A Simple Example

1. Load the package

library(SNFtool)

2. Set SNF hyperparameters

Three hyperparameters are introduced in this example: K, alpha (also referred to as sigma or eta in different documentations), and T. You can learn more about the significance of these hyperparameters in the original SNF paper (see references).

K <- 20
alpha <- 0.5
T <- 20

3. Load the data

The SNFtool package provides two mock dataframes titled Data1 and Data2 for this example. Data1 contains gene expression values of two genes for 200 patients. Data2 similarly contains methylation data for two genes for those same 200 patients.

data(Data1)
data(Data2)

Here’s what the mock data looks like:

library(ComplexHeatmap)

# gene expression data
gene_expression_hm <- Heatmap(
    as.matrix(Data1),
    cluster_rows = FALSE,
    cluster_columns = FALSE,
    show_row_names = FALSE,
    show_column_names = FALSE,
    heatmap_legend_param = list(
        title = "Gene Expression"
    )
)

gene_expression_hm

# methylation data
methylation_hm <- Heatmap(
    as.matrix(Data2),
    cluster_rows = FALSE,
    cluster_columns = FALSE,
    show_row_names = FALSE,
    show_column_names = FALSE,
    heatmap_legend_param = list(
        title = "Methylation"
    )
)

methylation_hm

The “ground truth” of how this data was generated was that patients 1 to 100 were drawn from one distribution and patients 101 to 200 were drawn from another. We don’t have access to that kind of knowledge in real data, but we do here.

true_label <- c(matrix(1, 100, 1), matrix(2, 100, 1))

4. Generate similarity matrices for each data source

We consider the two gene expression features in Data1 to contain information from one broader gene expression source and the two methylation features in Data2 to contain information from a broader methylation source.

The next step is to determine, for each of the sources we have, how similar all of our patients are to each other.

This is done by first determining how dissimilar the patients are to each other for each source, and then converting that dissimilarity information into similarity information.

To calculate dissimilarity, we’ll use Euclidean distance.

distance_matrix_1 <- as.matrix(dist(Data1, method = "euclidean"))
distance_matrix_2 <- as.matrix(dist(Data2, method = "euclidean"))

Then, we can use the affinityMatrix function provided by SNFtool to convert those distance matrices into similarity matrices.

similarity_matrix_1 <- affinityMatrix(distance_matrix_1, K, alpha)
similarity_matrix_2 <- affinityMatrix(distance_matrix_2, K, alpha)

Those similarity matrices can be passed into the SNF function to integrate them into a single similarity matrix that describes how similar the patients are to each other across both the gene expression and methylation data.

5. Integrate similarity matrices with SNF

fused_network <- SNF(
    list(similarity_matrix_1, similarity_matrix_2),
    K,
    T
)

6. Find clusters in the integrated matrix

If we think there are 2 clusters in the data, we can use spectral clustering to find 2 clusters in the fused network.

number_of_clusters <- 2
assigned_clusters <- spectralClustering(fused_network, number_of_clusters)

Sure enough, we are able to obtain the correct cluster label for all patients.

all(true_label == assigned_clusters)
#> [1] TRUE

1. Load the package

library(metasnf)

2. Store the data in a data_list

All the data we’re working with will get stored in a single object called the data_list. The data_list is made by passing in each dataframe into the generate_data_list function, alongside information about the name of the dataframe, the broader source (referred to in this package as a “domain”) of information that dataframe comes from, and the type of features that are stored inside that dataframe (can be continuous, discrete, ordinal, categorical, or mixed).

The data_list generation process also requires you to specify which column contains information about the ID of the patients. In this case, that information isn’t there, so we’ll have to add it ourselves. The added IDs span from 101 onwards (rather than from 1 onwards) purely for convenience: automatic sorting of patient names won’t result in patient 199 being placed before patient 2.

# Add "patient_id" column to each dataframe
Data1$"patient_id" <- 101:(nrow(Data1) + 100)
Data2$"patient_id" <- 101:(nrow(Data2) + 100)

data_list <- generate_data_list(
    list(
        data = Data1,
        name = "genes_1_and_2_exp",
        domain = "gene_expression",
        type = "continuous"
    ),
    list(
        data = Data2,
        name = "genes_1_and_2_meth",
        domain = "gene_methylation",
        type = "continuous"
    ),
    uid = "patient_id"
)

The first entries are all lists which contains the following elements:

1. The actual dataframe
2. A name for the dataframe (string)
3. A name for the domain of information the dataframe is representative of (string)
4. The type of feature stored in the dataframe (options are continuous, discrete, ordinal, categorical, and mixed)

Finally, there’s an argument for the uid (the column name that currently uniquely identifies all the subjects in your data).

Behind the scenes, this function is building a nested list that keeps track of all this information, but it is also:

• Converting the UID of the data into “subjectkey”
• Removing all observations that contain any NAs
• Removing all subjects who are not present in all input dataframes
• Arranging the subjects in all the dataframe by their UID
• Prefixing the UID values with the string “subject_” to help with cluster result characterization

Any rows containing NAs are removed. If you don’t want a bunch of your data to get removed because there are a few NAs sprinkled around here and there, consider using imputation. The mice package in R is nice for this.

Note that you do not need to name out every element explicitly. As long as you provide the objects within each list in the correct order (data, name, domain, type), you’ll get the correct result:

# Compactly:
data_list <- generate_data_list(
    list(Data1, "genes_1_and_2_exp", "gene_expression", "continuous"),
    list(Data2, "genes_1_and_2_meth", "gene_methylation", "continuous"),
    uid = "patient_id"
)

3. Store all the settings of the desired SNF runs in a settings_matrix

The settings_matrix is a dataframe where each row contains all the information required to convert the raw data into a final cluster solution. By varying the rows in this matrix, we can access a broader space of possible solutions and hopefully get closer to something that will be as useful as possible for our context.

In this case, we’re going to create only a single cluster solution using the same process outlined in the original SNFtool example above.

An explanation for the parameters in the settings_matrix can be found at the settings_matrix vignette.

settings_matrix <- generate_settings_matrix(
    data_list,
    nrow = 1,
    alpha_values = 0.5,
    k_values = 20,
    t_values = 20,
    dropout_dist = "none",
    possible_snf_schemes = 1
)

settings_matrix
#>   row_id alpha  k  t snf_scheme clust_alg cont_dist disc_dist ord_dist cat_dist
#> 1      1   0.5 20 20          1         2         1         1        1        1
#>   mix_dist inc_genes_1_and_2_exp inc_genes_1_and_2_meth
#> 1        1                     1                      1

The columns in this settings_matrix mean the following:

• row_id: A way to keep track of the different rows
• alpha, k, t: The hyperparameters seen above
• snf_scheme: Which “scheme” will be used to transform the inputs into a final fused network. We’ll discuss this in more detail in the next vignette.
• clust_alg: Which clustering algorithm will be applied to the final fused network. By default, one of two possible base options are randomly chosen. A value of 2 indicates that spectral clustering will be used and that the number of clusters will be determined by the rotation cost heuristic.
• Columns ending in “dist”: Which distance metric should be used. By default, 1 refers to Euclidean distance for continuous, discrete, and ordinal data, and 1 refers to Gower’s distance for categorical and mixed data.
• Columns starting with “inc”: Whether or not the corresponding dataframe will be included for this round of SNF.

More detailed descriptions on all of these columns can also be found in the settings_matrix vignette.

4. Run SNF

The batch_snf function will apply each row of the settings_matrix (in this case, just one row) to the data_list.

solutions_matrix <- batch_snf(
    data_list,
    settings_matrix
)

solutions_matrix[, 1:20] # it goes on like this for some time...
#>   row_id alpha  k  t snf_scheme clust_alg cont_dist disc_dist ord_dist cat_dist
#> 1      1   0.5 20 20          1         2         1         1        1        1
#>   mix_dist inc_genes_1_and_2_exp inc_genes_1_and_2_meth nclust subject_101
#> 1        1                     1                      1      2           1
#>   subject_102 subject_103 subject_104 subject_105 subject_106
#> 1           1           1           1           1           1

The solutions_matrix is essentially an augmented settings_matrix, where new columns have been added for each included patient. On each row, those new columns show what cluster that patient ended up in.

A friendlier format of the clustering results can be obtained:

cluster_solution <- get_cluster_df(solutions_matrix)

head(cluster_solution)
#>    subjectkey cluster
#> 1 subject_101       1
#> 2 subject_102       1
#> 3 subject_103       1
#> 4 subject_104       1
#> 5 subject_105       1
#> 6 subject_106       1

These cluster results are exactly the same as in the original SNF example:

identical(cluster_solution$"cluster", true_label)
#> [1] TRUE

Running batch_snf with the return_similarity_matrices parameter set to TRUE will let us also take a look at the final fused networks from SNF rather than just the results of applying spectral clustering to those networks:

batch_snf_results <- batch_snf(
    data_list,
    settings_matrix,
    return_similarity_matrices = TRUE
)

names(batch_snf_results)
#> [1] "solutions_matrix"    "similarity_matrices"

# The solutions_matrix
solutions_matrix <- batch_snf_results$"solutions_matrix"

# The first (and only, in this case) final fused network
similarity_matrix <- batch_snf_results$"similarity_matrices"[[1]]

The fused network obtained through this approach is also the same as the one obtained in the original example:

max(similarity_matrix - fused_network)
#> [1] 0

And now we’ve completed a basic example of using this package. The subsequent vignettes provide guidance on how you can leverage the settings_matrix to access a wide range of clustering solutions from your data, how you can use other tools in this package to pick a best solution for your purposes, and how to validate the generalizability.

Go give the less simple example a try!