acceptodds
Under review as a conference paper at ICLR 2027

Phenotype Driven Gene Prioritization and Fast Patient Retrieval with Rotational Knowledge Graph Embeddings

Abstract

The central challenge in variant interpretation for rare diseases is identifying the causal gene from a pool of candidates based on the patient's phenotypes. It is heavily dependent on subject domain expertise and is constrained to associations already documented in curated databases and the literature. In this study, we introduce a novel method - Patient Retrieval, Interpretation, and gene Scoring Model (PRISM), which is geometry aware, permutation invariant Set Transformer encoder that aligns variable size sets of patient phenotypes or knowledge graph (KG) entities, to candidate entities in rotational KG embedding (KGE) space. The KGE was built with the RotatE model over gene-phenotype and phenotype-phenotype relations from the Human Phenotype Ontology (HPO). Entity text descriptions were encoded with the pretrained biomedical language model, conferring inductive capability on PRISM. PRISM was trained on phenotypic profiles of approximately 22,000 rare disease patients, drawn from an in-house cohort of ∼30,000, to prioritize causal genes with a listwise multi-positive softmax loss function. On a held out cohort of more than 400 samples, integrated into the proprietary variant prioritization pipeline, PRISM recovered the causative pathogenic/likely pathogenic gene and variant within the Top-20 in 92.2% of cases and ranked the causative variant Top-1 in 57.8% of cases. Beyond the causal gene prioritization score, the same encoder embeds each patient's phenotypic profile in under half a second per inference, enabling retrieval of prior cases with similar phenotypes to aid variant interpretation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
permutation invariant representation learning, knowledge graph embeddings, RotatE, listwise ranking, phenotype driven gene prioritization, clinical genomics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_AyQcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PRISM, a phenotype-driven framework for causal gene prioritization. It combines frozen RotatE knowledge graph embeddings, biomedical text embeddings, and an information-content-aware Set Transformer to represent patients and rank candidate genes. The model is trained on approximately 22,000 clinical cases and evaluated on 410 held-out clinical patients and 42,224 simulated patients, with comparisons against SHEPHERD. Patient embeddings are also qualitatively assessed for similar-patient retrieval.

Critical strengths
  1. raining on a substantial real-world clinical cohort and separating patients by causal gene provide useful evidence about transfer to genes without positive training examples.
  2. integrating PRISM into the clinical pipeline improves the reported Top-1 performance, suggesting potential to reduce manual review effort.
Critical weaknesses
  1. There are no ablations testing the contributions of rotational scoring, frozen gene embeddings, text features, or information content. Moreover, PRISM is trained on the authors’ clinical cohort, whereas SHEPHERD uses released checkpoints. Differences in training data and knowledge resources therefore confound the architectural comparison.

  2. Test genes are absent as positive patient-training labels but remain present in the knowledge graph and scoring catalog, so the experiments establish a narrower form of generalization than scoring previously unknown genes.

Questions for authors
  1. How were the 410 evaluation patients selected, and how do they relate to the 2,460-patient test split? What is the evaluation unit in Figure 4(b)?

  2. Can the authors provide controlled ablations of the scoring heads and frozen gene embeddings, together with a comparison under matched clinical training data?

Reviewer_Yv1YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a phenotype-driven framework for causal gene prioritization and patient retrieval in rare disease diagnosis. PRISM first constructs a knowledge graph from Human Phenotype Ontology (HPO) and gene–phenotype associations, using RotatE to learn structured representations of genes and phenotypes. A Set Transformer then integrates patient phenotypes, incorporating biomedical text embeddings and ontology-derived information content to produce patient-level representations. The model employs rotational and cosine-based scoring heads with a listwise ranking objective to prioritize candidate causal genes. By freezing gene embeddings, PRISM aims to generalize to genes not observed as causal during training. Experiments on real clinical cohorts and simulated rare disease patients demonstrate improvements over SHEPHERD in causal gene and variant prioritization. The learned patient representations are also explored for patient clustering and retrieval.

Critical strengths
  1. The paper addresses an important and clinically relevant problem in rare disease diagnosis. The use of a relatively large real-world clinical cohort, rather than relying primarily on simulated patients, is a valuable aspect of the work. The gene-disjoint evaluation also reflects an effort to assess generalization beyond causal genes observed during training.

  2. The proposed framework provides a coherent integration of knowledge graph representations and patient-level phenotype modeling. Incorporating ontology-derived information content into a permutation-invariant Set Transformer is a reasonable approach to handling heterogeneous and variable-sized phenotype sets. The reported improvements over SHEPHERD, particularly in downstream variant prioritization, suggest potential practical value.

Critical weaknesses
  1. The claimed zero-shot generalization to unseen genes is not sufficiently justified. The paper employs a gene-disjoint split to evaluate generalization to unseen causal genes. However, RotatE embeddings are pretrained on a knowledge graph constructed from HPO and HPOA, which may already contain phenotype associations for the held-out genes. Although these genes do not appear as positive labels during patient-level training, their phenotype-related information may already be encoded in the pretrained embeddings. Consequently, it remains unclear whether the reported performance demonstrates generalization to genuinely novel gene–phenotype associations or primarily exploits existing knowledge graph information. A stricter evaluation excluding held-out gene–phenotype associations during KG pretraining would help clarify this issue.

  2. The methodological novelty and necessity of rotational knowledge graph embeddings remain insufficiently established. PRISM combines several established techniques, including RotatE, Set Transformer, information-content-weighted attention, and listwise ranking. Although preserving relational geometry is emphasized as a central contribution, the paper does not convincingly demonstrate why rotational scoring is preferable to conventional embedding similarity or other knowledge graph representation methods. It is therefore difficult to determine whether the improvements arise from the proposed geometric formulation, the pretrained knowledge graph representations, or the patient encoder.

  3. The experimental comparisons are not sufficiently comprehensive. The evaluation primarily focuses on SHEPHERD, while several relevant phenotype-driven gene prioritization methods, including Exomiser, Phen2Gene, Phrank, and other KG-based approaches, are not systematically compared. Although these methods have different input requirements, comparisons under appropriately controlled candidate gene sets would help establish the relative effectiveness of PRISM. Moreover, simpler alternatives based on phenotype semantic similarity or direct KG embedding matching are important for assessing whether the additional model complexity is necessary.

  4. The contributions of individual architectural components are not adequately validated. Several components, including the information-content-aware attention mechanism, frozen gene embeddings, and hybrid rotational–cosine scoring, are presented as important design choices. However, the paper lacks systematic ablation studies isolating their effects. In particular, the relative contributions of the rotational and cosine scoring heads, the necessity of freezing gene embeddings, and the benefit of ontology-derived information content remain unclear. These evaluations are important for understanding which components drive the reported improvements.

  5. The patient retrieval and interpretability claims lack sufficient quantitative support. Although patient retrieval and phenotype-level interpretability are emphasized as important capabilities, the evidence is primarily based on embedding visualization and patient clustering. The paper does not provide quantitative retrieval benchmarks against established phenotype similarity methods. Similarly, attention weights are used to explain phenotype importance without evaluating their faithfulness or clinical relevance. Consequently, these additional capabilities remain insufficiently validated.

Questions for authors

Questions for Authors

  1. How do the authors ensure that phenotype–gene associations involving held-out causal genes are not already encoded during RotatE pretraining? Could the authors evaluate PRISM under a stricter gene-edge-disjoint setting to distinguish genuine generalization from the reuse of existing KG knowledge?

  2. What is the specific advantage of preserving rotational relational geometry over conventional embedding similarity? Could the authors provide controlled comparisons with alternative KG embedding methods and non-rotational scoring mechanisms?

  3. Could the authors extend the evaluation to additional phenotype-driven gene prioritization methods, such as Exomiser, Phen2Gene, and Phrank, using comparable candidate gene sets and evaluation protocols?

  4. Could the authors provide systematic ablations of the RotatE and cosine scoring heads, information-content-aware attention, and frozen gene embeddings to establish the contribution of each component?

  5. Could the authors quantitatively evaluate patient retrieval using metrics such as Recall@K or MRR, and assess whether the attention-based phenotype explanations faithfully reflect the model's predictions?

Reviewer_84EuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces PRISM, a framework for phenotype-driven causal gene prioritisation in rare disease patients. The authors first construct a heterogeneous biomedical knowledge graph by integrating information from HPO and HPOA, together with additional relation types connecting genes, diseases, and phenotypes. PRISM operates in two stages. In the first stage, the curated knowledge graph is embedded using the RotatE knowledge graph embedding model, yielding representations for genes and other entities. In the second stage, patient phenotypes are represented through two complementary modules : an information-content gated aggregation of (i) phenotype embeddings learned from the knowledge graph, and (ii) semantic representations of phenotype descriptions obtained from the BioFormer-8L language model. The former is aligned with gene embeddings learned by RotatE, whereas the latter is aligned with gene representations projected into a PCA-derived latent space. The two similarity scores are subsequently combined to produce the final gene-ranking score used for causal gene prediction. Experimental results show that PRISM consistently outperforms the baseline SHEPHERD model on the reported evaluation benchmarks.

Critical strengths

The main strength of the paper lies in its representation learning framework. The authors construct a rich heterogeneous graph with bidirectional relations that improve information flow during embedding learning. They further align phenotype and gene representations into shared latent spaces, enabling meaningful comparisons. This design allows gene prioritisation to be performed through simple similarity-based retrieval, yielding a computationally efficient solution for both causal gene ranking and patient matching.

Critical weaknesses
  1. The contribution of the different representation spaces remains somewhat unclear. PRISM combines a RotatE-based phenotype-gene alignment branch with a second branch operating on PCA-derived gene representations, yet the individual impact of these components is not fully analysed. The paper would benefit from a more comprehensive ablation study that (i) removes the PCA branch while retaining only the RotatE-based retrieval objective, (ii) removes the RotatE branch while retaining only the PCA-aligned text-based objective, and (iii) evaluates the sensitivity of the model to the PCA dimensionality.
  2. The paper claims the ability to prioritise previously unseen genes. However, it is unclear whether this claim also applies to the RotatE-based branch. RotatE is inherently a transductive knowledge graph embedding method, meaning that entity embeddings are learned only for nodes observed during training. If unseen genes are included in the evaluation setting, how does the RotatE component handle them? If the RotatE branch is excluded from the unseen-gene scenario, how is the S_{rot} score computed for genes that were not present during training?
  3. The study compares the proposed approach only against SHEPHERD. Could the authors justify this choice and explain why other state-of-the-art baseline models were not considered?
  4. Figure 2 is difficult to read and interpret. Moreover, the overall quality and readability of the figures could be improved.
Questions for authors

Please refer to the Weaknesses section for details.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.