Artificial Intelligence in Bioscience Symposium 2

Berg Health, Slava Akmaev

Interrogative Biology

Looked at cancer, neurology, diabetes biologic models and disease pathways. Work primarily with human cell cultures invitro. Want to know what happens invivo. Does proteomics and genomics.

Difficult to analyse data using different statistics models. Says now based using bayesian networks and analyse topologies between environments. Developed algorithms to rank molecular targets or biomarkers. Have resources to do lab invitro validation. 

Says AI is some kind of software that may be able to develop reasoning. So not scientists deciding, just software. Just trying to automate research process in life sciences. Says bayesian networks can infer cause & effect hypothesis not associations and correlations.

Parkinsons – with Parkinsons Institute & Clinical Center, not much known re molecular pathways. Now collaborating with AstraZeneca. Looking at human fibroblasts.GBA linked to Parkinsons. Showing as potential top target but GBA well known. Wanted novel IP so looking at GBA near networks proteins. PIG 3 gene identified. Induced neurogeneration. Found expression associated with rotenone induced apoptosis in dopaminergic cell line.
UCL Cell & Development Biology, Dr Caswell Barry

Translational research not primary aim. Interested in problems do neuroscientists face and what can AI do. Mix of neuroscientists – some study single neuron, others study multiple e.g. using fMRI

Memory and brain. Experimental neuroscientists – place cells and grid cells – found through recording of human / animal movement. I

Problems aren’t computational power, can already simulate complex neural networks. Isn’t volume of data either (e.g.lower definition fMRI may use 2GB). Problem is understanding what is in the data and generating hypotheses and models. He refers toAI as deep learning e.g. networks that attempt to mimic architecture of brain. Computer scientists effective at borrowing ideas from brain.

Noisy neural data from stimuli, link to behaviour and try to predict. Idea is to try to train a neural network to do something a human brain might. Theory.

Examples?

2 from own lab

– working with Tartu Uni, Estonia have place cells from animals, can train betwork to predict where animal is then compare back? Stacked LSTMs  benchmark bayesian decoder. Network consistently better than bayesian decoder. Not actual neuroscience so examining nodes in network, when active, which active, where active. 

– with Trondheim uni, Norway – tracking location in virtual reality using human subjects in fmri. Feedforward network. Can do some decoding of location. Doing better than chance. Why not performing SVM?

If had more data could do more. So do DNNs  make good models of brain? Mentions  Marcel Van German – Jnl of Neuroscience 2 yrs ago. 

Does he / lab have a functional connectome?  Using test bed with worm brain. 

Important differences between brain and machine neural networks – no back propogation, spikes, brains only have positive firing rates, segregation of excitation/inhibition between neurons.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s