Source-linked AI summary
Drug Repurposing for COVID-19 via Knowledge Graph Completion
Rui Zhang, Dimitar Hristovski, Dalton Schutte, Andrej Kastrin, Marcelo Fiszman, Halil Kilicoglu
TL;DR
Effective COVID-19 therapies were not guaranteed, while developing new antivirals was risky, costly, and time-consuming. The paper combines literature-based discovery, knowledge graph completion, time-sliced evaluation, and discovery patterns to identify repurposing candidates and mechanistic hypotheses. It identifies known and novel candidate drugs and concludes that the approach can generalize beyond COVID-19.
Problem
Effective COVID-19 therapies were not guaranteed, and de novo antiviral development remained risky, costly, and time-consuming, motivating drug repurposing.
Method
The approach combines broad literature-derived biomedical knowledge, knowledge graph completion models, time slicing, and discovery patterns to identify candidate drugs and generate mechanistic explanations.
Results
The methods identified useful repurposing candidates, including known COVID-19-linked drugs and candidates not previously studied, while discovery patterns generated plausible mechanistic explanations.
Takeaways & Limitations
The approach can support computational drug repurposing for COVID-19 and can be generalized to other diseases and clinical questions.
Takeaways & Limitations
The approach depends on SemRep predication accuracy, and remaining extraction errors affected the knowledge graph completion task; model performance may also depend on hyperparameter tuning.
Abstract
from arXiv · showhide
Objective: To discover candidate drugs to repurpose for COVID-19 using literature-derived knowledge and knowledge graph completion methods. Methods: We propose a novel, integrative, and neural network-based literature-based discovery (LBD) approach to identify drug candidates from both PubMed and COVID-19-focused research literature. Our approach relies on semantic triples extracted using SemRep (via SemMedDB). We identified an informative subset of semantic triples using filtering rules and an accuracy classifier developed on a BERT variant, and used this subset to construct a knowledge graph. Five SOTA, neural knowledge graph completion algorithms were used to predict drug repurposing candidates. The models were trained and assessed using a time slicing approach and the predicted drugs were compared with a list of drugs reported in the literature and evaluated in clinical trials. These models were complemented by a discovery pattern-based approach. Results: Accuracy classifier based on PubMedBERT achieved the best performance (F1= 0.854) in classifying semantic predications. Among five knowledge graph completion models, TransE outperformed others (MR = 0.923, Hits@1=0.417). Some known drugs linked to COVID-19 in the literature were identified, as well as some candidate drugs that have not yet been studied. Discovery patterns enabled generation of plausible hypotheses regarding the relationships between the candidate drugs and COVID-19. Among them, five highly ranked and novel drugs (paclitaxel, SB 203580, alpha 2-antiplasmin, pyrrolidine dithiocarbamate, and butylated hydroxytoluene) with their mechanistic explanations were further discussed. Conclusion: We show that an LBD approach can be feasible for discovering drug candidates for COVID-19, and for generating mechanistic explanations. Our approach can be generalized to other diseases as well as to other clinical questions.
1. Introduction
COVID-19 created an urgent need for therapies, while drug repurposing offered a faster alternative to de novo antiviral development. The paper proposes integrating broad biomedical literature with COVID-19-focused knowledge to identify repurposing candidates.
- 1. Introduction: De novo antiviral development is risky, costly, and time-consuming, motivating efforts to repurpose existing drugs for COVID-19.Some repurposed drugs reached clinical trials, and remdesivir was approved by the FDA for hospitalized patients.
- 1. Introduction: Computational repurposing methods analyze biological and clinical data to generate hypotheses about approved or investigational drugs.Examples include drug signature matching, molecular docking, genome-wide association studies, and network analysis.
- 1. Introduction: Most computational approaches emphasize biological and SARS-CoV-2-specific data rather than the broader biomedical literature.The authors argue that effective and safe repurposing requires integrating literature-based knowledge with high-throughput biological data.
- 1. Introduction: The proposed approach treats COVID-19 drug repurposing as knowledge graph completion using PubMed, COVID-19 research literature, neural algorithms, and discovery patterns.The approach aims to identify known and novel candidate drugs while generating mechanistic explanations.
2. Related Work
Prior work has used biological data, networks, and literature-based discovery to prioritize COVID-19 drugs. This paper builds on these strands through literature-derived knowledge graphs and link prediction.
- 2. Related Work: Molecular docking studies screen compounds against host proteins in SARS-CoV-2 interactomes, while network studies use biological interaction and expression data.Reported candidates include ritonavir, ribavirin, remdesivir, and other compounds.
- 2. Related Work: Literature-based discovery automatically generates hypotheses by connecting concepts across otherwise separated research literatures.Its ABC model infers a potential A-C relationship from known A-B and B-C relationships; closed discovery instead proposes B to explain a known A-C relationship.
- 2. Related Work: Drug repurposing is a prominent biomedical application of literature-based discovery, including studies using semantic discovery patterns and SemMedDB relations.Previous work has identified potential therapeutic uses and candidate drugs for diseases including prostate cancer.
- 2. Related Work: Knowledge graph completion predicts unseen relations or entities from head-relation-tail triples using embedding, neural, and context-based models.The surveyed methods include translational models, semantic matching models, graph convolutional networks, and context-based encoders.
3. Materials and Methods
The workflow constructs a literature knowledge graph and applies knowledge graph completion and discovery-pattern methods to predict COVID-19 drug candidates and explain their relationships.
- 3. Materials and Methods: The methods section combines data preprocessing, knowledge graph completion, discovery patterns, and evaluation schemes in one workflow.The workflow is illustrated in Figure 1, and the source code and data are publicly available.
3.1. Data
The biomedical knowledge graph combines SemMedDB relations with SemRep-extracted relations from COVID-19 literature, expanding coverage beyond PubMed.
- 3.1. Data: SemMedDB provides semantic subject-predicate-object triples automatically extracted from biomedical literature using SemRep.Concepts are normalized to UMLS identifiers, enriched with semantic types, and linked to supporting articles and sentences.
- 3.1. Data: CORD-19 articles not included in PubMed were processed with SemRep to improve COVID-19 knowledge coverage.The study used the CORD-19 release dated 09/25/2020.
- 3.1. Data: 107 645 218 relations among 339 638 concepts came from SemMedDB, while processed CORD-19 data contributed 505 968 relations among 41 609 concepts.These sources differ substantially in both relation and concept counts.
3.2. Preprocessing
The preprocessing pipeline combines PubMed and CORD-19 semantic relations, removes generic or uninformative links, and retains an informative COVID-19-focused subset. It also improves relation precision with Transformer-based classifiers trained on annotated predications.
- The pipeline removed generic concepts, self-relations, and predicate types considered unsuitable for drug repurposing.
- 331 427 unique nodes and 20 017 236 relations formed the combined SemMedDB and CORD-19 knowledge graph.
- High-degree concepts and uninformative semantic relations were eliminated using network degree centrality and log-likelihood ratio.
- The final informativeness score combined normalized measures, with lower scores indicating more specific and informative relations.
- Transformer-based pretrained language models classified SemRep predications as correct or incorrect to improve relation precision.
- 6 492 predications were annotated and split into 80/10/10 training, validation, and test sets for classifier development.
3.3. Knowledge Graph Completion
The study compares translational, semantic-matching, and context-based knowledge graph completion models for ranking plausible triples. These models encode entities and relations in low-dimensional spaces or contextual representations and use model-specific scoring or classification objectives.
- Knowledge graph completion infers new triples by learning a scoring function that ranks true triples above false triples.
- Five models were explored: TransE, RotatE, DistMult, ComplEx, and STELP, spanning three knowledge graph completion classes.
- Translational models: TransE represents a triple through h + r ≈ t and scores plausibility with the distance ∥h + r − t∥.
- Translational models: RotatE was selected as a TransE counterpart because TransE cannot model one-to-many, many-to-one, and many-to-many relations effectively.
- Semantic matching models: DistMult models symmetric relations, whereas ComplEx extends it to complex embeddings that can represent asymmetric relations.
- Context-encoding models: STELP uses a Siamese BERT architecture with shared weights, combining triple classification and contrastive learning objectives.
3.4. Discovery patterns
Discovery patterns identify candidate drugs through constrained intermediate concepts and provide plausible mechanisms for drugs predicted by knowledge graph completion. The approach supports both open discovery and closed discovery.
- The study used discovery patterns to identify drugs for COVID-19 and to propose mechanisms for candidates generated by knowledge graph completion.
- The pattern links a drug to an intermediate concept and the intermediate concept to COVID-19, while excluding an existing drug-treats-COVID-19 relation.
- DrugA denotes a Pharmacologic Substance, COVIDConcept denotes selected UMLS COVID-19 concepts, and ConceptB may be any concept.
3.5. Evaluation
The evaluation combines time-sliced prediction testing, standard filtered link-prediction metrics, qualitative literature-based assessment, and comparisons with prior drug lists. It evaluates whether models can recover later COVID-19 drug proposals from earlier literature.
- Time slicing: Time slicing trained models on relations from publications dated 03/11/2020 or earlier and tested predictions against drugs proposed afterward.
- Link-prediction evaluation: The filtered evaluation corrupted each test triple by replacing its head and tail with other entities, excluding corrupted triples already present in the graph.
- Link-prediction evaluation: Performance was measured using mean rank, mean reciprocal rank, and Hits@k, with lower MR and higher MRR and Hits@k preferred.
- Qualitative evaluation: A medical-informatics author evaluated highly ranked drugs qualitatively through literature review and closed discovery, while open discovery directly evaluated discovery patterns.
- Comparative evaluation: Candidate lists were compared with one another and with drug lists from three prior studies, using top-ranked subsets from each method.
4. Results
PubMedBERT filtered semantic predications for knowledge graph construction, and TransE produced the strongest completion results. The resulting analysis identified plausible COVID-19 drug candidates and overlaps with other methods and published studies.
- Accuracy classifier: F1 = 0.854 was achieved by PubMedBERT, with recall = 0.895 and precision = 0.816, for semantic relation classification.
- Accuracy classifier: 1 907 717 of 2 558 935 predications, or 74.9%, were classified as correct and retained for downstream model training.
- Knowledge graph completion: The filtered graph contained 115 451 unique biomedical concepts and 1 907 717 relations.
- Knowledge graph completion: TransE outperformed all other models on average across every performance measure.The optimal configuration used k = 400 hidden dimensions, L1 norm, learning rate η = 0.01, and regularization coefficient λ = 2×10−8.
- Comparison of proposed drug lists: The t-SNE visualization showed relatively good co-localization of selected concepts, especially Suspected COVID-19 and paclitaxel.The visualization highlighted five drugs identified as potential new COVID-19 treatments and displayed the eight most frequent semantic types by color.
- Comparison of proposed drug lists: 33 of TransE’s top 150 drug predictions were judged plausible after manual analysis.Only estradiol overlapped with the drug list from Zeng et al.; TransE also predicted broader classes such as anthelmintics, mTOR inhibitors, and neuraminidase inhibitors.
- Comparison of proposed drug lists: TransE and STELP shared one drug class, RotatE and STELP shared five drugs, and DistMult and ComplEx shared no predictions with other methods.
- Comparison of proposed drug lists: Open discovery identified overlaps including estradiol with TransE, paclitaxel with RotatE, and hydrocortisone and indomethacin with Zeng et al.
5. Discussion
The discussion identifies model-performance trade-offs, computational differences, and candidate drugs whose literature-based patterns suggest mechanisms relevant to COVID-19.
- Knowledge graph completion models: TransE efficiently embedded the large-scale biomedical knowledge graph, while more complex models were not necessarily efficient for drug repurposing.Differences among DistMult, ComplEx, and RotatE were relatively small, with low MRR, Hits@1, and Hits@3 but moderate Hits@10.
- Knowledge graph completion models: STELP showed poor evaluation performance and mainly clustered ground-truth drugs into a few large groups.Its stronger performance on mixed corrupted triples suggests it learned predicate-level groupings rather than granular distinctions among subjects for treats-COVID-19.
- Knowledge graph completion models: TransE and STELP produced substantially different novel-relation rankings, with no significant correlation and only 47 common items among their top 1000.The Wilcoxon test reported p = 0.846, while Spearman’s rank correlation was −0.004.
- Knowledge graph completion models: TransE and RotatE trained in approximately 15 minutes, compared with 5 days for STELP, making the latter difficult to tune on limited resources.The authors note that the lighter models may enable exploration of larger graphs.
- Discovery patterns: The discovery-pattern approach identified five highly ranked drugs not previously discussed for COVID-19 and also ranked several drugs evaluated in clinical trials.The five novel candidates were paclitaxel, SB 203580, alpha 2-antiplasmin, pyrrolidine dithiocarbamate, and butylated hydroxytoluene.
- Discovery patterns: Discovery patterns supplied mechanistic hypotheses linking candidates to inflammatory, coagulation, and immune pathways relevant to COVID-19.Paclitaxel and SB 203580 were associated with potential inhibition of hyperinflammatory responses; alpha 2-antiplasmin was linked to plasmin, fibrinogen, and IgY mechanisms.
- Discovery patterns: Metoclopramide and oxymatrine were connected to potential anti-inflammatory mechanisms through cholinergic, TNF, cytokine, NF-kappa B, and TLR4 relationships.The proposed explanations concern suppression or regulation of inflammatory cytokine responses associated with COVID-19.
6. Conclusion
The study combines literature-based discovery with knowledge graph completion to identify COVID-19 drug candidates and generate plausible mechanistic explanations. The approach uses broader biomedical literature and can extend beyond COVID-19 to other diseases and clinical questions.
- The approach combines literature-based discovery and knowledge graph completion to identify candidate drugs for COVID-19.
- Knowledge graph models identify candidate drugs, while discovery patterns provide plausible mechanistic explanations.
- The approach incorporates a wider range of biomedical literature rather than focusing largely on COVID-19-specific knowledge.
- The approach can be used for drug repurposing in other diseases and for clinical questions such as drug-drug interactions or adverse effects.
- The approach is presented as an efficient computational method for leveraging existing biomedical knowledge during global health crises.
Funding
The study received support from the U.S. National Institutes of Health and the Slovenian Research Agency.
- Funding came from the U.S. National Institutes of Health and the Slovenian Research Agency through grants supporting the listed authors.