Structure-Aware Contrastive Learning for Biomedical Embeddings: Bridging the Gap Between HPO and Clinical Literature

1 Introduction

Accurate characterisation of patient phenotypes is crucial for diagnosing rare genetic diseases. After a clinical evaluation, the physician produces a medical report in which the phenotypes observed during that evaluation are noted and used as the basis to identify and suggest one or more possible genetic diseases (which is called differential diagnosis). A definitive genetic diagnosis is only obtained when a pathogenic variant associated with the corresponding disease and its related clinical phenotypes is identified in the genome. With more than 10,000 distinct rare disorders identified, clinicians rely heavily on standardized vocabularies, such as the Human Phenotype Ontology (HPO) [Robinson et al., 2008; K¨ohler and others, 2020], to bridge the gap between clinical observations and the genetic diagnostic [Genetic Alliance and others, 2010]. Therefore, the application of LLM-related technologies to biomedical tasks in general, and to genetic diagnostic in particular, must be sustained in a precise knowledge and understanding of the terminology linked to phenotypes and diseases.

Large Language Models (LLMs) pre-trained on biomedical corpora, such as BioBERT [Lee et al., 2019] and PubMedBERT [Gu et al., 2021], are the state-of-the-art models when applied to biomedical information retrieval tasks. However, a significant semantic gap remains when applying these models to deep phenotyping, i.e., the task of describing the disease as completely and precisely as possible. While LLMs excel at capturing distributional semantics from unstructured text, they often fail to comprehend the strict, biologically grounded hierarchical relationships encoded in expert ontologies [Hao et al., 2020]. For instance, a general-purpose biomedical LLM may recognize "cataract" and "glaucoma" as related eye diseases based on lexical co-occurrence, but it lacks the structural knowledge that links "iris coloboma" and "renal failure", two apparently unrelated phenotypes, as phenotypically associated manifestations of Lowe Syndrome [Lowe et al., 1952].

While existing biomedical LLMs largely focus on entity linking (recognizing synonyms), our work specifically targets the optimization of embeddings for phenotype-disease association relationships, essential for differential diagnosis. Current approaches to bridge this gap typically rely on fine-tuned embeddings using the ontology's graph structure [Garnot and Landrieu, 2020] or phenotype definition-based contrastive learning [Wen et al., 2025]. Yet, these methods face two critical limitations. First, ontologies suffer from data scarcity in terms of natural language: HPO terms are often defined by short, dictionary-like descriptions that fail to capture the lexical variability of real-world clinical reports. Second, standard topological similarity measures between pairs of terms in the ontology (e.g., Resnik [Resnik, 1995], Lin [Lin, 1998]) prioritize ontology graph distance over clinical co-occurrence information on observed phenotypes, misaligning the embedding space for diagnostic tasks where functional association is potentially more critical than taxonomic proximity.

In this work, we propose a novel Neuro-Symbolic Alignment Framework that injects both ontological structure and rich clinical context into the latent space of LLMs. Our approach diverges from previous work in three key aspects:

1. Literature-Mined Data Augmentation: Instead of relying solely on static HPO definitions, we automatically generate a large training corpus of abundant and diverse phenotype descriptions mined from scientific literature (PubMed), exposing the model to the pragmatic and real usage of terms.
2. Annotation-Based Supervision: We guide the contrastive learning process using a "Disease-Overlap" similarity metric (see section 3.2) derived from the Relative Best Pair strategy [Gong et al., 2018]. This forces the model to learn relationships between phenotypes based on their shared pathophysiology rather than mere graph distance.
3. Rank-Aware Optimization: We employ AnglE Loss [Li and Li, 2024] to optimize the vector space geometry, ensuring that phenotypes sharing a disease cause are pulled mathematically closer than those that merely sound alike, effectively translating clinical relevance into vector distance.

We evaluate our method against strong baselines, including SapBERT [Liu and others, 2021], on intrinsic semantic correlation tasks and extrinsic downstream applications such as Gene-Disease association prediction (GSC+) and literaturebased phenotype retrieval (JAX). All the code is available in a GitHub repository and our best model is available in Hugging Face.

2 RelatedWork

2.1 Biomedical Language Models

The adaptation of BERT-based architectures to the biomedical domain has become the standard for clinical NLP. BioBERT [Lee et al., 2019] demonstrated that pre-training on PubMed abstracts significantly improves performance on named entity recognition (NER) and relation extraction. PubMedBERT [Gu et al., 2021] further advanced this by pretraining from scratch on biomedical text, overcoming the vocabulary mismatch inherent in models initialized on generaldomain corpora. More recently, SapBERT [Liu and others, 2021] introduced a self-alignment pre-training scheme using the UMLS metathesaurus to cluster synonymous medical terms. While SapBERT successfully performs at entity linking (grouping synonyms), it does not explicitly optimize for the hierarchical or disease-association relationships required for differential diagnosis.

2.2 Semantic Similarity in Ontologies

Traditionally, semantic similarity between ontology terms has been quantified using Information Content (IC) approaches. Resnik [Resnik, 1995] defined similarity based on the IC of the Most Informative Common Ancestor (MICA). Lin [Lin,

1998] refines Resknik by applying a normalization to the measure. Although robust for graph analysis, these topological metrics ignore the extensional meaning of phenotypes, i.e., the set of diseases they annotate. Alternative approaches which account for this, such as those used in Phenomizer [K¨ohler and others, 2009] or LIRICAL [Robinson and others, 2020], utilize Bayesian frameworks or pre-computed statistical scores for diagnosis but do not generate embeddings that can be used in downstream tasks, e.g., model building pipelines. Our work bridges this traditionally independent approaches by using annotation-based similarity as a supervision signal for neural embeddings.

2.3 Knowledge-Enhanced Representation Learning

Recent efforts in "Neuro-Symbolic AI" seek to integrate Knowledge Graphs (KGs) into deep learning. Methods like TransE [Bordes and others, 2013] and Node2Vec [Grover and Leskovec, 2016] generate graph embeddings but discard the rich textual information of nodes. Conversely, text-based approaches like CODER [Yuan et al., 2020] use contrastive learning to align medical terms with KG relations. However, CODER and similar models [Cui et al., 2020] typically use the KG structure (parent-child) as the sole ground truth. Our approach is distinct in that we utilize literature-mined contexts as a data augmentation strategy, inspired by distant supervision techniques in phenotype recognition [Luo et al., 2021], and optimize for both clinical outcome (disease overlap) and taxonomy.

3 Methodology

Our approach consists of three main components: (1) generating a literature-based corpus to enrich phenotype contexts, (2) transforming to a pairs dataset with anchor-based sampling and disease-overlap similarity metric, and (3) finetuning the embedding model using a rank-based loss function. This pipeline is shown in Fig. 1.

3.1 Literature-Based Corpus Generation

Standard approaches to generate a corpus based on HPO typically use the single, short definition provided by the HPO as the basic element to compose the training dataset. We will demonstrate that this basic approach for learning robust representations can be improved. For such purpose, we developed a corpus generation pipeline using literature included in PubMed with the focus of providing abundant quality context for all terms in the ontology. The algorithm is shown in Alg. 1.

For each term t ∈ V (the set of HPO terms), we queried the PubMed database to retrieve scientific papers associated with t. Sentences containing the term or its synonyms provided by HPO were extracted. To ensure high-quality supervision, we applied a Dynamic Context Windowing strategy:

• Filtering: We discarded sentences containing explicit negations (e.g., "no signs of...") detected via dependency parsing, as well as enumeration lists containing more than three distinct phenotypes to avoid false cooccurrences.

• Windowing: For sentences exceeding 50 words, we extracted a window of ±25 words around the target term to maintain local semantic focus.

This results in a distribution of linguistic contexts s, for each phenotype t, denoted as St = {s1, s2, ..., sn}, rather than a single definition.

Algorithm 1 Context Extraction and Refinement

Require: Set of HPO terms V , PubMed Papers P Ensure: Training Corpus C = {(s, t) | s is sentence, t ∈ V } 1: for all paper p ∈ P do 2: Sraw ← SentenceSplitter(p) 3: for all sentence s ∈ Sraw do 4: Clean references and brackets from s 5: IdentifiedTerms ← FindAllHPOTerms(s, V ) 6: if |IdentifiedTerms| > 3 then 7: continue {Discard enumeration lists} 8: end if 9: for all term t ∈ IdentifiedTerms do 10: if DependencyParsing(s, t) indicates Negation then 11: continue {Discard negated contexts} 12: end if 13: if Length(s) > 50 words then 14: s′ ←ExtractWindow(s, center=t, window=25) 15: else 16: s′ ← s 17: end if 18: Add (s′ , t) to C 19: end for 20: end for 21: end for

3.2 Neuro-Symbolic Pairing Semantic Similarity Labeling (Ground Truth)

To supervise the model, we require a "gold standard" similarity score between phenotype pairs. Instead of relying solely on topological measures (like graph distance), we employed an annotation-based similarity measure that aims towards a better recapitulation of clinical knowledge, through capturing the real extensional meaning of phenotypes by shared association with diseases. This similarity is based on the RelativeBestPair (RBP) similarity function [Gong et al., 2018].

Let D(t) be the set of diseases annotated to a phenotype term t or any of its descendants (sourced from the OMIM database):

This definition integrates the ontology's topological structure via the True Path Rule: a term inherently carries the clinical semantics of its descendants. By basing similarity on these propagated sets, our metric implicitly captures the hierarchical information (similar to Resnik or Lin) but grounds it in extensional disease overlap. Thus, a separate topological term is not required, as the graph structure defines the composition of D(t).

Given a source phenotype ti and a target tj, we first compute a directional similarity score δ(ti, tj). This score measures the coverage of the target's disease set by the source, weighted by the specificity of the target:

Here, α is a saturation hyperparameter (set to 0.01). The term 1/|D(tj)| acts as an Information Content weighting. The final symmetric semantic similarity Sim(ti, tj), used as the label ysem, is the normalized mean of the bidirectional scores:

Anchor-Based Hard Sampling

Given the sparse topology of the HPO graph, uniform random sampling of pairs from the |V | ≈ 18, 000 terms results in a distribution heavily skewed towards negligible similarity. Since most random pairs share only the root or highly generic ancestors (e.g., 'Phenotypic abnormality'), they constitute "easy negatives" that provide minimal gradient signal for ranking optimization. To enable efficient learning, we employed an Anchor-Based Hard Sampling strategy. For each anchor term t, we generated pairs in three categories:

1. Positive Pairs (33%): (t, t) pairs of the same phenotype using distinct sentences from St. This enforces semantic invariance across different linguistic contexts.
2. Hard Negatives (33%): Pairs (t, thard) where thard is a sibling or cousin node with high semantic similarity (0.3 < Sim(t, thard) < 0.7). We empirically capped the upper threshold at 0.7 to avoid "False Negatives": pairs with similarity > 0.7 often represent near-synonyms or clinically indistinguishable phenotypes. Treating such highly similar terms as negatives would force the model to artificially separate concepts that should naturally cluster together, potentially destabilizing the training.
3. Random Negatives (33%): Pairs (t, trand) with low similarity to preserve global structure.

3.3 Fine-Tuning

We employ a bi-encoder architecture optimized using AnglE Loss (Angle-optimized Text Embeddings) [Li and Li, 2024]. Unlike traditional cosine-based losses (e.g., CoSENT [Huang et al., 2024]) which can result in vanishing gradients when the similarities approach 1, a frequent occurrence in ontology sub-branches, AnglE introduces an angular optimization in complex space.

We formulate the objective as a contrastive loss with an angular margin m, enhancing the discriminative power of the embeddings:

where θ+ is the angle between the positive pair (determined by high annotation similarity), θ- corresponds to negative pairs, and τ is the temperature parameter. We employ a Partial Freezing strategy.

4 Experiments

4.1 Experimental Setup

To ensure a comprehensive evaluation, we selected three distinct base models representing different paradigms: (1) PubMedBERT [Gu et al., 2021], a domain-specific model pretrained from scratch on biomedical text; (2) SapBERT [Liu and others, 2021], the current state-of-the-art for biomedical entity alignment initialized from PubMedBERT; (3) allminiLM-L6-v2 [Wang et al., 2020] and all-mpnet-base-v2, general-purpose efficient sentence transformers, included to assess domain transfer capabilities. For comparative analysis, we evaluated both the off-the-shelf versions (zero-shot) and versions fine-tuned using our proposed methodology.

Hyperparameter Optimization. We performed a Bayesian hyperparameter search using the Optuna framework [Akiba and others, 2019] to maximize the Spearman correlation on a held-out validation set of HPO pairs. The optimal configuration identified and used for all subsequent experiments was: 4 training epochs with a batch size of 64 and a maximum sequence length of 256. We utilized the AdamW optimizer with a weight decay of 0.05, a maximum learning rate of 7.87 × 10-5 , and a linear warmup over the first 6% of steps. We applied a Partial Freezing strategy for the bottom 6 layers of the encoder. Training was accelerated using Automatic Mixed Precision (AMP).

4.2 Components Study

To quantify the contribution of each component, we evaluated four configurations aside from the base models:

• Baseline: Topological Fine-Tuning (Defs Only + Lin Similarity).
• +Context: Using PubMed Corpus + Lin Similarity.
• +Metric: Using Defs Only + Disease-Overlap Similarity.
• Full Method: Using PubMed Corpus + DiseaseOverlap Similarity.

4.3 Evaluation Tasks

Task 1: Intrinsic Semantic Correlation. How well the LLM recapitulates HPO structure? We computed the Spearman rank correlation between the cosine similarity of the model's embeddings and the ground-truth semantic similarity (both Lin and Disease-Overlap) on a held-out test set of HPO pairs.

Task 2: Extrinsic Phenotype Association (GSC+ and JAX). How well the LLM associates HPO phenotypes to biomedical texts? We utilized the GSC+ corpus [Lobo et al., 2017], which consists of 228 manually annotated PubMed abstracts, with a total of 1933 annotations that cover 497 unique HPO concepts. We formulated this as a retrieval task: for each abstract, we ranked HPO terms and measured Recall@1, Recall@5 and Mean Reciprocal Rank (MRR). Additionally, we use the JAX corpus [Luo et al., 2021], which provides document-level annotations of HPO terms in 131 PMC fulltext articles, with a total of 988 unique HPO concepts.

Task 3: Synthetic Patient DiseaseRanking. Howdoes the LLM perform in modeling patients based on combination of phenotypes? To evaluate diagnostic reasoning, we simulated patients using the HPO annotation database. For a given disease D, we randomly sampled k ∈ {5, 10} associated phenotypes. We generated a "Patient Vector" by averaging the embeddings of the corresponding PubMed sentences extracted for each phenotype. The model was tasked with retrieving the correct disease D from the set of all OMIM diseases. We report Top-1 and Top-5 Accuracy and MRR for 1000 generated synthetic patients, where diseases are represented with a "Disease Vector" averaging the embeddings of the PubMed sentences extracted for all the phenotypes related to the disease.

Task 4: Real-World Phenopacket Prioritization. Asking the same question again as in Task 3, we evaluated performance on real-world clinical cases from the Phenopacket Store [Danis and others, 2025]. This repository offers cohorts of GA4GH Phenopackets that represent individuals with Mendelian diseases. We selected a subset of cases with confirmed genetic diagnoses. For each case, we extracted the observed phenotypes (excluding negated terms), converted them to embeddings using our model, and ranked candidate diseases the same way as the synthetic patients. This represents the most challenging and realistic evaluation setting. We report Top-1 and Top-5 Accuracy and MRR for a total of 6556 patients with at least 3 phenotypes in its description and a confirmed Omim disease within the HPO dataset.

Task 5. General Semantic Capabilities. Does our model present catastrophic forgetting? We evaluated the performance on three standard biomedical and STS benchmarks to evaluate the degradation of performance in general semantic tasks: BIOSSES (Biomedical) [So˘gancıo˘glu et al., 2017], STS-B (General) [Cer and others, 2017], and SICK-R (General) [Enevoldsen and others, 2025; Marelli and others, 2014; Muennighoff et al., 2022].

5 Results

5.1 Components Study

To disentangle the contributions of data augmentation (PubMed contexts) and the supervision signal (Semantic Metric), we compared four fine-tuning configurations (see section 4.2) against the pre-trained PubMedBERT baseline. Table 1 summarizes the performance on intrinsic semantic correlation (Spearman ρ on held-out HPO pairs) and extrinsic genedisease association retrieval (GSC+).

Comparing the models trained on standard definitions (Defs), we observe that shifting from topological similarity (Lin) to annotation-based similarity (RBP) improves the Top5 Accuracy on GSC+ from 0.30 to 0.31 and the MRR on JAX from 0.013 to 0.014. This suggests that even with limited textual context, forcing the model to respect disease cooccurrence patterns aligns the latent space better with downstream diagnostic tasks.

A crucial finding is the behavior of the PubMed + Lin configuration. Surprisingly, augmenting the model with rich literature contexts while supervising with a graph-based metric (Lin) degrades performance below the baseline (GSC+ Top-1 drops to 0.10). We hypothesize that this represents a neuro-symbolic mismatch: the rich functional descriptions in PubMed papers often imply clinical connections that contradict the rigid, graph-based distance of Lin similarity, introducing noise during training [Gong et al., 2018].

In contrast, our proposed PubMed + RBP configuration achieves a synergistic effect. By aligning rich clinical language (PubMed) with a clinically grounded metric (RBP), we achieve a remarkable improvement: GSC+ Top-1 Accuracy doubles compared to the baseline (0.13 → 0.26), and intrinsic Spearman correlation reaches 0.84. This confirms that data augmentation requires a commensurate increase in the sophistication of the supervision signal.

5.2 Comparative Analysis

We benchmarked our proposed fine-tuning strategy (using PubMed contexts and RBP metric) across four distinct model architectures: PubMedBERT (domain-specific base), SapBERT (state-of-the-art entity alignment), MiniLM (efficient distilled model) and Mpnet (best general model). Table 2 details the performance shifts from baseline to fine-tuned states.

Performance vs. Architecture

Both PubMedBERT and SapBERT demonstrated significant gains, validating the universality of our method. Notably, our fine-tuned PubMedBERT achieves a GSC+ MRR of 0.320, effectively bridging the gap with the base SapBERT (MRR 0.215) and surpassing it significantly. This indicates that our neuro-symbolic alignment can instill deep clinical knowledge into a standard biomedical model without requiring the massive entity-level pre-training of SapBERT.

When applying our method to SapBERT (FT-SapBERT), we achieved the highest overall performance across all tasks (GSC+ MRR 0.340). This suggests that our contribution is additive: SapBERT provides robust synonym resolution (entity linking), while our method injects hierarchical and pathophysiological reasoning. The combination yields a model that is superior at both recognizing terms and understanding their clinical implications.

Conversely, the distilled MiniLM model suffered a performance collapse upon fine-tuning (Spearman dropped to 0.748, GSC+ Top-1 halved to 0.087). We hypothesize that the reduced parameter space of MiniLM lacks the capacity to accommodate the drastic re-organization of the embedding space required by our loss function. Something similar occurs with the Mpnet model, where the performance does not upgrade with our fine-tuning. This model has a really strong spearman correlation (0.855) but the metrics are way worse than the fine-tuned models in the rest of experiments.

Analysis of Downstream Tasks

GSC+ (Gene-Disease): The improvement in Top-1 Accuracy is the most striking result. FT-PubMedBERT retrieves the correct phenotype for a gene description in the first position 26.1% of the time, compared to just 13.1% for the baseline. This doubling of precision significantly enhances the model's utility for practical applications. For instance, in automated curation (extracting structured data from literature) and variant prioritization (ranking candidate genes in genomic analysis), a high Top-1 accuracy directly minimizes the need for manual verification by experts.

JAX (Literature Retrieval): While absolute MRR scores remain low due to the extreme difficulty of the task (matching short phenotype queries against full-text noise), our method achieves a relative improvement of over 130% compared to baselines. This confirms that the model is learning a robust signal that persists even in high-noise environments, although specialized retrieval architectures (e.g., re-rankers) would be needed for practical deployment on full texts.

5.3 Diagnostic Reasoning

To rigorously evaluate the clinical utility of our aligned embeddings, we conducted a experiment ranging from controlled simulations to large-scale real-world patient data.

We generated 1000 synthetic patients by sampling N = {5, 10} phenotypes from disease definitions in HPO to generate a synthetic clinical case for the patient with a sentence per phenotype. Phenotypes were represented by random sentences mined from PubMed rather than canonical definitions. Additionally, we evaluated the models on the Phenopacket store dataset, comprising 6,556 curated clinical cases extracted from medical literature with and average of 8.56 phenotypes per patient. Unlike synthetic patients, these cases contain noise, incomplete phenotyping, and heterogeneous reporting styles. The phenotypes were also represented by random sentences from PubMed related to the phenotypes.

As shown in Table 3, baselines perform relatively well in controlled synthetic settings, with MiniLM even showing strong keyword-matching capabilities (Top-5 24.9% for N=5). However, this performance collapses in the real-world Phenopackets task, where base models achieve a Top-1 accuracy of only ∼3-4% and a Top-5 around 11%. They fail to bridge the gap between complex clinical narratives and structured disease entities. The performance of Mpnet is really similar, so we can intuit that general models do not have the necessary knowledge for this type of task, even with specific fine-tuning.

In contrast, our fine-tuned models exhibit remarkable robustness. FT-PubMedBERT achieves a Top-1 Accuracy of 17.5% and a Top-5 Accuracy of 34.1%, representing a 4-fold improvement in the Top-1 metric and a 3-fold improvement in the Top-5 retrieval zone compared to the baseline (11.4%). This means that in a real-world scenario, our model successfully places the correct diagnosis within the top 5 candidates for more than one-third of the complex cases, drastically reducing the search space for clinicians.

5.4 The Cost of Specialization

We evaluated performance on three standard biomedical and general STS benchmarks: BIOSSES (Biomedical), STS-B (General), and SICK-R. Table 4 presents the results.

We observe a divergent behavior depending on the training data source. When fine-tuning with HPO Definitions (FT Defs), the model retains or even improves general biomedical similarity (BIOSSES ρ = 0.85). This is likely because definitions preserve standard lexicographic structures.

However, our best diagnostic model (FT PubMed+RBP)

exhibits a significant performance drop on BIOSSES (ρ = 0.83 → 0.57). Rather than simple catastrophic forgetting, we interpret this as a necessary semantic shift. Standard benchmarks like BIOSSES reward lexical and topical similarity (e.g., "Eye pain" ≈ "Eye irritation"). In contrast, our model is optimized for functional similarity (e.g., "Eye pain" ≈ "Kidney failure" in the context of Lowe Syndrome).

This degradation in general benchmarks is the trade-off for the massive gains in Phenopacket retrieval (4x improvement). It confirms that the model has specialized effectively: it has ceased to be a general-purpose text encoder and has become a specialized diagnostic retriever.

6 Conclusion

In this work, we proposed a Neuro-Symbolic Alignment Framework to bridge the semantic gap between biomedical Large Language Models and expert knowledge graphs. By augmenting the Human Phenotype Ontology with literaturemined contexts and supervising the embedding space with a disease-overlap metric, we successfully instilled clinical reasoning capabilities into standard transformer architectures.

Our extensive experimental evaluation leads to three critical conclusions. The most significant finding is the dramatic performance gap in real-world scenarios. While standard baselines like PubMedBERT and SapBERT collapse when facing the noise and heterogeneity of real clinical reports (Phenopackets), our fine-tuned models maintain robust performance, achieving a 4-fold improvement in diagnostic retrieval (Top-1 Accuracy 17.5% vs 4.2%). This confirms that our method effectively aligns the latent space with pathophysiology, allowing the model to look beyond lexical surface forms.

We demonstrated that initializing with state-of-the-art weights (SapBERT) further boosts performance, but applying our methodology to a standard domain model (PubMedBERT) yields comparable or superior results to the original SOTA in diagnostic tasks. This suggests that how we align models (using disease logic) is as important as the data scale used for pre-training.

The decline of MiniLM highlight a boundary condition: deep semantic restructuring requires sufficient parameter capacity. Efficient models, while useful for keyword matching, appear too rigid to accommodate the complex, nontaxonomic relationships inherent in medical diagnosis.

While our neuro-symbolic alignment yields a 4-fold improvement over baselines in real-world scenarios, an absolute Top-1 accuracy of 17.5% (and Top-5 of 34.1%) on Phenopackets indicates that the problem of automated rare disease diagnosis via text is far from solved. The complexity of clinical variability, combined with the noise inherent in unstructured case reports, presents a significant barrier. Our results suggest that while embedding alignment is a necessary step, achieving clinical-grade accuracy likely requires integrating multimodal data (e.g., genomic variants) and perhaps reasoning capabilities beyond vector similarity. Consequently, current LLM-based phenotyping tools should be deployed as "human-in-the-loop" prioritization assistants rather than autonomous diagnostic agents.

Future work will explore extending this framework to other biomedical ontologies, such as the Gene Ontology (GO), and integrating these aligned embeddings into retrievalaugmented generation (RAG) pipelines for explainable rare disease diagnosis.

Acknowledgments

This work was supported by the Fundaci´on S´eneca-Agencia de Ciencia y Tecnolog´ıa de la Regi´on de Murcia (Spain)

(22308/FPI/23) and the Spanish Council of Science and Innovation through the Grant PID2022-136306OB-100 funded by MICIU/AEI/10.13039/501100011033 and FEDER/UE.

Contribution Statement

The authors Alejandro Cisterna-Garc´ıa and Juan A. Bot´ıa contributed equally.