Detecting Salient Changes in Gene Profiles

Detecting Salient Changes in Gene Profiles

Detecting Salient Changes in Gene Profiles

Sohei Okamoto University of Nevada, Reno

Dr. Tanveer Syeda-Mahmood IBM Almaden Research Center

Dr. George Bebis University of Nevada, Reno

Dr. Dwight Egbert University of Nevada, Reno

Sohei Okamoto University of Nevada, Reno

Dr. Tanveer Syeda-Mahmood IBM Almaden Research Center

Dr. George Bebis University of Nevada, Reno

Dr. Dwight Egbert University of Nevada, Reno

OUTLINEOUTLINE

• Motivation

• Approach

• Method

• Difficulty

• Future Work

• Motivation

• Approach

• Method

• Difficulty

• Future Work

MOTIVATIONMOTIVATION

• Salient changes in gene expressions indicate important events such as onset of disease.

• Salient changes in multiple gene profiles are similarities among them

• Salient changes in gene expressions indicate important events such as onset of disease.

• Salient changes in multiple gene profiles are similarities among them

• Salient changes are maxima in vector magnitude function

• Build scale-space bitmap, and find salient change contours

• Choose optimal smoothing scale by objective functions

• Salient changes are maxima in vector magnitude function

• Build scale-space bitmap, and find salient change contours

• Choose optimal smoothing scale by objective functions

APPROACHAPPROACH

• Finding maxima in vector magnitude function

• Salient change is inflection point which has maximum slope

• Finding maxima in vector magnitude function

• Salient change is inflection point which has maximum slope

METHOD – Salient ChangesMETHOD – Salient Changes

Maxima and minima of derivative

Maxima and minima of derivative

Maxima of vector magnitude

Maxima of vector magnitude

Points obtained in input signalPoints obtained in input signal

METHOD – Salient ChangesMETHOD – Salient Changes

METHOD – Scale-SpaceMETHOD – Scale-Space

• Scale-space visualizes how salient changes preserved over increasing scale of smoothing

• Gaussian smoothing with increasing standard deviation

• Scale-space visualizes how salient changes preserved over increasing scale of smoothing

• Gaussian smoothing with increasing standard deviation

METHOD – Scale-SpaceMETHOD – Scale-Space

Result of increasing smoothingResult of increasing smoothing

METHOD – Scale-SpaceMETHOD – Scale-Space

Input signalInput signal

Vector magnitudeVector magnitude

Derivative of vector magnitude

Derivative of vector magnitude

METHOD – Scale-SpaceMETHOD – Scale-Space

• Construct binary representation of vector magnitude derivative for each increasing scale

• Construct binary representation of vector magnitude derivative for each increasing scale

METHOD – Scale-SpaceMETHOD – Scale-Space

• Negative-going zero-crossing contours in scale-space

• Negative-going zero-crossing contours in scale-space

METHOD – Optimal Scale SelectionMETHOD – Optimal Scale Selection

• Select optimal smoothing scale to ignore non-significant salient change with minimum error

• Calculate objective functions and combine

• Select optimal smoothing scale to ignore non-significant salient change with minimum error

• Calculate objective functions and combine

METHOD – Optimal Scale SelectionMETHOD – Optimal Scale Selection

• Roughness – # of salient changes at each scale

• Mean Square Error – Error between input and smoothed signal at each scale:

• Combined objective function – take average of two functions

• Roughness – # of salient changes at each scale

• Mean Square Error – Error between input and smoothed signal at each scale:

• Combined objective function – take average of two functions

METHOD – Optimal Scale SelectionMETHOD – Optimal Scale Selection

• Find crossing point of two objective functions, which is also minimum of combined objective function

• Find crossing point of two objective functions, which is also minimum of combined objective function

METHOD – Optimal Scale SelectionMETHOD – Optimal Scale Selection

• Trace back salient change contours exist at optimal scale to lowest scale

• Trace back salient change contours exist at optimal scale to lowest scale

METHOD – Optimal Scale SelectionMETHOD – Optimal Scale Selection

• Salient changes found• Salient changes found

DIFFICULTYDIFFICULTY

• When there are relatively few salient change contours

• When there are relatively few salient change contours

DIFFICULTYDIFFICULTY

• when one contour branch out to two at some lower scale

• when one contour branch out to two at some lower scale

DIFFICULTYDIFFICULTY

• When time points are few

• Visualization of more than 3-dimensional signal

• When time points are few

• Visualization of more than 3-dimensional signal

FUTURE WORKFUTURE WORK

• Collect large amount of results using this method for validity

• Adding analysis functionality as an event mining tool

• Integration with other data mining tools

• Collect large amount of results using this method for validity

• Adding analysis functionality as an event mining tool

• Integration with other data mining tools

ACKNOWLEDGMENTACKNOWLEDGMENT

• University of Nevada, Reno, Computer Science Department

• UNR, Computer Vision Laboratory

• UNR-CRCD Program in Computer Vision

• IBM Almaden Research Center

• National Science Foundation

• University of Nevada, Reno, Computer Science Department

• UNR, Computer Vision Laboratory

• UNR-CRCD Program in Computer Vision

• IBM Almaden Research Center

• National Science Foundation