Discovering causal paths to diabetic nephropathy by combining computable biomedical knowledge with graph mining algorithms

Abstract

Purpose The present study aimed to discover the causal paths from diabetes to diabetic nephropathy (DN) from scientific texts by combining computable biomedical knowledge in SemMedDB with graph mining algorithms.

Methods A total of 12,662 triples were included in this study, containing 3,374 unique concepts and 44 semantic relations. We built a directed knowledge graph (KG) and then pruned it to a causal graph via word2vec word embeddings, semantic relations, and path length. Filtering thresholds were adjusted multiple times to find optimal causal paths and third variables. The paths and variables were validated by a nephrologist.

Results A path from diabetes to DN was sorted out, illustrating the key inducer of pathogenesis and two of the most noteworthy clinical outcomes. With the decrease of the directed causal score (S_di) from Quantile 95% to Quantile 75%, paths from diabetes to DN increased and third explicable variables and edges emerged additionally.

Conclusions This study developed an efficient causal path discovery approach to sort out the predominant path from pathogenesis to the manifestation of complex disorders.

Introduction

The biomedical literature contains rich information on potential causal mechanisms, and explanations of biomedical processes linking risk factors and diseases (1). Nonetheless, it is difficult to construct a causal graph relying on manually reading. Given the information overload of biomedical literature, there is an increasing need for computable biomedical knowledge buried in free text (2). Fortunately, the Semantic MEDLINE Database (SemMedDB) presented a comprehensive and computable dataset by extracting subject-predicate-object triples from PubMed titles and abstracts. The subject-predicate-object triple is the basic form for building a conceptual knowledge graph, which greatly reduced the abundancy of the unstructured text. In this paper, we focus on the causal relations by turning the conceptual knowledge graph into causal knowledge graph. A causal (cause-effect) relation is defined as an association between two entities (events) in which the first must occur before the second, and the first causes the second (3).

SemMedDB has been used to construct causal graphs by using manual selected causal predicates, such as causes, prevents, disrupts, and merge them with Electronic Health Records (EHR) for causal inference in observational research (4) (5). It also presents great potential for deriving known and unknown intermediates (1) and confounders (6) between the investigated exposure and outcome. Most recently, triples from SemMedDB are pruned gradually for generating causal associations between risk factors and chronic diseases (7), as well as extracting mechanistic associations that aggravate chronic conditions due to COVID-19 (8).

As one of the most common causes of morbidity and premature mortality, diabetes affects more than 440 million individuals worldwide and the number keeps increasing (9). With the progress of diabetes, multiple organs including cardiovascular system, foot, kidney, retina and peripheral nervous would be injured, and these complications are often reciprocal causation and further promote the injury of each other. For diabetes, there is a complex network of pathogenesis and manifestations, as well as numerous known and unknown intermediates and confounders influencing its progress to other complications. In other words, it is a representative case needing to systematically develop causal paths between diabetes and its complications.

In this paper, we define causal graph as a complex knowledge graph which consists of the header and tailor concepts and all the third variables linking them. Among the complications of diabetes, diabetic nephropathy (DN) is an important mediator, reflecting the diabetic microvascular damage, sharing the pathogenesis of retinopathy and strongly associated with the increased risk of cardiovascular events and mortality (9, 10). Hence, the present study focused on DN to discover the DN-related causal paths by combining computable biomedical knowledge from SemMedDB with graph mining algorithms.

Methods

1. Data collection and preprocessing

We used PubMed as the data repository and retrieved publications of DN associated risk factors by searching “(Diabetic Nephropathy) AND (Risk Factors)” (11). Next, SemRep (12), an automatic extracting tool to pull predicates from PubMed scientific texts, was utilized to extract triples and related provenance information of all retrieved abstract texts, including entities’ names and semantic types, unique PubMed article identifier (PMID), and source sentences of the triples. Also, entities and predicates were mapped to relevant information of Unified Medical Language System (UMLS) (13) semantic network. As a result, a total of 6,085 articles and 18,814 triples were acquired.

Firstly, 172 triples were removed because of duplicates and questionable triples (i.e., A-causes-A). Secondly, to integrate synonymous concepts in UMLS, we translated English concepts into Chinese concepts by using Google Translator and a correspondence table of Chinese and English clinical terms released by National Health Commission, Peoples republic of China. For concepts that are the same in Chinese and are in the same UMLS semantic type, we unified their English concept name with the most frequent entity name and replaced the corresponding UMLS concept unique id (CUI), which dropped 25 English concepts.

Thirdly, we removed 14 pregnancy related concepts (i.e., early pregnancy, pregnancy complications), as gestational diabetes is out of the scope of our study. Finally, a total of 12,662 triples were left for the present study, including 3,374 unique concepts and 44 UMLS semantic relations. Our dataset contains English subject, subject CUI, Chinese subject, predicate, English object, Chinese object, semantic type, source sentence of the triples, section of the triples’ location, and PMID.

2. Pruning a causal graph from knowledge graph

A directed KG was built from the cleaned triples, with subject concepts as source nodes, object concepts as target nodes, and semantic relations as edges connecting nodes. The network structure of KG was explored first by identifying network features (i.e., nodes’ count, edges’ count) and network density (i.e., subgraphs’ distribution) for optimal network prune strategies. Then we pruned the KG to a causal graph (Fig. 1).

Figure 1.

Workflow Chart

Edge We categorized predicates (edges) into causal and non-causal related predicates based on UMLS semantic network and relevant prior work (4) (7). Ultimately, we identified all functionally_related_to predicates, a category of predicates in the UMLS semantic network associated with functionality (i.e., causes, treats, process_of), together with augments, disrupts, inhibits, stimulates, and their neg_verbs as causal predicates, and the rest of predicates as non-causal predicates (i.e., is_a, degree_of).

Node Causal concepts were extracted based on the causal relatedness.

1. To be causal related, concepts ought to be related first. The word2vec neural network was used, which is an algorithm to calculate word embedding vectors with a textual corpus (source sentences) (14). Then, cosine similarity was calculated to measure the spatial distance between two embedding vectors of a triples’ subject-object pair, representing the semantic relatedness of terms. Specifically, a cosine similarity is computed by the cosine value of the angle between two vectors to identify whether two vectors have the same orientation (15). When two vectors point in the same directions with an included angle of 0, the cosine similarity is 1; when the included angle is 90 degrees, the cosine similarity is 0. Thus, the smaller the angles, the larger the cosine similarities implying closer semantic relatedness.

2. Beyond relatedness, concepts should be causal related. Given a specific subject-object pair, in which all triples have the same subject and object connected by multiple relations, the more causal predicates (edges), the more causal related they are. A direct relatedness score was calculated using Dr. Daowd’s direct relation computing method which integrate concepts’ similarity and the proportion of causal predicates among all predicates (7). The equation is shown as below.

   Specifically, S_direct may also be abbreviated as S_di, denoting the direct relatedness score between the source node (s) and the target node (t). Similarly, the similarity (s, t) represents the cosine similarity score between the pair of the source node (s) and target node (t). Overall, S_di is the product of two parts, the left part of the differential ratio and the right part of cosine similarity (s, t). For a given pair of source and target nodes (s, t), in the right position, the higher the similarity (s, t), the more semantic related the two concepts are. To go further to causal relatedness, on the left, for all edges of the given pair of nodes, Co() is a count function aiming at the difference count between causal edges and non-causal edges and the proportion ratio in the total count of causal edges. The higher the difference ratio, the more causal edges between two nodes. The equation integrates the association of two concepts from semantic relatedness to direct causal relatedness.

   $$Sdirect(s,t)=\frac{Co(\mathrm{causal})-Co(\mathrm{non}-causal)}{Co|(\mathrm{causal edges})}*(similarity(s,t))$$

3. Focusing on the pivotal third variables of the disease progression from diabetes mellitus (DM) to DN, we removed less-informative concepts such as disease and chronic disease.

Path A directed path is basically an ordered sequence of edges joining a sequence of distinct nodes (16). Path length indicates the number of edges in a path. To identify third intermediate variables, all paths from DM to DN were extracted. As previous experience suggested, the longer the path, the less informative it is (5). Hence, a path count distribution was plotted from 1 to 10 to define the appropriate path length. All paths were ranked in descending order in accordance with the path score, which was defined as the sum of direct relatedness score divided by the steps (edges) of the path.

Ultimately, the threshold of direct related score was adjusted for multiple times from Quantile 95% to Quantile 75% with 5% interval. Various sets of mediators were extracted from all filtered paths for comparison and validation.

Results

After data cleaning and preprocessing, an overall directed KG was built from 1932 unique nodes (concepts) and 9840 edges (triples). Similarities and S_di scores were then calculated. For example, the cosine similarity between cholesterol and triglycerides is up to 0.98, but the cosine similarity between cholesterol and kidney diseases is only 0.32. The S_di score between cholesterol and systolic blood pressure is 0.86, but the S_di score between cholesterol and microalbuminuria is 0.24. To extract causal paths from DM to DN, similarity threshold was set to 0.5 which indicates a good starting point of semantic relatedness, S_di ranged from Quantile 95% to Quantile 75% with the increment of 5% each time. Since 0 paths found with S_di of 95% or 90% quantile, we explored paths with S_di of Quantile 85 %, 80% and 75%. To select the path length, we explored the candidates of length via path count distribution from path length 1 to 10. As presented in Fig. 2, paths appeared until the length increasing to 4. Thus, we chose the median from 4 to 10, which was 7, as the path length in the following analyses, so as to achieve the feasible quantity of information.

Figure 2.

Paths count distribution (similarity > 0.5, S_di >= quantile 85%)

Generally, more intermediate nodes and paths emerged with the decrease of the levels of S_di.

Thresholds 1: similarity > 0.5, S_di >= Quantile 85% (0.63), length(path) =< 7

Intermediate nodes (nodes = 23, edges = 51, paths = 22): ‘Diabetes’, ‘Hyperglycemia’, ‘Blind Vision’, ‘Hospital admission’, ‘Disease of capillaries’, ‘Foot Ulcer, Diabetic’, ‘Diabetic Neuropathies’, ‘Complications of Diabetes Mellitus’, ‘Glycosylation End Products, Advanced’, ‘Injury’, ‘Lipids’, ‘Kidney Failure’, ‘Myocardial Infarction’, ‘Heart failure’, ‘Diabetic Nephropathy’, ‘End stage renal failure’, ‘Kidney Failure, Chronic’, ‘Hospitalization’, ‘Glomerulonephritis’, ‘Congestive heart failure’, ‘glomerulopathy’, ‘hypoglycemia’, ‘Kidney Diseases’

It is evident that “advanced glycosylation end products - cause (S_di = 0.93) - injury” is the strongest link in the causal path. However, no paths were detected with the threshold of 90% quantile of S_di, and the causal path was not formed until the S_di decreased to 85% quantile. This suggests the emergence of direct causal links of “diabetes causes (S_di = 0.67) hyperglycemia” and “myocardial infarction precedes (S_di = 0.69) DN” create the full path in Fig 3, which have S_di scores greater than 85% quantile (0.63) but smaller than 90% quantile (0.75). In other words, they are the weakest links in the path. This makes sense as “myocardial infarction precedes DN” is not convincing according to prior clinical knowledge and implies further optimization of the computing model.

Figure 3.

Causal path in thresholds 1

Thresholds 2: similarity > 0.5, S_di >= Quantile 80% (0.52), length(path) =< 7

New nodes found (nodes = 35, edges = 86, paths = 64): ‘Glucose’, ‘Nitric Oxide’, ‘Transforming Growth Factor beta’, ‘Reactive Oxygen Species’, ‘Genes’, ‘Angiotensin Receptor’, ‘Renal glomerular disease’, ‘IGA Glomerulonephritis’, ‘Contrast Media’, ‘NEPHROTOXICITY’, ‘Diabetes Mellitus, Insulin-Dependent’, ‘Obesity’

Thresholds 3: similarity > 0.5, S_di >= Quantile 75% (0.36), length(path) =< 7

New nodes found (nodes = 47, edges = 140, paths = 298): ‘follow-up’, ‘Proteinuria’, ‘Cardiovascular Diseases’, ‘Autonomic neuropathy’, ‘Vascular Diseases’, ‘Diabetic Retinopathy’, ‘receptor’, ‘Angiotensin II’, ‘C-reactive protein’, ‘Superoxides’, ‘Peptidyl-Dipeptidase A’, ‘Angiotensin I’

Discussion and Conclusion

This study developed an efficient causal path discovery approach to sort out the predominant path from pathogenesis to manifestation of the complex disorders. We successfully filtered the apparently causal relations from all triples and linked them together in accordance with the strength of causal relations and the length of causal paths. Recently, an evidence-based knowledge graph of the DM complications was constructed by manually extracting relevant disease, risk factors, risk outcomes, and other diabetes entities and the strength of the data for the odds ratio (OR) or relative risk (RR) correlations from clinical evidence (17). The approach in our study is automatic.

Although it is productive and fruitful to investigate the pathogenesis of diabetic complications and its strategy of management section by section, it is difficult to clarify the intricate network among the complications and generate a holistic view timely to care the patient and guide the next research. High quality systemic review and consensus of experts on clinical practice are classical way to aggregate the knowledge reliably. Nevertheless, because of the time- consuming process and the potential bias on literature selection, it is difficult to systematically keep up with the rapid increment and update of knowledge regarding the important and complicate issue, such as diabetes and its complications, especially in such a flourish era of biomedicine. In contrast, the present model in our manuscript efficiently aggregated the DN-centered knowledge and illustrated a holistic pathway from the key pathogenesis to the predominant relationship between DN and the fatal event.

It clearly pointed that glycation end production is an important inducer of the further injury in diabetes, and myocardial infarction is the most noteworthy clinical events among patients with DN and kidney failure. As validated by the clinical expert of kidney diseases, this pathway consists of the classical prior knowledge on the pathogenesis of diabetes and the progression and prognosis of DN, which validated the accuracy of the present model. Furthermore, results generated by the present model could help physicians and investigators to sort out the paramount issue from the complicate network of diabetes and its complications. By lowering the cut-off value of S_di, additional relationships appeared among the key nodes and all of them are explicable using priori biomedical knowledge. It provides additional information for physicians and investigators if they have additional focus surrounding the main pathway.

Collectively, the present informatic approach of causal paths discovery may also have scientific and applicatory implications in other complex diseases since in essence it is a complex causal knowledge graph by leveraging computable forms of biomedical knowledge.

Limitations

Though promising results have been detected, the present study has some limitations. First, due to restrictions of natural language processing techniques, SemRep is a broad-coverage cue words triggered extracting engine with 0.69 precision, 0.42 recall, and 0.52 F1 score in a relaxed evaluation (18). This implies SemRep can only extract limited biomedical relations compared to human-curated knowledge, and extracted results may contain some errors (i.e., trigger detection errors, 12.5%) (18). Second, evidence quality is not fully evaluated before detecting causal patterns. However, this effect caused by low-quality results may be diluted by mainstream results in our model. Third, science has inherent uncertainty, and some contradictory knowledge exists (19). For the next step, besides SemRep, we would combine other knowledge extraction engines (i.e., INDRA) (20) for comprehensive knowledge extraction. Further, we would identify contradictory knowledge by the same entities but contradict predicates and utilize argument mining techniques for evidence appraisal and contradictory knowledge interpretation.

Figures & Tables

Figure 4.

Causal path in thresholds 2

Figure 5.

Causal path in thresholds 3