CS Department Wins Top Prizes at COSE Science Fair

The following CS department projects won awards at recent COSE Science Fair:

 

First place for graduate project in physical sciences:

MACE: A Simple Algorithm for Lossless Compression of Microarray Images without Spot Segmentation

 

Team:
Robert Bierman - bierman@sfsu.edu
Dr. Rahul Singh - rsingh@cs.sfsu.edu

 

Faculty Advisor:
Dr. Rahul Singh - rsingh@cs.sfsu.edu

 

The widespread adoption of microarray technology coupled with the large volume of image-based data generated per experiment underlines the importance of microarray image compression. We present a simple and practical compression technique that resolves a key problem at the state-of-the-art, namely the dependence of compression methods on the complex and error-prone step of spot detection. In our method, the highly skewed intensity distribution of microarray images is utilized in combination with bit-plane decomposition to yield significant compression without either the need for spot demarcation or the risk of loosing biologically valuable data. Moreover, the method can be applied to microarrays with arbitrary spot layouts since it does not require spot gridding. Experiments on a wide class of microarray images, including a comprehensive set of benchmarks, clearly show the efficacy of the proposed algorithm.

 

Third place for graduate project in physical sciences:

Prediction of protein-ligand interactions using geometric models of active sites

 

Joanna Lipinski - lipinski@sfsu.edu
Carla Webster - cweb@sfsu.edu
Dr. Rahul Singh - rsingh@cs.sfsu.edu

 

Faculty Advisor:
Dr. Rahul Singh - rsingh@cs.sfsu.edu

 

Drug development is a very time consuming and costly process. One of the biggest challenges is predicting whether a drug, a ligand, will bind to the targeted protein. One of the key requirements for binding is that the ligand has to be "anchored" in a small "pocket" on the surface of the protein, called the active site, through hydrogen or covalent bonds, or interactions such as hydrophobic and hydrophilic ones. These stabilizing bonds and interactions can be formed only with the nearest atoms and only at very specific locations. For this reason, the spatial distribution of atoms in the active site is thought to play a key role in ligand-protein binding.

 

Based on this observation, we hypothesize that geometric properties of an active site might be good predictors of protein's affinity for a ligand. And hence, we are proposing to use a geometric hashing based approach to model known active sites and to use these models to predict ligand-protein interactions of novel proteins.

 

Honorable mention in graduate category, physical sciences:

Novel Bayesian Network Evaluation Algorithm for Discovering Gene Regulatory Pathways from Microarray Data

 

Students:
Arturo Flores (Computer Science/Bionformatics) aflores@sfsu.edu
Lala Motlhabi(Biology/Bioinformatics) lalam@hotmail.com
Rocco Varela(Computer Science/Bioinformatics) rocco408@yahoo.com
Elinor Velasquez(Computer Science/Bioinformatics) velasque@sfsu.edu

 

Faculty Advisors:
Frank Bayliss, (Biology) fbayl@sfsu.edu
Hui Yang, (Computer Science) huiyang@cs.sfsu.edu
Ilmi Yoon, (Computer Science) yoon@cs.sfsu.edu

 

Staff Advisor:
Mike Wong, (Center for Computing for Life Sciences) mikewong@sfsu.edu

 

Bayesian statistical thinking has been considered as a revolutionary force within genetics and bioinformatics. A Bayesian network is a graphical model that encodes probabilistic causal relationships among a set of variables. Bayesian networks are important for integrating biological data and for inferring cellular networks and pathways, and therefore are considered as a promising model for inferring gene regulatory pathways. Comparative approaches in bioinformatics include supervised learning (mining) of the data by clustering techniques. The method of Bayesian networks improves upon clustering techniques by providing a probabilistic and causal estimate of the relationships between the variables of interest. The state-of-the-art in Bayesian networks assumes a uniform distribution and binary discretization of the data, handles a dozen data points and does not make estimates as to the robustness of the resulting network. The state-of-the-art uses a scoring measure that takes into account a single parameter for inferring a regulatory pathway. Our results avoid overfitting of the data, handle large data sets, such as microarray data, permit a data-driven distribution, and are effective for extracting useful information from noisy data. Additionally our work allows for a more expressive data discretization, makes estimates as to the robustness of the resulting network and provides a many-parameter, time-dependent scoring criterion by which to learn the pathways. Our approach confirms known and predicts novel gene regulatory pathways in an organism, such as the budding yeast, Saccharomyces cerevisiae.

 

Honorable mention in undergraduate category, physical sciences:

Vision-Based Detection of Visually Dissimilar Objects

 

Taeil Goh - imbb@sfsu.edu
Ryan West - rawest@sfsu.edu
Dr. Kaz Okada - kazokada@sfsu.edu

 

Faculty Advisor:
Dr. Kaz Okada - kazokada@sfsu.edu

 

This research develops a novel and intuitive computer vision system which detect hard-to-find objects by automatically learning their geometrical contexts. The developed framework has been successfully applied to detecting facial parts towards realizing more reliable face and object recognition system.