of inputs and responding appropriately, with their connection "strengths" having been selected during evolution.1011  Eric Winfree and Hopfield view gene transcription networks as CTRNNs, pto.

Incomplete promotor region Sigmoidal activation function is sequence programmable but hard wired.

The New RNA World  John Mattick, Sean Eddy etc… are investigating RNA networks in Eukaryotes.  98% of RNAs in human cells is Intron or non- coding.  10x more RNA on Chromosome 21 and 22 is non- coding than exons.  Many types of novel RNA have been discovered capable of regulation of gene expression.  It may eventually explain the C-value paradox.  John Mattick, Sean Eddy etc… are investigating RNA networks in Eukaryotes.  98% of RNAs in human cells is Intron or non- coding.  10x more RNA on Chromosome 21 and 22 is non- coding than exons.  Many types of novel RNA have been discovered capable of regulation of gene expression.  It may eventually explain the C-value paradox.

1

What if RNA networks are really like little neural networks?  Current neural network metaphors for cellular networks have not included plasticity.  How difficult would it be to design (or evolve) a network capable of Hebbian learning?  What sorts of tasks could an intra-cellular Hebbian learning mechanism solve?  Current neural network metaphors for cellular networks have not included plasticity.  How difficult would it be to design (or evolve) a network capable of Hebbian learning?  What sorts of tasks could an intra-cellular Hebbian learning mechanism solve?

u 1 + W* 1 u 2 + W* 2 u 2 W* 2 u 1 W* 1 RNA W* 2 W W* 1 W RNA + u 1 RNAu 1 RNA + u 2 RNAu 2 Promotor Gene k1k1 k2k k1k1 k2k2

Time [Conc] [RNA] [TF 1 ] [TF 2 ] [RNATF 1 ] [RNATF 2 ] Stimulate with unit TF 2 Stimulate with unit TF 1 Experiment 1: Sensitization

Time [Conc] [W 1 *] [W 2 *] ‘Weight’ Change in Experiment 1

Associative Learning  UCS (e.g. glocuse) stimulates TF 1, which binds strongly to the promotor and produces RNA (innately).  CS (e.g. potassium permanganate, or NO or something else) stimulates TF 2 but these TFs bind only very weakly to the promotor.  Paired exposure to UCS and CS result in strengthening of the TF 2 binding to promotor, and a response to the ‘smell’ associated with glucose.  UCS (e.g. glocuse) stimulates TF 1, which binds strongly to the promotor and produces RNA (innately).  CS (e.g. potassium permanganate, or NO or something else) stimulates TF 2 but these TFs bind only very weakly to the promotor.  Paired exposure to UCS and CS result in strengthening of the TF 2 binding to promotor, and a response to the ‘smell’ associated with glucose.

Self-Organized Maps  Two TFs act on two genes  Now add lateral inhibition  Gene ‘receptive field’ changes  Gene A represents TF 1 level  Gene B represents TF 2 level  Two TFs act on two genes  Now add lateral inhibition  Gene ‘receptive field’ changes  Gene A represents TF 1 level  Gene B represents TF 2 level

Supervised Learning: Training a gene perceptron.

Applications  Engineering cells as biosensors in complex environments. Put cell in, then measure W* levels to obtain the recorded weights.  To train a cell to produce a protein under required conditions, without having to hard- wire the promotor-TF interaction perfectly.  Engineering cells as biosensors in complex environments. Put cell in, then measure W* levels to obtain the recorded weights.  To train a cell to produce a protein under required conditions, without having to hard- wire the promotor-TF interaction perfectly.

Thanks to  Dov Stekel  Jon Rowe  Bruce Shapiro  Sally Milwidsky  Dov Stekel  Jon Rowe  Bruce Shapiro  Sally Milwidsky