Skip to main content
Human Genomics logoLink to Human Genomics
. 2026 Aug 14;20:159. doi: 10.1186/s40246-026-01023-9

GenPhenia: using deep neural networks to accelerate rare-disease diagnosis

Franco Gino Brunello 1,✉, Gianlucca Colangelo 1, Ana Rius 1, Lorenzo Erra 1, Ana Clara Lugones 2, Nicolás Aguirre 2, Germán Biagioli 2, Jonathan Zaiat 1, Adrián Turjanski 2, Marcelo A Martí 1,✉
PMCID: PMC13757151  PMID: 43015840

Abstract

Background

Traditional molecular diagnosis of rare genetic diseases often struggles with complex genotype–phenotype relationships, such as pleiotropy and locus heterogeneity. Despite advances in next-generation sequencing, variant interpretation remains a major bottleneck, frequently requiring labor-intensive expert review to establish phenotypic specificity. To address this challenge, we developed GenPhenia, a deep learning–based framework designed to assist phenotype-driven gene prioritization using clinical features encoded with the Human Phenotype Ontology (HPO).

Results

GenPhenia leverages large-scale synthetic clinical history datasets and a graph neural network architecture with message-passing strategies combined with BERT-based phenotype embeddings to learn phenotype–gene associations. The model was validated using both ClinVar-derived datasets and real-world clinical cohorts. On the challenging MCRD benchmark dataset, GenPhenia achieved 60% Top-1 recall and 80% Top-10 recall, substantially outperforming previously reported methods, which achieved less than 30% Top-10 recall under the same conditions. Clinical case studies across endocrinology, cardiology, and immunology further demonstrate that the model effectively refines gene rankings as phenotypic specificity increases during the diagnostic process.

Conclusions

GenPhenia provides a scalable and highly sensitive computational approach for phenotype-driven gene prioritization that aligns with the iterative nature of clinical phenotyping and genomic diagnosis. The framework may facilitate more efficient variant interpretation and support the integration of artificial intelligence methods into clinical genomics workflows for rare disease diagnosis.

Graphical Abstract

graphic file with name 40246_2026_1023_Fig3_HTML.webp

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40246-026-01023-9.

Keywords: Human Phenotype Ontology, Gene Prioritization, Rare Disease Diagnosis, Graph Neural Networks, Deep Learning, Phenotype-Driven Analysis, Next-Generation Sequencing, Clinical Genomics

Background

Before the advent of next-generation sequencing (NGS), clinical interpretation of genetic variants relied heavily on phenotypic specificity. Gene selection was guided by clinical presentation, and Sanger sequencing analyzed candidate genes exon by exon (or variant by variant), a labor-intensive process constrained by the one-gene-one-disease paradigm [11]. When a candidate gene tested negative, subsequent testing focused on the next most phenotype-associated gene, even when, in many cases, no reliable alternatives were available in reference databases such as OMIM [1]. The phenotype-driven approach proved insufficient as evidence of complex genotype–phenotype relationships grew, including locus heterogeneity and pleiotropy [7, 16].

NGS transformed molecular diagnosis by enabling rapid, high-throughput sequencing of the whole exome or genome [2, 29]. Bioinformatic tools further allowed variant filtering based on molecular effect, population allele frequency and predicted pathogenicity, among other variant features. Despite initial optimism, variant interpretation remained challenging due to technical and clinical uncertainties. Inconsistent classifications across laboratories prompted the ACMG/AMP to issue standardized guidelines in 2015 [22], incorporating multiple lines of evidence. Among them, phenotypic specificity was included as a supporting criterion for pathogenicity (PP4, “Patient’s phenotype or family history is highly specific for a disease with a single genetic etiology”), limited by concerns about variable expressivity and specificity, emphasizing that variant interpretation should not rely too heavily on an individual patient’s clinical presentation. The practical implementation of PP4 raised questions regarding the required level of diagnostic sensitivity attributable to the implicated gene, the degree of overlap between the patient’s phenotype and the established disease spectrum, and other interpretative criteria. Zhang et al. proposed that the sole prerequisite for applying PP4 should be that the phenotype is attributable to a limited number of well-curated genetic etiologies [31]. Additionally, several authors suggested that expert review and context-specific interpretation could justify upgrading the level of evidence assigned to PP4. This elevated the role of clinical expertise in variant interpretation [23]. In 2024, the ClinGen Sequence Variant Interpretation Working Group (SVI) revised the PP4 criterion to integrate phenotypic specificity with familial segregation data [4]. Under this framework, when a patient’s presentation matches that of genes contributing ≥ 20% of molecular diagnoses for a disorder, variants in those genes receive increased pathogenicity weight. However, the broad application of this approach is limited by the scarce availability of curated gene–disease frequency profiles for many rare conditions.

In parallel to the PP4 discussion, in a clinical genomics analysis setting, phenotype-driven gene prioritization tools emerged as critical for efficient variant prioritization. Leveraging structured vocabularies of phenotypic abnormalities encountered in human disease such as the Human Phenotype Ontology (HPO) [13], these tools aim to associate a patient’s phenotypic profile with candidate genes. Methods such as Phenomizer and Phenolyzer rely on semantic similarity between patient-entered HPO terms and disease annotations, utilizing ontological distance metrics and Bayesian frameworks to prioritize gene-disease associations [18, 30]. VarElect and Phen2Gene implement knowledge-based prioritization strategies grounded in curated phenotype–gene associations [24, 32]. VarElect re-ranks variant-bearing genes based on semantic and literature-based phenotypic relevance, while Phen2Gene performs genome-wide ranking directly from HPO terms using precomputed, specificity-weighted gene associations.

More recently, other tools incorporated data-driven or machine learning strategies. GADO (GeneNetwork Assisted Diagnostic Optimization) leverages gene co-expression networks derived from large transcriptomic datasets to prioritize genes [10]. PCAN (Phenotype Consensus Analysis Network) integrates verdicts from multiple methods to compute consensus scores for gene-phenotype associations [15]. In contrast, CADA (Case-Annotated Disease gene Association) utilizes clinical case data to quantify phenotypic similarity between patients and known gene-disease profiles. PPAR (Prioritization of Potentially Associated Rare variants) combines phenotypic similarity with variant-level pathogenicity predictions to rank candidate genes directly from raw exome data [14, 20]. Unlike knowledge-based tools, these methods vary in their dependence on prior annotation. However, their performance often depends on the availability and quality of training data, which can limit applicability in ultra-rare or poorly characterized disorders.

To address these limitations, we present the clinical evaluation of GenPhenia, a previously developed graph neural network for phenotype-driven gene prioritization, in the context of rare disease diagnosis from exome sequencing [9]. While the underlying model architecture, training strategy, and implementation have been described previously, the objective of the present study is to assess its performance and practical utility in real-world clinical genomics practice. Specifically, we evaluate GenPhenia across multiple independent clinical and benchmark datasets, compare its performance with existing phenotype-driven gene prioritization methods, and investigate how gene prioritization evolves as phenotypic information is progressively incorporated during the diagnostic process. Together, these analyses illustrate the potential of GenPhenia to support phenotype-driven gene prioritization, facilitate variant interpretation, and ultimately improve molecular diagnosis in routine clinical genomics practice.

Computational methods

Datasets

In the present study, we used four datasets for model training and evaluation. The first dataset is a synthetic cohort generated exclusively for model training and consists of 25 million simulated cases. Rather than reproducing existing clinical records, synthetic cases were generated by sampling plausible phenotype combinations from the set of Human Phenotype Ontology (HPO) terms associated with each disease-causing gene. These gene–phenotype relationships were obtained from the HPO annotation table (release v2026-02-16), in which each gene is linked to the collection of phenotypes reported across reported affected individuals. Consequently, each annotation represents the phenotypic spectrum associated with a disease-causing gene rather than the phenotype of a single patient. Under this framework, each gene defines a characteristic phenotypic space determined by the combination and semantic relationships among its associated HPO terms. Synthetic cases were generated by randomly sampling phenotype subsets while constraining both the number and specificity of HPO terms to reproduce the distributions observed in real clinical records. Thus, the synthetic dataset is intended to sample the phenotypic space associated with each disease-causing gene rather than reproduce previously observed patient records. A comprehensive description of the synthetic dataset generation procedure is available in the previously published technical description of GenPhenia (20).

The second dataset was used for model validation and comprises more than 20,000 real-world cases derived from ClinVar and curated by ClinGen. The third dataset corresponds to a real-world clinical cohort comprising 129 patients previously reported by our research group [5, 6, 12, 26–28]. Finally, we also evaluated GenPhenia using publicly available benchmark datasets employed in phenotype-driven gene prioritization studies, namely the DDD and MCRD datasets from the Mayo Clinic.

In the synthetic dataset, each case corresponds to a simulated phenotype–gene pair generated for model training. In contrast, in all real-world datasets, each case corresponds to a single patient and consists of two complementary components: (i) a phenotypic profile represented by a set of HPO terms describing the patient's clinical manifestations, and (ii) the corresponding causative gene identified through prior molecular diagnosis. The causative gene was determined through previous clinical genomic analyses, in which pathogenic or likely pathogenic variant(s) explaining the patient's phenotype were identified and interpreted according to standard clinical practice. Thus, for every evaluated patient, both the phenotypic profile and the corresponding causative gene were available, enabling an objective assessment of phenotype-driven gene prioritization performance. A summary of all four datasets is provided in Table 1.

Table 1.

Phenotype-genotype datasets used in the present work

Name Description Number of cases References
Synthetic Synthetic dataset used to train the deep neural network (DNN) model 25 million [9]
ClinGen ClinGen-derived dataset of gene–disease relationships based on published literature and clinical data. These annotations are publicly released in structured formats and reflect community consensus under the ClinGen Gene-Disease Validity Curation framework  > 20,000 [8]
Clinical Previously reported real-world clinical cases from our research group 129 [5, 6, 12, 26–28]
MCRD Internally curated Mayo Clinic rare disease cohort comprising clinically diagnosed Mendelian cases annotated with HPO terms and confirmed causal genes. The data were collected over multiple years through the Mayo Clinic Center for Individualized Medicine under IRB-approved protocols 229 [14]
DDD Publicly available dataset from the Deciphering Developmental Disorders study, containing pediatric developmental disorder cases annotated with HPO terms and known causal genes. The version used here was frozen in November 2013 and released as part of the original DDD study 1048 [14]

GenPhenia model

The full implementation details are described in our previous work [9]. Here we summarize the key methodological aspects relevant to the present study.

Input representation

The model takes as input the phenotypic profile of a case, represented as a set of HPO terms. Each HPO term is represented as a graph node, and the hierarchical relationships defined by the HPO ontology are used to connect these nodes into an HPO subgraph that encodes an enriched and contextualized representation of the patient's phenotypic profile. Each graph node (HPO term) is represented by a 768-dimensional embedding vector generated using the fine-tuned gsarti/biobert-nli model. We selected the gsarti/biobert-nli embedding model since it produced the least collapsed cosine-similarity distribution across HPO definitions among the different encoders we tested. This is important since highly anisotropic embeddings would make many phenotypes appear artificially similar, reducing the information available to the GNN before message passing.

Model architecture

We use a Graph Isomorphism Network (GIN) classifier operating on the induced HPO subgraph for each case. Message passing is directed child-to-parent; reverse edges are not added. The network contains three message-passing layers with trainable epsilon parameters. The first layer maps 768-dimensional node features to a 512-dimensional hidden space, and the following two layers operate in the same 512-dimensional space. Each GIN layer updates node representations by applying a two-layer MLP to the sum of the aggregated representations of incoming neighbors and the node's own representation, the latter being scaled by a learnable (1 + ε) parameter. This learnable self-weight allows the network to balance preservation of the original phenotype representation against contextual information propagated through the ontology. Each message-passing block is followed by batch normalization, ReLU activation, and dropout with probability 0.3. After message passing, node embeddings are aggregated into a graph-level case representation using learned attentional pooling, with a gate network of size 512—> 256—> 1 and a softmax over nodes. The pooled 512-dimensional representation is then passed to a linear classifier over the 5229-gene output vocabulary, and gene rankings are obtained from the resulting logits.

Training procedure

GenPhenia was trained as a supervised multiclass gene-prioritization model. Each training example consists of a simulated case represented as an HPO graph together with its corresponding causal gene label, and the model is optimized using cross-entropy loss over the full gene vocabulary. We used AdamW optimization with learning rate 5e-4 and weight decay 5e-4. Checkpoint selection was based on validation ranking performance. The final checkpoint was selected at epoch 18. Training was performed on a single NVIDIA GeForce RTX 5080 GPU with 16 GB of VRAM and required 71.4 h of wall-clock time.

Evaluation

Model performance was assessed using the Recall@N metric (N = 1, 5, 10), defined as the proportion of cases in which the true causative gene was ranked within the Top-N predicted genes.

Case studies

To illustrate the behavior of GenPhenia during the diagnostic process, we selected three representative clinical cases from routine clinical practice spanning endocrinology, cardiology, and immunology (Table 2). These cases were selected because they illustrate how sequential incorporation of phenotypic information influences phenotype-driven gene prioritization in clinically relevant scenarios. Phenotypic information was extracted from the clinical descriptions and encoded using HPO terms. Recall was defined as the proportion of disease-associated genes present within the Top-N ranked genes, where N corresponds to the total number of genes included in the panel under consideration. The gene score represents the ranking score assigned by GenPhenia. The Top-5 cumulative score was calculated as the sum of the scores of the five highest-ranked genes produced by GenPhenia. The ranking position of the ultimately confirmed causal gene is indicated by the corresponding dot, with the ranking number displayed alongside it.

Table 2.

Summary of clinical cases used to illustrate phenotype-driven gene prioritization

Case study Clinical specialty Causative gene Phenotypic profile (HPO terms)
1 Endocrinology ANOS1 Delayed puberty (HP:0000823), Cryptorchidism (HP:0000028), Micropenis (HP:0000054), Decreased testicular size (HP:0008734), Hyposmia (HP:0004409), Renal agenesis (HP:0000104)
2 Cardiology LMNA Syncope (HP:0001279), Atrial fibrillation (HP:0005110), Dilated cardiomyopathy (HP:0001644), Atrioventricular block (HP:0001678), Ventricular tachycardia (HP:0004756), Heart failure (HP:0001635), Sudden cardiac death (HP:0001645)
3 Immunology STAT3 Recurrent respiratory infections (HP:0002205), Eczematoid dermatitis (HP:0000964), Increased circulating IgE concentration (HP:0003212), High palate (HP:0000218)

Results

GenPhenia model and performance

GenPhenia is a DNN model that takes as input a set of HPO terms encoding the patient's clinical phenotype and outputs a ranked list of genes annotated as disease-associated in OMIM. The method was trained using a large number of synthetic symptom-gene pairs in a supervised multilabel classification scheme, where each gene corresponds to a distinct label. Model performance was evaluated using the Recall@N metric, defined as the proportion of cases for which the true causative gene was ranked within the Top-N predicted genes. For example, if 100 cases are evaluated with N = 5, a Recall@5 value of 0.6 indicates that the correct gene appears among the Top-5 ranked genes in 60% of cases.

We first assessed GenPhenia’s performance across the datasets constructed for this study (Fig. 1A). These include a curated ClinGen-derived dataset and an independent real-world clinical cohort comprising 129 cases previously analyzed by our research group. These cases correspond to patients who underwent whole exome sequencing (WES) followed by variant prioritization using the clinical information provided by the referring physician (hereafter referred to as the observed phenotype). Only cases with a high-confidence molecular diagnosis were included, defined as the presence of one or two (in cases of compound heterozygosity) variants classified as (likely) pathogenic according to ACMG guidelines in a gene considered consistent with the observed phenotype during clinical interpretation. Performance on the challenging MCRD dataset from the Mayo Clinic was also evaluated.

Fig. 1.

Fig. 1

GenPhenia outperforms previous methods across multiple real-world datasets. Recall@N is shown for the top 1, 10, and 50 ranked genes. A Performance of GenPhenia on the MCRD, ClinGen, and internal clinical datasets. B Comparison of GenPhenia with Phen2Gene, CADA, PPAR, and PCAN using the MCRD dataset. C Equivalent comparison performed on the DDD dataset

For the ClinGen and MCRD datasets, GenPhenia achieved strong performance, reaching close to 80% recall at Top-10. In the internal real-world cohort, Top-10 recall was approximately 50%, increasing to 80% recall at Top-50, only slightly lower than that obtained for the ClinGen dataset.

Figure 1B, C compare GenPhenia with other phenotype-based gene prioritization approaches using the DDD and MCRD benchmark datasets, with comparative performance values obtained from previously published studies (19). Across both datasets, GenPhenia outperformed all competing methods, achieving 60% recall at Top-1 and 80% recall at Top-10. In the DDD cohort (Fig. 1C), all methods achieved good performance, exceeding 80% recall at Top-10, with GenPhenia providing only modest additional improvement. In contrast, in the more challenging MCRD cohort (Fig. 1B), competing methods performed poorly, with Top-10 recall values below 30%, whereas GenPhenia reached 80%, representing a substantial improvement. Notably, the performance curves for most approaches, including GenPhenia, plateaued between N = 20 and 30, suggesting that a relatively small subset of genes can be prioritized with high confidence in a clinical setting.

GenPhenia case studies

To illustrate GenPhenia’s performance in a clinically relevant context, we present representative examples of its application within a diagnostic workflow. The first case involves a proband with a clinical diagnosis of Kallmann Syndrome (KS), supported by the combination of reproductive and olfactory phenotypes. Typically, WES data processing, annotation, and variant interpretation using a quantitative ACMG framework yield, per patient, an average of 15 potentially relevant variants across 10 genes [5, 6]. Most of these variants generally display low to moderate Bayesian pathogenicity scores (0.1–0.5) and are therefore classified as variants of uncertain significance (VUS). Consequently, the key step for downstream variant interpretation and potential clinical reporting of a candidate variant lies in establishing the best genotype–phenotype correlation, which can justify the application of the ACMG evidence code PP4 (i.e., phenotype highly consistent with a disease with a single genetic etiology).

As shown in Fig. 2A, the progressive incorporation of the patient’s clinical features—encoded as HPO terms—into GenPhenia results in changes in gene prioritization that ultimately lead to the identification of a missense variant NM_000216.4:c.1540G > A p.(Glu514Lys) in ANOS1 as the likely cause of the disease. ANOS1 initially ranks below the Top-50 but progressively increases in rank as additional clinical features are incorporated, ultimately reaching rank 5, preceded by CCDC141, FEZF1, NDNF, and FGF17. All these genes are strongly associated with KS (see below). The final ANOS1 score, together with the absence of potentially causative variants in higher-ranked or similarly scored genes, provides strong support for applying the PP4 criterion to the corresponding variant.

Fig. 2.

Fig. 2

Progressive phenotype-driven gene prioritization using GenPhenia in three clinical scenarios. Added symptoms versus gene-related metrics are shown for three reported patients: (A) Kallmann syndrome, (B) congenital heart defects, and (C) inborn error of immunity. For each case, recall, Top-5 cumulative score, probability score, and ranking of the causal gene are displayed as phenotypic information is progressively added

A broader analysis of GenPhenia’s gene prioritization performance for this case can be conducted in the context of the clinical spectrum of KS. KS represents a specific form of congenital hypogonadotropic hypogonadism (CHH), a disorder of the gonadotropin-releasing hormone (GnRH) neuron development that leads to dysfunction of the reproductive endocrine axis. The condition arises from defects in the migration of these neurons from the nasal placode to the hypothalamus—a process that relies on vomeronasal and olfactory axons as scaffolds—and is therefore characterized by olfactory dysfunction. This contrasts with normosmic CHH (nCHH), which lacks olfactory symptoms. In Fig. 2A we compare GenPhenia’s gene rankings with two curated gene panels: a broad CHH-associated panel (49 genes) and a KS-specific panel (21 genes). To quantify performance, we computed GenPhenia’s recall as the fraction of panel genes recovered among the Top-N ranked genes (blue KS and red CHH curves). The resulting plot demonstrates that as additional HPO terms are iteratively incorporated—thereby increasing phenotype specificity—the proportion of correctly ranked KS and CHH genes increases accordingly.

The second case involves a proband presenting with a clinical diagnosis of dilated cardiomyopathy (DCM) and atrial fibrillation (AF). These diagnoses were established based on standard clinical assessments: echocardiographic documentation of left ventricular dilation and systolic dysfunction (DCM), and electrocardiographic confirmation of AF, coupled with a history of syncope. WES was performed, followed by variant annotation and prioritization aided by the GenPhenia gene-ranking output. The clinical presentation could potentially be attributed to multiple genes, including SCN5A, TTN, and LMNA. Diagnostic complexity in this scenario arises from the inherent heterogeneity of inherited cardiovascular disorders. Cardiogenetic gene panels are typically extensive, often encompassing up to 200 genes across multiple established subgroups (e.g., hypertrophic cardiomyopathies, channelopathies, and DCM). In addition, genotype–phenotype correlations frequently show substantial overlap, whereby genes classically associated with one subgroup may manifest phenotypes characteristic of another. Furthermore, structural abnormalities of the myocardium may give rise to conduction disease, whereas primary electrical disorders can secondarily lead to ventricular remodeling and dilation. This substantial phenotypic and genotypic heterogeneity significantly complicates the establishment of a definitive molecular diagnosis.

The patient carried a missense variant, NM_170707.4:c.575A > T p.(Asp192Val), located in the rod domain of LMNA. Missense variants within this region—upstream of the nuclear localization signal—have been consistently associated with a high-risk LMNA phenotype characterized by systolic dysfunction, atrial and ventricular arrhythmias, progressive conduction system disease, and an increased risk of ventricular arrhythmias and advanced heart failure [3]. As shown in Fig. 2B the LMNA gene rank increases steadily until reaching a plateau, a pattern expected in scenarios involving large cardiogenetic gene sets. Broad cardiomyopathy-related genes are frequently ranked highly, whereas genes associated with DCM and arrhythmogenic cardiomyopathy (ACM)—including LMNA—progressively emerge among the top-ranked positions. In this case, a substantial number of genes would need to be evaluated for the absence of potentially relevant variants before ultimately considering the LMNA missense variant and assessing the applicability of the PP4 criterion.

The third case involves a proband presenting with clinical suspicion of an Inborn Error of Immunity (IEI), specifically consistent with a Combined Immunodeficiency (CID). The patient’s clinical history was notable for recurrent infections, eczematoid dermatitis, markedly elevated IgE concentrations, and syndromic features such as a high-arched palate. As in the previous cases, WES was performed, followed by variant annotation and phenotype-guided gene prioritization using GenPhenia. Diagnostic complexity in this scenario arises from the large number of IEI-associated genes and their pleiotropic nature. According to the IUIS classification [21], this case could be categorized under CID with associated or syndromic features (Table 2, ~ 86 genes) or, more specifically, under the Hyper-IgE Syndromes (HIES, historically known as Job’s syndrome) category (Table 2, subtable 5, ~ 12 genes).

The proband harbored a variant in STAT3, a gene characterized by marked phenotypic heterogeneity depending on the underlying molecular mechanism. Whereas gain-of-function (GOF) variants are associated with early-onset autoimmunity, the clinical presentation of this patient—particularly the triad of eczema, recurrent infections, and elevated IgE levels—points toward an autosomal dominant loss-of-function (LOF) mechanism, classically associated with HIES [19]. As illustrated in Fig. 2C, the rank of STAT3 improves rapidly across successive phenotypic query stages. At the initial stage, when the phenotype is limited to nonspecific features such as recurrent respiratory infections, the CID gene set shows greater recall and outperforms the HIES-specific panel. However, as more informative phenotypic features characteristic of HIES (e.g., eczematoid dermatitis, high-arched palate) are incorporated, the HIES-specific metrics overtake the broader CID performance. Consequently, the STAT3 ranking climbs steeply, ultimately reaching a top-tier position (rank 7). This highlights the tool's ability to refine gene prioritization from a broad immunological category to a specific molecular diagnosis as increasingly detailed clinical information is integrated.

Making GenPhenia available to the clinical genomics community

To facilitate the use of GenPhenia in clinical genomics practice, we developed a dedicated web-based portal, freely available at https://www.genphenia.xyz, which provides direct access to the model through an intuitive interface. Recognizing that the diagnostic odyssey (the time elapsed between the onset of signs and symptoms and the establishment of a molecular diagnosis) experienced by many patients can be further complicated by the technical barriers associated with advanced bioinformatics tools, the platform allows clinicians and researchers to interact with the model without requiring programming expertise. By hosting GenPhenia as a secure web application, the system enables real-time phenotype-driven gene prioritization that can be readily incorporated into diagnostic workflows. Users can upload patient profiles encoded using HPO terms and obtain prioritized candidate gene lists within seconds. A detailed step-by-step tutorial illustrating the use of the platform in a real-world clinical scenario is provided as supplementary material [see Additional file 1]. Improving accessibility to phenotype-driven gene prioritization tools may help accelerate the identification of candidate disease genes and ultimately contribute to reducing the time required to reach a molecular diagnosis.

Discussion

Phenotype-driven gene prioritization tools have emerged as a central component for efficient variant prioritization in clinical genomics, particularly for the molecular diagnosis of rare diseases. Moreover, the advent of artificial intelligence (AI) technology has significantly improved the accuracy of these methods. In this context, GenPhenia contributes to the growing family of gene prioritization tools by introducing, to the best of our knowledge, an original approach based on the generation of large-scale synthetic clinical history data. Our goal was to develop a biologically sound and accurate method that also provides an intuitive user experience in the clinical setting. Our experience with GenPhenia indicates that the algorithm is highly sensitive to the phenotypic information provided as input. We observed a marked improvement in recall both when increasingly specific phenotypic features were incorporated and when phenotypes were added sequentially, reflecting the natural progression of disease. This behavior is particularly valuable for clinicians, who require tools that operate in alignment with the biological progression of a disorder and adapt to the evolving specificity of patient signs and symptoms—especially when such features are pathognomonic—. While overall performance was acceptable across all evaluated scenarios, a consistent pattern emerged: large and heterogeneous gene panels (e.g., congenital heart disease or inborn errors of immunity) tended to achieve lower recall at later stages compared to more narrowly defined panels (e.g., congenital hypogonadotropic hypogonadism). This observation may reflect an inherent baseline overlap among phenotypes associated with congenital disorders, which is difficult—and arguably undesirable—to fully resolve. Variable expressivity and phenotypic heterogeneity create clinical scenarios in which alternative diagnoses must remain under consideration, even in the presence of a presumptive clinical diagnosis. For example, optimizing the algorithm to achieve maximal recall in congenital heart disease could inadvertently suppress syndromic diagnoses in which cardiac manifestations represent only one component of a broader phenotype.

Although GenPhenia demonstrated promising performance in clinical applications, we acknowledge that it retains certain black-box characteristics that may limit full confidence among healthcare professionals. While users without specialized training in AI are not expected to understand the mathematical foundations of deep neural networks, they are naturally interested in the interpretability and structure of the model’s output. Given that GenPhenia reports an approximately 80% success rate when considering the Top-10 ranked genes, it is important to examine not only performance metrics but also the composition and scoring of these ranked outputs. For instance, in the congenital hypogonadotropic hypogonadism case, following the inclusion of all associated phenotypes, CCDC141 emerged as the top-ranked gene (score = 0.26), while ANOS1, the gene harboring the causative variant, ranked fifth (score = 0.06). Although CCDC141 lacks a fully curated genotype–phenotype association in OMIM or ClinGen, several studies have implicated it in Kallmann syndrome through its role in GnRH neuron migration and its recurrence in affected patients [17, 25]. This ranking reflects GenPhenia’s reliance on phenotypic training data derived from the Human Phenotype Ontology, where phenotype–gene associations are aggregated from diverse bibliographic sources beyond strictly curated databases. This behavior highlights an inherent trade-off between enabling the identification of less conventional or emerging genetic contributors and prioritizing historically established, high-confidence disease genes such as ANOS1. Incorporating a downstream gene prioritization step based on curated genotype–phenotype databases—potentially through score normalization or reweighting—may enhance clinical confidence and facilitate integration into routine diagnostic workflows by emphasizing well-documented and reportable molecular diagnoses.

Finally, it is noteworthy that across all three cases, the most substantial increases in gene ranking were driven not by pathognomonic features, but rather by the combination of highly prevalent and nonspecific phenotypes introduced at early stages. Although this may initially appear counterintuitive to clinicians, it represents a potential strength of the approach. In the context of the diagnostic odyssey, early suspicion of a genetic etiology based on common clinical features may be more valuable than reliance on late-emerging, potentially irreversible manifestations that delay molecular investigation and diagnosis.

Conclusions

GenPhenia represents a novel approach to phenotype-driven gene prioritization, providing a biologically grounded and accurate tool to support the molecular diagnosis of rare diseases. By leveraging graph neural networks and large-scale synthetic clinical datasets, the model addresses limitations associated with incomplete or unevenly curated genotype–phenotype knowledge bases and offers a scalable framework for phenotype-based gene prioritization in clinical genomics. Our results demonstrate that the algorithm is highly responsive to the progressive incorporation of phenotypic information, reflecting the natural evolution of a patient’s diagnostic trajectory. Although certain black-box characteristics remain inherent to deep learning models, GenPhenia’s capacity to capture phenotype–gene associations derived from diverse sources enables the identification of potential genetic contributors beyond strictly curated databases. Together, these features position GenPhenia as a useful complementary tool for phenotype-driven variant prioritization and may contribute to improving efficiency in the interpretation of genomic data in rare disease diagnostics. GenPhenia is available through a web-based portal, enabling clinicians and researchers to integrate phenotype-driven gene prioritization into genomic diagnostic workflows.

Supplementary Information

Below is the link to the electronic supplementary material.

Acknowledgements

The authors acknowledge the Federal High-Impact Network Project “Genómica clínica de enfermedades poco frecuentes (GeC-EPoF)” (CONVE 2023-100390147-APN-MCT), promoted by the Secretaría de Innovación, Ciencia y Tecnología, Argentina, within whose framework this work was developed.

Author contributions

MAM conceived the study, supervised the overall project and methodological development, and contributed substantially to manuscript writing. GC implemented the deep learning framework, including model development, training, and validation, and developed the web platform to enable clinical use of the tool. FGB contributed to the clinical validation analyses across multiple medical specialties and participated in manuscript preparation and the development of the user tutorial. AR contributed to the design and implementation of the platform interface, including the integration of Human Phenotype Ontology (HPO) term input. LE performed clinical validation analyses in immunology. ACL, NA, AT, and GB contributed to the evaluation of the tool and provided expert feedback based on their experience in clinical genomic diagnostics at the Bitgenia laboratory. JZ provided computational infrastructure and technical support for model training. All authors read and approved the final manuscript.

Funding

Not applicable.

Data availability

The synthetic dataset used for model training, the real-world clinical dataset comprising 129 anonymized cases previously reported by our research group, and the Jupyter Notebook used to generate Fig. 2—along with the underlying data and intermediate results—are publicly available at https://github.com/gianluccacolangelo/GraPhens. This repository also contains the complete pipeline used to generate the synthetic training dataset from the Human Phenotype Ontology (HPO) gene–phenotype annotation table (release v2026-02-16), available from the official HPO repository at https://github.com/obophenotype/human-phenotype-ontology/releases/tag/v2026-02-16. The ClinGen-derived dataset used for model validation was obtained from a publicly available resource and is referenced in the manuscript (ClinGen Consortium. Electronic address: splon@bcm.edu and ClinGen Consortium 2025). Public benchmark datasets, including the Deciphering Developmental Disorders (DDD) and Mayo Clinic Rare Disease (MCRD) cohorts, were obtained from previously published studies [14]. The benchmarking framework used to evaluate and compare phenotype-driven gene prioritization methods under a unified protocol is publicly available at https://github.com/gianluccacolangelo/phenogene-arena. The GenPhenia model and its implementation are described in detail in the corresponding methodological publication (Colangelo and Marti 2026).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Franco Gino Brunello, Email: fgbrunello@gmail.com.

Marcelo A. Martí, Email: marti.marcelo@gmail.com

References

  • 1.Amberger JS, Bocchini CA, Schiettecatte F, Scott AF, Hamosh A. OMIM.org: online mendelian inheritance in man (OMIM®), an online catalog of human genes and genetic disorders. Nucleic Acids Res. 2015;43(Database issue):D789–98. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Bainbridge MN, Wiszniewski W, Murdock DR, et al. Whole-genome sequencing for optimized patient management. Sci Transl Med. 2011;3(87):87re3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Bhaskaran A, Ben Yaou R, Helms AS, et al. Location of LMNA variants and clinical outcomes in cardiomyopathy. JAMA Cardiol. 2025;10(9):896–903. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Biesecker LG, Byrne AB, Harrison SM, et al. ClinGen guidance for use of the PP1/BS4 co-segregation and PP4 phenotype specificity criteria for sequence variant pathogenicity classification. Am J Hum Genet. 2024;111(1):24–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Brunello FG, Castro S, Zaiat J, et al. Updated compendium of genes and variants associated with congenital hypogonadotropic hypogonadism: systematic review, classification pipeline, and network analysis. Hum Mol Genet. 2026. 10.1093/hmg/ddag007. [DOI] [PubMed] [Google Scholar]
  • 6.Castro S, Brunello FG, Sansó G, et al. Clinical presentation of congenital hypogonadotropic hypogonadism in males with delayed puberty according to genetic etiology: a systematic review and meta-analysis after reclassification of gene variants. Human Reproduction (Oxford, England). 2025;40(5):904–18. [DOI] [PubMed] [Google Scholar]
  • 7.Cerrone M, Remme CA, Tadros R, Bezzina CR, Delmar M. Beyond the one gene-one disease paradigm: complex genetics and pleiotropy in inheritable cardiac disorders. Circulation. 2019;140(7):595–610. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.ClinGen Consortium. Electronic address: splon@bcm.edu; ClinGen Consortium. The Clinical Genome Resource (ClinGen): Advancing genomic knowledge through global curation. Genet Med. 2025 Jan;27(1):101228. doi: 10.1016/j.gim.2024.101228. Epub 2024 Oct 15. PMID: 39404758; PMCID: PMC11984750. [DOI] [PMC free article] [PubMed]
  • 9.Colangelo G, Martí M. Solving the diagnostic odyssey with synthetic phenotype data. In bioRxiv. BioRxiv. 2026. 10.64898/2026.03.19.712946. [DOI]
  • 10.Deelen P, van Dam S, Herkert JC, et al. Improving the diagnostic yield of exome- sequencing by predicting gene-phenotype associations using large-scale gene expression analysis. Nat Commun. 2019;10(1):2837. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Van Dijk EL, Auger H, Jaszczyszyn Y, Thermes C. Ten years of next-generation sequencing technology. Trends Genet. 2014;30(9):418–26. [DOI] [PubMed] [Google Scholar]
  • 12.Erra L, Colado A, Brunello FG, et al. T and NK cell functionality in a patient harboring heterozygous novel BCL11B p.Asp632fsAla∗91 and STX11 p. R129P mutations. Heliyon. 2025;11(4):e42636. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Gargano MA, Matentzoglu N, Coleman B, et al. The Human Phenotype Ontology in 2024: phenotypes around the world. Nucleic Acids Res. 2024;52(D1):D1333–46. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Gnanaolivu R, Oliver G, Jenkinson G, et al. A clinical knowledge graph-based framework to prioritize candidate genes for facilitating diagnosis of Mendelian diseases and rare genetic conditions. BMC Bioinformatics. 2025;26(1):82. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Godard P, Page M. PCAN: phenotype consensus analysis to support disease-gene association. BMC Bioinformatics. 2016;17(1):518. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Gratten J, Visscher PM. Genetic pleiotropy in complex traits and diseases: implications for genomic medicine. Genome Med. 2016;8(1):78. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Hutchins BI, Kotan LD, Taylor-Burds C, et al. CCDC141 mutation identified in anosmic hypogonadotropic hypogonadism (Kallmann syndrome) alters GnRH neuronal migration. Endocrinology. 2016;157(5):1956–66. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Köhler S, Schulz MH, Krawitz P, et al. Clinical diagnostics in human genetics with semantic similarity searches in ontologies. Am J Hum Genet. 2009;85(4):457–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Lodi L, Faletti LE, Maccari ME, et al. STAT3-confusion-of-function: beyond the loss and gain dualism. J Allergy Clin Immunol. 2022;150(5):1237-1241.e3. [DOI] [PubMed] [Google Scholar]
  • 20.Peng C, Dieck S, Schmid A, et al. CADA: phenotype-driven gene prioritization based on a case-enriched knowledge graph. NAR Genom Bioinform. 2021;3(3):lqab078. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Poli MC, Aksentijevich I, Bousfiha AA, et al. Human inborn errors of immunity: 2024 update on the classification from the International Union of Immunological Societies Expert Committee. J Hum Immunol. 2025. 10.70962/jhi.20250003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genet Med. 2015;17(5):405–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Smedley D, Robinson PN. Phenotype-driven strategies for exome prioritization of human mendelian disease genes. Genome Med. 2015;7(1):81. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Stelzer G, Plaschkes I, Oz-Levi D, et al. VarElect: the phenotype-based variation prioritizer of the GeneCards Suite. BMC Genomics. 2016;17(Suppl 2):444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Turan I, Hutchins BI, Hacihamdioglu B, et al. CCDC141 mutations in idiopathic hypogonadotropic hypogonadism. J Clin Endocrinol Metab. 2017;102(6):1816–25. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Urdinez L, Erra L, Palma AM, et al. Expanding spectrum, intrafamilial diversity, and therapeutic challenges from 15 patients with heterozygous CARD11-associated diseases: a single center experience. Front Immunol. 2022;13:1020927. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Vishnopolska SA, Mercogliano MF, Camilletti MA, et al. Comprehensive identification of pathogenic gene variants in patients with neuroendocrine disorders. J Clin Endocrinol Metab. 2021;106(7):1956–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Vishnopolska SA, Turjanski AG, Piñero MH, et al. Genetics and genomic medicine in Argentina. Mol Genet Genomic Med. 2018;6(4):481–91. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Wong LJ. Next generation sequencing based clinical molecular diagnosis of human genetic disorders. Springer; 2017. [Google Scholar]
  • 30.Yang H, Robinson PN, Wang K. Phenolyzer: phenotype-based prioritization of candidate genes for human diseases. Nat Methods. 2015;12(9):841–3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Zhang J, Yao Y, He H, Shen J. Clinical interpretation of sequence variants. Curr Protoc Hum Genet. 2020;106(1):e98. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Zhao M, Havrilla JM, Fang L, et al. Phen2gene: rapid phenotype-driven gene prioritization for rare diseases. NAR Genom Bioinform. 2020;2(2):lqaa032. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Data Availability Statement

The synthetic dataset used for model training, the real-world clinical dataset comprising 129 anonymized cases previously reported by our research group, and the Jupyter Notebook used to generate Fig. 2—along with the underlying data and intermediate results—are publicly available at https://github.com/gianluccacolangelo/GraPhens. This repository also contains the complete pipeline used to generate the synthetic training dataset from the Human Phenotype Ontology (HPO) gene–phenotype annotation table (release v2026-02-16), available from the official HPO repository at https://github.com/obophenotype/human-phenotype-ontology/releases/tag/v2026-02-16. The ClinGen-derived dataset used for model validation was obtained from a publicly available resource and is referenced in the manuscript (ClinGen Consortium. Electronic address: splon@bcm.edu and ClinGen Consortium 2025). Public benchmark datasets, including the Deciphering Developmental Disorders (DDD) and Mayo Clinic Rare Disease (MCRD) cohorts, were obtained from previously published studies [14]. The benchmarking framework used to evaluate and compare phenotype-driven gene prioritization methods under a unified protocol is publicly available at https://github.com/gianluccacolangelo/phenogene-arena. The GenPhenia model and its implementation are described in detail in the corresponding methodological publication (Colangelo and Marti 2026).


Articles from Human Genomics are provided here courtesy of BMC

RESOURCES