1. Machine Learning in Computational Biology to Accelerate High-Throughput Protein Expression
Anand Sastry, Jonathan Monk, Hanna Tegel, Mathias Uhlén,
Bernhard O. Palsson, Johan Rockberg, and Elizabeth Brunk
Presentation by Alex Joens

2. Background
The Human Protein Atlas (HPA), started in 2001, is a project aimed at exploring the entire human proteome
Proteome: The entire set, or composition, of proteins in a specific “thing” in a certain environment
Ex, the cell proteome when exposed to a certain hormone
You can also analyze an entire organism’s proteome
Specifically, attempts to store information on the expression and spatial location of all protein-coding genes
By 2017, its knowledge base encompasses 84% of the human proteome
Adding to this base requires an insane number of recombinant protein fragments to be expressed and sampled, which can be “difficult”

3. Background
Current process is to inject fragments of human DNA into E. Coli bacteria, and attempt to have it express large amount of heterologous protein fragments
Heterologous expression: expression of a gene or gene fragment in a host organism which does not naturally have that gene or gene fragment
Though not complete proteins, these fragments are still useful
This is even harder than it sounds, because not all fragments show high expression, which is required for analysis
Current approaches heavily rely on trial-and-error
New idea: use machine learning techniques to be able to predict a protein’s expression, based on the protein structure itself and other factors

4. Random Graphic

5. Machine Learning
Machine learning “computers the ability to learn without being explicitly programmed.” -- Arthur Samuel, 1959
These algorithms are designed to look for patterns in data, and can use these patterns to make predictions
Algorithms can have different performance on different problems, so picking algorithms is not an exact science!
For this paper, four algorithms were selected:
Linear Regression
Random Forest Decision Trees
Support Vector Machines (SVMs)
Neural Networks

6. Converting to Machine Learning Problem

7. Random Graphic

8. Results: Expression

9. Results: Expression

10. Results: Expression

11. Results: Solubility

12. Results: PrEST Selection

13. Results
Protein expression prediction rates is on par with the state of the art algorithms
Surprisingly, mRNA features had very little predictive power for expression
Protein solubility prediction is significantly better than previous state of the art algorithms
Using the learned classifiers reduced the number of needed experiments by 39%

14. Additional Results
The authors developed a simple and very easy to use framework to implement their experiments, all freely available online as IPython Notebook
Designed to work with any “omics” data, so can be applied to a wider variety of bioinformatics problems