Team: 98
School: Los Alamos High
Area of Science: Computational Biology
Proposal:
Rationale:
Protein-protein interaction(PPI) networks are useful data tables when inferring functional annotations. However, transforming this data into a function is a difficult problem. Here I use neural networks, trained to compress and reconstruct proteins, to construct lower dimensional representations of protein data for function prediction. Because PPI occur in every biological process, knowing protein function is elementary for understanding biology further. By knowing protein function one could examine a pathogens gene to determine its weakness, or discover human genes linking to cancer.
Research Questions:
Can we infer the function of other proteins based on their interaction with certain proteins?
Can we identify protein groups that perform larger functions which would be a good target for drug intervention?
Hypothesis:
A neural network will allow for accurate predictions of protein function.
Engineering Goals:
Build a working neural net which translates PPI data to labeled proteins.
Design system for high throughput on GPU, and good resiliency to noisy data.
Procedure:
Program a Neural Network in tensorflow which does dimensional compression and reconstruction.
Save data to tensorboard(visualization tool), to converge on a well working model.
Classify functions using the reduced dimensional tensor features
Visualize feature vectors by principal component analysis and look for possible interpretations
Data Analysis:
Network prediction accuracy: recall and precision
Visualization and dimensional reduction: PCA of input vectors and
tensors via tensorboard.
Mentors: Vladimir Gligorijevic, Richard A Bonneau
Team Members:
Sponsoring Teacher: NA