Quality Expression Profiling .


41 views
Uploaded on:
Description
Great Microarray Studies Have Clear Objectives. Class Comparison (quality finding)Find qualities whose expression varies among foreordained classesClass PredictionPrediction of foreordained class utilizing data from quality expression profileResponse versus no responseClass DiscoveryDiscover groups of examples having comparative expression profilesDiscover bunches of qualities having comparable expression profile
Transcripts
Slide 1

Quality Expression Profiling

Slide 2

Good Microarray Studies Have Clear Objectives Class Comparison (quality discovering) Find qualities whose expression varies among foreordained classes Class Prediction of foreordained class utilizing data from quality expression profile Response versus no reaction Class Discovery Discover groups of examples having comparative expression profiles Discover bunches of qualities having comparable expression profiles

Slide 3

Class Comparison and Class Prediction Not grouping issues Supervised strategies

Slide 4

Levels of Replication Technical repeats RNA test isolated into numerous aliquots Biological recreates Multiple subjects Multiple creatures Replication of the tissue culture try

Slide 5

Comparing classes or creating classifiers requires free organic duplicates. The force of factual techniques for microarray information relies on upon the quantity of organic imitates

Slide 6

Microarray Platforms Single mark clusters Affymetrix GeneChips Dual name exhibits Common reference outline Other plans

Slide 7

Common Reference Design A 1 A 2 B 1 B 2 RED R GREEN Array 1 Array 2 Array 3 Array 4 An i = i th example from class A B i = i th example from class B R = aliquot from reference pool

Slide 8

The reference for the most part serves to control variety in the measure of comparing spots on various clusters and variety in test dispersion over the slide. The reference gives a relative measure of expression for a given quality in a given specimen that is less factor than an outright measure. The reference is not the protest of correlation.

Slide 10

Dye swap specialized imitates of a similar two rna tests are a bit much with the regular reference plan For two-mark coordinate correlation outlines for looking at two classes, color predisposition is of concern and color swaps might be required.

Slide 11

Class Comparison Gene Finding

Slide 13

Controlling for Multiple Comparisons Bonferroni sort systems control the likelihood of making any false positive mistakes Overly traditionalist for the setting of DNA microarray ponders

Slide 14

Simple Procedures Control expected the quantity of false disclosures by testing every quality for differential expression between classes utilizing a stringent noteworthiness level expected number of false revelations in testing G qualities with criticalness limit p* is G p* e.g. To farthest point of 10 false revelations in 10,000 correlations, lead each test at p<0.001 level Control FDR Expected extent of false disclosures among the qualities announced differentially communicated Benjamini-Hochberg system FDR = G p*/#(p p*)

Slide 15

Additional Procedures Multivariate change tests Korn et al. Detail Med 26:4428,2007 SAM - Significance Analysis of Microarrays Advantages without distribution regardless of the possibility that they utilize t measurements Preserve/misuse relationship among tests by permuting each profile as a unit More compelling than univariate change tests particularly with predetermined number of tests

Slide 16

Randomized Variance t-test Wright G.W. what\'s more, Simon R. Bioinformatics19:2448-2455,2003 Pr(  - 2 =x) = x a-1 exp(- x/b)/(a)b a

Slide 18

Class Prediction

Slide 19

Components of Class Prediction Feature (quality) choice Which qualities will be incorporated into the model Select model sort E.g. Inclining direct discriminant examination, Nearest-Neighbor, … Fitting parameters (relapse coefficients) for model Selecting benefit of tuning parameters

Slide 20

Feature Selection Genes that are differentially communicated among the classes at an importance level  (e.g. 0.01) The  level is chosen just to control the quantity of qualities in the model For class examination false revelation rate is essential For class forecast, prescient exactness is vital

Slide 21

Complex Gene Selection Small subset of qualities which together give most precise expectations Genetic calculations Little proof that mind boggling highlight determination is valuable in microarray issues

Slide 22

Linear Classifiers for Two Classes

Slide 23

Linear Classifiers for Two Classes Fisher direct discriminant investigation Requires evaluating relationships among all qualities chose for model Diagonal straight discriminant investigation (DLDA) accept elements are uncorrelated Compound covariate indicator (Radmacher) and Golub\'s strategy are like DLDA in that they can be seen as weighted voting of univariate classifiers

Slide 24

Linear Classifiers for Two Classes Compound covariate indicator Instead of for DLDA

Slide 25

Linear Classifiers for Two Classes Support vector machines with inward item piece are direct classifiers with weights resolved to isolate the classes with a hyperplane that limits the length of the weight vector

Slide 26

Support Vector Machine

Slide 27

Other Linear Methods Perceptrons Principal segment relapse Supervised main segment relapse Partial slightest squares Stepwise calculated relapse

Slide 28

Other Simple Methods Nearest neighbor grouping Nearest k-neighbors Nearest centroid arrangement Shrunken centroid order

Slide 29

Nearest Neighbor Classifier To characterize a specimen in the approval set, decide it\'s closest neighbor in the preparation set; i.e. which test in the preparation set is its quality expression profile is most like. Comparability measure utilized depends on qualities chose as being univariately differentially communicated between the classes Correlation similitude or Euclidean separation for the most part utilized Classify the example as being in an indistinguishable class from it\'s closest neighbor in the preparation set

Slide 30

Nearest Centroid Classifier For a preparation set of information, select the qualities that are educational for recognizing the classes Compute the normal expression profile ( centroid ) of the enlightening qualities in each class Classify a specimen in the approval set in view of which centroid in the preparation set it\'s quality expression profile is most like.

Slide 31

When p>>n The Linear Model is Too Complex It is constantly conceivable to locate an arrangement of components and a weight vector for which the order blunder on the preparation set is zero. It might be unlikely to expect that there is adequate information accessible to prepare more unpredictable non-direct classifiers

Slide 32

Other Methods Top-scoring sets Claim that it gives exact forecast with few sets since sets of qualities are chosen to function admirably together Random Forest Very prevalent in machine learning group Complex classifier

Slide 33

Comparative reviews demonstrate that straight strategies and closest neighbor sort techniques regularly fill in also or superior to more mind boggling strategies for microarray issues since they stay away from over-fitting the information.

Slide 37

Evaluating a Classifier Fit of a model to similar information used to create it is no confirmation of forecast precision for free information Goodness of fit versus expectation exactness

Slide 39

Class Prediction A classifier is not an arrangement of qualities Testing whether examination of autonomous information brings about determination of a similar arrangement of qualities is not a suitable trial of prescient precision of a classifier The order of autonomous information ought to be precise. There are many reasons why the classifier might be unsteady. The grouping ought not be insecure.

Slide 42

Hazard proportions and factual centrality levels are not fitting measures of expectation precision A peril proportion is a measure of affiliation Large estimations of HR may relate to little change in forecast exactness Kaplan-Meier bends for anticipated hazard assembles inside strata characterized by standard prognostic factors give more data about change in forecast precision Time subordinate ROC bends inside strata characterized by standard prognostic components can likewise be valuable

Slide 44

Time-Dependent ROC Curve M(b) = twofold marker in view of edge b PPV = prob{S  T|M(b)=1} NPV = prob{S < T|M(b)=0} ROC Curve is Sensitivity versus 1-Specificity as an element of b Sensitivity = prob{M(b)=1|S  T} Specificity = prob{M(b)=0|S<T}

Slide 46

Validation of a Predictor Internal approval Re-substitution appraise Very one-sided Split-example approval Cross-approval Independent information approval

Slide 48

Split-Sample Evaluation Split your information into a preparation set and a test set Randomly (e.g. 2:1) By focus Training-set Used to choose highlights, select model sort, decide parameters and cut-off edges Test-set Withheld until a solitary model is completely indicated utilizing the preparation set. Completely determined model is connected to the expression profiles in the test-set to foresee class marks. Number of mistakes is excludeed

Slide 49

Leave-one Cross Validation Leave-one-out cross-approval reproduces the procedure of independently building up a model on one arrangement of information and foreseeing for a test set of information not utilized as a part of building up the model

Slide 50

Leave-one-out Cross Validation Omit test 1 Develop multivariate classifier without any preparation on preparing set with test 1 discarded Predict class for test 1 and record whether expectation is right

Slide 51

Leave-one-out Cross Validation Repeat examination for preparing sets with each single specimen overlooked each one in turn e = number of misclassifications dictated by cross-approval Subdivide e for estimation of affectability and specificity

Slide 52

Cross approval is just legitimate if the test set is not utilized at all in the advancement of the model. Utilizing the entire arrangement of tests to choose qualities abuses this presumption and nullifies cross-approval. With legitimate cross-approval, the model must be created without any preparation for each forget one preparing set. This implies include se

Recommended
View more...