Machine Learning for Life Sciences¶ – Definitions sure to annoy serious practitioners¶ – Opportunities in the Life Sciences¶

View Github Repository Open presentation in a new window

5x5x5x5

See all presentation from 5x5x5x5

Machine Learning for Life Sciences¶ – Definitions sure to annoy serious practitioners¶ – Opportunities in the Life Sciences¶

0 1

Machine_Learning_Life_Science

On Github 5x5x5x5 / Machine_Learning_Life_Science

Machine Learning for Life Sciences

Follow along at: 5x5x5x5.github.io/Machine_Learning_Life_Sciences

Definitions sure to annoy serious practitioners

Machine Learning - Using algorithms and computation to generalize from data

Life Sciences - Anything that wiggles from 20nm to 30m in length

Most of the subjects I will touch on are incredibly deep and worthy of their own talk. Thankfully, the Research Triangle Analysts have already given some of them.

Opportunities in the Life Sciences

  • Research frontier is relatively close
  • New developments aimed at reducing costs and increasing reproducibility
    • Science Exchange
    • Riffyn
    • Emerald Cloud Lab
    • Transcriptics
  • More data to feed the Data Scientists
  • New approaches open up new avenues for data analysis

Three classification problems with three techs

  • classification (QSAR) using scikit-learn in Python
  • next gen sequencing RNA-seq MLSeq classification example in R
  • diagnostic image analysis with ConvNets in Python using Theano

Remote Laboratory Services facilitated by Science Exchange

Protocol Standardization and Optimization by Riffyn

Quality systems engineering for research methods

Design of experiments analysis

Interview with the Founder of Riffyn

What to do with all that data?

Obligitory Data Science Triforce

Data Science to the rescue

Highway to the Danger Zone?

  • no statistical grounding
  • no basis for claims

Fortunately, it requires near willful ignorance to acquire hacking skills and substantive expertise without also learning some math and statistics along the way. As such, the danger zone is sparsely populated, however, it does not take many to produce a lot of damage. - Drew Conway

Three classification problems with three techs

  • classification (QSAR) using scikit-learn in Python
  • next gen sequencing (standard R processing pipeline) RNA-seq
  • diagnostic image analysis with ConvNets in Python using Theano

The emphasis here is on finding a common understanding of the vocabulary between life scientists and analysts; things like pipelines, dataframes and representations.

Classification in Machine Learning

A classic supervised learning problem.

Here is an excellent visualization of the process

1 Choose a representation

2 Train a classifier

  • Holdouts
  • Cross-validation
  • Classifier selection
  • Parameter search

3 Make predictions

4 Evaluation metrics

Quantitaive Structure Activity Relationship for Predictive Toxicology

Check out some chemicals

In [1]:
from rdkit import Chem
from rdkit.Chem import Draw
%matplotlib inline
In [2]:
m3 = Chem.MolFromSmiles('O=C1OC2=C(C=C1)C1=C(C=CCO1)C=C2')
fig3 = Draw.MolToMPL(m3)
In [4]:
smiles = ("O=C(NCc1cc(OC)c(O)cc1)CCCC/C=C/C(C)C", "CC(C)CCCCCC(=O)NCC1=CC(=C(C=C1)O)OC", "c1(C(=O)O)cc(OC)c(O)cc1")
mols = [Chem.MolFromSmiles(x) for x in smiles]
Draw.MolsToGridImage(mols)
Out[4]:

Next Gen Sequencing - RNA-Seq

NGS Woodchipper

Dealing with 30X genome sized datasets initially

  • lots of interesting statistics work around:
    • multiple hypothesis testing
    • assembly and annotation of nucleic acid sequences
    • error correction
    • not machine learning

Comparing RNA expression levels takes this from a big data problem back to another simple classification problem

R Classification exercise in Jupyter notebook

Diagnostic Image Analysis

Diabetic Retinopathy Competition recently from Kaggle.

The top 10 all used Convolutional Neural Nets

The winners in their own words

First place winner Ben Graham

Fourth place winners Julian De Wit & Daniel Hammack

Jeffrey De Fauw most lucid explination yet.

hotak

Ramon Peres got a PLOS publication out of his entry.

Introduce deep learning

Please do not get all Strong AI on me

Lends itself to parallel computation. GPUs are usefull for this.

Picture of simple net

Picture of architecture

Example of python code with Theano

New opportunities come from tying together multiple models.

Tools for speaking the same language

Use AWS free tier

Groundhog Day for computing

yes, it's slow, but it's free

scripts coming to github

Thanks for listening. Next steps.

  • For the hackers

    • Stay away from anything making claims that would involve the FDA or HIPAA or humans in general.
    • There is enough going on in your yard or compost pile or maybe even your fridge.
    • Check out TriDIYbio

  • For the employed

    • run a Software or Data Carpentry workshop within your school or research org.

  • For the enthusiast

    • Help me start the Research Triangle BioCoders
    • Try to find public biological datasets and publish better tutorials and example analyses.
In [ ]:
!jupyter nbconvert --to slides MLforLS.ipynb --post serve

[NbConvertApp] Converting notebook MLforLS.ipynb to slides
[NbConvertApp] Writing 202636 bytes to MLforLS.slides.html
[NbConvertApp] Redirecting reveal.js requests to https://cdn.jsdelivr.net/reveal.js/2.6.2
Serving your slides at http://127.0.0.1:8000/MLforLS.slides.html
Use Control-C to stop this server
Created new window in existing browser session.
WARNING:tornado.access:404 GET /custom.css (127.0.0.1) 0.79ms
WARNING:tornado.access:404 GET /favicon.ico (127.0.0.1) 0.47ms

Podcasts Talking Machines podcast starting after 10 minutes a16z - breathless, but not all hype