This manuscript (permalink) was automatically generated from greenelab/connectivity-search-manuscript@9726869 on July 5, 2021.
Daniel S. Himmelstein
0000-0002-3012-7446
·
dhimmel
·
dhimmel
Department of Systems Pharmacology and Translational Therapeutics, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America; Related Sciences
· Funded by GBMF4552
Michael Zietz
0000-0003-0539-630X
·
zietzm
·
ZietzMichael
Department of Systems Pharmacology and Translational Therapeutics, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America; Department of Biomedical Informatics, Columbia University, New York, New York, United States of America
David N. Nicholson
0000-0003-0002-5761
·
danich1
Department of Systems Pharmacology and Translational Therapeutics, Perelman School of Medicine University of Pennsylvania, Philadelphia PA, USA
· Funded by The Gordon and Betty Moore Foundation (GBMF4552); The National Institutes of Health (T32 HG000046)
Casey S. Greene
0000-0001-8713-9213
·
cgreene
·
GreeneScientist
Department of Systems Pharmacology and Translational Therapeutics, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America; Department of Biochemistry and Molecular Genetics, University of Colorado School of Medicine, Aurora, Colorado, United States of America; Center for Health AI, University of Colorado School of Medicine, Aurora, Colorado, United States of America
· Funded by the National Human Genome Research Institute (R01 HG010067); the National Cancer Institute (R01 CA237170); the Gordon and Betty Moore Foundation (GBMF 4552)
Hetnets, short for “heterogeneous networks,” contain multiple node and relationship types and offer a way to encode biomedical knowledge. For example, Hetionet connects 11 types of nodes — including genes, diseases, drugs, pathways, and anatomical structures — with over 2 million edges of 24 types. Previously, we trained a classifier to repurpose drugs using features extracted from Hetionet. The model identified types of paths between a drug and disease that occurred more frequently between known treatments.
For many applications however, a training set of known relationships does not exist; Yet researchers would still like to learn how two nodes are meaningfully connected. For example, users may be curious not only how metformin is related to breast cancer, but also how the GJA1 gene might be involved in insomnia. Therefore, we developed hetnet connectivity search to propose the most important paths between any two nodes.
The algorithm behind connectivity search identifies types of paths that occur more frequently than would be expected by chance (based on node degree alone). We implemented the method on Hetionet and provide an online interface at https://het.io/search. Several optimizations were required to precompute significant instances of node connectivity at scale. We provide an open source implementation of these methods in our new Python package named hetmatpy.
To validate the method, we show that it identifies much of the same evidence for specific instances of drug repurposing as the previous supervised approach, but without requiring a training set.
A network (also known as a graph) is a conceptual representation of a group of entities — called nodes — and the relationships between them — called edges. Typically, a network has only one type of node and one type of edge. But in many cases, it is necessary to be able to distinguish between different types of entities and relationships.
A hetnet (short for heterogeneous information network [1]) is a network where nodes and edges have type. The ability to differentiate between different types of entities and relationships allows a hetnet to accurately describe more complex data. Hetnets are particularly useful in biomedicine, where it is important to capture the conceptual distinctions between various concepts, such as genes and diseases, or upregulation and binding.
The types of nodes and edges in a hetnet are defined by a schema, referred to as a metagraph. The metagraph consists of metanodes (types of nodes) and metaedges (types of edges). Note that the prefix meta is used to refer to type (e.g. compound), as opposed to a specific node/edge/path itself (e.g. acetaminophen).
Hetionet is a knowledge graph of human biology, disease, and medicine, integrating information from millions of studies and decades of research. Hetionet v1.0 combines information from 29 public databases. The network contains 47,031 nodes of 11 types (Table 1) and 2,250,197 edges of 24 types (Figure 1A).
Metanode | Abbr | Nodes | Description |
---|---|---|---|
Anatomy | A | 402 | Anatomical structures, excluding structures that are known not to be found in humans. From Uberon. |
Biological Process | BP | 11381 | Larger processes or biological programs accomplished by multiple molecular activities. From Gene Ontology. |
Cellular Component | CC | 1391 | The locations relative to cellular structures in which a gene product performs a function. From Gene Ontology. |
Compound | C | 1552 | Approved small molecule compounds with documented chemical structures. From DrugBank. |
Disease | D | 137 | Complex diseases, selected to be distinct and specific enough to be clinically relevant yet general enough to be well annotated. From Disease Ontology. |
Gene | G | 20945 | Protein-coding human genes. From Entrez Gene. |
Molecular Function | MF | 2884 | Activities that occur at the molecular level, such as “catalysis” or “transport.” From Gene Ontology. |
Pathway | PW | 1822 | A series of actions among molecules in a cell that leads to a certain product or change in the cell. From WikiPathways, Reactome, and Pathway Interaction Database. |
Pharmacologic Class | PC | 345 | “Chemical/Ingredient,” “Mechanism of Action,” and “Physiologic Effect” FDA class types. From DrugCentral. |
Side Effect | SE | 5734 | Adverse drug reactions. From SIDER/UMLS. |
Symptom | S | 438 | Signs and Symptoms (i.e. clinical abnormalities that can indicate a medical condition). From the MeSH ontology. |
Hetionet provides a foundation for building hetnet applications. It unifies data from several different, disparate sources into a single, comprehensive, accessible, common-format network. The database is publicly accessible without login at https://neo4j.het.io. The Neo4j graph database enables querying Hetionet using the Cypher language, which was designed to interact with networks where nodes and edges have both types and properties.
One limitation that restricts the applicability of Hetionet is incompleteness. In many cases, Hetionet v1.0 includes only a subset of the nodes from a given resource. For example, the Disease Ontology contains over 9,000 diseases [2], while Hetionet includes only 137 diseases [3]. Nodes were excluded to avoid redundant or overly specific nodes, while ensuring a minimum level of connectivity for compounds and diseases. See the Project Rephetio methods for more details [4]. Nonetheless, Hetionet v1.0 remains one of the most comprehensive and integrative networks that consolidates biomedical knowledge into a manageable number of node and edge types. Other integrative resources, some still under development, include Wikidata [5], SemMedDB [6,7,8], SPOKE, and DRKG.
Project Rephetio is the name of the study that created Hetionet and applied it repurpose drugs [4]. This project predicted the probability of drug efficacy for 209,168 compound–disease pairs. The approach learned which types of paths occur more or less frequently between known treatments than non-treatments (Figure 1B). To train the model, Rephetio created PharmacotherapyDB, a physician-curated catalog of 755 disease-modifying treatments [9].
Project Rephetio was able to successfully predict treatments, including those under investigation by clinical trail. However, two challenges limit the applicability of the Rephetio approach, which this study aims to address. First, Rephetio required known labels (i.e. treatment status) to train a model. Hence, the approach cannot be applied to domains where training labels do not exist. Second, the DWPC metric used to assess connectivity is sensitive to node degree. The Rephetio approach was incapable of detecting whether a high DWPC score indicated meaningful connectivity above the level expected by the background network degrees. Here we propose Hetnet connectivity search, which defines a null distribution for DWPCs that accounts for degree and enables detecting meaningful hetnet connectivity without training labels.
Existing research provides methods for determining whether two nodes are related, although primarily focuses on homogeneous networks (without type). Early approaches detected related nodes by measuring neighborhood overlap or path similarity between two nodes [11,12]. These approaches predicted node relatedness with success. However, they are difficult to scale as a network grow in size or semantic richness (i.e. type) [11].
More recently, focus has shifted to graph embeddings to determine if two nodes are related, specifically in the context of knowledge graphs, which are often semantically rich and include type [13,14,15,16,17]. These types of methods involve mapping nodes and sometimes edges to dense vectors via a neural network model [18,19,20], matrix factorization [21], or by translational distance models [22]. Once these dense vectors have been produced, quantitative scores that measure node relatedness can be generated via a machine learning model [14,23,24] or by selected similarity metrics [13,15,25,26,27]. These approaches have been quite successful in determining node relatedness. Yet, they only state whether two nodes are related and fail to provide an explanation on why two nodes are related.
Explaining why two nodes are related is a non-trivial task because approaches are required to output more information than a simple similarity score. The first group of approaches output a list of ranked paths are most relevant between two nodes [28,29,30,31]. For example, Ghazimatin et al. constructed a tool that provides an explanation for why items appear on a user’s social media feed [29]. The authors constructed an interaction graph, which is a heterogenous network of users and content classes (e.g. categories, user posts, songs). From this graph, they generated paths based on content timestamps and generated various features for each path. Using these generated features, the authors used a learn to rank model [32] to highlight the most relevant path between a user and the content of interest [29]. Besides providing a list of paths, another way to explain how two nodes are related is to provide a listing of sub-graphs for a given network [33]. However, this approach requires a weighted network to generate results [33].
TODO: touch on supervised versus unsupervised. TODO: differentiate hetnet connectivity search
TODO: integrate this content into Related Works.
https://github.com/greenelab/hetmech/issues/56
Network embeddings edge2vec [20] (cited above), metapath2vec [19] (cited above), HINE [34].
[35] training node pairs to important metapaths (Forward Stagewise Path Generation). MetaExp [30] user selects two sets of nodes. MetaExp detects metapaths and interacts with the user to progressively refine metapaths.
We created the connectivity search webapp available at https://het.io/search/. The tool is free to use, without any login or authentication. The purpose is let users quickly explore how any two nodes in Hetionet v1.0 might be related. The workflow is based around showing the user the most important metapaths and paths for a pair of query nodes.
The design guides the user through selecting a source and target node (Figure 2A). The webapp returns metapaths, scored by whether they occurred more than expected based on network degree (Figure 2B). Users can proceed by requesting the specific paths for each metapath, which are placed in a unified table sorted according to their path score (Figure 2C). Finally, the webapp produces publication-ready visualizations containing user-selected paths (Figure 2D).
We created the hetmatpy Python package, available on GitHub and PyPI under the permissive BSD-2-Clause Plus Patent License. This package provides a matrix-based utilities for hetnets.
TODO: improve flow and cohesion between software methods and results.
To assess connectivity between a source and target node, we use the DWPC (degree-weighted path count) metric. The DWPC is similar to path count (number of paths between the source and target node along a given metapath), except that it downweights paths through high degree nodes. Rather than using the raw DWPC for a source-metapath-target combination, we transform the DWPC across all source-target node pairs for a metapath to yield a distribution that is more compact and amenable to modeling [36].
Previously, we had no technique for detecting whether a DWPC value was exceptional. One possibility is to evaluate the DWPCs for all pairs of nodes and select the top scores (e.g. the top 5% of DWPCs). Another possibility is to pick a transformed DWPC score as a cutoff. The shortcomings of these methods are twofold. First, neither the percentile nor absolute value of a DWPC has inherent meaning. To select transformed DWPCs greater than 6, or alternatively the top 1% of DWPCs, is arbitrary. Second, comparing DWPCs between node pairs fails to account for the situation where high-degree node pairs are likely to score higher, solely on account of their degree (TODO: figure).
To address these shortcomings, we developed a method to compute the right-tail p-value of a DWPC. p-values have a broadly understood interpretation — in our case, the probability that a DWPC equal to or greater than the observed DWPC could occur under a null model. By tailoring the null distribution for a DWPC to the degree of its source and target node, we account for degree effects when determining the significance of a DWPC.
TODO: write this section
Figure 3 shows the information used to compute p-value for enriched metapaths. The table includes the following columns:
TODO: write this section
The paths webapp panel includes the following information (Figure 2C):
TODO: discuss path score and ranking by path score.
TODO: write this section.
TODO: determine what further use cases and examples we’d like to explore.
TODO: Assess ability to predict paths in https://github.com/SuLab/DrugMechDB
STUB: Contributions of this work:
STUB: Future work:
At the core of the hetmatpy package is the HetMat data structure for storing and accessing the network.
HetMats are stored on disk as a directory, which by convention uses a .hetmat
extension.
A HetMat directory stores a single heterogeneous network, whose data resides in the following files.
metagraph.json
file stores the schema, defining which types of nodes and edges comprise the hetnet.
This format is defined by the hetnetpy Python package.
Hetnetpy was originally developed with the name hetio during prior studies
[4,37],
but we renamed it to hetnetpy for better disambiguation from hetmatpy.nodes
directory containing one file per node type (metanode) that defines each node.
Currently, .tsv
files where each row represents a node are supported.edges
directory containing one file per edge type (metadata) that encodes the adjacency matrix.
The matrix can be serialized using either the Numpy dense format (.npy
) or SciPy sparse format (.sparse.npz
).For node and edge files, compression is supported as detected from .gz
, .bz2
, .zip
, and .xz
extensions.
This structure of storing a hetnet supports selectively reading nodes and edges into memory.
For example, a certain computation may only require access to a subset of the node and edge types.
By only loading the required node and edge types, we reduce memory usage and read times.
Additional subdirectories, such as path-counts
and permutations
, store data generated from the HetMat.
By using consistent paths for generated data, we avoid recomputing data that already exists on disk.
A HetMat directory can be zipped for archiving and transfer.
Users can selectively include generated data in archives.
Since the primary application of HetMats is to generate computationally demanding measurements on hetnets, the ability to share HetMats with precomputed data is paramount.
The HetMat
class implements the above logic.
A hetmat_from_graph
function creates a HetMat object and directory on disk from the pre-existing hetnetpy.hetnet.Graph
format.
We converted Hetionet v1.0 to HetMat format and uploaded the hetionet-v1.0.hetmat.zip
archive to the Hetionet data repository.
Prior to this study, we used two implementations for computing DWPCs.
The first is a pure Python implementation available in the hetnetpy.pathtools.DWPC
function [37].
The second uses a Cypher query, prepared by hetnetpy.neo4j.construct_dwpc_query
, that is executed by the Neo4j database [4,38].
Both of these implementations require traversing all paths between the source and target node.
Hence, they are computationally cumbersome despite optimizations [39].
Since our methods only require degree-weighted counts, not fully enumerated paths, adjacency matrix multiplication presents an alternative approach. Multiplication alone, however, counts walks rather than paths, meaning paths traversing a single node multiple times are counted. When computing network-based features to quantify the relationship between a source and target node, we would like to exclude traversing duplicate nodes (i.e. paths, not trails nor walks) [40]. To benefit from the speed advantages of only counting paths, we developed a suite of algorithms to compute true path counts and DWPCs using matrix multiplication.
Our implementation begins by categorizing a metapath according to the pattern of its repeated metanodes, allowing DWPC computation using a specialized order of operations. For example, the metapath DrDtCrC is categorized as a set of disjoint repeats, while DtCtDpC is categorized as repeats of the form BABA. Many complex repeat patterns can be represented piecewise as simpler patterns, allowing us to compute DWPC for most metapaths up to length 5 and many of length 6 and beyond without enumerating individual paths. For example, disjoint groups of repeats like DrDtCrC can be computed as the matrix product of DWPC matrices for DrD and CrC. Randomly-inserted non-repeated metanodes (e.g. G in DrDaGaDrD) require no special treatment, and are included in DWPC with a simple matrix multiplication.
After metapath categorization, we segment metapaths according to their repeat pattern, following our order of operations. By segmenting and computing recursively, we can evalute DWPC efficiently on highly complex metapaths, using simple patterns as building-blocks for higher-level patterns. Finally, our specialized DWPC functions are applied to individual segments, the results are combined, and final corrections are made to ensure no repeated nodes are counted. The recursive, segmented approach we developed allowed us additionally to implement a caching strategy that improved speed by avoiding duplicate DWPC computations. In summary, the functionality we developed resulted in greater than a 175-fold reduction in compute time, allowing us to compute millions of DWPC values across Hetionet [41].
DWPC computation requires us to remove all duplicate nodes from paths. We used three repeat patterns as the building blocks for DWPC computation: short repeats (AAA), nested repeats (BAAB), and overlapping repeats (BABA). Let \(\mathrm{D}(\mathit{XwXyZ})\) denote the DWPC matrix for metapath XwXyZ. Under this notation, \(\mathrm{D}(\mathit{XyZ})\) is the degree-weighted (bi)adjacency matrix for metaedge XyZ. Additionally, let \(\mathrm{diag}(A)\) represent a diagonal matrix whose entries are the diagonal elements of \(A\).
For the case of short (< 4) repeats for a single metanode, XaXbX (e.g. GiGdG), we simply subtract the main diagonal.
\[\mathrm{D}(\mathit{XaXbX}) = \mathrm{D}(\mathit{XaX}) \mathrm{D}(\mathit{XbX}) - \mathrm{diag}(\mathrm{D}(\mathit{XaX}) \mathrm{D}(\mathit{XbX}))\]
Nested repeats XaYbYcX (e.g. CtDrDtC), are treated recursively, with both inner (YY) and outer (XX) repeats treated as separate short repeats.
\[ \mathrm{D}(\mathit{XaYbYcX}) = \mathrm{D}(\mathit{XaY}) \mathrm{D}(\mathit{YbY}) \mathrm{D}(\mathit{YcX}) - \mathrm{diag}(\mathrm{D}(\mathit{XaY}) (\mathrm{D}(\mathit{YbY}) \mathrm{D}(\mathit{YcX})) \]
Overlapping repeats XaYbXcY (e.g. CtDtCtD) require several corrections (\(\odot\) denotes the Hadamard product).
\[\begin{align} \mathrm{D}(\mathit{XaYbXcY}) =\ &\mathrm{D}(\mathit{XaY})\ \mathrm{D}(\mathit{YbX})\ \mathrm{D}(\mathit{XcY}) \\ &- \mathrm{diag}(\mathrm{D}(\mathit{XaY})\ \mathrm{D}(\mathit{YbX}))\ \mathrm{D}(\mathit{XcY}) \\ &- \mathrm{D}(\mathit{XaY})\ \mathrm{diag}(\mathrm{D}(\mathit{YbX})\ \mathrm{D}(\mathit{XcY})) \\ &+ \mathrm{D}(\mathit{XaY})\ \odot \mathrm{D}(\mathit{YbX})^T\ \odot \mathrm{D}(\mathit{XcY}) \end{align}\]
Most paths of length six—and many even longer paths—can be represented hierarchically using these patterns. For example, a long metapath pattern of the form CBABACXYZ can be segmented as (C(BABA)C)XYZ using patterns for short and overlapping repeats and can be computed using the tools we developed. In addition to these matrix routines—which advantageously count rather than enumerate paths—we implemented a general matrix method for any metapath type. The general method is important for patterns such as long (≥ 4) repeats, or complex repeat patterns (e.g. of the form ABCABC), but it requires path enumeration and is therefore slower. As an alternative approach for complex paths, we developed an approximate DWPC method that corrects repeats in disjoint simple patterns but only corrects the first repeat in complex patterns (e.g. ≥ length four repeat). Mayers et al. developed an alternative approximation, which subtracts the main diagonal at every occurrence of the first repeated metanode [42]. All our matrix methods were validated against existing implementations involving explicit path enumeration to ensure consistent results.
In order to generate a null distribution for a DWPC, we rely on DWPCs computed from permuted hetnets. We derive permuted hetnets from the unpermuted network using the XSwap algorithm [43]. XSwap randomizes edges while preserving node degree. Therefore, it’s ideal for generating null distributions that retain general degree effects, but destroy the actual meaning of edges. We adapt XSwap to hetnets by applying it separately to each metaedge [4,44,45].
Project Rephetio created 5 permuted hetnets [4,44],
which were used to generate a null distribution of classifier performance for each metapath-based feature.
Here, we aim to create a null distribution for individual DWPCs, which requires vastly more permuted values to estimate with accuracy.
Therefore, we generated 200 permuted hetnets (archive).
More recently, we also developed the xswap
Python package, whose optimized C/C++ implementation will enable future research to generate even larger sets of permuted networks [45].
For each of the 200 permuted networks and each of the 2,205 metapaths, we computed the entire DWPC matrix (i.e. all source nodes × target nodes). Therefore, for each actual DWPC value, we computed 200 permuted DWPC values. Because permutation preserves only node degree, DWPC values among nodes with the same source and target degrees are equivalent to additional permutations. We greatly increased the effective number of permutations by grouping DWPC values according to node degree, affording us a superior estimation of the DWPC null distribution.
We have applied this degree-grouping approach previously when calculating the prior probability of edge existence based on the source and target node degrees [45,46]. But here, we apply degree-grouping to null DWPCs. The result is that the null distribution for a DWPC is based not only on permuted DWPCs for the corresponding source–metapath–target combination, but instead on all permuted DWPCs for the source-degree–metapath–target-degree combination.
The “# DWPCs” column in Figure 3 illustrates how degree-grouping inflates the sample size of null DWPCs. The p-value for the DaGiGpPW metapath relies on the minimum number of null DWPCs (200), since no other disease besides Alzheimer’s had 196 associates edges (source degree) and no other pathway besides circadian rhythm had 201 participates edges (target degree). However, for other metapaths with over 5,000 null DWPCs, degree-grouping increased the size of the null distribution by a factor of 25. In general, source–target node pairs with lower degrees receive the largest sample size multiplier from degree-grouping. This is convenient since low degree nodes also tend to produce the highest proportion of zero DWPCs, by virtue of low connectivity. Consequently, degree-grouping excels where it is needed most.
One final benefit of degree-grouping is that reduces the disk space required to store null DWPC summary statistics. For example, with 20,945 genes in Hetionet v1.0, there exists 438,693,025 gene pairs. Gene nodes have 302 distinct degrees for interacts edges, resulting in 91,204 degree pairs. This equates to an 4,810-fold reduction in the number of summary statistics that need to be stored to represent the null DWPC distribution for a metapath starting and ending with a Gene–interacts–Gene metaedge.
We store the following null DWPC summary statistics for each metapath–source-degree–target-degree combination: total number of null DWPCs, total number of nonzero null DWPCs, sum of null DWPCs, sum of squared null DWPCs, and number of permuted hetnets. These values are sufficient to estimate the p-value for a DWPC, by either fitting a gamma-hurdle null distribution or generating an empiric p-value. Furthermore, these statistics are additive across permuted hetnets. Their values are always a running total and can be updated incrementally as statistics from each additional permuted hetnet become available.
We are interested in identifying source and target nodes whose connectivity exceeds what typically arises at random. To identify such especially-connected nodes, we compare DWPC values to the distribution of permuted network DWPC values for the same source and target nodes. While a single DWPC value is not actually a test statistic, we use a framework akin to classical hypothesis testing to identify outliers.
Two observations led us to the quasi significance testing framework we developed. First, a sizable fraction of permuted DWPC values are often zero, indicating that the source and target nodes are not connected along the metapath in the permuted network. Second, we observed that non-zero DWPC values for any given source and target nodes are reasonably approximated as following a gamma distribution. Motivated by these observations, we parametrized permuted DWPC values using a zero-inflated gamma distribution, which we termed the gamma-hurdle distribution. We fit a gamma-hurdle distribution to each combination of source node, target node, and metapath. Finally, we estimate the probability of observing a permuted DWPC value greater than DWPC computed in the unpermuted network, akin to a one-tailed p-value. These quasi significance scores (‘p-values’) allow us to identify outlier node pairs at the metapath level (see examples in Figure 4).
Let X be a gamma-hurdle random variable with parameters λ, α, and β.
\[ X \sim \Gamma_H(\lambda, \alpha, \beta) \]
The probability of a draw from the distribution is
\[\begin{align} P(X = 0) &= 1 - \lambda \\ P(X \in A; A \subseteq (0, \infty)) &= \frac{\lambda \beta^\alpha}{\Gamma(\alpha)} \int _{x \in A} \bigg( x^{\alpha - 1} e^{-\beta x} \bigg) \end{align}\]
We estimate all three parameters using the method of moments (using Bessel’s correction to estimate the second moment). As a validation of our method, we compared our method of moments parameter estimates to approximate maximum likelihood estimates (gamma distribution parameters do not have closed-form maximum likelihood estimates) and found excellent concordance between the methods. Let N be the number of permuted DWPC values, and n the number of nonzero values.
\[\begin{align} \hat{\lambda} &= \frac{n}{N} \\ \hat{\alpha} &= \frac{(n - 1) \sum x_i}{n \sum (x_i^2) - (\sum x_i)^2} \\ \hat{\beta} &= \frac{n - 1}{n} \frac{(\sum x_i)^2}{n \sum (x_i)^2 - (\sum x_i)^2} \end{align}\]
Finally, we compute a p-value for each DWPC value, t.
\[ p = P(X ≥ t) = \frac{\beta^\alpha}{\Gamma(\alpha)} \int_t^\infty x^{\alpha - 1} \exp(-\beta x) dx \]
We calculate an empirical p-value for special cases where the gamma-hurdle model cannot be applied. These cases include when the observed DWPC is zero or when the null DWPC distribution is all zeroes or has only a single distinct nonzero value. The empirical p-value (pempiric) equals the proportion of null DPWCs ≥ the observed DWPC.
Since we don’t store all null DWPC values, we apply the following criteria to calculate pempiric from summary statistics:
We decided to compute DWPCs and their significance for all source–target node pairs for metapaths with length ≤ 3. On Hetionet v1.0, there are 24 metapaths of length 1, 242 metapaths of length 2, and 1,939 metapaths of length 3. The decision to stop at length 3 was one of practicality, as length 4 would have added 17,511 metapaths.
For each of the 2,205 metapaths, we computed the complete path count matrix and DWPC matrix (notebook). In total, we computed 137,786,767,964 path counts (and the same number of DWPCs) on the unpermuted network, of which 11.6% were nonzero.
The DWPC has a single parameter, called the damping exponent (w), which controls how much paths through high-degree nodes are downweighted [37]. When w = 0, the DWPC is equivalent to the path count. Previously, we found w = 0.4 was optimal for predicting disease-associated genes [37]. Here, we use w = 0.5, since taking the square root of degrees has more intuitive appeal.
We selected data types for matrix values that would allow for high precision. We used 64-bit unsigned integers for path counts and 64-bit floating-point numbers for DWPCs. We considered using 16-bits or 32-bits per DWPC to reduce memory/storage size, but decided against it in case certain applications required greater precision.
We used SciPy sparse for path count and DWPC matrices with density < 0.7, serialized to disk with compression and a .sparse.npz
extension.
This format minimizes the space on disk and load time for the entire matrix, but does not offer read access to slices.
We used Numpy 2D arrays for DWPC matrices with density ≥ 0.7, serialized to disk using Numpy’s .npy
format.
We bundled the path count and DWPC matrix files into HetMat archives by metapath length and deposited the archives to Zenodo [47].
The archive for length 3 DWPCs was the largest at 131.7 GB.
We also generated null DWPC summary statistics for the 2,205 metapaths,
which are also available by metapath length from Zenodo as HetMat archives consisting of .tsv.gz
files [47].
Due to degree-grouping, null DWPCs summary statistic archives are much smaller than the DWPC archives.
The archive for length 3 null DWPCs summary statistics was 733.1 MB.
However, the compute required to generate null DWPCs is far greater, because there are multiple permuted hetnets (in our case 200).
As a result, computing and saving all DWPCs took 6 hours,
whereas computing and saving the null DWPC summary statistics took 361 hours.
Including null DWPCs and path counts, the Zenodo deposit totals 185.1 GB and contains the results of computing ~28 trillion DWPCs — 27,832,927,128,728 to be exact.
When a user applies hetnet connecitivity search to identify enriched metapaths between two nodes, many metapaths are evaluated for significance. Due to multiple testing of many DWPCs, low p-values are likely to arise by chance. Therefore, we devised a multiple testing correction.
For each combination of source metanode, target metanode, and length, we counted the number of metapaths. For Disease…Pathway metapaths, there are 0 metapaths of length 1, 3 metapaths of length 2, and 24 metapaths of length 3. We calculated adjusted p-values by applying a Bonferroni correction based on the number of metapaths of the same length between the source and target metanode. Using Figure 3 as an example, the DdGpPW p-value of 5.9% was adjusted to 17.8% (multiplied by a factor of 3).
Bonferroni controls familywise error rate, which corresponds here to incorrectly finding that any metapath of a given length is enriched. As a result, our adjusted p-values are conservative. We would prefer to adjust p-values for false discovery rate [48], but these methods often require access to all p-values at once (impractical here) and assume a uniform distribution of p-values when there is no signal (not the case here when most DWPCs are zero).
Storing DWPCs and their significance in the database (as part of the PathCount
table in Figure 5) enables the connectivity search webapp to provide users with enriched metapaths between query nodes in realtime.
However, storing ~15.9 billion rows (the total number of nonzero DWPCs) in the database’s PathCount
table would exceed a reasonable disk quota.
An alternative would be to store all DWPCs in the database whose adjusted p-value exceeded a universal threshold (e.g. p < 5%).
But we estimated this would still be prohibitively expensive.
Therefore, we devised a metapath-specific threshold.
For metapaths with length 1, we stored all nonzero DWPCs,
assuming users always want to be informed about direct edges between the query nodes, regardless of significance.
For metapaths with length ≥ 2, we chose an adjusted p-value threshold of
5 × (nsource × ntarget)−0.3,
where nsource and ntarget are the node counts for the source and target metanodes (i.e. “Nodes” column in Table 1).
Notice that metapaths with large number of possible source–target pairs (large DWPC matrices) are penalized.
This decision is based on practicality, since otherwise the majority of the database quota would be consumed by a minority of metapaths between plentiful metanodes (e.g. Gene…Gene metapaths).
Also we assume that users will search nodes at a similar rate by metanode (e.g. they’re more likely to search for a specific disease than a specific gene).
The constants in the threshold formula help scale it.
The multiplier of 5 relaxes the threshold to saturate the available database capacity.
The −0.3 exponent applies the large DWPC-matrix penalty.
Users can still evaluate DWPCs that are not stored in the database, using either the webapp or API. These are calculated on-the-fly, delegating DWPC computation to the Neo4j database. Unchecking “precomputed only” on the webapp shows all possible metapaths for two query nodes. For some node pairs, the on-the-fly computation is quick (less than a second). Othertimes, computing DWPCs for all metapaths might take more than a minute.
We created a backend application using Python’s Django web framework.
The source code is available in the connectivity-search-backend
repository.
The primary role of the backend is to manage a relational database and provide an API for requesting data.
We define the database schema using Django’s object-relational mapping framework (Figure 5).
We import the data into a PostgreSQL database.
Populating the database for all 2,205 metapaths up to length 3 was a prolonged operation, taking over 3 days.
The majority of the time is spent populating the DegreeGroupedPermutation
(37,905,389 rows) and PathCount
(174,986,768 rows) tables.
The database is located at search-db.het.io
with public read-only access available.
We host a public API instance at https://search-api.het.io.
Version 1 of the API exposes several endpoints that are used by the connectivity search frontend including queries for
node details (/v1/node
),
node lookup (/v1/nodes
),
metapath information (/v1/metapaths
),
and path information (/v1/paths
).
The endpoints return JSON payloads.
Producing results for these queries relies on internal calls to the PostgreSQL relational database as well as the pre-existing Hetionet v1.0 Neo4j graph database.
They were designed to power the hetnet connectivity search webapp,
but are also available for general research use.
Todo:
The database only stores a single orientation of a metapath. For example, if GpPpGaD is stored between the given source and target node, DaGpPpG would not also be stored. Therefore, both orientations of a metapath are searched against the PathCount table.
TODO: write this section.
This study was conducted entirely in the open via public GitHub repositories. We used GitHub Issues for discussion, leaving a rich online history of the scholarly process. Furthermore, most additions to the analyses were performed by pull request, whereby a contributor proposes a set of changes. This provides an oppertunity for other contributors to review changes before they are officially accepted. For example, in greenelab/hetmech#156 @zietzm proposed a notebook to visualize parameters for null DWPC distributions. After @zietzm addressed @dhimmel’s comments, the pull request was approved and merged into the project’s main branch.
The manuscript for this study was written using Manubot, which allows authors to collaboratively write manuscripts on GitHub [49]. The Manubot-rendered manuscript is available at https://greenelab.github.io/connectivity-search-manuscript/. We encourage readers with feedback or questions to comment publicly via GitHub Issues.
This study primarily involves the following repositories:
The hetmech and hetionet repositories contain datasets related to this study. Large datasets were compressed and tracked with Git LFS (Large File Storage). GitHub LFS had a max file size of 2 GB. Datasets exceeding this size, along with other essential datasets, are available from Zenodo [47].