---
title: "A meta analysis on effects of Parkinson's Disease using Gemma.R"
author:
- name: B. Ogan Mancarci
  affiliation: Michael Smith Laboratories, University of British Columbia, Vancouver, Canada
package: gemma.R
output:
    BiocStyle::html_document
header-includes:
vignette: >
    %\VignetteIndexEntry{A meta analysis on effects of Parkinson's Disease using Gemma.R}
    %\VignetteEncoding{UTF-8}
    %\VignetteEngine{knitr::rmarkdown}
editor_options: 
  markdown: 
    wrap: 72
  chunk_output_type: console
---

```{r, include = FALSE}
# Prevent certificate issues for GitHub actions
options(gemma.SSL = FALSE,
        datatable.print.trunc.cols = FALSE)

# temporary options
# options(gemma.API = "https://dev.gemma.msl.ubc.ca/rest/v2/")
# options(gemma.memoised = TRUE)

knitr::opts_chunk$set(
    comment = "",
    cache = FALSE
)
```

```{r setup, message = FALSE}
library(gemma.R)
library(dplyr)
library(poolr)
library(ggplot2)
library(stringr)
```

```{r, include = FALSE}
gemma.R:::setGemmaPath('prod')
```

Gemma.R contains a large number of datasets representing a wide array of
conditions. These datasets are manually annotated by our curators to facilitate
discovery and categorization making it a useful tool for meta-analysis.

In this example we search Gemma for datasets comparing healthy
controls and patients with Parkinson's Disease. Our curators use ontology terms
to annotate datasets when possible. This allows us to use the MONDO Disease Ontology term
"MONDO:0005180" to search for datasets and samples annotated with the term to make sure
we are accessing datasets relevant for our purposes with minimal effort.

# Identifying what to search for

Gemma allows searching for datasets using plain text queries. However for a
structured analysis, it is usually more helpful to use ontology terms to 
make sure we are getting the right datasets. To identify potentially useful ontology
terms, we can use search_annotations

```{r}
annots <- search_annotations("parkinson's")
annots %>% head %>% gemma_kable()
```

This will return matching ontology terms within every ontology loaded to Gemma.
To identify which ones are actually used to annotate our datasets, you can do

```{r}
annot_counts <- annots$value.URI %>% sapply(\(x){
    attributes(get_datasets(uris = x))$totalElements
})

annots$counts = annot_counts

annots %>% filter(counts>0) %>%
    arrange(desc(counts)) %>% gemma_kable()

```

# Querying datasets of interest

The `get_datasets` function can be used to find datasets of interest,
either using ontology terms or plain text. Since Parkinson's Disease has an unambiguous
ontology term in the Disease Ontology, we will be using it to avoid acquiring
datasets that tangentially mention the disease in their descriptions. Here we also
limit our results to only include human samples since we are not interested in 
models from other species and samples that are explicitly from the brain using the
ontology term for brain. Here we will searching within "allCharacteristics" however other filter properties are available. You can use `filter_properties` to
access all available properties for anything filterable in Gemma. 

Some noteable ones for our purposes would be "allCharacteristics", which searches within every annotation on a dataset and its samples while we could have also used "experimentalDesign.experimentalFactors.factorValues.characteristics" to specifically look within experimental factors. Instead we will be showing how to
limit for relevant contrasts in the next steps for this example.

```{r}
filter_properties()$dataset %>% head %>% gemma_kable
```

```{r}
# getting all resulting datasets using limit and offset arguments
results <- get_datasets(filter = "allCharacteristics.valueUri = http://purl.obolibrary.org/obo/MONDO_0005180 and allCharacteristics.valueUri = http://purl.obolibrary.org/obo/UBERON_0000955",
                       taxa ='human') %>% get_all_pages()



results %>% select(experiment.shortName,experiment.name) %>% head
```


# Filtering the datasets for suitability

While we know that all the resulting datasets were annotated for the
term for Parkinson's Disease, we currently do not know how many of them
are comparisons of healthy controls and patients with Parkinson's
Disease. We also do not know if the datasets have batch effects that may affect
our findings.

For this example, we decided we should not include data sets that have a batch confound (though that is up to the user). Gemma internally handles batch correction if batch
information is available for the dataset. We will be looking at
`experiment.batchEffect` column. As explained in the `get_datasets`
documentation, this column will be set to -1 for datasets where batch
confound is detected, 0 for datasets without available batch information
and to 1 if the data is clear of batch confounds.

```{r}
results <- results %>% filter(experiment.batchEffect == 1)
```


We now want to ensure that the differential expressions we analyze
compare control and Parkinson's Disease patients. This information is
available via `get_dataset_differential_expression_analyses` which
returns the experimental groups for differential expression analyses
performed for the dataset. The columns we are primarily interested in
are `baseline.factors` which typically
records the control group of the differential expression analysis and
`experimental.factors` which typically
records the test case

```{r}
experiment_contrasts <- results$experiment.shortName %>% 
    lapply(function(x){
        out = get_dataset_differential_expression_analyses(x)
        }) %>% do.call(rbind,.)

```


The factor we are interested in is Parkinson's Disease in
`experimental.factorValue` or
"<http://purl.obolibrary.org/obo/MONDO_0005180>" in `experimental.factors`'s `value.URI`.
We also need to make sure that the baseline that the sample is compared against is 
a control experiment, samples annotated with "reference subject role" or "<http://purl.obolibrary.org/obo/OBI_0000220>"

```{r}
parkin_contrasts <- experiment_contrasts %>% 
    filter(factor.category == 'disease') %>% 
    filter(sapply(experimental.factors,function(x){
        "http://purl.obolibrary.org/obo/MONDO_0005180" %in% x$value.URI
        }) &
           sapply(baseline.factors,function(x){
               "http://purl.obolibrary.org/obo/OBI_0000220" %in% x$value.URI
               }))

```

Note that while this example attempts to show ways you can automate this process, you should take care to perform some manual examination of the data and determine how independent the contrasts you selected actually are. For instance in our selection, there are two contrasts from the same experiment where differential expression is performed on different brain regions.


```{r}
dup_exp <- parkin_contrasts$experiment.ID %>% 
    table %>% {.[.>1]} %>% names

dup_exp


parkin_contrasts %>% filter(experiment.ID %in%  dup_exp)
```

We should at least make sure that these two contrasts don't share samples between them, one thing to check is the subsetFactor which might explain why these two
separate contrasts exist

```{r}
parkin_contrasts %>% filter(experiment.ID %in%  dup_exp) %>% {.$subsetFactor}
```

And we can see that these two contrasts are created to split the samples between different brain regions, therefore do not actually share samples between them.

<!-- 
Expose some version of subset_factorValues to make this process more streamlined
-->

In the next steps we will be downloading expression data in bulk. If you are here to just
try the code out, you can speed the process a bit by limiting the number of experiments
to deal with

```r
# arbitrarily select a few datasets
# not executed in this example
parkin_contrasts <- parkin_contrasts[1:5,]
```

<!-- In the next steps we will be downloading differential expression data in bulk to -->
<!-- perform our meta-analysis. Since downloading and analyzing all these datasets -->
<!-- takes quite some time, in this example we will limit ourselves to a smaller selection. -->
<!--
```{r}
# We are using the first 5 contrasts to keep this analysis short
# sanity_sets = c('GSE8397','GSE7621','GSE20168','GSE20291','GSE20292')
# sanity_ids = results %>% filter(experiment.Accession %in% sanity_sets) %>%  .$experiment.ID

# parkin_contrasts <- parkin_contrasts %>% filter(experiment.ID %in% sanity_ids)
# parkin_contrasts <- parkin_contrasts[1:10,]
```
-->

Now that we have our relevant contrasts, we can download them using
`get_differential_expression_values`. This function can be used to download
differential expression fold change and p values, either using the
experiment name/ids or more specifically using the result.IDs

```{r}
differentials <- parkin_contrasts$result.ID %>% lapply(function(x){
    # take the first and only element of the output. the function returns a list 
    # because single experiments may have multiple resultSets. Here we use the 
    # resultSet argument to directly access the results we need
    get_differential_expression_values(resultSets = x)[[1]]
})

# some datasets might not have all the advertised differential expression results
# calculated due to a variety of factors. here we remove the empty differentials
missing_contrasts <- differentials %>% sapply(nrow) %>% {.==0}
differentials <- differentials[!missing_contrasts]
parkin_contrasts <- parkin_contrasts[!missing_contrasts,]
```

# Getting the p-values for the condition comparison

`differentials` is now a list of data frames containing the differential
expression information. To run a simple meta-analysis, we need the p values for
the genes from the relevant contrasts.

```{r}

condition_diffs <- seq_along(differentials) %>% lapply(function(i){
    # iterate over the differentials
    diff = differentials[[i]]
    # get the contrast information about the differential
    contrast = parkin_contrasts[i,]
    p_vals = diff[[paste0('contrast_',contrast$contrast.ID,"_pvalue")]]
    log2fc = diff[[paste0('contrast_',contrast$contrast.ID,"_log2fc")]]
    genes = diff$GeneSymbol
    
    data.frame(genes,p_vals,log2fc)
})

# we can use result.IDs and contrast.IDs to uniquely name this. 
# we add the experiment.id for readability
names(condition_diffs) = paste0(parkin_contrasts$experiment.ID,'.',
                                parkin_contrasts$result.ID,'.',
                                parkin_contrasts$contrast.ID)

condition_diffs[[1]] %>% head
```

# Combining the acquired p values

Now that we have acquired the values we need for a meta-analysis from Gemma,
we can proceed with any methodology we deem suitable for our analysis. 

In this example we will use a very simple approach, [Fisher's combined p-value test](https://en.wikipedia.org/wiki/Fisher%27s_method). 
This is implemented in the `fisher` function from `poolr` package. Fisher's method has the advantage that it operates on the p-values,
which are in the processed differential expression results in Gemma. It has some 
disadvantages like ignoring the direction of the expression change (Gemma's p-values are two-tailed) and is very sensitive
to a single 'outlier' p-value, but for this demonstration we'll go with it.

The first step is to identify which genes are available in our results:

```{r}
all_genes <- condition_diffs %>% lapply(function(x){
    x$genes %>% unique
}) %>% unlist %>% table

# we will remove any gene that doesn't appear in all of the results
# while this criteria is too strict, it does help this example to run 
# considerably faster
all_genes <- all_genes[all_genes==max(all_genes)]
all_genes <- names(all_genes)

# remove any probesets matching multiple genes. gemma separates these by using "|"
all_genes <- all_genes[!grepl("|",all_genes,fixed = TRUE)]

# remove the "". This comes from probesets not aligned to any genes
all_genes <- all_genes[all_genes != ""]
all_genes %>% head
```

Now we can run the test on every gene, followed by a multiple testing correction.

```{r, eval = FALSE}
fisher_results <- all_genes %>% lapply(function(x){
    p_vals <- condition_diffs %>% sapply(function(y){
        # we will resolve multiple probesets by taking the minimum p value for
        # this example
        out = y[y$genes == x,]$p_vals
        if(length(out) == 0 ||all(is.na(out))){
            return(NA)
        } else{
            return(min(out))
        }
    })
    
    fold_changes <- condition_diffs %>% sapply(function(y){
        pv = y[y$genes == x,]$p_vals
        if(length(pv) == 0 ||all(is.na(pv))){
            return(NA)
        } else{
            return(y[y$genes == x,]$log2fc[which.min(pv)])
        }
    })
    
    median_fc = fold_changes %>% na.omit() %>% median
    names(median_fc) = 'Median FC'
    
    combined = p_vals %>% na.omit() %>% fisher() %>% {.$p}
    names(combined) = 'Combined'
    c(p_vals,combined,median_fc)
}) %>% do.call(rbind,.)
fisher_results <- as.data.frame(fisher_results)
rownames(fisher_results) = all_genes

fisher_results[,'Adjusted'] <- p.adjust(fisher_results[,'Combined'],
                                        method = 'fdr')

fisher_results %>%
    arrange(Adjusted) %>% 
    select(Combined,Adjusted,`Median FC`) %>% 
    head
```

```{r fisher, include = FALSE}
library(parallel)
cores = detectCores()
if (Sys.info()['sysname'] == 'Windows'){
    cores = 1
}

fisher_results <- all_genes %>% mclapply(function(x){
    p_vals <- condition_diffs %>% sapply(function(y){
        # we will resolve multiple probesets by taking the minimum p value for
        # this example
        out = y[y$genes == x,]$p_vals
        if(length(out) == 0 ||all(is.na(out))){
            return(NA)
        } else{
            return(min(out))
        }
    })
    
    fold_changes <- condition_diffs %>% sapply(function(y){
        pv = y[y$genes == x,]$p_vals
        if(length(pv) == 0 ||all(is.na(pv))){
            return(NA)
        } else{
            return(y[y$genes == x,]$log2fc[which.min(pv)])
        }
    })
    
    median_fc = fold_changes %>% na.omit() %>% median
    names(median_fc) = 'Median FC'
    
    combined = p_vals %>% na.omit() %>% fisher() %>% {.$p}
    names(combined) = 'Combined'
    c(p_vals,combined,median_fc)
},mc.cores = ceiling(cores/2)) %>% do.call(rbind,.)
fisher_results <- as.data.frame(fisher_results)
rownames(fisher_results) = all_genes

fisher_results[,'Adjusted'] <- p.adjust(fisher_results[,'Combined'],
                                        method = 'fdr')

fisher_results %>%
    arrange(Adjusted) %>% 
    select(Combined,Adjusted,`Median FC`) %>% 
    head
```

We end up with quite a few differentially expressed genes in the meta-analysis (Fisher's method is very sensitive)

```{r}
sum(fisher_results$Adjusted<0.05) # FDR<0.05
nrow(fisher_results) # number of all genes
```

Next we look at markers of dopaminergic cell types and how they rank compared to other genes.
Parkinson's Disease is a neurodegenerative disorder, leading to death of dopaminergic cells. We should expect them to show up in our
results.

```{r}
# markers are taken from https://www.eneuro.org/content/4/6/ENEURO.0212-17.2017
dopa_markers <-  c("ADCYAP1", "ATP2B2", "CACNA2D2", 
"CADPS2", "CALB2", "CD200", "CDK5R2", "CELF4", "CHGA", "CHGB", 
"CHRNA6", "CLSTN2", "CNTNAP2", "CPLX1", "CYB561", "DLK1", "DPP6", 
"ELAVL2", "ENO2", "GABRG2", "GRB10", "GRIA3", "KCNAB2", "KLHL1", 
"LIN7B", "MAPK8IP2", "NAPB", "NR4A2", "NRIP3", "HMP19", "NTNG1", 
"PCBP3", "PCSK1", "PRKCG", "RESP18", "RET", "RGS8", "RNF157", 
"SCG2", "SCN1A", "SLC12A5", "SLC4A10", "SLC6A17", "SLC6A3", "SMS", 
"SNCG", "SPINT2", "SPOCK1", "SYP", "SYT4", "TACR3", "TENM1", 
"TH", "USP29")

fisher_results %>% 
    arrange(Combined) %>% 
    rownames %>%
    {.%in% dopa_markers} %>%
    which %>% 
    hist(breaks=20, main = 'Rank distribution of dopaminergic markers')
```

In agreement with our hypothesis, we can see that the dopaminergic markers tend to have high ranks in our results.

# Acquiring the expression data for a top gene

Now that we have our results, we can take a look at how the expression of one of our top
genes in these experiments. To do this we will use the `get_dataset_expression_for_genes` function
to get the expression data. Once we get the expression data for our genes of interest, we will use 
the `get_dataset_samples` function to identify which samples belongs to which experimental group.

> Note: as of this writing (early 2023), the Gemma.R methods used in this section are not yet released in Bioconductor; install via devtools to try it out.

For starters, lets look at our top pick and the its p values in the individual datasets, on a log scale:

```{r}
# the top gene from our results
gene <- fisher_results %>% arrange(Adjusted) %>% .[1,] %>% rownames

p_values <- fisher_results %>%
    arrange(Adjusted) %>% 
    .[1,] %>% 
    select(-Combined,-`Median FC`,-Adjusted) %>% 
    unlist %>%
    na.omit()

#sum(p_values<0.05)
#length(p_values)

# p values of the result in individual studies
p_values %>% log10() %>% 
    data.frame(`log p value` = .,dataset = 1:length(.),check.names = FALSE) %>% 
    ggplot(aes(y = `log p value`,x = dataset)) +
    geom_point() + 
    geom_hline(yintercept = log10(0.05),color = 'firebrick') + 
    geom_text(y = log10(0.05), x = 0, label = 'p < 0.05',vjust =-1,hjust = -0.1) + 
    theme_bw() + ggtitle(paste("P-values for top gene (", gene, ") in the data sets")) +
    theme(axis.text.x = element_blank())
```
This shows that the gene is nominally significant in some but not all of the data sets.

To examine this further, our next step is to acquire gene expression data from the 
relevant datasets using `get_dataset_object`. We will use both dataset ids and resultSet ids
when using this function since some of the returned analysis are only performed on a subset
of the data. Providing resultSet ids allows us to harmonize the differential expression
results with expression data by returning the relevant subset.


```{r}


# we need the NCBI id of the gene in question, lets get that from the original
# results
NCBIid <- differentials[[1]] %>% filter(GeneSymbol == gene) %>% .$NCBIid %>% unique
#NCBIid


expression_frame <- get_dataset_object(datasets = parkin_contrasts$experiment.ID,
                                       resultSets = parkin_contrasts$result.ID,
                                       contrasts = parkin_contrasts$contrast.ID,
                                       genes = NCBIid,type = 'tidy',consolidate = 'pickvar')

# get the contrast names for significance markers 
signif_contrasts <- which(p_values < 0.05) %>% names
```

Finally, we'll make a plot showing the expression of the gene of interest in all the data sets.
This helps us see the extent to which there is evidence of consistent differential expression.

```{r}
expression_frame <- expression_frame %>%
    filter(!is.na(expression)) %>% 
    # add a column to represent the contrast 
    dplyr::mutate(contrasts = paste0(experiment.ID,'.',
                                    result.ID,'.',
                                    contrast.ID)) %>%
    # simplify the labels
    dplyr::mutate(disease = ifelse(disease == 'reference subject role','Control','PD'))

# for adding human readable labels on the plot
labeller <- function(x){
    x %>% mutate(contrasts = contrasts %>% 
                     strsplit('.',fixed = TRUE) %>%
                     purrr::map_chr(1) %>%
                     {expression_frame$experiment.shortName[match(.,expression_frame$experiment.ID)]})
}

# pass it all to ggplot
expression_frame %>% 
    ggplot(aes(x = disease,y = expression)) + 
    facet_wrap(~contrasts,scales = 'free',labeller = labeller) + 
    theme_bw() + 
    geom_boxplot(width = 0.5) + 
    geom_point() + ggtitle(paste("Expression of", gene, " per study")) + 
    geom_text(data = data.frame(contrasts = signif_contrasts,
                             expression_frame %>%
                             group_by(contrasts) %>%
                             summarise(expression = max(expression)) %>%
                             .[match(signif_contrasts,.$contrasts),]),
              x = 1.5, label = '*',size=5,vjust= 1)

```


# Session info

```{r}
sessionInfo()
```