extended Exploration

tutorials/exploration.md

@@ -24,3 +24,124 @@ remove outliers, thus obtaining a proper dataset for differential expression ana
 You can either continue the session from the previous module, or download
 GCT file from [here]({{ site.baseurl }}{% link tutorials/preparing.files/GSE53986_norm.gct%}) 
 and load it into Phantasus.
+
+## PCA Plot
+
+Principal Component Analysis (PCA) is an essential way to asses quality of the dataset.
+One way to look at it is a as a linear dimension reduction method, that allows
+to go from 10000-dimension space (by a number of genes) to something smaller,
+like two or three dimensions, which can be displayed on a plane.
+Importantly, PCA assumes Gaussian-distributed noise, so it should be
+run on log-scaled gene expression values.
+
+PCA plot can be done using _Tools/Plots/PCA Plot_ menu.
+Each point on a PCA plot represents one sample.
+You can customize color, size and labels of points on the chart 
+using values from the annotation. For example one can set the color 
+to come from _treatment_ annotation.
+
+<img src="exploration.files/pcaplot_tool_finished.png" width="550px" />
+
+From this plot we can see that our samples more or less group together,
+which is good, meaning that the samples from the same condition correlate 
+well with each other and no so much with other conditions.
+However, there are first replicates which seems to be outliers in three
+of four conditions.
+
+Notable property of PCA, as opposed to other dimension reduction techniques like
+MDS or t-SNE, is that it is linear and thus principle components are 
+interpretable.
+In our case we can see that PC1 corresponds to LPS treatment (all LPS-treated
+samples are on the right side, and non-LPS are on the left, note that
+PCA can't distinguish positive from negative direction, so in your 
+case the sides can be swapped) and PC2 corresponds to IFNg treatment.
+Sometimes, sample labels can be accidentally mixed up and PCA analysis
+can be used to catch this.
+
+It is also can be useful too look at percent of variation explained by each component.
+For example, in our case LPS treatment seems to have bigger effect compared
+to IFNg treatment.
+
+
+
+## K-means clustering
+
+Another useful dataset exploration tool is k-means clustering.
+It is a general clustering method, that can group genes into _k_
+clusters based on their correlation (technically, Euclidean 
+distance for z-score expression values is used). 
+
+To apply k-means clusting use _Tools/Clustering/k-means_ menu.
+There is no definite answer on the best value of _k_ to be used,
+but the general heuristic is that the number of clusters scales
+with the number of biological conditions in the dataset.
+For the example dataset we will be clustering into 16 cluster:
+
+<img src="exploration.files/kmeans_tool.jpg" width="500px" />
+
+Afterwards, rows can be sorted by _clusters_ column. By using 
+menu _View/Fit to window_ one can get a "bird's-eye view" on the dataset,
+which usually helps understanding the structure of the dataset.
+In particular, it easy to see that the first replicates in each 
+of the four conditions looks quite different from the other samples in the group.
+
+<img src="exploration.files/kmeans_result.jpg" width="600px" />
+
+## Hierarchical clustering
+
+Similarly, clustering can applied to samples to better display 
+inter- and intra- group correlations. Hierarchical clustering
+is perfect for such analysis. Briefly it works by starting
+from a set of cluster consisting of individual samples,
+and on each step two closest clusters are joined, 
+until a single cluster is left.
+
+Let us use _Tool/Hierarchical clustering_ menu to apply hierarchical clustering.
+There is a number of parameters, but they can be left as is.
+
+<img src="exploration.files/hierarchical_tool.jpg" width="450px" />
+
+A dendrogram showing how clusters were joins is added on top of the heatmap.
+We can again see that the first replicates are clustering outside of their groups.
+
+<img src="exploration.files/hierarchical_results.png" width="500px" />
+
+
+## Filtering outliers
+
+From our exploration we can conclude that while overall the dataset
+looks good and samples in the groups correlate well with each other,
+there are outliers which can be removed.
+You can select the good samples 
+and extract them into another heatmap (by clicking _Tools/New Heat Map_ or pressing _Ctrl+X_).
+
+<img src="exploration.files/good_samples.png" width="600px" />
+
+Finally, let us save the filtered dataset as "GSE53986_filtered.gct" using 
+_File/Save Dataset_ menu. For the reference, the result should be similar
+to the following file: 
+[GSE53986_filtered.gct]({{ site.baseurl }}{% link tutorials/exploration.files/GSE53986_filtered.gct%}).
+
+There is a couple of points that should be kept in mind about filtering outliers:
+* There is no definite way to distinguish outliers from biological variation,
+so one has to be cautious about them for the process not to turn 
+into cherry-picking. For this particular dataset we can see that, for example,
+*untreated 1* sample is quite different from pretty tight group of other untreated samples.
+Moreover, the fact that all outliers are the first replicates indicate that there
+was a systematic batch problem, probably related to plate layout during sample
+preparation step.
+* Sometimes outliers can be very different from other samples. In this case
+the analysis has to be redone from scratch after removing the sample. 
+Otherwise, during quantile normalization step the bad samples were taken into
+account as well, thus making data in the good samples more noisy.
+
+## Summary
+
+To summarize this and the previous modules:
+1. For the most types of analysis gene expression data should be log-scaled. 
+Sometimes the data can be already log-scaled, some time it is not.
+It is important to check this and make it log-scale if necessary.
+2. Gene expression data should be normalized, so that the samples can be properly
+compared. One of the most straightforward way to do it is quantile normalization.
+3. The data has to be quality checked. For this you can look at marker genes,
+do a PCA analysis and clustering.