Overview of the GINI approach

GINI first extracts the gene expression pattern from each image using a computer version driven image analysis pipeline [15]. These expression patterns are spatially aligned and normalized to enable spatial comparison of gene expression across multiple images. Next, the expression patterns are represented by suitable standardized features through a process called “triangulation”, followed by feature normalization and selection. Since each gene may have a different number of images in the data, each gene can now be represented by a “bag” or a set of feature vectors - one feature vector per image. Thus, our task is to estimate the gene network, given bags of images per gene (Figure 1). We cast the problem of estimating a gene interaction network as the task of estimating the graph structure of a Markov random field (MRF) over the genes. The underlying graph encodes conditional independence assumptions between the genes, that is, two genes are said to not interact in the network if their gene expressions are conditionally independent of each other, conditioned on the expression of all other genes in the network. We employ multi-instance kernel technique using different order statistics to compute similarity between bags of images. We then estimate a sparse network of gene interactions by modeling the data as a multi-variate multi-attribute Gaussian, and estimating the sparse inverse covariance matrix of the model. A schematic diagram of the system pipeline can be seen in Figure 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. GINI schematic.

The schematic shows an outline of the overall system to reverse engineer gene networks from ISH data. Sample output of each step is shown on top of the box corresponding to that step.

https://doi.org/10.1371/journal.pcbi.1003227.g002

Contributions

GINI is a bioimage informatics system based on a computer vision pipeline for ISH micro-image processing and a statistical learning algorithm for network inference. The main contributions of this work are summarized below.

First, the image analytic pipeline used by GINI offers a rigorous and universal approach to extract a standardized representation of spatial patterns of gene expressions. Comparing to the popular SIFT features [24], which is based on detecting interest points with heavy assumptions on object shape, texture, and other physical properties originally meant for natural objects, our approach is more suitable for ISH staining in Drosophila embryos which do not resemble natural objects and require preservation of overall spatial shape and overall intensity information in a canonical way for intra-gene normalization and inter-gene comparison.

Second, GINI infers a network that enjoys the Markov network property: it gives globally consistent conditional-independency interpretation for every edge, and therefore is biologically more meaningful. It is known that marginal correlation (as often used in estimating an ad hoc network), which is computed for every gene-pair in isolation (i.e., ignoring all other genes in the system), confounds direct and indirect dependencies, and could result in a clique-like dense graph or subgraph among genes that are not directly dependent, but have a long-distance interaction. Studying conditional independencies in a network allows us to predict interactions between a pair of genes in the context of other genes, allowing a distinction to be made between direct and indirect relationships between the genes, and reducing false positives.

Third, our formulation based on Gaussian Markov random field and multi-instance kernel for the GINI network is convex, hence the globally optimal estimator of the network is computed, no approximations are involved. Furthermore, under suitable conditions, our graphical model learning algorithm is sparsistent, i.e. as the amount of available data increases, the algorithm is statistically guaranteed to predict the correct interactions between the genes. While Bach et.al. [27] have previously proposed learning the structure of graphical models from data using Mercer kernels, their approach is based on a non-convex local greedy search to find edges in the graph. Our approach represents the first work that uses Mercer kernels and Gaussian Graphical Models to predict kernelized graphical models using a convex formulation.

Finally, with the GINI system, we were able to systematically perform a genome-scale network learning and analysis of the genes expressed during 2 time points of Drosophila embryogenesis recorded by ISH imaging from BDGP [16]. In both time points, we find that the GINI networks are modular and scale free, which are properties predicted to hold true in gene interaction networks. Further, different regions of the networks are enriched for spatial annotations, thus GINI is able to cluster spatially similar genes. The hubs of the networks, i.e., the genes with the largest number of predicted interactions are functionally enriched for important cellular functions. We demonstrate that the networks predicted by analyzing microarray data does not have either spatial or functional enrichment, thus these results could not have been obtained by analyzing microarray data.

To the best of our knowledge, GINI represents one of the first efforts to reverse engineering gene networks from ISH image data. In both extensive simulation studies and empirical biological analysis, we demonstrate the effectiveness of GINI in predicting networks, and show that the statistical assumptions behind GINI are reasonable, and the biological analysis enabled by GINI merits close examination and further exploration.

Network inference from “one image per gene” ISH data

We first show how GINI estimates a gene network, when each gene has only one image. The next subsection extends the GINI algorithm to deal with multiple images per gene.

Let denote the set of genes being studied, so that is the gene, where , and is the number of features extracted per image. Each feature represents the gene expression in a spatial location of the embryo.

Note that algorithms that analyze microarray data typically treat samples drawn from different time points as independent samples [28], even though expressions of the same gene across time is expected to be auto correlated. We similarly assume that the different spatial features are independent of each other. The spatial independence assumption has also been implicitly made by [29], [30] while modeling transcription networks in Drosophila embryos. In the results section, we use simulated data to demonstrate that this assumption does not affect the accuracy of the algorithm significantly.

By modeling the gene interactions as invariant across the spatial locations in the embryo, we can assume that each feature is independently and identically drawn (i.i.d.) from the same distribution. Inferring gene interactions is then equivalent to modeling the dependence between the expression values of different genes at the same spatial location. Expression of the genes in each spatial location is assumed to be drawn from some (multi-variate) distribution, independent of all other spatial locations. Each spatial feature () may be modeled as a vector of length , with capturing the expression value of the gene in this location . This gives us independent samples with which the parameters of the underlying distribution may be learned. Formally, let each spatial location be drawn independently from a multi-variate Gaussian , where is the mean vector, and is the positive semi-definite covariance matrix between the genes.

In a multivariate Gaussian distribution, the entry of the inverse covariance matrix is zero if and only if the corresponding genes are conditionally independent given the rest of the graph. Thus, the non-zero entries of the inverse covariance matrix correspond to edges in the corresponding Gaussian Markov random field, giving rise to the gene interaction network. The Gaussian Markov random field is also known as a Gaussian graphical model (GGM) [31]. Since we expect a small number of interactions per gene, the estimated graph must be sparse, i.e. the number of non-zero entries of the inverse covariance matrix must be small.

Thus, the gene interaction network may be estimated by learning a Gaussian distribution from the observed images, such that the inverse covariance matrix is sparse. The mean of the Gaussian is estimated by the observed sample mean,(1)

Then, the inverse covariance matrix can be estimated by minimizing the negative log-likelihood of the data, over all possible positive semi-definite matrices. To enforce sparsity, the norm of , which counts the number of non-zero elements, is added to the negative log likelihood. Since optimizing the norm is non-convex and NP hard, the norm is used as a convex relaxation to the norm. The norm of a matrix is the sum of the absolute values of the elements of the matrix, and also enforces sparsity in the solution. Adding the norm regularization also ensures that the minimizer of the objective function exists, and is well defined. Thus, our objective function is(2)where is the second moment matrix about the mean(3) is a tuning parameter, by which we determine the strength of the penalty. As we increase the value of , we increase the penalty on the absolute values of , and hence, the graph induced by becomes more sparse. The edges in the graphical model are then estimated as(4)

Network inference from “multiple images per gene” ISH data

Multiple images of the same gene at the same time point should have the same gene expression pattern. However, in practice, the expression patterns in different images may differ considerably, for three main reasons.

Firstly, there is a wide interval of time considered as a single time point while collecting such data. For instance, the BDGP data divides embryonic development into 6 time stages. The last stage 13–16 corresponds to development of the embryo 9.3 to 15 hours after fertilization, which represents more than a third of the time taken for embryonic development. Hence, the true gene expression pattern may be dynamic within the time period of a single development stage, and the gene expressions captured for the same gene at the same time may not look similar to each other. Secondly, we might expect that for any organism for which ISH data is collected, there will necessarily be some ambiguity in how the development stage of the organism is labeled by human annotators. Finally, noise in the expression patterns due to excessive staining, lighting conditions and similar other reasons will also be observed. For all of the above reasons, any network-learning algorithm should leverage the existence of multiple images per gene per time point in improving its estimates of gene similarity.

The problem of multiple images per gene is reminiscent of multi-instance learning [35], [36]. Multi-instance learning is a form of supervised learning, where instead of labeling each instance, a bag of instances is labeled. A popular solution to the multi-instance problem is to define a multi-instance kernel, that can compute the similarity between bags of instances. Let be a collection of order statistics of the set , for example, mean, median, minimum, maximum etc. In dimensions, is computed on each dimension independently, to form a vector of order statistics. If we use order statistics, then the length of will be . The similarity between gene with a set of images and gene with images can then be computed as(5)where is an appropriate kernel function between vectors and . Such a kernel is called the statistic kernel.

The choice of the order statistics used in the kernel depends on the data collection procedure of the ISH. One concern in ISH data is that images may be overstained. In such a scenario, the median may be an appropriate choice of order statistic. If over-staining is not a concern, the maximum statistic may be more appropriate to ensure that information about presence of gene expression is not lost.

For the BDGP data, we use the covariance kernel , and the mean statistic . The choice of using a single statistic to represent information from multiple images was due to the presence of noisy images in the data set. Thus,(6)

Thus, our choice of kernel is equivalent to computing the mean similarity of all pairs of images in bags and . This specific kernel is also known as the normalized set kernel, and has been shown to perform very well in multi-instance classification [37].

Any kernel function may be written as the dot product in some higher dimensional feature space, i.e. [38]. Hence, if we assume that the data is drawn from a distribution such that is a zero-mean Gaussian, we can learn the gene interaction network by treating as the sample covariance matrix. Since estimating the inverse covariance matrix by solving equation 2 requires only the sample covariance matrix and not the data itself, we can kernelize it by using the kernel matrix defined in equation 6 as the required sample covariance matrix. Thus, the objective function is(7)which can be solved as discussed in the previous section.

Image processing

We convert the ISH images into canonical feature vectors suitable for analysis by our algorithm described above in a three-step manner. First, the precise expression pattern found in each image is extracted and aligned spatially to make all images spatially comparable. Next, each image is represented by a feature vector using Delaunay triangulation. Finally, features are normalized and feature selection is performed to extract meaningful features, that can be then used to compute the multi-set kernels to obtain gene similarity and learn the gene network.

Feature representation via Delaunay triangulation.

While makes the images of different genes alignable spatially, and therefore directly comparable, the expression patterns must still be converted into an appropriate feature representation. One commonly used method for feature representation is to use the SIFT feature descriptor [40] in either a grid of points spaced uniformly through the image, with principal component analysis (PCA) used for dimensionality reduction [15], or via interest point detection and codebook generation [24]. Such techniques work well for supervised tasks like image annotation where a weight can be learned for each direction computed by PCA or for each codeword in the generated codebook. However, for unsupervised tasks, where weights cannot be learned, we wish to extract features that explicitly take into account the spatial distribution of the gene expression. A pixel level feature representation on the other hand, allows us to capture spatial information, but has high redundancy due to the correlation between neighboring pixels.

To reduce redundancy while capturing spatial gene expression information, we choose to overlay a fixed triangular mesh on top of the standardized embryo. The gene expression pattern for each image may then be represented as the median gene expression present in each triangle in this mesh. A mesh of 311 equilateral triangles was produced by using the Delaunay triangulation algorithm [41], and aligning the mesh to the standardized embryo, as described in [23]. Each image can then be represented as a feature vector of length 311, with each feature representing the median gene expression expressed in a specific location on the embryo, which is fixed across all images. For example, triangle 1 may correspond to the head in all images, and so on. Modeling the spatial locations in a lower dimensional space via triangulation helps in approximating the independence assumption made in the GINI algorithm, analogous to using coarse time definitions while making microarray measurements.

Figure 3 shows examples of ISH images converted into the triangulated gene expression patterns. As can be seen, triangulating the output captures the key features of the gene expression location and strengths. Thus, triangulation enables reducing the dimensionality of the feature space, while retaining explicit spatial information about the gene expression, which other dimensionality reduction techniques would not be able to capture.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Triangulation.

Examples of how ISH images are converted into low-dimensional triangulated representations, for efficient feature representation.

https://doi.org/10.1371/journal.pcbi.1003227.g003

Summary of the GINI system

Putting everything together, we conclude the method section with a summary of the GINI system for network inference from ISH images. Each ISH image is converted into a standardized expression pattern using , and then triangulated to extract a low-dimensional spatial feature vector. Next, feature values are normalized, uninformative features are removed, and genes with insufficient information available are rejected. Finally, the multi-set kernel is used to compute the similarity between the bags of image vectors available for each gene, and the gene network is estimated using Equation 7. The algorithm is summarized in Algorithm 2.

Validation of the GINI assumptions: Independent spatial data

GINI assumes that the gene expression in each triangle can be assumed to be independently drawn from a multi-variate Gaussian. However, the true gene expression in adjacent spatial locations is correlated and not independent. To verify that this dependence of adjacent samples does not affect the accuracy of the estimated network, we simulate synthetic data where the underlying network is known, but the data points are not independent of each other, and test whether GINI can recover the correct network in such a scenario. The data samples depend on each other via a parameter that captures degree of dependence between data samples. When , all data samples are drawn i.i.d. from the known distribution. As increases, data samples are drawn from the same distribution, but they depend on the adjacent samples.

Data generation.

For dimensions, the true inverse covariance matrix was constructed by using the AR(1) model from [42]. That is, , and , with all other elements being zero. Dependent samples are generated from a zero mean Gaussian having the above known inverse covariance matrix, as explained below.

Let be the fractional overlap between adjacent samples. The first sample is sampled independently from the above specified Gaussian. The sample is generated from the sample as follows. Pick random features , and copy the value of the previous sample for these features : . Now, can be partitioned into the “known” features and the remaining features which still need to be sampled, conditioned on . If we partition asthen conditioned on can be shown to be Gaussian with mean and covariance , which can be computed as below, and can be sampled from it.(8)(9)

We ran two experiments. In the first, a fixed number of samples() were used to learn the network. In the second, as increases, more samples () are available for learning the network. In both experiments, for each value, we randomly sample data points using the method outlined above, estimate the matrix, and compare it to the known matrix, to compute precision and recall. Results are averaged over 50 runs of the experiment.

Figure 4(a) shows that as increases, the precision (fraction of correct interactions among all inferred ones) and recall (fraction of correct interaction among all true interactions) stay constant for small values for . Only when the amount of dependence increases beyond half, do we see a small reduction in accuracy. Thus, we can conclude that even if there is a large spatial dependence in gene expression, the result is equivalent to a slight reduction in performance. Futher, in Figure 4(b), we see that if we can increase the number of data points as we increase , the performance remains the same as using i.i.d. data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. GINI assumptions are reasonable.

Even if the data are not independent draws from the Gaussian, the network can still be learned with high precision and recall. (a) For a fixed number of data points, as increases beyond 0.5, the precision and recall reduces. (b) If we allow the number of data points to increase as increases, the precision and accuracy of the method is not affected. The standard deviation at each point in both results is approximately 0.09.

https://doi.org/10.1371/journal.pcbi.1003227.g004

GINI explains the ISH data well

For a high-dimensional distribution, it is not feasible to test if the data is truly Gaussian. However, a consequence of Gaussianity is that for each gene, the gene expression can be expressed as a weighted linear sum of the expression values of a few other genes, which form the edges of the network. To test if this assumption holds true in ISH data, for each gene, we fit a linear regression between the gene and its neighbors found by GINI and look at the absolute value of the error i.e. the mean absolute difference between the predicted and the known gene expression. When the maximum expression value is 1, for more than 90% of the genes we looked at, the absolute error was less than 0.02; 99.5% of all genes had absolute error less than 0.05, confirming that the GINI generative model explains the ISH data.

We also confirm that the prediction error is not systematic with respect to the spatial location. For each gene, we compute the prediction error (residue) when the gene is predicted by regressing it on its neighbors. For each spatial location, we plot the mean residue at that location for all genes. As can be seen in Figure 5, there is no systematic bias in the spatial positions that are hardest to predict for any gene.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. GINI error analysis.

Locations where gene expression cannot be predicted easily. Red color indicates that the true gene expression was higher than predicted by the regression, while blue indicates that the true gene expression was lower than predicted by the regression. Note that since the difference in the true and predicted gene expressions is very small, the mean residue values were multiplied by 10 to improve the contrast of the image for visualization purposes. Thus, there is no systematic bias in the spatial locations where expression is hard to predict.

https://doi.org/10.1371/journal.pcbi.1003227.g005

Network on limited data

Before running our algorithm on a large sized dataset, we construct an artificial small data set to verify the results. We input 12 images, shown in Figure 6(a) from 6 genes to the GINI algorithm (each gene has 1–3 images in the data set). With , 4 edges are predicted in the network, shown in Figure 6(b). As can be seen, the three genes hunchback(hb), four-jointed(fj), and Blimp-1, which are expressed in the dorsal, ventral and procephalic ectoderm, are connected in a single cluster. Similarly, the genes organic anion transporting polypeptide 74D(Oatp74D) and bicoid(bcd) are connected by an edge, since both show expression in the foregut and the anterior endoderm. Finally, the expression of sloppy paired-1(slp1) was considered to be sufficiently different from the other genes, hence it is not connected to any other gene in the network.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 6. Example network on small data.

(a) 12 images input to the GINI system, and (b) network of genes learnt by it, with each gene represented by one image.

https://doi.org/10.1371/journal.pcbi.1003227.g006

Thus, the gene interaction network found by GINI can be verified to be reasonable for the above small data set.

Network on the whole BDGP data

We now turn our attention to the ISH images from the Berkeley Drosophila Genome Project data set. We have obtained around 67400 ISH images of 3509 protein-coding genes from the BDGP data released in September 2009, captured at key development stages of embryonic development. Each image captures embryonic gene expression of a single gene using RNA in-situ hybridization. Each image was labeled manually with the age of the embryo, categorized into six distinct embryonic stages : 1–3, 4–6, 7–8, 9–10, 11–12, and 13–16. Genes are also annotated with ontology terms from a controlled vocabulary of around 295 terms, describing the unique embryonic structures in which gene expression is observed during the various stages of embryonic development. analyzes these image automatically, rejecting unsuitable images, to produce 51593 expression patterns of 3347 genes.

As proof of concept, we focus on images viewed from a lateral perspective from two development stage ranges of this data : 9–10 and 13–16. For the stage 9–10, we have 2869 expression patterns of 2609 genes, and for stage 13–16, we have 6350 expression patterns of 3258 genes. We extracted features as described in the methods section. For each development stage, we ran a separate analysis.

Using a value of 0.775 for stage 9–10, we ran GINI and obtained a network having 258 genes, and 516 interactions (edges) between them. For the development stage 13–16, we used , and obtained a network with 1202 genes and 3666 interactions between them. The value was selected for each network by running GINI for 21 values between 0.5 and 1, and picking a value such that the mean-degree for the network is reasonable (approximately 2–3) - see Supplementary Figure S1 for a plot that shows how the number of edges in the network decreases as increases.

Some of the interactions predicted by GINI have already been reported in the literature. For example, in the network for stage 9–10, GINI predicts that DCP-1 (CG5370), an effector caspase which is involved in apoptosis, will interact with the thread gene (CG12284), a known inhibitor of apoptosis protein [43]. GINI also predicts that Snf5- related 1(CG1064) interacts with echinoid (CG12676), both of which are known to be involved in epidermis development, muscle organ development, as well as imaginal disc-derived wing vein morphogenesis. In the 13–16 development network, GINI predicts that the capping protein beta gene (CG17158) interacts with the Glycogen phosphorylase gene (CG7254), and Tpc1 (CG6608) interacts with CG2812, which has been previously reported in [44].

The next five subsections do a detailed analysis of the 2 networks.

Scale free network

A network is said to be scale free if its degree distribution asymptotically follows a power law. That is, the fraction of genes that have at least interactions with other genes is(10)where is the scale free parameter, and is the normalization constant. It has been hypothesized that gene regulatory networks are scale free [10]. We looked at the characteristic of our interaction networks by plotting the number of interactions per gene (Figure 7), and found that the networks found by GINI are scale free. The parameter obtained is 2.3 and 2.5 for the 9–10 and 13–16 networks respectively, which corresponds well to the values observed for a large variety of power law graphs. The scale free nature of the network was found to be independent of the tuning parameter of the algorithm.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 7. Scale-free network.

Connectivity properties of the reconstructed network for time stage 9–10, and 13–16. The scale free nature of the plot can be observed for both networks. The plot for stage 9–10 has fewer points since the network constructed has fewer nodes and edges.

https://doi.org/10.1371/journal.pcbi.1003227.g007

Unlike the gene regulatory network obtained for Human-B cells [10], we found that the scale-free nature of the gene network we obtain has a good fit, without observing a deviation from the expected at low connectivity values. However, this could be a side-effect of the larger number of genes they analyzed.

The BDGP networks are modular

Using spectral clustering, we construct 12 regions or clusters within each network, and visualize the five biggest clusters of each of the networks in Figure 8. All 12 clusters in both networks are very well separated. The ratio of within-cluster edges to total number of edges is 70% and 87% for the 9–10 and 13–16 development stage networks respectively, indicating that the estimated networks are highly modular. From a biological perspective, different parts of gene networks may be responsible for different pathways or biological functional components of the cell, thus modularity is a good prediction for real interaction networks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 8. Modular network.

A global view of the networks constructed by our algorithm for development stage 9–10, and 13–16, visualized for 5 of the 12 clusters in the network. The nodes of each cluster in the network are represented by different colors. Red edges are edges between nodes in the same cluster, while green edges are edges between nodes in different clusters. Each cluster is represented by one or two spatial annotation terms enriched in the cluster.

https://doi.org/10.1371/journal.pcbi.1003227.g008

Hub analysis

Given the scale-free nature of the network, a small number of the genes have a large number of interactions. We analyze the Gene Ontology functions of the genes having the largest number of interactions, i.e. the hubs of the network. The question we wish to address is: if we pick the top 5% of the genes having the maximum connectivity with other genes, what kind of functional enrichment do these genes have? Our background population is of the 2609 and 3258 genes for which we have at least one ISH image describing its expression for the 9–10 and 13–16 stages respectively. We use the hypergeometric test, with Bonferroni correction used to correct for multiple hypothesis tests [45]. As can be seen in Table 1, we observe enrichment of a wide variety of functions that are essential to cell growth and functioning, including metabolic processes, cellular respiration, transport of electrons and ions, protein modification, ribosome biogenesis etc.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. GO functional analysis for the gene hubs of the GINI network.

https://doi.org/10.1371/journal.pcbi.1003227.t001

Next, we examine a few high-degree hubs in the two networks in detail, along with their neighborhood genes in the networks. Figure 9 shows the hub neighborhood for two genes in the 9–10 development stage network. CG3969 is a Activated Cdc42 kinase-like gene known to be involved in protein phosphorylation [46] and cell death [47], and CG9984 (TH1) is known to be involved in regulation of biosynthetic process [48] and nervous system development [49]. Both genes interact with many genes having functions related to the primary metabolic process, and single-organism cellular process. In stage 13–16, we examine the hub neighborhood of CG5904 and CG6501. The mitochondrial ribosomal protein CG5904 has been previously predicted to be a structural constituent of ribosome [50], and we find that it interacts with many genes involved in the ribosome biogenesis. Gene CG6501 (Ns2) has been previously predicted to be involved in phagocytosis, engulfment [51], and ribosome biogenesis [46]; CG6501's neighborhood has multiple genes that are also involved in ribosome biogenesis and single-organism cellular process.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 9. Hubs of the GINI network.

A look into the neighborhoods of a few hubs from the GINI networks for stage 9–10 and 13–16. A few enriched GO groups are highlighted in the subnetworks as shown.

https://doi.org/10.1371/journal.pcbi.1003227.g009

Enrichment of annotation terms

Each gene in the BDGP data has been labeled manually by annotations describing the spatial gene expression, using 295 annotation terms. We expect that since the gene interaction network is constructed via spatial similarity, genes that are connected to each other in the network will have similar spatial annotation terms.

To test this, we cluster the gene network using spectral clustering [52] into 12 clusters, and analyze the enrichment of each cluster for annotation terms using the hypergeometric test, with Bonferroni correction used to correct for multiple hypothesis tests. In the gene network for the 9–10 stage, 11 of the 12 clusters are enriched for 63 total annotation terms (Figure 10). The only cluster not showing any enrichment in the 9–10 stage network is also the smallest cluster, having only 4 genes. For example, in cluster 8, 92% of the genes have expression in the ventral nerve cord primordium P3 , while only 8% of the genes in the data have expression in this region. Similarly, 73% of the genes in cluster 11 have expression in the trunk mesoderm primordium, while only 16% of the genes in the data have expression in this region. For the 13–16 stage network, all 12 clusters are enriched for a total of 81 enrichments, a part of which is visualized in Figure 10. Tables S1 and S2 in the supplementary material report the complete enrichment analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 10. Spatial annotations.

Enrichment analysis for clusters in the gene interaction networks found by GINI. A green dot indicates enrichment with a P-value.

https://doi.org/10.1371/journal.pcbi.1003227.g010

Triangulation improves quality of result.

Previous work on image processing for ISH images has focused on using SIFT features, and constructing a codebook that contain all the embryonic structures that the system is expected to annotate [24]. In this section, we show that triangulation produces more interesting networks over such a SIFT feature representation. We use the gene expression patterns, and represent them by constructing SIFT features of the expression pattern over a grid. These grid SIFT features are then represented with a codebook of 2000 dictionary features, as described in [24]. We then use these dictionary features instead of the triangulated features to learn the GINI network. Figure 11 shows that the resulting networks are not as richly enriched as the ones derived from the triangulation features in Figure 10. The total number of enrichments in the SIFT codebook network is 42 and 21 for the 9–10 and 13–16 development stage networks respectively. In contrast, the triangulated GINI networks had 63 and 81 enrichments for the 9–10 and 13–16 stage networks. Figure 12 shows that this result is independent of the number of clusters selected for the analysis, for both triangulated networks as well as SIFT codebook networks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 11. Enrichment analysis on networks learned from SIFT dictionary features instead of triangulation features.

The network for development stage 9–10 has only 7 enriched clusters of the 12 clusters in the network. For the stage 13–16 network, only 3 of the 12 clusters are enriched for spatial annotations.

https://doi.org/10.1371/journal.pcbi.1003227.g011

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 12. SIFT codebook features do not perform as well as triangulated features on ISH data.

Percentage of clusters enriched for spatial annotations in networks predicted by GINI as a function of number of clusters for data from development stage 9–10 and 13–16. As can be seen, using triangulated features produces networks with more enriched clusters than using SIFT-codeword features, independent of the number of clusters selected for the analysis. Further, the enrichment of the GINI network clusters does not significantly vary as the number of clusters are varied.

https://doi.org/10.1371/journal.pcbi.1003227.g012

Sensitivity to the tuning parameter .

Supplementary Figure S1 shows how the number of edges in the network decreases as the tuning parameter() of the GINI algorithm increases. To confirm that the enrichment results are not sensitive to the choice of , we obtained 21 predicted networks by varying the value uniformly from 0.5 to 1. For each network, we repeated the clustering and enrichment analysis, and found that the enrichment for term annotations is not highly sensitive to choice of ( Figure 13). The enrichment results are also not dependent on the number of clusters - we get high enrichment, independent of the number of clusters chosen while running the clustering algorithm (Figure 12).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 13. tuning. Percentage of clusters enriched for spatial annotations in networks predicted by GINI as a function of tuning parameter for data from development stage 9–10 and 13–16. As we increase , the number of edges predicted in the network decrease, however, the enrichment of the different clusters stays almost constant. Thus, qualitative analysis of the network does not seem to be sensitive to choice of .

https://doi.org/10.1371/journal.pcbi.1003227.g013

Comparison with microarray network

We learn a network from microarray data collected by the BDGP project over 12 time points in embryonic development [16], over the same genes that are being studied in the 9–10 and 13–16 networks, using covariance between the microarray expression as the kernel. We find that the overlap in edges between the 2 networks is very small, only 1% of the edges are common to both networks. If we assume that spatial expression annotations are a proxy for functional enrichment, then we can check if the microarray network is enriched for the spatial annotation terms. Figure 14 shows that the percentage of enriched clusters in the microarray network is small, independent of the number of clusters analyzed.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 14. Microarray v/s ISH data.

The percentage of clusters that are enriched for spatial term annotations using networks learned from ISH and microarray data.

https://doi.org/10.1371/journal.pcbi.1003227.g014

We can also test functional GO enrichment of the hubs of the network. Table 2 shows that the hubs of the microarray network for stage 13–16 are enriched for only a single function, where 4 of the 145 hub genes are involved in the “aromatic compound catabolic process”, while the microarray data network for stage 9–10 has no enrichments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. GO functional analysis for the gene hubs of the microarray network.

https://doi.org/10.1371/journal.pcbi.1003227.t002

Thus, we find that the network learned from ISH images is clearly different from a network learned from microarray data. The ISH image network is enriched for spatial annotation terms, as well as functional enrichment of the hubs of the network, which does not hold true for the microarray network. This suggests that analyzing ISH images could support different scientific conclusions, which should be studied in greater detail.