Knowledge-Informed Gene Embeddings

• Knowledge-informed gene embeddings are vector representations of genes shaped by integrating omics data, literature-derived text, and structured knowledge graphs.
• They fuse multiple modalities to capture complementary biological signals, enabling improved prediction of gene interactions and disease associations.
• These embedding techniques offer robust frameworks for gene regulatory network inference, pathway analysis, and contextualized clinical predictions.

Knowledge-informed gene embeddings are vector representations of genes whose geometry is shaped by structured biological information rather than by expression matrices alone. In the recent literature, this includes multimodal molecular embeddings that integrate omics, literature, and knowledge graphs; ontology-grounded embeddings for gene–disease association; embeddings derived from protein–protein interaction topology; graph-derived priors for gene regulatory network inference; and patient- or cell-level models in which gene-level knowledge provides structural bias or auxiliary context (Zheng et al., 10 Jul 2025). A crucial boundary condition is that not every gene-aware model is a gene-embedding model in the strict sense: Dual Aspect Embedding, for example, uses data-driven gene–gene interaction structure but ultimately produces a 100-dimensional cell embedding, not a standalone gene embedding (Goudarzi et al., 1 Sep 2025).

1. Knowledge sources and representational scope

Knowledge-informed gene embeddings are heterogeneous in their upstream knowledge sources. One line of work explicitly distinguishes three major sources: omics experimental data, literature-derived text data, and knowledge graph-based representations. In "Platform for Representation and Integration of multimodal Molecular Embeddings" (Zheng et al., 10 Jul 2025), nine embedding sources are collected across these categories and reformatted to 512-dimensional vectors, with identifiers standardized to HGNC symbols using MyGene.info, Ensembl REST API, and UniProt resources. The omics and molecular-data-driven group includes Gene2Vec, Omics, Geneformer, and ProtTrans; the text modalities are GenePT and BioLinkBERT; and the structured-knowledge modalities are struc2vec, Know2BIO+TransE, and Know2BIO+MurE (Zheng et al., 10 Jul 2025).

This categorization clarifies a central distinction within the field. Literature-derived embeddings encode contextual, functional, and disease-related information expressed in biomedical text, whereas KG-derived embeddings encode structured biomedical relations across curated resources. In the PRISME study, the strongest nontrivial overlap in adjusted SVCCA occurs between GenePT and BioLinkBERT, with SVCCA $=0.46$ and adjusted SVCCA $=0.27$, while the two Know2BIO KG embeddings show SVCCA $=1.0$, reflecting shared input graph structure (Zheng et al., 10 Jul 2025). The broader empirical pattern is that most embedding pairs show low correlation and low adjusted similarity, which the authors interpret as evidence that existing molecular embeddings capture largely non-overlapping signals (Zheng et al., 10 Jul 2025).

A separate literature-derived regime appears in "Predicting drug-gene relations via analogy tasks with word embeddings" (Yamagiwa et al., 2024). There, genes and drugs are embedded as normalized biomedical concepts from PubMed abstracts, either using BioConceptVec skip-gram or a newly trained skip-gram model over about 35 million PubMed abstracts. In that setting, knowledge enters through concept normalization with PubTator and through pathway-conditioned post hoc structure from KEGG. This suggests that literature-trained gene vectors can be knowledge-informed in a weaker sense even when the embedding objective itself is purely distributional (Yamagiwa et al., 2024).

The scope of the term also depends on whether external prior knowledge is used. Some methods are knowledge-informed in the classic sense because they encode curated ontologies, pathways, or KGs. Others are better described as data-driven relational embeddings. Dual Aspect Embedding is explicit on this point: it infers gene–gene interactions directly from the same scRNA-seq dataset using GENIE3-style random forests, then incorporates that structure through a Cell-Leaf Graph and LINE, without using curated pathways, ontologies, or prior interaction networks (Goudarzi et al., 1 Sep 2025).

2. Ontology-grounded and knowledge-graph embeddings

A major strand of research learns gene embeddings from ontologies and heterogeneous biomedical KGs. "Neuro-symbolic representation learning on biological knowledge graphs" (Alshahrani et al., 2016) is an early formulation: genes are directly represented as graph nodes in a heterogeneous knowledge graph containing genes/proteins, diseases, chemicals, ontology classes from GO, HPO, and Disease Ontology, and annotation instances. The graph is deductively closed with ELK over OWL 2 EL, expanding an initial graph of 7,855,737 triples by 5,664,387 inferred triples, and embeddings are then learned from edge-labeled random walks using skip-gram with embedding size 512, walk length 20, number of walks 100, and context size 10 (Alshahrani et al., 2016). The resulting gene embeddings are informed by direct annotations, interactions, disease links, ontology hierarchy, and inferred semantics.

Ontology integration becomes more explicit in "Predicting Gene-Disease Associations with Knowledge Graph Embeddings over Multiple Ontologies" (Nunes et al., 2021). That paper constructs three KG variants—HP, HP+GO, and HP+LD+GO—combining phenotype ontology, function ontology, annotations, and logical-definition bridges. Five KG embedding methods are evaluated, all producing 200-dimensional embeddings. The best reported result is 0.745 WAF using HP+GO+LD + OPA2Vec + Random Forest, and the paper highlights that employing KGEs combined with the Hadamard operator reaches an AUC score of 0.821 (Nunes et al., 2021). The gains from multi-ontology integration are explicitly described as comparatively small, and the paper attributes this in part to there being only 351 LDs linking HP and GO (Nunes et al., 2021).

"A Systematic Evaluation of Knowledge Graph Embeddings for Gene-Disease Association Prediction" (Canastra et al., 11 Apr 2025) extends this ontology-grounded perspective by comparing link prediction and node-pair classification on KGs built from GO, HP, Disease Ontology, and GO–HP semantic bridges via logical definitions and ontology mappings. The largest graph, G+H+D+L+M, contains 8,881 genes, 36,028 diseases, and 5,732 gene-disease pairs, with ontology statistics reaching 325,221 classes depending on the graph version (Canastra et al., 11 Apr 2025). A key empirical result is that adding cross-ontology links often produces a greater impact than merely adding DO, and that link prediction methods better exploit semantic richness encoded in the KG (Canastra et al., 11 Apr 2025). This supports a recurring conclusion: semantic bridges can matter more than isolated ontology expansion.

"Knowledge Graph Completion based on Tensor Decomposition for Disease Gene Prediction" (Wang et al., 2023) takes a different KG route. KDGene builds a biological KG with 48,340 entities and 1,231,084 triples from DisGeNet, SymMap, STRING, GO resources, and KEGG, then learns entity and relation embeddings through an interactional tensor decomposition model. The task is tail prediction for triples of the form $(\text{disease}, \text{disease\_gene}, ?)$, and genes are embedded through disease associations, protein–protein interactions, GO annotations, pathway memberships, and disease–symptom context (Wang et al., 2023). Compared with the direct tensor-decomposition parent model CP-N3, KDGene improves HR@1/3/10/50 by 39.86%, 47.35%, 52.12%, and 31.81%, and MAP@1/3/10/50 by 39.85%, 46.30%, 47.52%, and 45.39%, averaging 43.77% improvement (Wang et al., 2023). The same paper also reports a cautionary result: among six KG variants, KG$_2$, which adds disease–symptom information, performs best, while adding PPI and GO/pathway relations does not consistently improve performance and may degrade it (Wang et al., 2023). This directly challenges the misconception that more prior biological knowledge is automatically better.

A more recent hierarchy-aware formulation appears in "Graph Neural Network based Hierarchy-Aware Embeddings of Knowledge Graphs: Applications to Yeast Phenotype Prediction" (Kronström et al., 5 May 2026). There, a heterogeneous Saccharomyces cerevisiae KG is built from SGD, BioCyc, GO, APO, ChEBI, INO, MI, and RO, and a heterogeneous GraphSAGE model is regularized with box-based semantic losses derived from ontology hierarchies. Over 10-fold cross validation, the best predictive model improves from $R^2=0.360$ with prior box embeddings to $R^2=0.377$ with distance-based semantic loss, and reaches $R^2=0.415$ on triple knockouts without retraining the architecture (Kronström et al., 5 May 2026). The paper is explicit that simply adding subClassOf links to the graph yields little gain relative to semantic regularization, which is a stronger statement about how ontology knowledge should be injected (Kronström et al., 5 May 2026).

3. Learning paradigms for knowledge injection

The literature uses several distinct mechanisms to inject knowledge into gene representations. One family learns embeddings directly from graph walks. RDF2Vec-based workflows exemplify this pattern. "Multi-dataset and Transfer Learning Using Gene Expression Knowledge Graphs" (Sousa et al., 26 Mar 2025) constructs a gene expression KG linking patient nodes, gene nodes, and GO terms, then learns 500-dimensional RDF2Vec node embeddings with maximum number of walks per entity 500, maximum walk depth 4, skip-gram Word2Vec, window size 5, epochs 5, negative samples 5, and learning rate 0.025 (Sousa et al., 26 Mar 2025). Expression values are transformed into binary patient–gene links using the rule $GE_{P_i,G_k}^{(z)} > 1$, while GO and GOA provide the domain-specific knowledge backbone (Sousa et al., 26 Mar 2025). This allows integration of multiple datasets even when measured gene sets differ.

A second family uses embeddings as structural priors inside downstream neural architectures. "Using ontology embeddings for structural inductive bias in gene expression data analysis" (Trębacz et al., 2020) constructs a gene–gene KNN graph from DL2Vec ontology embeddings over GO, UBERON, and MP, using cosine distance to connect each gene to its $K$-nearest neighbors. The graph then sparsifies a GCN for patient classification from METABRIC gene expression (Trębacz et al., 2020). With 24368 genes as input, OntoGCN reaches 72.3 ± 2.9 accuracy on PAM50 with training size 100 and 81.2 ± 0.8 with training size 1500, outperforming MLP and slightly exceeding GeneMANIA and STRINGdb graph variants on the same task (Trębacz et al., 2020). The technical role of the embedding here is not to provide a final gene vector for external reuse, but to impose semantic locality on graph convolution.

A third family uses embeddings as priors in generative inference. "InfoSEM: A Deep Generative Model with Informative Priors for Gene Regulatory Network Inference" (Cui et al., 6 Mar 2025) defines a directional pair prior using BioBERT gene embeddings $=0.27$0 and $=0.27$1:

$=0.27$2

This makes the prior mean over regulatory effects a learned linear function of textual gene embeddings, rather than a zero-centered sparsity prior (Cui et al., 6 Mar 2025). In InfoSEM-BC, the model further decomposes $=0.27$3, using embeddings to regularize effect magnitudes $=0.27$4 and observed labels to regularize edge-probability logits $=0.27$5 (Cui et al., 6 Mar 2025). This is a particularly explicit example of knowledge-informed gene embeddings functioning as Bayesian side information rather than as ordinary predictive covariates.

A fourth family performs multimodal integration at the embedding level. PRISME concatenates aligned gene embeddings from all modalities and compresses them with an autoencoder whose encoder has hidden sizes 1024 and 512, with Leaky ReLU and a weighted MSE objective (Zheng et al., 10 Jul 2025). The final integrated representation is a 512-dimensional multimodal molecular embedding (Zheng et al., 10 Jul 2025).

Finally, some systems combine knowledge-informed graph encoders with perturbation models. PREDIKTOR constructs a patient-specific graph $=0.27$6 over a 7,800-gene universe, where edges come from DysRegNet-inferred dysregulated TF–target interactions and DrugBank drug–target links, then encodes the graph with a two-layer GCN (Bang et al., 6 Jul 2026). In parallel, a frozen condition-specific gene–gene attention model pretrained on LINCS L1000 predicts a simulated post-perturbation transcriptomic profile over 978 landmark genes. The two patient-drug views are aligned in a shared latent space via a CLIP-style contrastive objective (Bang et al., 6 Jul 2026). This suggests that knowledge-informed gene embeddings increasingly appear as components of multi-view representation systems rather than as isolated vector tables.

4. Gene-centric versus cell- and patient-centric formulations

A recurring source of confusion is the difference between gene-centric embeddings and models that merely use gene structure to improve another entity’s representation. Dual Aspect Embedding is the clearest example. The method begins from the expression matrix $=0.27$7, retains the top 2000 highly variable genes in a working matrix $=0.27$8, constructs a Cell-Leaf Graph from random-forest leaf memberships, fuses it with a KNN graph over cells, and uses LINE to generate a 100-dimensional cell embedding (Goudarzi et al., 1 Sep 2025). The paper explicitly states that the final feature-importance ranking and regulatory network extraction is not performed, and that the “ultimate cell embedding serves as a fresh latent representation of our cells” (Goudarzi et al., 1 Sep 2025). Gene–gene interactions are therefore implicit and auxiliary, not the primary learned object.

By contrast, GT-GRN is gene-centric in its final prediction target. It constructs three per-gene modalities—VAE expression embeddings, BERT-derived global embeddings from multiple inferred GRNs, and Laplacian positional encodings—projects each to a shared space, and sums them: $=0.27$9 These fused node embeddings are then processed by a graph transformer for GRN link prediction (Teji et al., 23 Apr 2025). In the PBMC ablation study, the global embeddings only condition reaches 0.8860 AUROC, higher than gene expression embeddings only at 0.8693 and structural positional encodings only at 0.8480, while the full trimodal fusion reaches 0.8877 (Teji et al., 23 Apr 2025). The numerical pattern indicates that prior structural knowledge from multiple inferred networks is the dominant contributor in that setting.

Patient-centric contextualization appears in the gene expression KG framework of (Sousa et al., 26 Mar 2025) and in PREDIKTOR (Bang et al., 6 Jul 2026). In the former, RDF2Vec jointly embeds patients, genes, and GO terms, and the resulting patient embeddings are used by MLP and GCN classifiers. In the latter, graph-derived gene node embeddings exist internally, but the deployed representation is a drug-centric patient-drug embedding concatenated with a perturbation-derived transcriptomic vector (Bang et al., 6 Jul 2026). The paper nevertheless shows that the network-module embeddings are biologically richer at the gene level: for GO Molecular Function prediction, perturbation-module embeddings achieve $=1.0$0, AUPRC $=1.0$1, whereas network-module embeddings reach $=1.0$2, AUPRC $=1.0$3 (Bang et al., 6 Jul 2026).

This suggests a useful conceptual taxonomy. Some methods output reusable gene vectors; some output contextualized node states inside a downstream model; and some use gene-level knowledge only to regularize cell or patient embeddings. Treating these as equivalent obscures the actual object being represented.

5. Applications and empirical performance

Knowledge-informed gene embeddings have been evaluated in gene–gene interaction prediction, PPI prediction, gene–disease association, GRN inference, patient diagnosis, phenotype prediction, and pathway analysis. PRISME reports the highest performance in gene–gene interaction prediction with accuracy 0.77 and AUC 0.85, and in PPI prediction with accuracy 0.76 and AUC 0.83 (Zheng et al., 10 Jul 2025). These results support the broader claim that integrating text- and KG-based modalities with omics-derived embeddings yields robust, task-agnostic molecular representations (Zheng et al., 10 Jul 2025).

In disease gene prediction, KDGene reaches HR@1 0.126, HR@3 0.243, HR@10 0.416, HR@50 0.620, MAP@1 0.406, MAP@3 0.365, MAP@10 0.361, and MAP@50 0.370, outperforming both classical disease-gene baselines and KGC baselines (Wang et al., 2023). The ontology-grounded gene–disease prediction work based on multiple ontologies reports the best overall WAF of 0.745 with HP+GO+LD + OPA2Vec + Random Forest and highlights that cosine similarity alone underuses embedding information relative to supervised pairwise operators such as Hadamard (Nunes et al., 2021).

For gene regulatory network inference, GT-GRN reports strong link-prediction results on BEELINE datasets, including AUROC 0.9483, AUPRC 0.8990 on mESC-1000, AUROC 0.9402, AUPRC 0.8853 on mESC-500, AUROC 0.8793, AUPRC 0.5932 on hESC-500, and AUROC 0.8784, AUPRC 0.8604 on hESC-1000 (Teji et al., 23 Apr 2025). InfoSEM addresses the same general problem from a different angle and reports that textual-embedding priors improve GRN inference by 38.5% across four datasets, with an additional 11.1% boost when labeled data are integrated as priors (Cui et al., 6 Mar 2025). On unseen-gene evaluation with cell-type-specific ChIP-seq targets, InfoSEM-B improves hESC AUPRC from 0.265 to 0.331, and InfoSEM-BC raises hESC Hit@1% from 0.419 to 0.585 relative to DeepSEM (Cui et al., 6 Mar 2025).

Knowledge-informed embeddings also support phenotype prediction and clinical prediction. The hierarchy-aware yeast KG model reaches $=1.0$4 on double-knockout growth prediction with distance-based semantic loss and $=1.0$5 on trigenic deletions (Kronström et al., 5 May 2026). In patient diagnosis from gene expression KGs, the RDF2Vec-based framework improves single-dataset performance substantially; for example, on diabetes GSE184050, baseline F1 is 0.222, KG+MLP F1 is 0.675, and KG+GCN F1 is 0.757, while on breast cancer GSE86374, baseline F1 is 0.525, KG+MLP reaches 0.903, and KG+GCN 0.906 (Sousa et al., 26 Mar 2025). PREDIKTOR, which aligns a knowledge-graph network view with a perturbation view, achieves AUROC 0.837 ± 0.0338 and AUPRC 0.829 ± 0.0491 in patient-split TCGA evaluation, and AUROC 0.661 in zero-shot transfer to I-SPY2, improving AUROC by 5.6% over the next-best baseline (Bang et al., 6 Jul 2026).

Pathway analysis provides a different downstream use. Gene Set Proximity Analysis embeds genes from a STRING v11 PPI network with 12,396 proteins and 324,152 interactions filtered at confidence $=1.0$6, then expands curated gene sets by proximity in the latent space before applying a GSEA-like statistic (Cousins et al., 2022). Across 33 disease-matched datasets with known disease-associated KEGG pathways, GSPA outperforms GSEA with $=1.0$7 and NGSEA with $=1.0$8, winning against GSEA on 25/33 datasets and against NGSEA on 21/33 datasets (Cousins et al., 2022). This indicates that knowledge-informed gene geometry can improve pathway-level inference even when no end-to-end supervised model is trained.

6. Misconceptions, limitations, and current directions

Several recurring misconceptions are addressed directly by the literature. First, knowledge-informed does not always mean external prior knowledge. Dual Aspect Embedding uses only data-driven gene–gene interactions inferred from the same scRNA-seq dataset and is therefore better described as an interaction-informed cell-embedding method than as a canonical prior-knowledge-informed gene embedding framework (Goudarzi et al., 1 Sep 2025). Second, graph-based does not always mean graph neural network. The DAE abstract refers to “a graph neural network,” but the method section actually uses LINE, with first- and second-order proximity losses and no message-passing equations of the form $=1.0$9 (Goudarzi et al., 1 Sep 2025).

A further misconception is that adding more structured biology always improves representation quality. KDGene’s KG-content study shows that disease–symptom information is especially helpful, while adding PPI and GO/pathway relations can act as noise (Wang et al., 2023). The systematic KGE comparison likewise finds that enriching disease semantics with DO gives only slight improvement, whereas additional cross-ontology links often have greater impact (Canastra et al., 11 Apr 2025). These results suggest that relation quality, semantic alignment, and graph wiring are at least as important as raw graph size.

Bias and portability remain significant concerns. The PRISME paper notes that literature embeddings can reflect publication bias and summary-writing conventions, while KG embeddings reflect curation decisions and graph incompleteness (Zheng et al., 10 Jul 2025). GSPA inherits limitations of STRING-based PPI knowledge, including likely underrepresentation of some disease areas and the absence of tissue-specific rewiring (Cousins et al., 2022). InfoSEM emphasizes that standard supervised GRN benchmarks can be dominated by gene-specific bias and class imbalance, and shows that one-hot logistic regression can perform similarly to recent supervised methods on conventional seen-gene benchmarks (Cui et al., 6 Mar 2025).

The current direction of the field is therefore not a single architectural consensus but a convergence on several principles. The papers collectively point toward multimodal fusion with explicit handling of complementarity (Zheng et al., 10 Jul 2025), ontology or hierarchy injection as a geometric regularizer rather than mere extra edges (Kronström et al., 5 May 2026), stronger cross-ontology linking (Nunes et al., 2021), improved treatment of partial modality availability and missing data (Zheng et al., 10 Jul 2025), context-aware transfer across datasets and therapies (Sousa et al., 26 Mar 2025), and stricter evaluation on unseen genes, unseen drugs, unseen tissues, or zero-shot clinical cohorts (Cui et al., 6 Mar 2025). A plausible implication is that the most durable formulations of knowledge-informed gene embeddings will be those that separate the biological roles of different priors—function, hierarchy, interaction, regulation, and perturbation—rather than collapsing them into a single undifferentiated vector space.