HPO2GO: prediction of human phenotype ontology term associations for proteins using cross ontology annotation co-occurrences

Mutual annotation-based prediction of protein domain functions with Domain2GO

Identifying unknown functional properties of proteins is essential for understanding their roles in both health and disease states. The domain composition of a protein can reveal critical information in this context, as domains are structural and functional units that dictate how the protein should act at the molecular level. The expensive and time-consuming nature of wet-lab experimental approaches prompted researchers to develop computational strategies for predicting the functions of proteins. In this study, we proposed a new method called Domain2GO that infers associations between protein domains and function-defining gene ontology (GO) terms, thus redefining the problem as domain function prediction. Domain2GO uses documented protein-level GO annotations together with proteins' domain annotations. Co-annotation patterns of domains and GO terms in the same proteins are examined using statistical resampling to obtain reliable associations. As a use-case study, we evaluated the biological relevance of examples selected from the Domain2GO-generated domain-GO term mappings via literature review. Then, we applied Domain2GO to predict unknown protein functions by propagating domain-associated GO terms to proteins annotated with these domains. For function prediction performance evaluation and comparison against other methods, we employed Critical Assessment of Function Annotation 3 (CAFA3) challenge datasets. The results demonstrated the high potential of Domain2GO, particularly for predicting molecular function and biological process terms, along with advantages such as producing interpretable results and having an exceptionally low computational cost. The approach presented here can be extended to other ontologies and biological entities to investigate unknown relationships in complex and large-scale biological data. The source code, datasets, results, and user instructions for Domain2GO are available at https://github.com/HUBioDataLab/Domain2GO. Additionally, we offer a user-friendly online tool at https://huggingface.co/spaces/HUBioDataLab/Domain2GO, which simplifies the prediction of functions of previously unannotated proteins solely using amino acid sequences.

CoMentG: comprehensive retrieval of generic relationships between biomedical concepts from the scientific literature

The CoMentG resource contains millions of relationships between terms of biomedical interest obtained from the scientific literature. At the core of the system is a methodology for detecting significant co-mentions of concepts in the entire PubMed corpus. That method was applied to nine sets of terms covering the most important classes of biomedical concepts: diseases, symptoms/clinical signs, molecular functions, biological processes, cellular compartments, anatomic parts, cell types, bacteria and chemical compounds. We obtained more than 7 million relationships between more than 74 000 terms, and many types of relationships were not available in any other resource. As the terms were obtained from widely used resources and ontologies, the relationships are given using the standard identifiers provided by them and hence can be linked to other data. A web interface allows users to browse these associations, searching for relationships for a set of terms of interests provided as input, such as between a disease and their associated symptoms, underlying molecular processes or affected tissues. The results are presented in an interactive interface where the user can explore the reported relationships in different ways and follow links to other resources. Database URL: https://csbg.cnb.csic.es/CoMentG/.

ASCARIS: Positional feature annotation and protein structure-based representation of single amino acid variations

Background

Genomic variations may cause deleterious effects on protein functionality and perturb biological processes. Elucidating the effects of variations is critical for developing novel treatment strategies for diseases of genetic origin. Computational approaches have been aiding the work in this field by modeling and analyzing the mutational landscape. However, new approaches are required, especially for accurate representation and data-centric analysis of sequence variations.

Method

In this study, we propose ASCARIS (Annotation and StruCture-bAsed RepresentatIon of Single amino acid variations), a method for the featurization (i.e., quantitative representation) of single amino acid variations (SAVs), which could be used for a variety of purposes, such as predicting their functional effects or building multi-omics-based integrative models. ASCARIS utilizes the direct and spatial correspondence between the location of the SAV on the sequence/structure and 30 different types of positional feature annotations (e.g., active/lipidation/glycosylation sites; calcium/metal/DNA binding, inter/transmembrane regions, etc.), along with structural features and physicochemical properties. The main novelty of this method lies in constructing reusable numerical representations of SAVs via functional annotations.

Results

We statistically analyzed the relationship between these features and the consequences of variations and found that each carries information in this regard. To investigate potential applications of ASCARIS, we trained variant effect prediction models that utilize our SAV representations as input. We carried out an ablation study and a comparison against the state-of-the-art methods and observed that ASCARIS has a competing and complementary performance against widely-used predictors. ASCARIS can be used alone or in combination with other approaches to represent SAVs from a functional perspective. ASCARIS is available as a programmatic tool at https://github.com/HUBioDataLab/ASCARIS and as a web-service at https://huggingface.co/spaces/HUBioDataLab/ASCARIS.

How to approach machine learning-based prediction of drug/compound-target interactions

The identification of drug/compound-target interactions (DTIs) constitutes the basis of drug discovery, for which computational predictive approaches have been developed. As a relatively new data-driven paradigm, proteochemometric (PCM) modeling utilizes both protein and compound properties as a pair at the input level and processes them via statistical/machine learning. The representation of input samples (i.e., proteins and their ligands) in the form of quantitative feature vectors is crucial for the extraction of interaction-related properties during the artificial learning and subsequent prediction of DTIs. Lately, the representation learning approach, in which input samples are automatically featurized via training and applying a machine/deep learning model, has been utilized in biomedical sciences. In this study, we performed a comprehensive investigation of different computational approaches/techniques for protein featurization (including both conventional approaches and the novel learned embeddings), data preparation and exploration, machine learning-based modeling, and performance evaluation with the aim of achieving better data representations and more successful learning in DTI prediction. For this, we first constructed realistic and challenging benchmark datasets on small, medium, and large scales to be used as reliable gold standards for specific DTI modeling tasks. We developed and applied a network analysis-based splitting strategy to divide datasets into structurally different training and test folds. Using these datasets together with various featurization methods, we trained and tested DTI prediction models and evaluated their performance from different angles. Our main findings can be summarized under 3 items: (i) random splitting of datasets into train and test folds leads to near-complete data memorization and produce highly over-optimistic results, as a result, should be avoided, (ii) learned protein sequence embeddings work well in DTI prediction and offer high potential, despite interaction-related properties (e.g., structures) of proteins are unused during their self-supervised model training, and (iii) during the learning process, PCM models tend to rely heavily on compound features while partially ignoring protein features, primarily due to the inherent bias in DTI data, indicating the requirement for new and unbiased datasets. We hope this study will aid researchers in designing robust and high-performing data-driven DTI prediction systems that have real-world translational value in drug discovery.

Tissue registration and exploration user interfaces in support of a human reference atlas

Seventeen international consortia are collaborating on a human reference atlas (HRA), a comprehensive, high-resolution, three-dimensional atlas of all the cells in the healthy human body. Laboratories around the world are collecting tissue specimens from donors varying in sex, age, ethnicity, and body mass index. However, harmonizing tissue data across 25 organs and more than 15 bulk and spatial single-cell assay types poses challenges. Here, we present software tools and user interfaces developed to spatially and semantically annotate ("register") and explore the tissue data and the evolving HRA. A key part of these tools is a common coordinate framework, providing standard terminologies and data structures for describing specimen, biological structure, and spatial data linked to existing ontologies. As of April 22, 2022, the "registration" user interface has been used to harmonize and publish data on 5,909 tissue blocks collected by the Human Biomolecular Atlas Program (HuBMAP), the Stimulating Peripheral Activity to Relieve Conditions program (SPARC), the Human Cell Atlas (HCA), the Kidney Precision Medicine Project (KPMP), and the Genotype Tissue Expression project (GTEx). Further, 5,856 tissue sections were derived from 506 HuBMAP tissue blocks. The second "exploration" user interface enables consortia to evaluate data quality, explore tissue data spatially within the context of the HRA, and guide data acquisition. A companion website is at https://cns-iu.github.io/HRA-supporting-information/ .

Machine learning-based prediction of drug approvals using molecular, physicochemical, clinical trial, and patent-related features

BACKGROUND

Drug development productivity has been declining lately due to elevated costs and reduced discovery rates. Therefore, pharmaceutical companies have been seeking alternative ways to determine and evaluate drug candidates.

RESEARCH DESIGN AND METHODS

In this work, we proposed a new computational approach to directly predict the regulatory approval of drug candidates, and implemented it as a method called 'DrugApp.' To accomplish this task, we employed multiple types of features including molecular and physicochemical properties of drug candidates, together with clinical trial and patent-related features, which are then processed by random forest classifiers to train our disease group-specific approval prediction models.

RESULTS

Our evaluations indicated DrugApp has a high and robust prediction performance. Within a use-case study, we showed our method can predict phase IV trial drugs that are later withdrawn from the market due to severe side effects. Finally, we used DrugApp models to forecast the approval of drug candidates that are currently in phases I/II/III of clinical trials.

CONCLUSIONS

We hope that our study will aid the research community in terms of evaluating and improving the process of drug development. The datasets, source code, results, and pre-trained models of DrugApp are freely available at https://github.com/HUBioDataLab/DrugApp.

SLPred: a multi-view subcellular localization prediction tool for multi-location human proteins

SUMMARY

Accurate prediction of the subcellular locations (SLs) of proteins is a critical topic in protein science. In this study, we present SLPred, an ensemble-based multi-view and multi-label protein subcellular localization prediction tool. For a query protein sequence, SLPred provides predictions for nine main SLs using independent machine-learning models trained for each location. We used UniProtKB/Swiss-Prot human protein entries and their curated SL annotations as our source data. We connected all disjoint terms in the UniProt SL hierarchy based on the corresponding term relationships in the cellular component category of Gene Ontology and constructed a training dataset that is both reliable and large scale using the re-organized hierarchy. We tested SLPred on multiple benchmarking datasets including our-in house sets and compared its performance against six state-of-the-art methods. Results indicated that SLPred outperforms other tools in the majority of cases.

AVAILABILITY AND IMPLEMENTATION

SLPred is available both as an open-access and user-friendly web-server (https://slpred.kansil.org) and a stand-alone tool (https://github.com/kansil/SLPred). All datasets used in this study are also available at https://slpred.kansil.org.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Investigation of Genetic Causes in Patients with Congenital Heart Disease in Qatar: Findings from the Sidra Cardiac Registry

Congenital heart disease (CHD) is one of the most common forms of birth defects worldwide, with a prevalence of 1–2% in newborns. CHD is a multifactorial disease partially caused by genetic defects, including chromosomal abnormalities and single gene mutations. Here, we describe the Sidra Cardiac Registry, which includes 52 families and a total of 178 individuals, and investigate the genetic etiology of CHD in Qatar. We reviewed the results of genetic tests conducted in patients as part of their clinical evaluation, including chromosomal testing. We also performed whole exome sequencing (WES) to identify potential causative variants. Sixteen patients with CHD had chromosomal abnormalities that explained their complex CHD phenotype, including six patients with trisomy 21. Moreover, using exome analysis, we identified potential CHD variants in 24 patients, revealing 65 potential variants in 56 genes. Four variants were classified as pathogenic/likely pathogenic based on the American College of Medical Genetics and Genomics and the Association for Molecular Pathology (ACMG/AMP) classification; these variants were detected in four patients. This study sheds light on several potential genetic variants contributing to the development of CHD. Additional functional studies are needed to better understand the role of the identified variants in the pathogenesis of CHD.

Quantitative dissection of multilocus pathogenic variation in an Egyptian infant with severe neurodevelopmental disorder resulting from multiple molecular diagnoses

Genomic sequencing and clinical genomics have demonstrated that substantial subsets of atypical and/or severe disease presentations result from multilocus pathogenic variation (MPV) causing blended phenotypes. In an infant with a severe neurodevelopmental disorder, four distinct molecular diagnoses were found by exome sequencing (ES). The blended phenotype that includes brain malformation, dysmorphism, and hypotonia was dissected using the Human Phenotype Ontology (HPO). ES revealed variants in CAPN3 (c.259C > G:p.L87V), MUSK (c.1781C > T:p.A594V), NAV2 (c.1996G > A:p.G666R), and ZC4H2 (c.595A > C:p.N199H). CAPN3, MUSK, and ZC4H2 are established disease genes linked to limb-girdle muscular dystrophy (OMIM# 253600), congenital myasthenia (OMIM# 616325), and Wieacker-Wolff syndrome (WWS; OMIM# 314580), respectively. NAV2 is a retinoic-acid responsive novel disease gene candidate with biological roles in neurite outgrowth and cerebellar dysgenesis in mouse models. Using semantic similarity, we show that no gene identified by ES individually explains the proband phenotype, but rather the totality of the clinically observed disease is explained by the combination of disease-contributing effects of the identified genes. These data reveal that multilocus pathogenic variation can result in a blended phenotype with each gene affecting a different part of the nervous system and nervous system-muscle connection. We provide evidence from this n = 1 study that in patients with MPV and complex blended phenotypes resulting from multiple molecular diagnoses, quantitative HPO analysis can allow for dissection of phenotypic contribution of both established disease genes and novel disease gene candidates not yet proven to cause human disease.

HPODNets: deep graph convolutional networks for predicting human protein-phenotype associations

MOTIVATION

Deciphering the relationship between human genes/proteins and abnormal phenotypes is of great importance in the prevention, diagnosis and treatment against diseases. The Human Phenotype Ontology (HPO) is a standardized vocabulary that describes the phenotype abnormalities encountered in human disorders. However, the current HPO annotations are still incomplete. Thus, it is necessary to computationally predict human protein-phenotype associations. In terms of current, cutting-edge computational methods for annotating proteins (such as functional annotation), three important features are (i) multiple network input, (ii) semi-supervised learning and (iii) deep graph convolutional network (GCN), whereas there are no methods with all these features for predicting HPO annotations of human protein.

RESULTS

We develop HPODNets with all above three features for predicting human protein-phenotype associations. HPODNets adopts a deep GCN with eight layers which allows to capture high-order topological information from multiple interaction networks. Empirical results with both cross-validation and temporal validation demonstrate that HPODNets outperforms seven competing state-of-the-art methods for protein function prediction. HPODNets with the architecture of deep GCNs is confirmed to be effective for predicting HPO annotations of human protein and, more generally, node label ranking problem with multiple biomolecular networks input in bioinformatics.

AVAILABILITY AND IMPLEMENTATION

https://github.com/liulizhi1996/HPODNets.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Protein domain-based prediction of drug/compound-target interactions and experimental validation on LIM kinases

Predictive approaches such as virtual screening have been used in drug discovery with the objective of reducing developmental time and costs. Current machine learning and network-based approaches have issues related to generalization, usability, or model interpretability, especially due to the complexity of target proteins' structure/function, and bias in system training datasets. Here, we propose a new method "DRUIDom" (DRUg Interacting Domain prediction) to identify bio-interactions between drug candidate compounds and targets by utilizing the domain modularity of proteins, to overcome problems associated with current approaches. DRUIDom is composed of two methodological steps. First, ligands/compounds are statistically mapped to structural domains of their target proteins, with the aim of identifying their interactions. As such, other proteins containing the same mapped domain or domain pair become new candidate targets for the corresponding compounds. Next, a million-scale dataset of small molecule compounds, including those mapped to domains in the previous step, are clustered based on their molecular similarities, and their domain associations are propagated to other compounds within the same clusters. Experimentally verified bioactivity data points, obtained from public databases, are meticulously filtered to construct datasets of active/interacting and inactive/non-interacting drug/compound-target pairs (~2.9M data points), and used as training data for calculating parameters of compound-domain mappings, which led to 27,032 high-confidence associations between 250 domains and 8,165 compounds, and a finalized output of ~5 million new compound-protein interactions. DRUIDom is experimentally validated by syntheses and bioactivity analyses of compounds predicted to target LIM-kinase proteins, which play critical roles in the regulation of cell motility, cell cycle progression, and differentiation through actin filament dynamics. We showed that LIMK-inhibitor-2 and its derivatives significantly block the cancer cell migration through inhibition of LIMK phosphorylation and the downstream protein cofilin. One of the derivative compounds (LIMKi-2d) was identified as a promising candidate due to its action on resistant Mahlavu liver cancer cells. The results demonstrated that DRUIDom can be exploited to identify drug candidate compounds for intended targets and to predict new target proteins based on the defined compound-domain relationships. Datasets, results, and the source code of DRUIDom are fully-available at: https://github.com/cansyl/DRUIDom.

Deep semi-supervised learning ensemble framework for classifying co-mentions of human proteins and phenotypes

Background

Identifying human protein-phenotype relationships has attracted researchers in bioinformatics and biomedical natural language processing due to its importance in uncovering rare and complex diseases. Since experimental validation of protein-phenotype associations is prohibitive, automated tools capable of accurately extracting these associations from the biomedical text are in high demand. However, while the manual annotation of protein-phenotype co-mentions required for training such models is highly resource-consuming, extracting millions of unlabeled co-mentions is straightforward.

Results

In this study, we propose a novel deep semi-supervised ensemble framework that combines deep neural networks, semi-supervised, and ensemble learning for classifying human protein-phenotype co-mentions with the help of unlabeled data. This framework allows the ability to incorporate an extensive collection of unlabeled sentence-level co-mentions of human proteins and phenotypes with a small labeled dataset to enhance overall performance. We develop PPPredSS, a prototype of our proposed semi-supervised framework that combines sophisticated language models, convolutional networks, and recurrent networks. Our experimental results demonstrate that the proposed approach provides a new state-of-the-art performance in classifying human protein-phenotype co-mentions by outperforming other supervised and semi-supervised counterparts. Furthermore, we highlight the utility of PPPredSS in powering a curation assistant system through case studies involving a group of biologists.

Conclusions

This article presents a novel approach for human protein-phenotype co-mention classification based on deep, semi-supervised, and ensemble learning. The insights and findings from this work have implications for biomedical researchers, biocurators, and the text mining community working on biomedical relationship extraction.

Computational Methods for Prediction of Human Protein-Phenotype Associations: A Review

Deciphering the relationship between human proteins (genes) and phenotypes is one of the fundamental tasks in phenomics research. The Human Phenotype Ontology (HPO) builds upon a standardized logical vocabulary to describe the abnormal phenotypes encountered in human diseases and paves the way towards the computational analysis of their genetic causes. To date, many computational methods have been proposed to predict the HPO annotations of proteins. In this paper, we conduct a comprehensive review of the existing approaches to predicting HPO annotations of novel proteins, identifying missing HPO annotations, and prioritizing candidate proteins with respect to a certain HPO term. For each topic, we first give the formalized description of the problem, and then systematically revisit the published literatures highlighting their advantages and disadvantages, followed by the discussion on the challenges and promising future directions. In addition, we point out several potential topics to be worthy of exploration including the selection of negative HPO annotations and detecting HPO misannotations. We believe that this review will provide insight to the researchers in the field of computational phenotype analyses in terms of comprehending and developing novel prediction algorithms.

DeepPheno: Predicting single gene loss-of-function phenotypes using an ontology-aware hierarchical classifier

Predicting the phenotypes resulting from molecular perturbations is one of the key challenges in genetics. Both forward and reverse genetic screen are employed to identify the molecular mechanisms underlying phenotypes and disease, and these resulted in a large number of genotype–phenotype association being available for humans and model organisms. Combined with recent advances in machine learning, it may now be possible to predict human phenotypes resulting from particular molecular aberrations. We developed DeepPheno, a neural network based hierarchical multi-class multi-label classification method for predicting the phenotypes resulting from loss-of-function in single genes. DeepPheno uses the functional annotations with gene products to predict the phenotypes resulting from a loss-of-function; additionally, we employ a two-step procedure in which we predict these functions first and then predict phenotypes. Prediction of phenotypes is ontology-based and we propose a novel ontology-based classifier suitable for very large hierarchical classification tasks. These methods allow us to predict phenotypes associated with any known protein-coding gene. We evaluate our approach using evaluation metrics established by the CAFA challenge and compare with top performing CAFA2 methods as well as several state of the art phenotype prediction approaches, demonstrating the improvement of DeepPheno over established methods. Furthermore, we show that predictions generated by DeepPheno are applicable to predicting gene–disease associations based on comparing phenotypes, and that a large number of new predictions made by DeepPheno have recently been added as phenotype databases.

Gene–phenotype associations can help to understand the underlying mechanisms of many genetic diseases. However, experimental identification, often involving animal models, is time consuming and expensive. Computational methods that predict gene–phenotype associations can be used instead. We developed DeepPheno, a novel approach for predicting the phenotypes resulting from a loss of function of a single gene. We use gene functions and gene expression as information to prediction phenotypes. Our method uses a neural network classifier that is able to account for hierarchical dependencies between phenotypes. We extensively evaluate our method and compare it with related approaches, and we show that DeepPheno results in better performance in several evaluations. Furthermore, we found that many of the new predictions made by our method have been added to phenotype association databases released one year later. Overall, DeepPheno simulates some aspects of human physiology and how molecular and physiological alterations lead to abnormal phenotypes.

Systematic identification of genetic systems associated with phenotypes in patients with rare genomic copy number variations

Copy number variation (CNV) related disorders tend to show complex phenotypic profiles that do not match known diseases. This makes it difficult to ascertain their underlying molecular basis. A potential solution is to compare the affected genomic regions for multiple patients that share a pathological phenotype, looking for commonalities. Here, we present a novel approach to associate phenotypes with functional systems, in terms of GO categories and KEGG and Reactome pathways, based on patient data. The approach uses genomic and phenomic data from the same patients, finding shared genomic regions between patients with similar phenotypes. These regions are mapped to genes to find associated functional systems. We applied the approach to analyse patients in the DECIPHER database with de novo CNVs, finding functional systems associated with most phenotypes, often due to mutations affecting related genes in the same genomic region. Manual inspection of the ten top-scoring phenotypes found multiple FunSys connections supported by the previous studies for seven of them. The workflow also produces reports focussed on the genes and FunSys connected to the different phenotypes, alongside patient-specific reports, which give details of the associated genes and FunSys for each individual in the cohort. These can be run in "confidential" mode, preserving patient confidentiality. The workflow presented here can be used to associate phenotypes with functional systems using data at the level of a whole cohort of patients, identifying important connections that could not be found when considering them individually. The full workflow is available for download, enabling it to be run on any patient cohort for which phenotypic and CNV data are available.

Expansion of the Human Phenotype Ontology (HPO) knowledge base and resources

The Human Phenotype Ontology (HPO)—a standardized vocabulary of phenotypic abnormalities associated with 7000+ diseases—is used by thousands of researchers, clinicians, informaticians and electronic health record systems around the world. Its detailed descriptions of clinical abnormalities and computable disease definitions have made HPO the de facto standard for deep phenotyping in the field of rare disease. The HPO’s interoperability with other ontologies has enabled it to be used to improve diagnostic accuracy by incorporating model organism data. It also plays a key role in the popular Exomiser tool, which identifies potential disease-causing variants from whole-exome or whole-genome sequencing data. Since the HPO was first introduced in 2008, its users have become both more numerous and more diverse. To meet these emerging needs, the project has added new content, language translations, mappings and computational tooling, as well as integrations with external community data. The HPO continues to collaborate with clinical adopters to improve specific areas of the ontology and extend standardized disease descriptions. The newly redesigned HPO website (www.human-phenotype-ontology.org) simplifies browsing terms and exploring clinical features, diseases, and human genes.

HPOLabeler: improving prediction of human protein–phenotype associations by learning to rank

Motivation

Annotating human proteins by abnormal phenotypes has become an important topic. Human Phenotype Ontology (HPO) is a standardized vocabulary of phenotypic abnormalities encountered in human diseases. As of November 2019, only <4000 proteins have been annotated with HPO. Thus, a computational approach for accurately predicting protein–HPO associations would be important, whereas no methods have outperformed a simple Naive approach in the second Critical Assessment of Functional Annotation, 2013–2014 (CAFA2).

Results

We present HPOLabeler, which is able to use a wide variety of evidence, such as protein–protein interaction (PPI) networks, Gene Ontology, InterPro, trigram frequency and HPO term frequency, in the framework of learning to rank (LTR). LTR has been proved to be powerful for solving large-scale, multi-label ranking problems in bioinformatics. Given an input protein, LTR outputs the ranked list of HPO terms from a series of input scores given to the candidate HPO terms by component learning models (logistic regression, nearest neighbor and a Naive method), which are trained from given multiple evidence. We empirically evaluate HPOLabeler extensively through mainly two experiments of cross validation and temporal validation, for which HPOLabeler significantly outperformed all component models and competing methods including the current state-of-the-art method. We further found that (i) PPI is most informative for prediction among diverse data sources and (ii) low prediction performance of temporal validation might be caused by incomplete annotation of new proteins.

Availability and implementation

http://issubmission.sjtu.edu.cn/hpolabeler/.

Contact

zhusf@fudan.edu.cn

Supplementary information

Supplementary data are available at Bioinformatics online.

HPO2Vec+: Leveraging heterogeneous knowledge resources to enrich node embeddings for the Human Phenotype Ontology

BACKGROUND

In precision medicine, deep phenotyping is defined as the precise and comprehensive analysis of phenotypic abnormalities, aiming to acquire a better understanding of the natural history of a disease and its genotype-phenotype associations. Detecting phenotypic relevance is an important task when translating precision medicine into clinical practice, especially for patient stratification tasks based on deep phenotyping. In our previous work, we developed node embeddings for the Human Phenotype Ontology (HPO) to assist in phenotypic relevance measurement incorporating distributed semantic representations. However, the derived HPO embeddings hold only distributed representations for IS-A relationships among nodes, hampering the ability to fully explore the graph.

METHODS

In this study, we developed a framework, HPO2Vec+, to enrich the produced HPO embeddings with heterogeneous knowledge resources (i.e., DECIPHER, OMIM, and Orphanet) for detecting phenotypic relevance. Specifically, we parsed disease-phenotype associations contained in these three resources to enrich non-inheritance relationships among phenotypic nodes in the HPO. To generate node embeddings for the HPO, node2vec was applied to perform node sampling on the enriched HPO graphs based on random walk followed by feature learning over the sampled nodes to generate enriched node embeddings. Four HPO embeddings were generated based on different graph structures, which we hereafter label as HPOEmb-Original, HPOEmb-DECIPHER, HPOEmb-OMIM, and HPOEmb-Orphanet. We evaluated the derived embeddings quantitatively through an HPO link prediction task with four edge embeddings operations and six machine learning algorithms. The resulting best embeddings were then evaluated for patient stratification of 10 rare diseases using electronic health records (EHR) collected at Mayo Clinic. We assessed our framework qualitatively by visualizing phenotypic clusters and conducting a use case study on primary hyperoxaluria (PH), a rare disease, on the task of inferring relevant phenotypes given 22 annotated PH related phenotypes.

RESULTS

The quantitative link prediction task shows that HPOEmb-Orphanet achieved an optimal AUROC of 0.92 and an average precision of 0.94. In addition, HPOEmb-Orphanet achieved an optimal F1 score of 0.86. The quantitative patient similarity measurement task indicates that HPOEmb-Orphanet achieved the highest average detection rate for similar patients over 10 rare diseases and performed better than other similarity measures implemented by an existing tool, HPOSim, especially for pairwise patients with fewer shared common phenotypes. The qualitative evaluation shows that the enriched HPO embeddings are generally able to detect relationships among nodes with fine granularity and HPOEmb-Orphanet is particularly good at associating phenotypes across different disease systems. For the use case of detecting relevant phenotypic characterizations for given PH related phenotypes, HPOEmb-Orphanet outperformed the other three HPO embeddings by achieving the highest average P@5 of 0.81 and the highest P@10 of 0.79. Compared to seven conventional similarity measurements provided by HPOSim, HPOEmb-Orphanet is able to detect more relevant phenotypic pairs, especially for pairs not in inheritance relationships.

CONCLUSION

We drew the following conclusions based on the evaluation results. First, with additional non-inheritance edges, enriched HPO embeddings can detect more associations between fine granularity phenotypic nodes regardless of their topological structures in the HPO graph. Second, HPOEmb-Orphanet not only can achieve the optimal performance through link prediction and patient stratification based on phenotypic similarity, but is also able to detect relevant phenotypes closer to domain expert's judgments than other embeddings and conventional similarity measurements. Third, incorporating heterogeneous knowledge resources do not necessarily result in better performance for detecting relevant phenotypes. From a clinical perspective, in our use case study, clinical-oriented knowledge resources (e.g., Orphanet) can achieve better performance in detecting relevant phenotypic characterizations compared to biomedical-oriented knowledge resources (e.g., DECIPHER and OMIM).