Package 'genewalkR'

Description

Constructs the graph Laplacian and runs the spectral decomposition, then applies the chosen kernel transform. The result is stored as an external pointer for subsequent calls to diffuse().

Usage

build_kernel(
  object,
  kernel = c("regularised_laplacian", "commute_time", "diffusion", "inverse_cosine",
    "pstep"),
  kernel_params = params_kernel(),
  normalised = TRUE,
  strategy = "truncated",
  k = 100L,
  .verbose = TRUE
)

Arguments

Value

The DiffusionScores object with kernel_pointer and params populated.

Description

Aligns the embedding of graph 2 onto the reference embedding of graph 1 via orthogonal Procrustes, then computes per-node cosine similarities as a measure of context drift. Nodes shared between both graphs receive a cosine similarity in ⁠[-1, 1]⁠. Nodes exclusive to graph 1 are assigned -1.1 and nodes exclusive to graph 2 are assigned +1.1 as sentinel values indicating absence rather than a computed similarity.

Usage

calculate_drift(object)

Arguments

Value

The EmbedDrift object with aligned_embd and stats populated. stats is a data.table with columns node, cosine_similarity, and node_status ("shared", "g1_only", "g2_only").

Description

Calculates statistical significance for gene-pathway associations using the GeneWalk algorithm. For each gene-pathway pair, this function computes p-values and FDR corrections (both globally across all pairs and per-gene) across multiple network permutations. The final output includes mean values and 95% confidence intervals for all statistics.

The function processes all permutations in Rust for optimal performance, calculating cosine similarities between gene and pathway embeddings, determining p-values based on null distributions from randomised networks, and applying Benjamini-Hochberg FDR corrections.

Usage

calculate_genewalk_stats(object, .verbose = TRUE)

Arguments

Value

The input object with an added stats property containing a data.table with the following columns for each gene-pathway pair:

• gene - Gene identifier

• pathway - Pathway identifier

• avg_pval - Mean p-value across permutations

• pval_ci_lower, pval_ci_upper - 95% CI for p-values

• avg_global_fdr - Mean global FDR across permutations

• global_fdr_ci_lower,global_fdr_ci_upper - 95% CI for global FDR

• avg_gene_fdr - Mean gene-wise FDR across permutations

• gene_fdr_ci_lower,gene_fdr_ci_upper - 95% CI for gene-wise FDR

References

Ietswaart, et al., Genome Biol, 2021

Description

Distances between two sets of stored profiles, e.g. drugs against diseases. Lower means more similar. Ruiz et al. rank drug-disease pairs by the correlation distance on the multiscale interactome and by Canberra when comparing against gene expression signatures.

Usage

calculate_profile_distances(
  object,
  from,
  to,
  metric = c("correlation", "canberra", "l1", "l2", "cosine")
)

Arguments

Details

Every profile sums to 1 and the seed keeps roughly 1 - alpha of the mass on itself. What each metric weights:

• "correlation": Pearson on the full vector. Driven by the large entries, i.e. the immediate neighbourhood of the seed and hubs.

• "cosine": Without centring. Since all profiles sum to 1 and are mostly near zero, it ranks almost identically to "correlation".

• "l1": Twice the total variation distance between the two distributions. Less dominated by the top entries than the above.

• "l2": Dominated by the largest entries; favours big, well-connected nodes.

• "canberra": Relative differences per node, so the many near-zero entries count as much as the large ones. Useful when the profile is compared against a signature on a different scale.

Value

A numeric matrix of ⁠length(from) x length(to)⁠ distances.

References

Ruiz, Zitnik and Leskovec, Identification of disease treatment mechanisms through the multiscale interactome, Nat Commun 2021.

Description

Shows the degrees of the different edge types in the GeneWalk network. Lack of PPI interactions for example will make it difficult for the method to identify patterns.

Usage

check_degree_distribution(object)

Arguments

Description

Collapses duplicate edges into one. The constrained PageRank keeps parallel edges, so a pair listed twice (e.g. the same gene pair from two interaction sources, or A-B plus B-A in an undirected graph) gets twice the transition weight. Run this before DiffusionProfiles() unless that is what you want.

Usage

dedup_edges(
  graph_dt,
  node_dt = NULL,
  directed = FALSE,
  weight_agg = c("max", "sum", "mean"),
  .verbose = TRUE
)

Arguments

Value

A data.table with the columns "from", "to" and, if present in the input, "weight". All other columns are dropped. For undirected graphs each edge is stored once with from <= to (lexicographically). Endpoints are returned as character.

Description

Creates a pair of graphs with known topological differences to verify context-aware differential embedding analysis. The two graphs share a common backbone but differ in defined regions:

• Community 1 is a stable negative control, identical across both graphs.

• Community 2 contains a hub node demoted to a peripheral node in graph 2.

• Two bridge nodes span communities 2 and 3 in graph 1 but are fully embedded within community 3 in graph 2.

• A set of exclusive nodes appears in only one graph.

Usage

differential_graph_test_data(
  n_stable = 50L,
  n_comm2 = 50L,
  n_comm3 = 50L,
  n_exclusive = 5L
)

Arguments

Value

A named list containing:

• g1_edges - data.table with columns from, to for graph 1.

• g1_nodes - data.table with columns node, cluster for graph 1.

• g2_edges - data.table with columns from, to for graph 2.

• g2_nodes - data.table with columns node, cluster for graph 2.

• data_info - data.table with columns node, is_differential, node_status ("shared", "g1_only", "g2_only").

Description

Diffuses input scores over the precomputed kernel. Requires build_kernel() to have been called first.

Usage

diffuse_scores(
  object,
  input,
  background = NULL,
  method = "raw",
  n_perm = 1000L,
  seed = 42L
)

Arguments

Value

The DiffusionScores object with scores populated as a data.table with columns node and one score column per input.

Description

S7 class for the diffusion profiles of Ruiz et al.: one constrained personalised PageRank per seed over a heterogeneous graph. The seed is a source; every other node of a sink type absorbs the walker. With type weights, the walker at node i first picks a neighbouring node type proportional to its weight, then a neighbour of that type.

Usage

DiffusionProfiles(
  graph_dt,
  node_dt,
  sink_types,
  type_weights = NULL,
  directed = FALSE
)

Arguments

Value

An initialised DiffusionProfiles object.

Properties

graph_dt

data.table. The edge table with from, to and optionally weight.

node_dt

data.table. The node table with id and type.

sink_types

Character vector. Node types that act as sinks.

type_weights

Named numeric vector of weights per node type, or NULL for the plain random walk.

profiles

Numeric matrix of nodes x seeds. NULL until generate_profiles() is called.

params

Named list of the parameters used.

References

Ruiz, Zitnik and Leskovec, Identification of disease treatment mechanisms through the multiscale interactome, Nat Commun 2021.

Description

S7 class wrapping the Rust spectral diffusion backend. Holds the input graph, the precomputed kernel (as an external pointer), and the resulting per-node diffusion scores.

Usage

DiffusionScores(graph)

Arguments

Value

An initialised DiffusionScores object.

Properties

graph

An igraph object. The input graph.

kernel_pointer

External pointer to the Rust DiffusionKernel. NULL until build_kernel() is called.

scores

A data.table of diffusion results. Empty until diffuse() is called.

params

Named list of parameters used during kernel construction and diffusion.

Description

Download database from GitHub release

Usage

download_database()

Description

S7 class for context-aware differential node embedding analysis between two graphs. Holds both input graphs, their embeddings, the Procrustes-aligned embedding, and the resulting cosine similarity statistics.

Usage

EmbedDrift(graph_dt_1, graph_dt_2)

Arguments

Value

An initialised EmbedDrift object.

Properties

graph_dt_1

A data.frame containing the edges of graph 1. Must have columns from, to, and optionally weight.

graph_dt_2

A data.frame containing the edges of graph 2. Must have columns from, to, and optionally weight.

embd_1

Numeric matrix. Node embedding for graph 1, rows are nodes.

embd_2

Numeric matrix. Node embedding for graph 2, rows are nodes.

aligned_embd

Numeric matrix. Procrustes-aligned embedding of graph 1 onto graph 2, restricted to shared nodes.

stats

A data.frame holding per-node cosine similarity scores and node metadata after running the differential analysis.

params

Named list of parameters used during embedding and alignment.

Description

Compares embedding similarities against ground truth clusters.

Usage

evaluate_node2vec_test(embeddings, node_labels)

Arguments

Value

A list with:

• within_cluster_sim - Mean similarity within clusters.

• between_cluster_sim - Mean similarity between clusters.

• within_dist - The within cluster similarities as vector.

• between_dist - The between cluster similarities as vector.

Description

This function will generate the initial node2vec embedding with the provided parameters. In brief, it will generate random walks over the graph and fit a Skip Gram model for the number of defined epochs to generate on a per node basis an embedding vector. These will be subsequently used for the calculation of test statistics. For more details, please refer to Ietswaart, et al.

Usage

generate_initial_emb(
  object,
  embd_dim = 8L,
  n_graph = 3L,
  genewalk_params = params_genewalk(),
  directed = FALSE,
  seed = 42L,
  .verbose = TRUE
)

Arguments

Value

The object with the added initial embeddings.

References

Ietswaart, et al., Genome Biol, 2021

Description

Runs node2vec independently on both graphs using identical hyperparameters, storing the resulting embedding matrices in the object. Graph 1 serves as the reference for subsequent Procrustes alignment. For details on the node2vec algorithm, see Grover & Leskovec. For the approach generally speaking, see Amir, et al.

Usage

generate_initial_embeddings(
  object,
  embd_dim = 64L,
  node2vec_params = params_node2vec(),
  directed = FALSE,
  seed = 42L,
  .verbose = TRUE
)

Arguments

Value

The EmbedDrift object with embd_1, embd_2, and params populated.

References

Grover & Leskovec, arXiv, 2016
Amir, et al., Cancer Cell, 2025

Description

This function will generate generate permuted subgraphs that show the same node degree and run the node2vec algorithm on top of these subgraphs. The cosine distance between the real edges will be calculated and stored in the object. These will be used subsequently for statistical significance testing. For more details, please refer to Ietswaart, et al. The function will return a warning and the class as is if you have not run generate_initial_emb().

Usage

generate_permuted_emb(object, n_perm = 3L, seed = 42L, .verbose = TRUE)

Arguments

Details

The function will be using the same parameters for the node2vec training as were used for generate_initial_emb().

Value

The object with the added random embeddings for statistical testing.

References

Ietswaart, et al., Genome Biol, 2021

Description

Runs one constrained personalised PageRank per seed (Ruiz et al.). The graph is built once in Rust and the seeds run in parallel. Each call replaces the stored profiles, so pass drugs and diseases together if you want to compare them afterwards.

Usage

generate_profiles(
  object,
  seeds,
  diffusion_profile_params = params_diffusion_profiles(),
  .verbose = TRUE
)

Arguments

Value

The DiffusionProfiles object with profiles populated as a numeric matrix of nodes x seeds. Each column sums to 1.

Description

Class that stores keeps all of the important data for the Gene Walk algorithm, see Ietswaart, et al. The class can be subsequently used for to apply the Gene Walk algorithm.

Usage

GeneWalk(graph_dt, gene_to_pathway_dt, gene_ids, pathway_ids)

Arguments

Value

Returns the initialised GeneWalk initialised object for subsequent analysis.

Properties

graph_dt

A data.table that contains the initial graph.

gene_to_pathway_dt

A data.table that contains the gene to pathway connections.

gene_ids

The genes represented in your gene walk list.

pathway_ids

The pathways represented in your gene walk network.

embds

The initial embedding matrices generated by node2vec on top of the graph. Depends on the n_graph parameter you choose.

permuted_embd

A list containing cosine distances of permuted graphs..

stats

A data.table containing the information on the statistical testing across the permutations.

params

A list in which used parameters will be stored.

References

Ietswaart, et al., Genome Biol, 2021

Description

R6 class to build Gene Walk networks from internal database sources. Supports building a master network once and subsetting to multiple gene sets.

Public fields

Methods

Public methods

• GeneWalkGenerator$new()

• GeneWalkGenerator$print()

• GeneWalkGenerator$add_pathways()

• GeneWalkGenerator$add_ppi()

• GeneWalkGenerator$build()

• GeneWalkGenerator$create_for_genes()

• GeneWalkGenerator$reset_choices()

• GeneWalkGenerator$return_full_network_dt()

• GeneWalkGenerator$clone()

GeneWalkGenerator$new()

Initialise the generator

Usage

GeneWalkGenerator$new()

GeneWalkGenerator$print()

Print for GeneWalkGenerator

Usage

GeneWalkGenerator$print(...)

Arguments

GeneWalkGenerator$add_pathways()

Add pathway sources to the network

Usage

GeneWalkGenerator$add_pathways(
  source = c("go", "reactome"),
  go_namespace = c("biological_process", "molecular_function", "cellular_component")
)

Arguments

GeneWalkGenerator$add_ppi()

Add PPI sources to the network

Usage

GeneWalkGenerator$add_ppi(
  source = c("combined", "string", "signor", "reactome", "intact"),
  string_threshold = NULL,
  intact_threshold = NULL,
  intact_physical_only = FALSE
)

Arguments

GeneWalkGenerator$build()

Build the full network from selected sources

Usage

GeneWalkGenerator$build(.verbose = TRUE)

Arguments

GeneWalkGenerator$create_for_genes()

Create a gene-specific GeneWalk object

Usage

GeneWalkGenerator$create_for_genes(genes)

Arguments

Returns

Returns the initialised GeneWalk.

GeneWalkGenerator$reset_choices()

Resets the internal choices and erases any stored network data.

Usage

GeneWalkGenerator$reset_choices()

GeneWalkGenerator$return_full_network_dt()

Returns the full network data.table.

Usage

GeneWalkGenerator$return_full_network_dt()

Returns

data.table with the full internal network.

GeneWalkGenerator$clone()

The objects of this class are cloneable with this method.

Usage

GeneWalkGenerator$clone(deep = FALSE)

Arguments

Description

Get the aligned embedding

Usage

get_aligned_embd(object)

Arguments

Value

A numeric matrix of the Procrustes-aligned embedding of graph 1 onto graph 2, restricted to shared nodes.

Description

Downloads the DuckDB if not available and unzips it.

Usage

get_db_connection()

Value

The DuckDB connection to the internal data

Description

Get the list of embeddings

Usage

get_embedding(object)

Arguments

Value

If found, returns the embedding list.

Description

Get node embeddings

Usage

get_embeddings(object, which = NULL)

Arguments

Value

A matrix, or a named list with elements embd_1 and embd_2.

Description

Get the gene main information

Usage

get_gene_info(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the gene information

Description

Get the Gene Ontology

Usage

get_gene_ontology_hierarchy(head_only = FALSE)

Arguments

Details

The data is extracted from the obo file downloaded from GO OBO Foundry. Last download: ⁠March '26⁠.

Value

data.table with the GO to GO edges with edge type info.

Description

Get the Gene Ontology main information

Usage

get_gene_ontology_info(head_only = FALSE)

Arguments

Details

The data is extracted from the obo file downloaded from GO OBO Foundry. Last download: ⁠March '26⁠.

Value

data.table with the Gene Ontology information

Description

Get the gene to pathway data.table

Usage

get_gene_pathway_dt(object)

Arguments

Value

Returns the gene to pathway data.table stored in the object

Description

Get the Gene Ontology to gene connections

Usage

get_gene_to_go(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Gene Ontology to gene edges

Description

Get the Reactome to gene connections

Usage

get_gene_to_reactome(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Reactome to gene edges

Description

Get the edge data.table

Usage

get_graph_dt(object)

Arguments

Value

Returns the edge data.table stored in the object

Description

Get graph edge tables

Usage

get_graph_dts(object, which = NULL)

Arguments

Value

A data.frame, or a named list with elements graph_dt_1 and graph_dt_2.

Description

Get the full GW data stored in the class

Usage

get_gw_data(x)

## S3 method for class 'DataBuilder'
get_gw_data(x)

Arguments

Value

A data.table with the full data stored in the class.

Description

Get the filtered GW data for a list of genes

Usage

get_gw_data_filtered(x, gene_ids)

## S3 method for class 'DataBuilder'
get_gw_data_filtered(x, gene_ids)

Arguments

Value

A list with the following elements

• gwn - The final GeneWalk network.

• genes_to_pathways - The genes to pathway information.

• represented_genes - All included genes.

• represented_pathways - All included pathways.

Description

This returns a combined network akin to the workflow from Barrio-Hernandez et al. based on high quality edges from STRING, physical interactions from Intact and the full sets from SIGNOR, Reactome and Pathway Commons interactions.

Usage

get_interactions_combined(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the combined, collapsed network data.

References

Barrio-Hernandez et al., Nat Gen, 2023

Description

Get the Intact network

Usage

get_interactions_intact(
  threshold = NULL,
  physical_interactions = FALSE,
  head_only = FALSE
)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Intact network data.

Description

Get the pathway commons network

Usage

get_interactions_pc(head_only = FALSE)

Arguments

Details

This is version 14 of Pathway Commons. For details, check Rodchenkov, et al., Nucleic Acids Res., 2019

Value

data.table with the Pathway Commons network data.

Description

Get the Reactome network

Usage

get_interactions_reactome(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Reactome network data.

Description

Get the SIGNOR network

Usage

get_interactions_signor(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the SIGNOR network data.

Description

Get the STRING network

Usage

get_interactions_string(threshold = NULL, head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the String network data.

Description

Get diffusion profiles

Usage

get_profiles(object)

Arguments

Value

A numeric matrix of nodes x seeds.

Description

Get the Reactome pathway hierarchy

Usage

get_reactome_hierarchy(
  relationship = c("child_of", "descendant_of", "parent_of", "ancestor_of"),
  head_only = FALSE
)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Reactome to Reactome edges with edge type info.

Description

Get the Reactome main information

Usage

get_reactome_info(head_only = FALSE)

Arguments

Details

The data is extracted from the OpenTargets release 26.03

Value

data.table with the Reactome information

Description

Get scores

Usage

get_scores(object)

Arguments

Value

A data.table of scores.

Description

Get the statistical results

Usage

get_stats(object)

Arguments

Value

If found, the statistics results within the class

Description

Runs metapath2vec over a heterogeneous graph. Random walks are constrained to follow a metapath schema over node types, e.g. c("gene", "pathway", "gene"), and a skip-gram model is trained on them. With metapath_plus = TRUE, negative samples are drawn from the context node's own type (metapath2vec++).

Walks start only on nodes of the metapath's first type. Nodes no walk ever reaches keep their random initialisation, so by default they are removed and listed in the unvisited_nodes attribute. The function warns when more than 10% of walks are dropped or walks reach less than half the requested length on average: the schema does not fit the graph. Details are in the walk_stats attribute. A metapath over a single type warns as well; that is DeepWalk on a subgraph, so use node2vec().

Usage

metapath2vec(
  graph_dt,
  node_dt,
  metapath,
  embd_dim = 8L,
  metapath2vec_params = params_metapath2vec(),
  metapath_plus = FALSE,
  filter_unvisited = TRUE,
  directed = FALSE,
  seed = 42L,
  .verbose = TRUE
)

Arguments

Value

A numeric matrix of ⁠n_nodes x embd_dim⁠ with the node ids as rownames. Rows are ordered by node type. Carries three attributes:

• node_type - Named factor with the type of each row.

• unvisited_nodes - Character vector with the ids of the nodes no walk visited. Removed from the matrix if filter_unvisited = TRUE.

• walk_stats - List with start_nodes, attempted, truncated, dropped, mean_length and walk_length.

References

Dong, Chawla and Swami, metapath2vec: Scalable Representation Learning for Heterogeneous Networks, KDD 2017.

Description

The MYC target genes from the MSigDB (Hallmarks V1).

Usage

myc_genes

Format

myc_genes

A data.table with 200 rows and 3 columns:

ensembl_gene

Ensembl identifier

gene_symbol

Gene Symbol

gene_set

Name of the gene set

References

Liberzon, et al., Cell Syst, 2015

Description

Number of nodes represented

Usage

n_nodes(x)

## S3 method for class 'DataBuilder'
n_nodes(x)

Arguments

Value

Number of unique nodes stored in the class.

Description

Generate a new DataBuilder helper class

Usage

new_data_builder(ppis, gene_to_pathways, pathway_hierarchy)

Arguments

Value

Returns the initialised DataBuilder object.

Description

Runs the Node2Vec algorithm

Usage

node2vec(
  graph_dt,
  embd_dim = 8L,
  node2vec_params = params_node2vec(),
  directed = FALSE,
  seed = 42L,
  .verbose = TRUE
)

Arguments

Description

Creates graphs with known structure to verify node2vec is working correctly. The following options exists:

• "barbell" - Two dense clusters connected by single edge.

• "cavemen" - Multiple isolated cliques with sparse inter-connections.

• "stochastic_block" - Communities with controlled edge probabilities.

Usage

node2vec_test_data(
  test_type = c("barbell", "cavemen", "stochastic_block"),
  n_nodes_per_cluster = 10L,
  n_clusters = 3L,
  p_within = 0.8,
  p_between = 0.05,
  seed = 42L
)

Arguments

Value

A list containing:

Description

Parameters for generate_profiles(), i.e. the constrained personalised PageRank behind the diffusion profiles of Ruiz et al.

Usage

params_diffusion_profiles(alpha = 0.85, max_iter = 100L, tol = 1e-06)

Arguments

Value

A named list with the following elements:

• alpha - Numeric. Probability of continuing the walk instead of restarting at the seed. Defaults to 0.85.

• max_iter - Integer. Maximum number of power iterations. Defaults to 100L.

• tol - Numeric. Convergence threshold on the L1 change between two iterations. Defaults to 1e-06.

Description

Wrapper function for the GeneWalk-specific node2vec parameters

Usage

params_genewalk(
  p = 1,
  q = 1,
  walks_per_node = 100L,
  walk_length = 10L,
  batch_size = 256L,
  n_epochs = 5L,
  n_negatives = 5L,
  window_size = 1L,
  lr = 0.01,
  num_workers = 1L
)

Arguments

Value

A named list with the following elements:

• p - Numeric. Return parameter for biased random walks. Defaults to 1.0.

• q - Numeric. In-out parameter for biased random walks. Defaults to 1.0.

• walks_per_node - Integer. Number of random walks per node. Defaults to 100L.

• walk_length - Integer. Length of each random walk. Defaults to 10L.

• num_workers - Integer. Number of worker threads. Anything above 1L makes the fitting much faster, at the cost of reproducibility. Defaults to 1L.

• batch_size - Integer. Batch size for training. Defaults to 256L.

• n_epochs - Integer. Number of training epochs. Defaults to 5L.

• n_negatives - Integer. Number of negative samples. Defaults to 5L.

• window_size - Integer. Context window size. For GeneWalk you want embeddings from direct neighbours. Defaults to 1L.

• lr - Numeric. Learning rate. Defaults to 0.01.

Description

GeneWalk synthetic data params

Usage

params_genewalk_data(
  n_signal_genes = 150L,
  n_noise_genes = 150L,
  n_roots = 10L,
  depth = 3L,
  branching = 3L,
  p_lateral = 0.1,
  p_ppi = 0.05,
  min_annotations = 10L,
  max_annotations = 25L,
  min_noise_subtrees = 15L
)

Arguments

Value

A named list with the following elements:

• n_signal_genes - Integer. Number of signal genes. Defaults to 150L.

• n_noise_genes - Integer. Number of noise genes. Defaults to 150L.

• n_roots - Integer. Number of ontology root nodes. Defaults to 10L.

• depth - Integer. Depth of each ontology subtree. Defaults to 3L.

• branching - Integer. Base branching factor per ontology node. Defaults to 3L.

• p_lateral - Numeric. Probability of a lateral edge between sibling nodes. Defaults to 0.1.

• p_ppi - Numeric. Probability of a PPI edge between any two genes in the same group. Defaults to 0.05.

• min_annotations - Integer. Minimum ontology term annotations per gene. Defaults to 10L.

• max_annotations - Integer. Maximum ontology term annotations per gene. Defaults to 25L.

• min_noise_subtrees - Integer. Minimum distinct ontology subtrees a noise gene must span. Defaults to 15L.

Description

These are parameters for the graph diffusion kernels. Not all parameters are used by all kernels, please see Details.

Usage

params_kernel(sigma2 = 1, add_diag = 1, a = 3, p = 5L)

Arguments

Details

Parameter usage by kernel type:

• regularised_laplacian: sigma2, add_diag

• commute_time: none

• diffusion: sigma2

• inverse_cosine: none

• pstep: a, p

Value

A named list with the following elements:

• sigma2 - Numeric. Bandwidth parameter controlling the extent of diffusion. Used by regularised_laplacian and diffusion kernels. Defaults to 1.0.

• add_diag - Numeric. Regularisation term added to the diagonal. Used by regularised_laplacian. Must be positive for every connected component to be invertible. Defaults to 1.0.

• a - Numeric. Regularisation parameter for the pstep kernel. Must be ⁠>= 2⁠, since the eigenvalues of the normalised Laplacian sit in ⁠[0, 2]⁠. Defaults to 3.0.

• p - Integer. Number of steps for the pstep kernel. Must be positive. Defaults to 5L.

Description

Parameters for metapath2vec(). There is no p or q: metapath walks are first-order and follow the schema, not a biased return/explore rule.

Usage

params_metapath2vec(
  walks_per_node = 40L,
  walk_length = 40L,
  n_epochs = 20L,
  n_negatives = 5L,
  window_size = 2L,
  lr = 0.01,
  sample = 0.001,
  num_workers = NULL
)

Arguments

Value

A named list with the following elements:

• walks_per_node - Integer. Number of random walks per node of the metapath's starting type. Defaults to 40L.

• walk_length - Integer. Length of each random walk. Rounded up so the walk closes on a full number of schema cycles. Defaults to 40L.

• num_workers - Integer or NULL. Number of worker threads. If kept to NULL, it resolves to ⁠available cores - 2 (min 1)⁠. Defaults to NULL.

• n_epochs - Integer. Number of training epochs. Defaults to 20L.

• n_negatives - Integer. Number of negative samples. Defaults to 5L.

• window_size - Integer. Context window size. Defaults to 2L.

• lr - Numeric. Learning rate. Defaults to 0.01.

• sample - Numeric. Subsampling threshold for frequent nodes. Defaults to 0.001.

Description

These are general node2vec parameters, designed for speed. They will default to ⁠available cores - 2 (min 1)⁠.

Usage

params_node2vec(
  p = 1,
  q = 1,
  walks_per_node = 40L,
  walk_length = 40L,
  batch_size = 256L,
  n_epochs = 20L,
  n_negatives = 5L,
  window_size = 2L,
  lr = 0.01,
  num_workers = NULL
)

Arguments

Value

A named list with the following elements:

• p - Numeric. Return parameter for biased random walks. Defaults to 1.0.

• q - Numeric. In-out parameter for biased random walks. Defaults to 1.0.

• walks_per_node - Integer. Number of random walks per node. Defaults to 40L.

• walk_length - Integer. Length of each random walk. Defaults to 40L.

• num_workers - Integer or NULL. Number of worker threads. If kept to NULL, it resolves to ⁠available cores - 2 (min 1)⁠. Defaults to NULL.

• batch_size - Integer. Batch size for training. Defaults to 256L.

• n_epochs - Integer. Number of training epochs. Defaults to 20L.

• n_negatives - Integer. Number of negative samples. Defaults to 5L.

• window_size - Integer. Context window size. Defaults to 2L.

• lr - Numeric. Learning rate. Defaults to 0.01.

Description

Plots the GeneWalk results. Creates an interactive plotly plot to explore the results interactively.

Usage

plot_gw_results(object, fdr_treshold = 0.25)

Arguments

Value

ggplot2 object with the plot for the gene

Description

Plots the cosine similarity distributions of the permuted embeddings alongside the actual values as density histograms.

Usage

plot_similarities(object)

Arguments

Value

A ggplot2 object with the similarity distributions.

Description

Clean the cache and download the DB again

Usage

reload_db()

Description

Generate synthetic GeneWalk data with controlled signal structure

Usage

rs_build_synthetic_genewalk(
  n_signal_genes,
  n_noise_genes,
  n_roots,
  depth,
  branching,
  p_lateral,
  p_ppi,
  min_annotations,
  max_annotations,
  min_noise_subtrees,
  seed
)

Arguments

Value

A list with: ontology_edges, gene_ont_edges, ppi_edges, signal_genes, noise_genes.

Description

Computes cosine similarity between two numeric vectors.

Usage

rs_cosine_sim(a, b)

Arguments

Value

Cosine similarity between the two vectors.

Description

Produces two graphs sharing the same backbone but differing in defined regions. Community 1 is a stable negative control, identical across both graphs. Community 2 contains a hub node demoted to a peripheral node in graph 2. Two bridge nodes span communities 2 and 3 in graph 1 but are fully embedded within community 3 in graph 2. Exclusive nodes appear in only one graph.

Usage

rs_differential_graph_data(n_stable, n_comm2, n_comm3, n_exclusive)

Arguments

Value

A named list with: g1_edges (from, to), g1_nodes (node, cluster), g2_edges (from, to), g2_nodes (node, cluster), data_info (shared_nodes, is_differential, g1_only, g2_only).

Description

Run diffusion scoring on a precomputed kernel

Usage

rs_diffuse(
  kernel,
  scores,
  n_bkgd,
  n_inputs,
  bkgd_indices,
  method,
  n_perm,
  seed
)

Arguments

Value

Scores per given node

Description

Takes the CSC slots directly from R's dgCMatrix (igraph output).

Usage

rs_diffusion_kernel(
  i,
  p,
  x,
  n,
  kernel,
  kernel_params,
  normalised,
  strategy,
  k,
  node_names,
  verbose
)

Arguments

Value

External pointer to DiffusionKernel

Description

Constrained personalised PageRank over a heterogeneous graph (Ruiz et al., 2021). The graph is built once and one profile per seed is computed in parallel. The seed acts as a source; all other nodes of a sink type absorb mass.

Usage

rs_diffusion_profiles(
  node_types,
  from,
  to,
  weights,
  type_weight_names,
  type_weight_values,
  sink_types,
  seeds,
  directed,
  diffusion_profile_params
)

Arguments

Value

Numeric matrix of n_nodes x n_seeds. Columns sum to 1.

Description

Trains node2vec on the original GeneWalk network n_graph times with different random seeds, returning a list of embedding matrices. This provides the variance across embedding reps needed for the final log_stats aggregation.

Usage

rs_gene_walk(
  from,
  to,
  weights,
  gene_walk_params,
  n_graph,
  embd_dim,
  directed,
  seed,
  verbose
)

Arguments

Value

A list of n_graph numeric matrices, each n_nodes x embedding dim.

Description

For each permutation, generates a random graph via the configuration model (matching the original degree distribution), trains node2vec on it, then collects cosine similarities between each node and its unique neighbours in the random graph. This matches the original Python GeneWalk procedure.

Usage

rs_gene_walk_perm(
  from,
  to,
  weights,
  gene_walk_params,
  n_perm,
  embd_dim,
  directed,
  seed,
  verbose
)

Arguments

Value

A list of n_perm numeric vectors of null cosine similarities.

Description

Pools all null distributions into one, then for each observed embedding rep: computes p-values, per-gene FDR, and global FDR. Finally aggregates across reps with geometric mean and 95\

Matches the paper: "we repeat the above-described network representation learning and significance testing procedures of the GWN nreps_graph times and provide the mean and 95\

Usage

rs_gene_walk_test(
  gene_embds_list,
  pathway_embds_list,
  null_similarities,
  connected_pathways,
  verbose
)

Arguments

Value

A list with per-pair statistics.

Description

Pull out the node names from the kernel

Usage

rs_kernel_node_names(kernel)

Arguments

Value

A string vector from R.

Description

Runs metapath-constrained random walks over a heterogeneous graph and trains the skip-gram model on them (metapath2vec, or metapath2vec++ with per-type negative sampling).

Usage

rs_metapath2vec(
  node_ids,
  node_types,
  from,
  to,
  weights,
  metapath,
  metapath_plus,
  metapath2vec_params,
  embd_dim,
  directed,
  seed,
  verbose
)

Arguments

Value

A list with:

• embedding - Matrix of n_nodes x embd_dim, rows in node_names order.

• node_names - Node identifiers in row order.

• node_types - Node type per row.

• visited - Logical per row. FALSE if no surviving walk touched the node, i.e. its row is the random initialisation.

• walk_stats - Named list with the walk generation statistics.

Description

Trains node2vec on the original provided graph.

Usage

rs_node2vec(
  from,
  to,
  weights,
  node2vec_params,
  embd_dim,
  directed,
  seed,
  verbose
)

Arguments

Value

An embedding of dimension n_nodes x embedding dim.

Description

Generate synthetic data for node2vec

Usage

rs_node2vec_synthetic_data(
  test_data,
  n_nodes_per_cluster,
  n_clusters,
  p_within,
  p_between,
  seed
)

Arguments

Value

A list with edges and nodes.

Description

Aligns embd1 onto embd2 via orthogonal Procrustes (SVD) and returns cosine similarities per node. Both matrices must have identical dimensions and matching row order, i.e. row i in embd1 corresponds to row i in embd2.

Usage

rs_procrustes_align(embd1, embd2)

Arguments

Value

Named list with aligned (k x d matrix, embd1 rotated onto embd2) and cosine_similarities (numeric vector of length k).

Description

SIMD-accelerated distances between every column of mat_a and every column of mat_b.

Usage

rs_profile_distances(mat_a, mat_b, metric)

Arguments

Value

Numeric matrix of ncol(mat_a) x ncol(mat_b).

Description

Generates a synthetic network that mimics real GeneWalk input data for testing and benchmarking. The network consists of three components: a hierarchical ontology (tree structure with lateral edges), gene-to-ontology annotations, and protein-protein interactions (PPI).

Signal genes are annotated exclusively within a single ontology subtree and form PPI edges among themselves, creating coherent graph neighbourhoods. Noise genes receive the same number of annotations but are forced to span multiple subtrees, and their PPI edges are random across the noise group. Both groups are degree-matched so that significance is driven by annotation coherence rather than connectivity alone.

Usage

synthetic_genewalk_data(data_params = params_genewalk_data(), seed = 42L)

Arguments

Value

A list with the following items:

full_data

A data.table with columns from and to containing all edges (ontology hierarchy, gene-to-pathway, and PPI).

gene_to_pathways

A data.table with columns from (gene) and to (pathway) for gene-to-ontology annotation edges only.

gene_ids

Character vector of all gene identifiers (signal and noise).

pathway_ids

Character vector of all ontology term identifiers.