Daisy Z. Wang -- Machine Learning for Bio-informatics

Constructing Regulatory Network for Yeast

Background

One of the leading challenges in contemporary biology is to discover and understand the regulation mechanisms of genes. Biological experiments for this purpose are typically costly and time consuming. As an example, it usually takes a year for biologist to verify a regulatory pathway between two genes and one knockout experiment to find the effect of deleting one gene from DNA is 100 dollars, but there are 6000 genes in the even simplest organism - yeast. On the other hand, a rich and diverse collection of high-throughput one biological data sources is currently available for elucidating transcriptional regulatory mechanisms from, for example, expression microarrays experiments. Thus, the corresponding challenge in the biocomputing sphere is (I) how to make use of existing data to infer documented and un-documented protein-protein, protein-gene causal relationships, (II) to provide biologist with hints on designing new experiments.

Project Objective

The main objective of this project is to solve challenge (I), as described above, for yeast. Because there only have been very few attempts to address the challenge so far, there is no fixed algorithms we can use. Our approach is based on experience from previous studies, and can be divided into several steps.

(1) Data Pre-processing: solving naming conventiona and normalization for different datasets

(2) Representation: using Regulatory Network

(3) Inference: clustering and combining different networks built from different datasets

(4) Verification: checking knock-out experiment result or annotated knowledge

(5) Demonstration: developing a visualization tool for resulting regulatory networks

Possible Extension

Because datasets are incomplete, we will end up with a network, which has multiple sets of plausible "functional annotations" (e.g causal directions, increase/inhibit) for some interactions, requiring further experiments to refine it. In order to address challenge (II): to provide biologist with hints to design new, information-rich experiments, which cast perturbance to the system to make the predicted responses of alternative interpretations of the network as different as possible, we will seek to find novel ranking algorithms for possible experiments.

Documents

On-Going Work

Papers
- Nir Friedman, Inferring Cellular Networks Using Probabilistic Graphical Models, Science VOL 303, 6 Feb. 2004
- Kevin P. Murphy, An introduction to graphical models, 10 May 2001
- C-H. Yeang, T. Jaakkola, Physical Network Models and Multi-source Data Integration
Datasets Collection
- 106 protein data
- knock-out data
Preprocessed Data (in matlab)
- knock-out data with preprocessing script