PhenoGnet: Graph Contrastive Learning for Disease Similarity

Updated 12 July 2026

PhenoGnet is a graph-based contrastive learning framework that integrates molecular gene networks and phenotype ontologies to compute disease similarities.
It employs dual graph encoders—a GCN for molecular data and a GAT for ontology data—aligned via a shared MLP using bidirectional contrastive loss.
Evaluation shows gene-only embeddings outperform combined representations, underscoring the discriminative power of molecular data in rare disease analysis.

PhenoGnet is a graph-based contrastive learning framework for predicting disease similarity by jointly modeling molecular and phenotypic evidence. It integrates a gene functional interaction network with the Human Phenotype Ontology (HPO) and aligns the two modalities through known gene–phenotype associations in a common latent space. Diseases are then represented by the mean embeddings of their associated genes and/or phenotypes, and disease–disease similarity is computed with cosine similarity. The framework is motivated by the observation that clinically or mechanistically similar diseases often share causal genes, pathways, phenotypic manifestations, biomarkers, or therapeutic responses, so reliable disease similarity estimates can support diagnosis, rare disease research, drug repurposing, and precision medicine (Baminiwatte et al., 17 Sep 2025).

1. Scope and conceptual basis

PhenoGnet addresses a specific limitation of prior disease-similarity methods: many use only one information source at a time, such as molecular networks, ontology semantics, or phenotype annotations, and therefore do not exploit complementary structure across modalities. Its central premise is that gene-level and phenotype-level evidence should not be modeled independently when known gene–phenotype associations already provide a natural cross-view supervisory signal. The model is therefore organized around two graph views, one molecular and one phenotypic, with alignment learned from curated gene–phenotype links (Baminiwatte et al., 17 Sep 2025).

The framework distinguishes among several graph objects. The two principal views are the weighted gene graph $G^g$ , derived from HumanNet v3, and the phenotype graph $G^p$ , derived from the HPO hierarchy. In the broader data construction it also uses three association graphs: gene–phenotype $G^{gp}$ , disease–phenotype $G^{dp}$ , and disease–gene $G^{dg}$ . The training signal comes from $G^{gp}$ : known gene–phenotype associations are treated as positive cross-view pairs, while unlinked pairs serve as negatives. This design makes disease similarity a downstream quantity induced from entity embeddings rather than a directly supervised disease-pair classifier (Baminiwatte et al., 17 Sep 2025).

A common misconception would be to treat PhenoGnet as a generic disease classifier. It is not presented that way. The paper states that disease-pair similarity is scored directly with cosine similarity on disease embeddings, and although F1, sensitivity, and specificity are reported, the threshold selection procedure for binary prediction is not specified. This suggests that the primary object of interest is a ranking or scoring function over disease pairs rather than a separately learned decision boundary.

2. Graph construction and data model

The molecular view is an undirected, weighted gene graph built from HumanNet v3, where nodes are genes and edges represent functional interactions. HumanNet edge scores are normalized from the minimum and maximum log-likelihood similarities before graph convolution is applied. Initial gene features are one-hot vectors, and the final learned gene embedding matrix is denoted $\mathbf{F}^G$ . The figure caption states that the gene encoder uses two graph convolution layers and maps to a 32-dimensional embedding space (Baminiwatte et al., 17 Sep 2025).

The phenotype view is an HPO graph $G^p$ in which nodes are phenotype terms and edges encode the ontology hierarchy. The ontology is augmented using the true-path rule, so each phenotype term is connected to all ancestor terms, and edges are unit-weight. Initial phenotype features are semantic text embeddings derived from HPO term descriptions using Sentence-BERT with pretrained all-mpnet-base-v2, without fine-tuning, yielding $E^p \in \mathbb{R}^{n_p \times 768}$ . The phenotype encoder also uses two graph layers and yields 32-dimensional embeddings (Baminiwatte et al., 17 Sep 2025).

The paper reports the following graph sizes.

Graph	Size
$G^p$	19,034 HPO terms, 46,784 edges
$G^p$ 0	18,459 genes, 977,495 edges
$G^p$ 1	18,459 genes, 19,034 HPO terms, 848,284 edges
$G^p$ 2	30,170 diseases, 19,034 HPO terms, 171,013 edges
$G^p$ 3	30,170 diseases, 18,459 genes, 807,292 edges

Gene–phenotype and disease–phenotype associations come from HPO, disease–gene associations come from DisGeNET, and HPO term descriptions were taken from HPO version 2025-01-16. For evaluation, the model uses a curated benchmark of 1,100 similar and 866 dissimilar rare disease pairs, all with known gene and HPO associations. The exact label-generation protocol for those pairs is inherited from prior studies and is not fully re-described in the paper (Baminiwatte et al., 17 Sep 2025).

3. Architecture and cross-view contrastive alignment

PhenoGnet has two main components: an intra-view model and a cross-view model. In the intra-view stage, the gene graph is encoded with a GCN, while the HPO graph is encoded with a GAT. The choice is asymmetric: the gene network uses weighted graph convolution over functional interactions, whereas the HPO view uses attention so that different ontology neighbors can receive different importance during aggregation. This architectural asymmetry is directly tied to the difference between a weighted interaction network and a directed acyclic ontology with true-path augmentation (Baminiwatte et al., 17 Sep 2025).

In the cross-view stage, gene and phenotype embeddings are projected into a common latent space with a single-hidden-layer shared-weight MLP. The purpose is to extract joint features from known gene–phenotype associations and maximize agreement between paired entities. The paper does not provide an explicit layer equation for this MLP. A plausible implication is that the shared projection enforces modality alignment at the representation level before disease embeddings are constructed.

The alignment objective is a bidirectional cross-view contrastive loss: $G^p$ 4 where $G^p$ 5 balances the gene-to-phenotype and phenotype-to-gene directions. The gene-side term is

$G^p$ 6

and the phenotype-side term is

$G^p$ 7

The similarity function is cosine similarity: $G^p$ 8

Here $G^p$ 9 when gene $G^{gp}$ 0 and phenotype $G^{gp}$ 1 are linked in $G^{gp}$ 2, and $G^{gp}$ 3 otherwise. The paper explicitly states that each true gene–phenotype pairing constitutes a positive example, while all other pairings serve as negatives. It does not report margin-based loss, hard-negative mining, negative subsampling, class weighting, or explicit regularization such as dropout or $G^{gp}$ 4 weight decay. That omission is consequential for reproducibility, but it also clarifies that the method’s central inductive bias comes from graph structure and bidirectional cross-view alignment rather than auxiliary training heuristics (Baminiwatte et al., 17 Sep 2025).

4. Disease representation and similarity scoring

After training, disease embeddings are constructed by average pooling over associated entity embeddings. For a disease $G^{gp}$ 5 with associated gene set $G^{gp}$ 6 and phenotype set $G^{gp}$ 7, the gene-based and phenotype-based disease embeddings are

$G^{gp}$ 8

The fused disease embedding is formed by weighted concatenation: $G^{gp}$ 9 where $G^{dp}$ 0 controls the relative contribution of the two views (Baminiwatte et al., 17 Sep 2025).

The paper evaluates three disease representations: gene-only, phenotype-only, and combined. Disease-pair similarity is scored directly by cosine similarity on these disease embeddings: $G^{dp}$ 1 No learned classifier is placed on top of these similarities. This is important because it means that improvements are attributable to representation quality rather than a more expressive decision layer. At the same time, the use of simple mean pooling implies a structural limitation: diseases with heterogeneous subtypes or very sparse annotations may be poorly summarized by an average of associated entities. This limitation is partly methodological inference from the reported design rather than an explicit author statement.

5. Optimization protocol and empirical evaluation

The labeled disease-pair split is unusual. The paper reports 210 training positives, 183 training negatives, 890 validation positives, and 683 validation negatives, with training and validation sets stated to be disjoint. On the training set, the authors use 5-fold cross-validation, training on 80% and validating on 20% of that training split in each fold. Hyperparameters are optimized with Optuna, using 30 trials and the Tree-structured Parzen Estimator (TPE) sampler. The optimization objective is the harmonic mean of AUROC and AUPRC, and tuned parameters are $G^{dp}$ 2, $G^{dp}$ 3, $G^{dp}$ 4, learning rate, and number of training epochs. Optimization uses RMSProp with an adaptive learning rate (Baminiwatte et al., 17 Sep 2025).

Several implementation details remain unreported: final selected hyperparameter values, batch size, early stopping rule, random seed, hardware, framework library, runtime, and code release. For a model positioned as scalable, this absence is notable. It does not invalidate the reported results, but it constrains exact replication.

The main validation results are as follows.

Representation or baseline	AUPRC / AUROC
Gene-based embeddings	0.9012 / 0.8764
Combined embeddings	0.8855 / 0.8678
HPO-based embeddings	0.7668 / 0.7383
CoGO baseline	0.8928 / 0.8594

The full metric profiles reported in the paper are: gene-based embeddings with F1 0.8254, sensitivity 0.8843, specificity 0.6633; combined embeddings with F1 0.8229, sensitivity 0.9112, specificity 0.6047; HPO-based embeddings with F1 0.7670, sensitivity 0.9393, specificity 0.3353; and the CoGO baseline with F1 0.7983, sensitivity 0.9247, specificity 0.4890 (Baminiwatte et al., 17 Sep 2025).

These results support two specific conclusions stated or directly implied by the paper. First, gene-based disease embeddings perform best, even better than the fused representation. Second, HPO-only embeddings have very high sensitivity but poor specificity, which is consistent with the authors’ interpretation that phenotype-only similarity captures broad disease relatedness but is less selective. A common expectation in multimodal representation learning is that fusion should dominate single-view models; in this study it does not. That result should be read as an empirical property of this particular alignment and fusion design rather than as a general statement against multimodal disease modeling.

6. Latent biological signal, interpretability, and limitations

PhenoGnet argues that it captures latent biological relationships beyond direct overlap. The evidence provided is targeted rather than fully exhaustive. Among the top 100 most similar disease pairs, the average direct overlap was only 11.41% of genes and 7.19% of HPO terms. After removing overlapping genes and HPO terms, the top pairs still retained significantly higher maximum similarity scores than 100 random pairs. Within the scope of the paper, this supports the claim that the model is learning indirect or higher-order structure through graph encoding and cross-view alignment rather than merely scoring shared annotations (Baminiwatte et al., 17 Sep 2025).

The interpretability claim is correspondingly modest. A disease embedding can be traced back to its associated genes and phenotypes, and the two-view design preserves biologically meaningful modalities rather than collapsing everything into an opaque classifier. This suggests a form of structural interpretability. However, the model does not provide pathway-level mechanistic explanations, attention rationales over disease pairs, or counterfactual explanations. It is therefore more accurate to describe PhenoGnet as interpretable at the level of inputs and latent biological relationships than as mechanistically transparent.

The limitations are mostly implicit. Performance depends heavily on the quality and completeness of curated association graphs, especially gene–phenotype links in $G^{dp}$ 5, which define the contrastive supervision. The labeled disease-pair benchmark is relatively small, especially in its training portion. The combined representation does not outperform the gene-only representation, suggesting that the phenotype branch may still be noisier or less well aligned than intended. The paper also omits several reproducibility-critical details and does not include a full component-wise ablation study such as removing contrastive learning, replacing GAT with GCN on HPO, or removing true-path augmentation. These limitations do not negate the empirical contribution, but they narrow what can be concluded about which individual design choices are indispensable.

The stated future directions are to incorporate additional modalities such as protein–protein interaction networks, gene expression networks, and additional ontologies, and to apply the framework to larger biobank-scale resources such as UK Biobank. This suggests a natural path from two-view alignment toward broader multimodal disease representation learning.

PhenoGnet belongs to a broader methodological family of phenotype-aware graph learning, but its endpoint is specifically disease similarity prediction. A nearby example is PhenoLinker, which predicts missing phenotype–gene associations using a heterogeneous graph neural network and feature attribution, rather than disease–disease similarity (Andreu et al., 2024). Another adjacent lineage is Phen-Gen, which combines symptoms and sequencing data to prioritize causal genes for rare disorder diagnosis, again targeting gene prioritization rather than disease similarity scoring (Javed et al., 2015).

This positioning matters because PhenoGnet’s task is downstream of entity-level representation learning. It first learns aligned embeddings for genes and HPO terms from graph structure and cross-view supervision, then composes disease embeddings from those entities. That distinguishes it from phenotype extraction systems, phenotype–gene link predictors, and patient-specific causal-gene ranking systems. A plausible implication is that PhenoGnet occupies an intermediate level in the phenotype–genotype modeling stack: more structured than ontology-only disease similarity measures, but less patient-specific than diagnostic gene prioritization frameworks.

Within that niche, the paper’s strongest claim is narrow and clear. PhenoGnet is a two-view graph representation learning framework that combines a GCN over a HumanNet gene interaction graph and a GAT over the HPO phenotype ontology, aligns the views with a shared-weight MLP under a bidirectional supervised contrastive objective, and uses mean-pooled entity embeddings plus cosine similarity to score disease pairs. On the reported benchmark of 1,100 similar and 866 dissimilar rare disease pairs, its strongest configuration is the gene-based disease embedding, with AUPRC 0.9012 and AUROC 0.8764, exceeding the reported CoGO baseline and supporting the view that latent molecular structure is the most discriminative signal in this formulation (Baminiwatte et al., 17 Sep 2025).