cg

changeset 29:5e2e4732b647

.
author bshanks@bshanks.dyndns.org
date Mon Apr 13 03:43:51 2009 -0700 (16 years ago)
parents 01c118d1074b
children 6ec3230fe1dc
files grant.html grant.odt grant.pdf grant.txt
line diff
1.1 --- a/grant.html Mon Apr 13 03:31:42 2009 -0700 1.2 +++ b/grant.html Mon Apr 13 03:43:51 2009 -0700 1.3 @@ -56,7 +56,7 @@ 1.4 a mapping from instances to labels, and the training data consists of a set of 1.5 instances (voxels) for which the labels (subregions) are known. 1.6 Each gene expression level is called a feature, and the selection of which 1.7 - genes to include is called feature selection. Feature selection is one component 1.8 + genes1 to include is called feature selection. Feature selection is one component 1.9 of the task of learning a classifier. Some methods for learning classifiers start 1.10 out with a separate feature selection phase, whereas other methods combine 1.11 feature selection with other aspects of training. 1.12 @@ -66,9 +66,11 @@ 1.13 case, a dynamic procedure may be used in which features are added and sub- 1.14 tracted from the selected set depending on how much they raise the score. Such 1.15 procedures are called “stepwise” or “greedy”. 1.16 +__________________________ 1.17 + 1Strictly speaking, the features are gene expression levels, but we’ll call them genes. 1.18 + 2 1.19 + 1.20 Although the classifier itself may only look at the gene expression data within 1.21 - 2 1.22 - 1.23 each voxel before classifying that voxel, the learning algorithm which constructs 1.24 the classifier may look over the entire dataset. We can categorize score-based 1.25 feature selection methods depending on how the score of calculated. Often 1.26 @@ -81,41 +83,38 @@ 1.27 Key questions when choosing a learning method are: What are the instances? 1.28 What are the features? How are the features chosen? Here are four principles 1.29 that outline our answers to these questions. 1.30 - Principle 1: Combinatorial gene expression 1.31 - Above, we defined an “instance” as the combination of a voxel with the 1.32 - “associated gene expression data”. In our case this refers to the expression level 1.33 - of genes within the voxel, but should we include the expression levels of all 1.34 - genes, or only a few of them? 1.35 - It is too much to hope that every anatomical region of interest will be iden- 1.36 - tified by a single gene. For example, in the cortex, there are some areas which 1.37 - are not clearly delineated by any gene included in the Allen Brain Atlas (ABA) 1.38 - dataset. However, at least some of these areas can be delineated by looking 1.39 - at combinations of genes (an example of an area for which multiple genes are 1.40 - necessary and sufficient is provided in Preliminary Results). 1.41 + Principle 1: Combinatorial gene expression It is too much to hope 1.42 + that every anatomical region of interest will be identified by a single gene. For 1.43 + example, in the cortex, there are some areas which are not clearly delineated 1.44 + by any gene included in the Allen Brain Atlas (ABA) dataset. However, at 1.45 + least some of these areas can be delineated by looking at combinations of genes 1.46 + (an example of an area for which multiple genes are necessary and sufficient 1.47 + is provided in Preliminary Results). Therefore, each instance should contain 1.48 + multiple features (genes). 1.49 Principle 2: Only look at combinations of small numbers of genes 1.50 - When the classifier classifies a voxel, it is only allowed to look at the expres- 1.51 - sion of the genes which have been selected as features. The more data that is 1.52 - available to a classifier, the better that it can do. For example, perhaps there 1.53 - are weak correlations over many genes that add up to a strong signal. So, why 1.54 - not include every gene as a feature? The reason is that we wish to employ 1.55 - the classifier in situations in which it is not feasible to gather data about every 1.56 - gene. For example, if we want to use the expression of marker genes as a trigger 1.57 - for some regionally-targeted intervention, then our intervention must contain a 1.58 - molecular mechanism to check the expression level of each marker gene before 1.59 - it triggers. It is currently infeasible to design a molecular trigger that checks 1.60 - the level of more than a handful of genes. Similarly, if the goal is to develop a 1.61 - procedure to do ISH on tissue samples in order to label their anatomy, then it 1.62 - is infeasible to label more than a few genes. Therefore, we must select only a 1.63 - few genes as features. 1.64 + When the classifier classifies a voxel, it is only allowed to look at the expression of 1.65 + the genes which have been selected as features. The more data that is available 1.66 + to a classifier, the better that it can do. For example, perhaps there are weak 1.67 + correlations over many genes that add up to a strong signal. So, why not include 1.68 + every gene as a feature? The reason is that we wish to employ the classifier in 1.69 + situations in which it is not feasible to gather data about every gene. For 1.70 + example, if we want to use the expression of marker genes as a trigger for some 1.71 + regionally-targeted intervention, then our intervention must contain a molecular 1.72 + mechanism to check the expression level of each marker gene before it triggers. 1.73 + It is currently infeasible to design a molecular trigger that checks the level of 1.74 + more than a handful of genes. Similarly, if the goal is to develop a procedure to 1.75 + do ISH on tissue samples in order to label their anatomy, then it is infeasible 1.76 + to label more than a few genes. Therefore, we must select only a few genes as 1.77 + features. 1.78 Principle 3: Use geometry in feature selection 1.79 When doing feature selection with score-based methods, the simplest thing 1.80 to do would be to score the performance of each voxel by itself and then com- 1.81 bine these scores (pointwise scoring). A more powerful approach is to also use 1.82 information about the geometric relations between each voxel and its neighbors; 1.83 - 3 1.84 - 1.85 this requires non-pointwise, local scoring methods. See Preliminary Results for 1.86 evidence of the complementary nature of pointwise and local scoring methods. 1.87 + 3 1.88 + 1.89 Principle 4: Work in 2-D whenever possible 1.90 There are many anatomical structures which are commonly characterized in 1.91 terms of a two-dimensional manifold. When it is known that the structure that 1.92 @@ -154,12 +153,12 @@ 1.93 special type of clustering task because we have an additional constraint on 1.94 clusters; voxels grouped together into a cluster must be spatially contiguous. 1.95 In Preliminary Results, we show that one can get reasonable results without 1.96 - 4 1.97 - 1.98 enforcing this constraint, however, we plan to compare these results against 1.99 other methods which guarantee contiguous clusters. 1.100 Perhaps the biggest source of continguous clustering algorithms is the field 1.101 of computer vision, which has produced a variety of image segmentation algo- 1.102 + 4 1.103 + 1.104 rithms. Image segmentation is the task of partitioning the pixels in a digital 1.105 image into clusters, usually contiguous clusters. Aim 2 is similar to an image 1.106 segmentation task. There are two main differences; in our task, there are thou- 1.107 @@ -200,17 +199,17 @@ 1.108 Clustering genes rather than voxels 1.109 Although the ultimate goal is to cluster the instances (voxels or pixels), one 1.110 strategy to achieve this goal is to first cluster the features (genes). There are 1.111 - 5 1.112 - 1.113 two ways that clusters of genes could be used. 1.114 Gene clusters could be used as part of dimensionality reduction: rather than 1.115 have one feature for each gene, we could have one reduced feature for each gene 1.116 cluster. 1.117 + 5 1.118 + 1.119 Gene clusters could also be used to directly yield a clustering on instances. 1.120 This is because many genes have an expression pattern which seems to pick 1.121 out a single, spatially continguous subregion. Therefore, it seems likely that an 1.122 anatomically interesting subregion will have multiple genes which each individ- 1.123 - ually pick it out1. This suggests the following procedure: cluster together genes 1.124 + ually pick it out2. This suggests the following procedure: cluster together genes 1.125 which pick out similar subregions, and then to use the more popular common 1.126 subregions as the final clusters. In the Preliminary Data we show that a num- 1.127 ber of anatomically recognized cortical regions, as well as some “superregions” 1.128 @@ -240,17 +239,17 @@ 1.129 Significance 1.130 The method developed in aim (1) will be applied to each cortical area to find 1.131 a set of marker genes such that the combinatorial expression pattern of those 1.132 + genes uniquely picks out the target area. Finding marker genes will be useful 1.133 + for drug discovery as well as for experimentation because marker genes can be 1.134 + used to design interventions which selectively target individual cortical areas. 1.135 __________________________ 1.136 - 1This would seem to contradict our finding in aim 1 that some cortical areas are combina- 1.137 + 2This would seem to contradict our finding in aim 1 that some cortical areas are combina- 1.138 torially coded by multiple genes. However, it is possible that the currently accepted cortical 1.139 maps divide the cortex into subregions which are unnatural from the point of view of gene 1.140 expression; perhaps there is some other way to map the cortex for which each subregion can 1.141 be identified by single genes. 1.142 6 1.143 1.144 - genes uniquely picks out the target area. Finding marker genes will be useful 1.145 - for drug discovery as well as for experimentation because marker genes can be 1.146 - used to design interventions which selectively target individual cortical areas. 1.147 The application of the marker gene finding algorithm to the cortex will 1.148 also support the development of new neuroanatomical methods. In addition to 1.149 finding markers for each individual cortical areas, we will find a small panel 1.150 @@ -292,18 +291,18 @@ 1.151 anatomy through computational methods. 1.152 [?] describes an analysis of the anatomy of the hippocampus using the ABA 1.153 dataset. In addition to manual analysis, two clustering methods were employed, 1.154 - 7 1.155 - 1.156 a modified Non-negative Matrix Factorization (NNMF), and a hierarchial bifur- 1.157 cation clustering scheme based on correlation as the similarity score. The paper 1.158 yielded impressive results, proving the usefulness of such research. We have run 1.159 + 7 1.160 + 1.161 NNMF on the cortical dataset and while the results are promising (see Prelim- 1.162 - inary Data), we think that it will be possible to find a better method2 (we also 1.163 + inary Data), we think that it will be possible to find a better method3 (we also 1.164 think that more automation of the parts that this paper’s authors did manually 1.165 will be possible). 1.166 and [?] describes AGEA. todo 1.167 __________________________ 1.168 - 2We ran “vanilla” NNMF, whereas the paper under discussion used a modified method. 1.169 + 3We ran “vanilla” NNMF, whereas the paper under discussion used a modified method. 1.170 Their main modification consisted of adding a soft spatial contiguity constraint. However, 1.171 on our dataset, NNMF naturally produced spatially contiguous clusters, so no additional 1.172 constraint was needed. The paper under discussion mentions that they also tried a hierarchial 1.173 @@ -320,14 +319,14 @@ 1.174 delineate some cortical areas 1.175 Here we give an example of a cortical area which is not marked by any 1.176 single gene, but which can be identified combinatorially. according to logistic 1.177 - regression, gene wwc13 is the best fit single gene for predicting whether or not a 1.178 + regression, gene wwc14 is the best fit single gene for predicting whether or not a 1.179 pixel on the cortical surface belongs to the motor area (area MO). The upper-left 1.180 picture in Figure shows wwc1’s spatial expression pattern over the cortex. The 1.181 lower-right boundary of MO is represented reasonably well by this gene, however 1.182 the gene overshoots the upper-left boundary. This flattened 2-D representation 1.183 does not show it, but the area corresponding to the overshoot is the medial 1.184 surface of the cortex. MO is only found on the lateral surface (todo). 1.185 - Gnee mtif24 is shown in figure the upper-right of Fig. . Mtif2 captures MO’s 1.186 + Gnee mtif25 is shown in figure the upper-right of Fig. . Mtif2 captures MO’s 1.187 upper-left boundary, but not its lower-right boundary. Mtif2 does not express 1.188 very much on the medial surface. By adding together the values at each pixel 1.189 in these two figures, we get the lower-left of Figure . This combination captures 1.190 @@ -339,17 +338,17 @@ 1.191 information 1.192 To show that local geometry can provide useful information that cannot be 1.193 detected via pointwise analyses, consider Fig. . The top row of Fig. displays the 1.194 - 3 genes which most match area AUD, according to a pointwise method5. The 1.195 + 3 genes which most match area AUD, according to a pointwise method6. The 1.196 bottom row displays the 3 genes which most match AUD according to a method 1.197 - which considers local geometry6 The pointwise method in the top row identifies 1.198 + which considers local geometry7 The pointwise method in the top row identifies 1.199 __________________________ 1.200 - 3“WW, C2 and coiled-coil domain containing 1”; EntrezGene ID 211652 1.201 - 4“mitochondrial translational initiation factor 2”; EntrezGene ID 76784 1.202 - 5For each gene, a logistic regression in which the response variable was whether or not a 1.203 + 4“WW, C2 and coiled-coil domain containing 1”; EntrezGene ID 211652 1.204 + 5“mitochondrial translational initiation factor 2”; EntrezGene ID 76784 1.205 + 6For each gene, a logistic regression in which the response variable was whether or not a 1.206 surface pixel was within area AUD, and the predictor variable was the value of the expression 1.207 of the gene underneath that pixel. The resulting scores were used to rank the genes in terms 1.208 of how well they predict area AUD. 1.209 - 6For each gene the gradient similarity (see section ??) between (a) a map of the expression 1.210 + 7For each gene the gradient similarity (see section ??) between (a) a map of the expression 1.211 of each gene on the cortical surface and (b) the shape of area AUD, was calculated, and this 1.212 was used to rank the genes. 1.213 9 1.214 @@ -389,11 +388,11 @@ 1.215 SVM on all genes at once 1.216 In order to see how well one can do when looking at all genes at once, we 1.217 ran a support vector machine to classify cortical surface pixels based on their 1.218 - gene expression profiles. We achieved classification accuracy of about 81%7. 1.219 + gene expression profiles. We achieved classification accuracy of about 81%8. 1.220 As noted above, however, a classifier that looks at all the genes at once isn’t 1.221 practically useful. 1.222 ____________ 1.223 - 75-fold cross-validation. 1.224 + 85-fold cross-validation. 1.225 11 1.226 1.227 The requirement to find combinations of only a small number of genes limits 1.228 @@ -489,7 +488,7 @@ 1.229 app2 has examples of genetic targeting to specific anatomical regions 1.230 — 1.231 note: 1.232 - do we need to cite: no known markers? impressive results? 1.233 + do we need to cite: no known markers, impressive results? 1.234 14 1.235 1.236
2.1 Binary file grant.odt has changed
3.1 Binary file grant.pdf has changed
4.1 --- a/grant.txt Mon Apr 13 03:31:42 2009 -0700 4.2 +++ b/grant.txt Mon Apr 13 03:43:51 2009 -0700 4.3 @@ -31,7 +31,7 @@ 4.4 4.5 In the machine learning literature, this sort of procedure may be thought of as a __supervised learning task__, defined as a task in which the goal is to learn a mapping from instances to labels, and the training data consists of a set of instances (voxels) for which the labels (subregions) are known. 4.6 4.7 -Each gene expression level is called a __feature__, and the selection of which genes to include is called __feature selection__. Feature selection is one component of the task of learning a classifier. Some methods for learning classifiers start out with a separate feature selection phase, whereas other methods combine feature selection with other aspects of training. 4.8 +Each gene expression level is called a __feature__, and the selection of which genes\footnote{Strictly speaking, the features are gene expression levels, but we'll call them genes.} to include is called __feature selection__. Feature selection is one component of the task of learning a classifier. Some methods for learning classifiers start out with a separate feature selection phase, whereas other methods combine feature selection with other aspects of training. 4.9 4.10 One class of feature selection methods assigns some sort of score to each candidate gene. The top-ranked genes are then chosen. Some scoring measures can assign a score to a set of selected genes, not just to a single gene; in this case, a dynamic procedure may be used in which features are added and subtracted from the selected set depending on how much they raise the score. Such procedures are called "stepwise" or "greedy". 4.11 4.12 @@ -41,14 +41,10 @@ 4.13 4.14 4.15 \vspace{0.3cm}**Principle 1: Combinatorial gene expression** 4.16 - 4.17 -Above, we defined an "instance" as the combination of a voxel with the "associated gene expression data". In our case this refers to the expression level of genes within the voxel, but should we include the expression levels of all genes, or only a few of them? 4.18 - 4.19 -It is too much to hope that every anatomical region of interest will be identified by a single gene. For example, in the cortex, there are some areas which are not clearly delineated by any gene included in the Allen Brain Atlas (ABA) dataset. However, at least some of these areas can be delineated by looking at combinations of genes (an example of an area for which multiple genes are necessary and sufficient is provided in Preliminary Results). 4.20 +It is too much to hope that every anatomical region of interest will be identified by a single gene. For example, in the cortex, there are some areas which are not clearly delineated by any gene included in the Allen Brain Atlas (ABA) dataset. However, at least some of these areas can be delineated by looking at combinations of genes (an example of an area for which multiple genes are necessary and sufficient is provided in Preliminary Results). Therefore, each instance should contain multiple features (genes). 4.21 4.22 4.23 \vspace{0.3cm}**Principle 2: Only look at combinations of small numbers of genes** 4.24 - 4.25 When the classifier classifies a voxel, it is only allowed to look at the expression of the genes which have been selected as features. The more data that is available to a classifier, the better that it can do. For example, perhaps there are weak correlations over many genes that add up to a strong signal. So, why not include every gene as a feature? The reason is that we wish to employ the classifier in situations in which it is not feasible to gather data about every gene. For example, if we want to use the expression of marker genes as a trigger for some regionally-targeted intervention, then our intervention must contain a molecular mechanism to check the expression level of each marker gene before it triggers. It is currently infeasible to design a molecular trigger that checks the level of more than a handful of genes. Similarly, if the goal is to develop a procedure to do ISH on tissue samples in order to label their anatomy, then it is infeasible to label more than a few genes. Therefore, we must select only a few genes as features. 4.26 4.27 4.28 @@ -317,4 +313,4 @@ 4.29 4.30 note: 4.31 4.32 -do we need to cite: no known markers? impressive results? 4.33 +do we need to cite: no known markers, impressive results?