Discovering disease-gene associations is an essential but challenging task in modern medicine. Within all the data-driven approaches targeting at this issue, literature-based knowledge discovery widely extends the discovering boundaries and uncovers implicit knowledge from unstructured textual data. However, most of the current literature-based methods require the involvement of specific expertise or prior knowledge. In this paper, we propose an adaptable and transferable methodology to 1) identify crucially genetic factors for a specific disease and 2) predict emerging genetic associations for the disease. Specifically, biomedical entities including diseases, chemicals, genes and genetic variations are extracted from literature data, then a heterogenous co-occurrence network is constructed and a semantic adjacency matrix is generated using the idea of Word2Vec. Following this, key genes and genetic variats are identified through centrality measurement on the network; emerging disease-gene associations are captured via a link prediction approach enhanced by the semantic matrix. We applied the proposed methodology to a literature dataset containing 54,219 scientific articles of atrial fibrillation (AF) to demonstrate its reliability. The results yielded a) crucial biomedical entities for AF highlighting five key gene groups and one potentially associated protein mutation; b) a list of emerging AF-genetic factors pairs that are worth in-depth exploration.
• Network algorithms • Social and professional topics
In modern medicine, deciphering disease-associated genes plays a
vital role in the diagnosis, treatment and prevention of diseases.
However, apart from the handful revealed molecular mechanisms
and disease pathogenesis, there is still a substantial amount of
disorders and abnormalities with their causes remaining
underneath the tip of the iceberg, especially for the process and
factors related to the inheritance and genetic basis. In the past few
years, many efforts have been addressed on exploring the genetic
basis of diseases. Although in the biomedical domain, genetic
linkage analysis [
In the past decades, researchers established various medical
ontologies and curated molecular networks to analyze and infer
molecular interactions for diseases based on accumulated
experimental and clinical experience [
The explosively increasing biomedical literature and thriving text
mining techniques provide a more open, real-time and economic
pathway to solve those issues [
The research framework of the proposed method is given in
polymorphism (SNP): DNA mutation refers to the permanent
change of a DNA sequence, protein mutation is the protein
encoded by a gene with mutation and SNP [
Working under the hypothesis that sentence-level occurrence indicates a strong association between entity pairs, we treated the extracted entities as nodes and assigned edges aligning with their sentence-level co-occurrence. In this way all the entities and their co-occurring relationships are transferred into a heterogeneous network, with its nodes representing entities and edges representing
sentence-level co-occurrence as the following adjacency matrix explains: = { ( , ) ( 0 − ) #(1) ( , )refers to the record frequency of sentence-level cowhere represents the th node in the th category, occurrence of and ; The graph representation of the heterogeneous network is: = ( , (+1 )) #(2)
2 where is the set of categories of entity nodes and is the set of ( + 1)/2
types of edges connecting different categories of nodes. 2.2
Enlightened by the idea of the Word2Vec model [
model which can transfer word into vectors by projecting word its one-hot representation into a lower dimension and largely reserve their semantic meanings. In our case, biomedical entities are regarded as words and their consecutive sequences form our training corpus. We select the Skip-Gram as our training algorithm since it offer better fit on small datasets, the training process of Skip-Gram can be concluded as: given an entity () in a corpus, the probabilities of its nearby entities in a certain window size w will be calculated based on the probability of the given entity ()
[
∑ −≤≤,≠0 log2 ( ( + )| ( ))) #(3) Through the Bioentity2Vec training, each entity would be represented as a fixed dimensional vector; we could then calculate the pairwise similarity of entities via cosine similarity and generate an adjacency similarity matrix : = cos ( , ) = ·
where is the corresponding vector of entity node .
2.3
2.3.1 Centrality Measurement
Centralities are a set of measurements evaluating nodes’ position
and importance within the network [
| | − 1 ∑=1 | | ∑ =1 #(6) where
is the topological distance from node to node ; Betweenness Centrality (BC): the betweenness centrality of a target node is the ratio of shortest paths between other node pairs that pass through the target node, it indicates the node’s potential to bridge other nodes in the network. In our network, a higher betweenness centrality reflects that the entity is highly likely to be an important connector or transmitter: 2 ∑,=1 | |
)is the number of all the shortest paths from node to and (
pass through node among all of them.
) is the number of paths that
To further comprehensively measure the entities’ importance
using the three centralities, non-dominated sorting is used to
combine the three centrality rankings for each entity category.
Non-dominated sorting is a
multiple-objective
optimization
method which re-rank the multi-dimensional scalable individuals
by dominating relationships of one individual against another,
after non-dominated sorting the individuals will be divided into
several consecutive Pareto fronts according to their domination
counts [
≠ ): presented in Figure 2.
for in : for node in : (
if ( )≥ ( )= [ ] = 0 ( )
( )
Output: Sorting results for categories
# Initialize the Domination Counts ( )≥ ( )
≥ = ( )
( )= ( ))==
[ ] += 1 ( ) : counts where ( )denotes the set of neighbor nodes of and ( , )is the co-occurring frequency of and , denotes the semantic similarity of and .
Using the genetic factors that haven’t co-occurred with the target disease before as our input, we could generate another final output through our link prediction approach: a ranking list of genetic factors with their corresponding predicting scores. The highly ranked ones are worth being validated by further biomedical experiments. 3
Atrial fibrillation (AF) is the most common form of cardiac
arrhythmia. The progress of AF is closely related to atrial size and
the extent of atrial fibrosis, both of which are affected by genetic
factors. Though several gene groups and mutations have been
linked to AF, clinical evidence and mechanistic explanations are
still far from being enough to integrate the knowledge of genetic
risk factors into clinical practice [
PubMed is a biomedical literature search engine which comprises more than 30 million citations from MEDLINE database, PMC citations and other online book resources. We used “atrial fibrillation” as the searching term in PubMed and refined the search results by restricting the fields “species” to “humans”, MeSH searching was adopted to promise precise AF-related search results and no restriction was applied to publication date. In all, 54,219 records were retrieved from the exact searching query: “("Atrial Fibrillation"[Mesh] AND Humans[Mesh])” 3.2
We exploited the Pubtator [
Using Pubtator API and we extracted 577,809 raw biomedical concepts from the 54,219 records. With the aid of MeSH and NCBI gene dictionary, we cleaned noise concepts and consolidated all the synonyms, generating
biomedical entities; furtherly we excluded those entities that never co-occurred with others before (i.e., the isolated nodes that are not connected to any other nodes in the network) and ended up with 5,838 nodes. The stepwise results are given in Table 1. 3.3
To identify the crucial biomedical entities in the AF progress, we respectively calculated the degree, closeness and betweenness centralities for all the nodes in our heterogeneous network, the gene entities in the top 20 of each centrality ranking list are given in Table 2. From the observation of their centrality characteristics, we classified the gene nodes into 5 groups and analyzed their topological features and node composition with the aid of NCBI gene database3 and biomedical literature investigation:
High Degree & Closeness & Betweenness Centralities: These topological features reflect the dominating positions of 3 More information could be found at https://www.ncbi.nlm.nih.gov/gene/
these nodes in the AF heterogeneous network. CRP, IL6, AGT, ACE, F2 and F10 are genes belong to this group; they are all early-discovered genes and have broad functioning ranges and massive interactions with other entities. Specifically, C-reactive protein (CRP) and interleukin 6 (IL6) are genes that function in inflammation reaction and the immune-related activities, with their encoding product’s levels associated with the prediction of a wide variety of cardiovascular events including AF; Angiotensin I converting enzyme (ACE) and angiotensinogen (AGT) are two chain functioning genes with the former encoding preangiotensinogen which would be cleaved by the angiotensin I converting enzyme encoded by the later, the product angiotensin II from this process is a significant protein in controlling the blood pressure (BP) and fluid-electrolyte balance, so both BP related symptoms like hypertension and electrolyte adjustment chemicals are associated with the two genes; coagulation factor II (F2) and coagulation factor X (F10) are genes that encode major coagulation factors to intermediate blood clotting and hemorrhagic conditions related to AF. Conclusively, activities of genes in this group may not directly result in AF but their functions engage the most primary and foundational molecular mechanisms in the progress of AF.
High Degree & Betweenness Centralities but low Close Centrality: From a graph theory perspective, we can interpret this group of gene nodes as the critical but localized controllers. This group of genes includes KCNA5 and SCN5A, both of which are also seed genes for AF with their associations with AF already revealed: KCNA5 and SCN5A respectively encode proteins for potassium and sodium voltage-gated channels, the loss or alternation of those channels’ function have a direct influence on the action potential and electrical activity of cardiomyocytes which may further lead to AF. For this group, gene nodes bridle the ion channel-related entities including symptoms and chemicals but one node majorly only covers one certain type of ion channel.
High Degree & Close Centralities but low Betweenness Centrality: This group of gene nodes are inclined to be central nodes in their sub-components with high independence. NPPB, FGB, COX8A, VWF and INS are genes in this group, in which NPPB encodes the cardiac hormone with its blood concentration indicating the heart failure; FGB encodes the beta component of fibrinogen with whose deficiency or mutation leading to afibrinogenemia; COX8A encodes the terminal enzyme of the respiratory chain related to ATP synthesis and cardiomyopathy; VWF encodes a glycoprotein involved in hemostasis; INS is the gene in charge of insulin synthesis which is the critical chemical in diabetes. We could conclude that genes in this group are most directly associated with the most common AF risk factors or complications rather than AF itself.
High Betweenness Centrality only: This group of genes including TBX5, SOX5 and PITX2 are less correlated to AF
compare with the aforementioned ones but the high betweenness centrality indicates their potential to connect AF with other entities. For example, TBX5 encodes transcription factor that is associated with heart developmental process and its mutation may result in a heart-affecting developmental disorder named Holt-Oram syndrome. High Closeness Centrality only: Topologically this group of gene it’s not the core nodes but still globally associated with the other AF entities. BID is the only gene in this group and it regulates cell’s apoptosis which is not a particular biological process for AF but generally correlated with other entities.
Moreover, with applying non dominated sorting to all the node
categories, we re-ranked the nodes for each category and
generated Table 3 to further identify the other crucial entities.
Apart from the genes which were evaluated before, we examined
the reliabilities of other core entities respectively by looking
through ClinVar [
The disease list presents the most common risk factors (Hypertension and Diabetes Mellitus), symptoms (Inflammation and Thrombosis), complications (Stroke and Heart Failure) and other correlating diseases (Hemorrhage, Fibrosis and Cerebral Infarction etc.) of AF, which are frequently reported in clinical cases. The chemical list highlights the regular treating drugs (Warfarin, Amiodarone, Digoxin, Verapamil, Quinidine, Aspirin, Apixaban, Sotalol, Heparin and Propafenone), the known pathological molecular mechanisms (Calcium, Sodium, Potassium and Magnesium channels) and risk factors (Ethanol and Glucose) of AF; Nitrogen and Oxygen are two other leading elements, resulting from the role of blood oxygen concentration as an indicating index and NOx as a risk factor in the progress of AF. To sum up, we regard the whole centrality measurement and nondominated sorting as a carding process for the AF-related biomedical entities, through which we not only captured the comprehensively critical biomedical entities but also gained clues on some potential AF-associated entities. 3.4
Before implementing the modified link prediction to our heterogeneous network, we performed a validating experiment on the rolled-back data to verify our algorithm’s usefulness on disease-gene association prediction. Two other link prediction algorithms were selected as our baselines: 1. 2.
The original resource allocation (RA): the original version of resource allocation index mentioned before; the co-occurring frequency (CF) weighted version of resource allocation: this version of RA adopted the same assumption with RA but uses weight ratio instead of degree proportion to calculate the resource diffusion, in our study edge weight is by the entity’s co-occurring frequency; The validation experiment was designed as follows: We rolled back our dataset by a five-year gap and constructed a corresponding network (i.e., rolled-back network), and the newly researched AF-linked genes or SNPs in the latest five years were collected as true labels. Then, two baselines, as well as our modified version, were applied to the rolled-back network to predict emerging links. The predictive results are a mixed ranking list of genes and SNPs with their corresponding predictive scores. If a gene or SNP in true labels was correctly predicted in the top ( is a selectable threshold according to predictive requirements, initially it was set as ) predicting list, it would count a true positive (TP), otherwise, it would constitute a false negative (FN). The outcomes are provided in Table 4.
= + From the results we can see that the modified link prediction approach method beat the other two baselines. This experiment fully proved the efficiency of our proposed method. Briefly, in the top 200 list, our algorithm successfully captured 74% of the correct genetic factors which would appear in the following five years, from an applying standpoint it largely reduces the necessary human workload by using the predictive shortlist to select candidate genetic factors. 3.5
We applied our proposed method to the heterogeneous network
and obtained a list of disease-gene pairs listed in Table 5. Then we
empirically searched evidence from literature to validate our
predication and found that the top 10 could all be linked according
to literature review. The predicted SNPs are attracting more
attention in the latest research of AF [
SNP | rs337711 SNP | rs11264280
Predicting Score Identifying disease-gene associations is a critical part of modern medicine. Increasing evidence reveals the strong links between genetics and human health, predicting those associations that will provide effective decision support for medical and clinical researchers. Compared to the earlier published version, we involved more categories of biomedical entities and centrality measurement approach to modify the research framework, through which we respectively added genetic variants into our predicting scope and identified critical genetic factors that function in the pathogenesis of the target disease.
This paper proposed a hybrid method for predicting the associations between diseases and genes from the standpoint of bibliometrics, based on the co-occurrence and semantic similarities of genetic factors and diseases. The advantages of our method are that we made use of text data in the latest published papers and took semantic relationships into consideration. By designing a cross-validation experiment, we compared our hybrid method with other classical ones and found that our method shows better performance. Furthermore, the predicted relationships are identified in the latest studies, proving the credibility and validity of our proposed method.
Apart from this case study on gene-related atrial fibrillation, the proposed method is expected to be applied to a broad range of investigations on discovering the relationships between genes and specific diseases. Such efforts could be expected to provide extensive and objective insights from global scientific articles to support decision making in related medical research and clinical practices, such as helping researchers identify unknown relationships between specific genes and diseases and propose effective treatments from globally published research, studies, and cases. There are also several limitations of this paper that may require further investigation in future studies: 1) Although we exploited the developed Pubtator as our biomedical entity extraction tool, there exist inevitably some false positives in the entity extraction process beyond our control, how to avoid the impact of using those toolkits is one of our concerns in the future; 2) we emphasized on the completeness of collecting strong association by adopting co-occurrence analysis, but this process would include some negative associations like “A is not associated with B”, in the future studies, sentiment analysis would be a promising approach to reduce the false positives. 3) Limited expertise is employed here to validate and explain our predicting results. In the future study, we plan to establish a cardiovascular specialist panel to provide expert instructions and interpretations of our outcomes.
An early version of this work has been published in the Proceedings of the 2020/2021 Portland International Center for Management of Engineering and Technology. This work is supported by the Australian Research Council under Discovery Early Career Researcher Award DE190100994.