Adam Wible
Junior Independent Work
Advisor: Olga Troyanskaya
Title: Machine Learning for Gene Function Prediction


The technology enabling biologists to map genomes has revolutionized the field.  New 
methods of genetic analysis, known as high throughput methods, can generate large 
quantities of genetic data, but no one method is perfect.  Different methods suffer from 
problems with specificity, selectivity, or both.  Combining these methods in a 
probabilistic Bayesian framework can increase the accuracy of the analysis of this data.  
Magic, a program written by Olga Troyanskaya and her associates at Stanford, does just 
this.  The goal of this project is to use machine learning techniques to meet or exceed the 
current level of accuracy, without requiring surveys of experts in the field to form the 
initial weights for the network.  This algorithm will work off of a subset of genes already 
annotated with a high confidence level, and attempt to predict the classifications for the 
rest of the set.  This project will be successful if the percentage of correct classifications 
is comparable to the percentage attained by the network constructed by experts.  There is 
a risk that the data will be overfitted, and it will be difficult to apply the method to other 
genomes (some of the methods are specific to the baker's yeast genome, which we'll be 
using).