Semantic embedding of variant effect annotations enables rapid and accurate pathogenicity prediction with VUS.Life Authors: Jiawei Wu, Marissa Stutzman, Michael Muriello, Joy Lincoln, Donald G. Basel, Xiaowu Gai Published: 2026 ABSTRACT Interpreting the pathogenicity of genetic variants remains a critical bottleneck in genomic medicine. Millions of variants of uncertain significance (VUS) hinder the clinical application of genetic findings. Traditional computational approaches often rely on hand-engineered features and fail to capture the complexity of multidimensional genomic annotations fully. We developed VUS.Life, a multi-modal framework that synergizes semantic text embeddings of biological and clinical annotations with protein language modeling. We transformed variant annotations from Variant Effect Predictor (VEP) into natural language descriptions which are then converted into vector embeddings via established Large Language Models (LLMs), namely all- mpnet-base-v2, MedEmbed-large-v0.1, and text-embedding-004. Pathogenicity of a variant of interest is predicted by its proximity in the vector embedding space with variants of known pathogenicity. We further extended VUS.Life by employing residue-level delta embeddings from the ESMC-600 M model to capture both clinical context and biophysical constraints. We evaluated the framework on > 10,000 variants across eight ACMG Tier 1 disease genes (BRCA1, BRCA2, FBN1, ATM, PALB2, MYH7, USH2A, and PAH), achieving Leave-One-Out Cross-Validation (LOO-CV) MCC of 0.895–0.989 and F1 ≥ 0.94 across all genes evaluated. Additionally, our unsupervised FBN1 structural analysis using ESMC-600M revealed that delta embeddings disentangled distinct pathogenic mechanisms, topologically separating disulfide bond disruptions from calcium- binding defects. These structural clusters correlated strongly with Zero-Shot Log-Likelihood Ratio (LLR) scores, validating evolutionary fitness as a proxy for pathogenicity. An ablation study removing all pre-computed pathogenicity scores demonstrated that MPNet embeddings retain full discriminative power, confirming that the classifier captures biological signal independent of existing scoring tools. This semantic embedding framework, VUS.Life, accurately captures pathogenicity-relevant features from complex variant annotations, enabling robust automated classification across eight ACMG Tier 1 disease genes and three embedding models. The approach generalizes beyond well-curated genes and supports scalable, interpretable, and representation-based classification of VUS. It holds significant promise for alleviating the variant interpretation bottleneck in clinical genomics.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-54844-y. Introduction In just two decades, genomic medicine has transformed from a futuristic vision into clinical reality. What once required years and millions of dollars—sequencing a human genome—can now be accomplished in hours for under a thousand dollars1. This revolution has enabled genetic testing to guide cancer treatment decisions, predict disease risk, and enable precision therapies. Yet paradoxically, as our capacity to identify genetic variants has increased, our ability to interpret their clinical significance has become the rate-limiting step in realizing the potential of genomic medicine. The magnitude of this challenge is staggering. Current databases contain hundreds of millions of human genetic variants, yet only a fraction carry established clinical classifications2–4. For genes like BRCA1 and BRCA2, where accurate interpretation can mean the difference between prophylactic surgery and routine screening, this uncertainty carries profound implications. Numerous computational tools like SIFT5, PolyPhen-26, and CADD7 often referred to as in silico predictors, have been developed to help classify the pathogenicity of genetic variants by leveraging evolutionary conservation, protein structural predictions, and statistical modeling. Table 1 lists some of the most popular and commonly used general-purpose predictors. These tools are widely used as a form of evidence in variant interpretation and are supported by clinical guidelines from the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP)8. Beyond these general-purpose predictors, significant research efforts have been invested in addressing challenges specific to a variant type or disease. For example, VEST-indel, focusing on insertions and deletions (indels), achieved sufficient accuracy to aid in clinical classification9,10. PathoPredictor, on the other hand, is a disease-specific ensemble classifier trained on data from distinct patient cohorts, demonstrating the value of tailored models over general-purpose predictors11.Table 1The summary of conventional methods of pathogenicity prediction.ToolCategoryBrief descriptionSIFT5Foundational(Sequence Conservation)Predicts if an amino acid substitution affects protein function based on sequence homology. It is an early, probabilistic-based methodPolyPhen-26Foundational (Sequence & Structure)Predicts the effect of an amino acid substitution using sequence homology, Pfam annotations, and 3D structures. It is also a probabilistic-based methodCADD7Meta-predictorIntegrates multiple annotations, including conservation and functional information, into a single score for deleteriousnessGERP18Foundational(Sequence Conservation)Identifies evolutionarily constrained regions to score how tolerant a position is to substitution. Its scores are often used as features in other toolsMutationTaster19FoundationalA general-purpose tool for predicting the impact of various variant typesREVEL20Meta-predictor / EnsembleAn ensemble method that combines scores from multiple tools to predict the pathogenicity of rare missense variantsMetaLR & MetaSVM21Meta-predictor / EnsembleEnsemble methods that integrate multiple deleteriousness scores and allele frequency data using logistic regression or a support vector machineClinPred22Meta-predictor / EnsembleA prediction tool that incorporates conservation scores, other prediction scores, and allele frequencies to identify disease-relevant variants Table 1The summary of conventional methods of pathogenicity prediction.ToolCategoryBrief descriptionSIFT5Foundational(Sequence Conservation)Predicts if an amino acid substitution affects protein function based on sequence homology. It is an early, probabilistic-based methodPolyPhen-26Foundational (Sequence & Structure)Predicts the effect of an amino acid substitution using sequence homology, Pfam annotations, and 3D structures. It is also a probabilistic-based methodCADD7Meta-predictorIntegrates multiple annotations, including conservation and functional information, into a single score for deleteriousnessGERP18Foundational(Sequence Conservation)Identifies evolutionarily constrained regions to score how tolerant a position is to substitution. Its scores are often used as features in other toolsMutationTaster19FoundationalA general-purpose tool for predicting the impact of various variant typesREVEL20Meta-predictor / EnsembleAn ensemble method that combines scores from multiple tools to predict the pathogenicity of rare missense variantsMetaLR & MetaSVM21Meta-predictor / EnsembleEnsemble methods that integrate multiple deleteriousness scores and allele frequency data using logistic regression or a support vector machineClinPred22Meta-predictor / EnsembleA prediction tool that incorporates conservation scores, other prediction scores, and allele frequencies to identify disease-relevant variants The summary of conventional methods of pathogenicity prediction. Foundational (Sequence Conservation) Foundational (Sequence Conservation) PathoPredictor is also one of many recent predictors that have leveraged machine learning (ML) and deep learning (DL) to capture complex genomic patterns and improve predictive power. Another prominent example of the group is AlphaMissense, which adapts the state-of-the-art AlphaFold protein structure model to predict the functional impact of missense variants12. GNN-MAP, on the other hand, uses graph neural networks (GNNs) to model the relationships between different variants13. This approach enables GNN-MAP to predict pathogenicity for multiple variant types (missense, stop-gain, frameshift) and has shown strong performance on rare variants. However, these afore-mentioned methods face inherent limitations: they typically focus on individual variant types or molecular features, require extensive domain expertise to engineer appropriate inputs, and struggle to integrate the increasingly complex and heterogeneous landscape of genomic annotations into a unified prediction framework. Critically, most tools predict variant deleteriousness—the degree of functional disruption—rather than clinical pathogenicity, which depends not only on molecular impact but also on inheritance pattern, penetrance, and disease context. These are distinct quantities and conflating them has been a persistent source of misclassification in clinical variant interpretation8. In the last few years, Large Language Models (LLMs) have demonstrated unprecedented ability to capture semantic relationships within textual data, leading to breakthroughs in machine translation, question answering, and document understanding14. They have also been successfully applied in the field of genomic medicine. One application is to apply language models to the clinical narratives associated with variants. For instance, ClinVar-BERT is a model fine-tuned on millions of free-text clinical summaries from the ClinVar database15. By training a biomedical-specific BERT model to discern patterns in the evidence described in these reports, they successfully re-classified VUS in a way that correlated with experimental functional data. This work demonstrated that unstructured clinical text contains rich, learnable signals of pathogenicity that can be captured by language models. Another powerful approach treats biological sequences themselves as a language. Protein language models (pLMs) like ESM-1b are pre-trained on hundreds of millions of protein sequences and learn the fundamental “grammar” of protein structure and function16. Building on this, Lin et al. developed VariPred, which uses embeddings from a pLM to predict the impact of missense variants17. By comparing the embeddings of the wild-type and mutant protein sequences, VariPred achieves state-of-the-art performance using only sequence information as input, effectively bypassing the need for explicit structural or evolutionary features. Similarly, Brandes et al. showed that the ESM1b model could be used directly, without fine-tuning, to predict the effects of all possible missense variants, in-frame indels, and even isoform-specific effects, showcasing the remarkable generalizability of these models16. Despite these advances, existing LLM-based methods address either the clinical text or the protein sequence in isolation: ClinVar-BERT processes free-text evidence summaries but ignores biophysical sequence constraints, while pLM-based tools capture evolutionary fitness but discard clinical context entirely. A unified framework that integrates both dimensions—clinical annotations and molecular biophysics—within a single, scalable representation has not yet been established. New paradigm: semantic embeddings for variant interpretation This study introduces VUS.Life, a framework that treats comprehensive genomic annotations as a structured clinical narrative rather than a set of discrete numerical features. We name the framework VUS.Life. Our framework transforms variant annotations into natural language descriptions, then leverages pre-trained language models to generate semantic embeddings that capture nuanced relationships between different types of genomic evidence. With VUS.Life, we imagine each genetic variant existing within a vast semantic space, where variants with similar functional consequences and clinical implications naturally cluster together. In this space, the pathogenicity of a newly discovered variant can be inferred from its semantic similarity to variants with established clinical significance, revealing how expert clinical geneticists integrate multiple lines of evidence and recognize patterns from experience. Unlike most, if not all, afore-mention-ed tools that focus on predicting the deleteriousness of a variant, which is not equivalent to pathogenicity, VUS.Life is a framework for pathogenicity prediction directly. However, biological and clinical narratives alone cannot fully capture the thermodynamic consequences of genetic variants. Alongside VUS.Life, we present a complementary structural analysis using the ESMC-600M protein language model. By calculating the vector difference between wild-type and mutant residue embeddings— the residue-level delta embedding—we demonstrate that this approach can isolate the specific “structural insult” of each variant, enabling unsupervised discrimination between mechanistically distinct pathogenic subtypes such as disulfide bond disruptions and calcium-binding defects. In this study, we demonstrate that VUS.Life achieves leave-one-out cross-validation (LOO-CV) MCC of 0.895–0.989 and F1 ≥ 0.94 across eight ACMG Tier 1 disease genes, and that its performance is independent of pre-computed pathogenicity scores from ablation test. A complementary residue-level analysis using ESMC-600M independently disentangles distinct pathogenic mechanisms in FBN1 without access to structural labels or functional annotations. Together, these results establish VUS.Life as a scalable, interpretable, and annotation-driven framework for clinical pathogenicity prediction. Methods Data acquisition and preprocessing Variant datasets To develop and validate the VUS.Life framework, we analyzed variants with established pathogenicity classifications across eight ACMG Tier 1 actionable disease genes sourced from BRCA Exchange23 and ClinVar24 (Table 2). For ClinVar entries, we retained only variants with a two-star or higher review status classified as pathogenic/likely pathogenic or benign/likely benign. The held-out test set was derived independently from ClinGen expert panel curations; no variant appears in both sets.Table 2Pathogenicity composition of the training (ClinVar) and held-out test (ClinGen expert panel) datasets per gene. Path = pathogenic/likely pathogenic; Ben = benign/likely benign; %Path = percentage of pathogenic variants.GeneTraining set (ClinVar)Test set (ClinGen)NPathBen (%Path)NPathBen (%Path)BRCA13,3232,2421,081 (67.5%)594910 (83.1%)BRCA24,0812,6451,436 (64.8%)533617 (67.9%)FBN11,532768764 (50.1%)89836 (93.3%)ATM4,4361,7152,721 (38.7%)11101 (90.9%)PALB21,632772860 (47.3%)110 (100%)MYH71,9162171,699 (11.3%)1085454 (50.0%)USH2A812626186 (77.1%)291712 (58.6%)PAH819430389 (52.5%)29928316 (94.6%)Total18,5519,4159,136649533116 Table 2Pathogenicity composition of the training (ClinVar) and held-out test (ClinGen expert panel) datasets per gene. Path = pathogenic/likely pathogenic; Ben = benign/likely benign; %Path = percentage of pathogenic variants.GeneTraining set (ClinVar)Test set (ClinGen)NPathBen (%Path)NPathBen (%Path)BRCA13,3232,2421,081 (67.5%)594910 (83.1%)BRCA24,0812,6451,436 (64.8%)533617 (67.9%)FBN11,532768764 (50.1%)89836 (93.3%)ATM4,4361,7152,721 (38.7%)11101 (90.9%)PALB21,632772860 (47.3%)110 (100%)MYH71,9162171,699 (11.3%)1085454 (50.0%)USH2A812626186 (77.1%)291712 (58.6%)PAH819430389 (52.5%)29928316 (94.6%)Total18,5519,4159,136649533116 Pathogenicity composition of the training (ClinVar) and held-out test (ClinGen expert panel) datasets per gene. Path = pathogenic/likely pathogenic; Ben = benign/likely benign; %Path = percentage of pathogenic variants. Dataset construction and independence The training set consists of ClinVar-derived variants with two-star or higher review status, retaining only those carrying a definitive classification of pathogenic/likely pathogenic or benign/likely benign. The held-out test set was derived entirely from ClinGen expert panel curations and is fully independent of the training data. The two datasets differ not only in provenance but also in character: ClinVar aggregates broad laboratory consensus across well-characterized variants, whereas ClinGen expert panels specifically curate contested or uncertain cases, making the test set structurally harder than the training distribution. The dataset split is performed at the variant level. Within a gene, variants share transcript-level annotations (gene identifier, canonical transcript) as necessary biological context, but each variant’s embedding is uniquely determined by its full annotation record—consequence type, allele frequency, SpliceAI scores, co-located variant identifiers, and genomic coordinates—such that two variants in the same gene with different consequences occupy distinct positions in the embedding space. No variant appears in both the training and test sets. Annotation fetch and processing Variants were annotated using the Variant Effect Predictor (VEP) with an extensive suite of annotation sources and prediction tools25. The VEP configuration was designed to capture a wide range of functional, evolutionary, and predictive information relevant to variant pathogenicity. Key annotation categories included: Core annotations: Canonical transcript identification, predicted consequence (e.g., missense, frameshift), protein-level effects, and splicing predictions (e.g., SpliceAI26). Pathogenicity prediction tools: Scores from established predictors such as CADD7, REVEL20, AlphaMissense12, SIFT5, and PolyPhen-26. Regulatory and functional annotations: Gene Ontology (GO) terms, pathway information, protein domain annotations, tissue-specific expression data, and loss-of-function (LoF) predictions. To prepare the data for language model processing, we systematically converted the structured VEP JSON annotations into a consistent, semi-structured textual format. This was achieved using a custom template that serializes the nested JSON data into a human-readable, key-value report. The template organizes information into logical sections, such as variant identifiers, transcript-level consequences, population frequencies, and scores from various prediction tools (e.g., CADD, SpliceAI, AlphaMissense). This ensures uniform representation of all annotation fields. By converting numerical scores, flags, and identifiers into a unified textual string, we create a comprehensive and uniform input format that enables the embedding model to learn the semantic relationships between disparate types of genomic evidence. An abbreviated example of the generated text for an intronic variant is shown below: Variant ID: NC_000017.11:g.43046253G > A. Variant class: SNV. Most severe consequence: intron_variant. Transcript consequence #1: The transcript id is NM_007294.4; The gene symbol is BRCA1; The gene id is 672; The gene symbol source is EntrezGene; The biotype is protein_coding. SpliceAI scores: The SpliceAI SYMBOL is BRCA1; The SpliceAI DS AG is 0.0; The SpliceAI DS AL is 0.0. Population frequencies: Allele: A. The frequency in gnomadg_asj is 0.02244; The frequency in amr is 0.0043; The frequency in gnomadg_ami is 0.008216; The frequency in gnomadg_afr is 0.002173. Semantic embedding models and generation Three distinct embedding models were selected to capture different aspects of semantic representation:all-mpnet-base-v2 (MPNet)28: A Transformer-based model pre-trained using masked and permuted language modeling. With 768 dimensions and a 512-token limit, it excels at semantic similarity tasks, balancing sentence structure understanding.MedEmbed-large-v0.1 (MedEmbed)29: A domain-specific Transformer trained on biomedical literature, offering 1,024 dimensions and a 512-token limit. Its specialization provides a nuanced understanding of medical and genomic terminology.text-embedding-004 (Google Embeddings)30: Google’s state-of-the-art general-purpose embedding model, generating 768-dimensional embeddings with a 2,048-token limit, performing exceptionally well on diverse semantic tasks. all-mpnet-base-v2 (MPNet)28: A Transformer-based model pre-trained using masked and permuted language modeling. With 768 dimensions and a 512-token limit, it excels at semantic similarity tasks, balancing sentence structure understanding. all-mpnet-base-v2 (MPNet)28: A Transformer-based model pre-trained using masked and permuted language modeling. With 768 dimensions and a 512-token limit, it excels at semantic similarity tasks, balancing sentence structure understanding. MedEmbed-large-v0.1 (MedEmbed)29: A domain-specific Transformer trained on biomedical literature, offering 1,024 dimensions and a 512-token limit. Its specialization provides a nuanced understanding of medical and genomic terminology. MedEmbed-large-v0.1 (MedEmbed)29: A domain-specific Transformer trained on biomedical literature, offering 1,024 dimensions and a 512-token limit. Its specialization provides a nuanced understanding of medical and genomic terminology. text-embedding-004 (Google Embeddings)30: Google’s state-of-the-art general-purpose embedding model, generating 768-dimensional embeddings with a 2,048-token limit, performing exceptionally well on diverse semantic tasks. text-embedding-004 (Google Embeddings)30: Google’s state-of-the-art general-purpose embedding model, generating 768-dimensional embeddings with a 2,048-token limit, performing exceptionally well on diverse semantic tasks. The embedding generation process transforms each variant’s description into vector representations using these models. These embeddings are stored in ChromaDB, a vector database optimized for similarity search and retrieval. Accompanying meta- data includes variant identifiers, genomic coordinates, pathogenicity classifications, confidence scores, gene annotations, functional consequences, and timestamps of processing. Vector indices are tuned for cosine similarity search, with parameters adjusted to each model’s dimensionality. Evaluation metrics and implementation Dimensionality reduction and visualization To analyze the high-dimensional embeddings, we applied three complementary dimensionality reduction techniques: Principal Component Analysis (PCA)31, t-Distributed Stochastic Neighbor Embedding (t-SNE)32, and Uniform Manifold Approximation and Projection (UMAP)33. PCA projects data into a visualizable space by preserving major variance components. t-SNE emphasizes local neighbor- hood relationships, while UMAP balances local and global structure preservation. Visualizations revealed distinct clusters of variants with similar pathogenicity, suggesting that the embedding models effectively capture pathogenicity-related features. The discriminative power was quantified by correlating embedding distances with pathogenicity differences, confirming tighter clustering of variants with similar clinical impacts. This structure supports the k-NN classification approach for predicting pathogenicity in variants of uncertain significance. k-NN pathogenicity prediction A k-NN classification approach was developed to predict variant pathogenicity using semantic embeddings. Performance was evaluated across multiple neighborhood sizes (k = 1, 5, 10, 15, 20) to identify gene-specific optimal k values. Cosine similarity was employed as the distance metric due to its effectiveness in high-dimensional spaces and its robustness to vector magnitude variations. For each query variant, the algorithm identifies the k most similar variants with known pathogenicity, excluding the query variant from the training set to prevent data leakage. Predictions are based on majority voting, with confidence scores reflecting the proportion of votes supporting the predicted class. The embedding space’s ability to separate variants by pathogenicity was assessed by computing a clustering accuracy metric for known variants. For each variant with established pathogenicity, its k nearest neighbors were identified, and the percentage of neighbors sharing the same classification was calculated. The classification criteria are outlined in Table 3.Table 3Criteria for classifying pathogenicity based on the proportion of pathogenic and benign neighbors among the top-k nearest neighbors.ClassificationCriteria (among k neighbors)Confidence ScorePotentially Pathogenic ≥ 90% pathogenic neighborsPathogenic ratioLikely Pathogenic ≥ 60% pathogenic neighbors (but < 90%)Pathogenic ratioPotentially Benign ≥ 90% benign neighborsBenign ratioLikely Benign ≥ 60% benign neighbors (but < 90%)Benign ratioUncertainMixed or no clear majoritymax(benign, pathogenic) ratio Table 3Criteria for classifying pathogenicity based on the proportion of pathogenic and benign neighbors among the top-k nearest neighbors.ClassificationCriteria (among k neighbors)Confidence ScorePotentially Pathogenic ≥ 90% pathogenic neighborsPathogenic ratioLikely Pathogenic ≥ 60% pathogenic neighbors (but < 90%)Pathogenic ratioPotentially Benign ≥ 90% benign neighborsBenign ratioLikely Benign ≥ 60% benign neighbors (but < 90%)Benign ratioUncertainMixed or no clear majoritymax(benign, pathogenic) ratio Criteria for classifying pathogenicity based on the proportion of pathogenic and benign neighbors among the top-k nearest neighbors. Distance ratio analysis To quantitatively validate the clustering structure, we introduced the Distance Ratio metric. For any given variant, we calculate the median cosine distance to its k nearest neighbors of the same class (Dsame) versus the opposite class (Dopposite).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Distance Ratio }} = D_{same} /D_{opposite}$$\end{document} A ratio < 1.0 indicates that the variant resides within a semantically consistent neighborhood (correct clustering), while a ratio > 1.0 indicates separation failure. Implementation workflow The pathogenicity prediction pipeline targets novel and variants of uncertain significance (VUS) using a modular, efficient workflow. Unknown variants are retrieved from ChromaDB using pathogenicity status filters, processed in batches for computational efficiency, and analyzed for their k (default: 20) nearest known variants (benign or pathogenic) via cosine similarity. Predictions are generated using a confidence- weighted voting system, with confidence scores reflecting prediction reliability. Visualizations employing PCA, UMAP, and t-SNE illustrate spatial relationships between known and unknown variants, aiding genetic counselors and clinicians in interpreting results (see the workflow in Fig. 1).Fig. 1End-to-End architecture of VUS.Life. The pipeline separates offline indexing from online inference. In Phase 1, genomic data is transformed from structured JSON (VEP) into natural language, vectorized by Large Language Models, and stored in a vector database. In Phase 2, the system performs a semantic search for an unknown variant against this index. The diagram highlights the critical “Masking” step (C) ensures that the model learns from biological context rather than explicit labels, and the k-NN consensus mechanism (H) that provides the final confidence score. End-to-End architecture of VUS.Life. The pipeline separates offline indexing from online inference. In Phase 1, genomic data is transformed from structured JSON (VEP) into natural language, vectorized by Large Language Models, and stored in a vector database. In Phase 2, the system performs a semantic search for an unknown variant against this index. The diagram highlights the critical “Masking” step (C) ensures that the model learns from biological context rather than explicit labels, and the k-NN consensus mechanism (H) that provides the final confidence score. Ablation study design To quantify the independent contribution of semantic embeddings relative to pre- computed pathogenicity scores, we constructed an ablation variant of the VEP annotation text prior to embedding. As summarized in Table 4, Ablation 1 removes all pre-computed predictor scores (CADD, REVEL, AlphaMissense, ClinPred, and EVE) from the annotation text. All other variant features genomic coordinates, consequence type, SpliceAI scores, population allele frequencies, and co-located variant information are retained.Table 4Summary of ablation conditions. Scores removed from VEP annotation text prior to embedding; all other variant features are retained.ConditionScores RemovedBaselineNone (full annotation)Ablation 1CADD, REVEL, AlphaMissense, ClinPred, EVE Table 4Summary of ablation conditions. Scores removed from VEP annotation text prior to embedding; all other variant features are retained.ConditionScores RemovedBaselineNone (full annotation)Ablation 1CADD, REVEL, AlphaMissense, ClinPred, EVE Summary of ablation conditions. Scores removed from VEP annotation text prior to embedding; all other variant features are retained. Each ablation variant was independently re-embedded using the same three embedding models and evaluated under the identical leave-one-out cross-validation protocol at k = 5. Performance change is reported as ∆MCC = MCCablation − MCCbaseline; negative values indicate that removing those features degraded performance relative to the full-annotation baseline. Protein language model analysis To capture the thermodynamic and evolutionary constraints of missense variants beyond what text annotations alone convey, we integrated the ESMC-600M (600 million parameters) protein language model34. We focused this analysis on FBN1, implementing a specific pipeline to handle its large protein size (2,871 amino acids) and mechanistically heterogeneous pathogenic variants. Compared with full-sequence embeddings, residue-level delta embeddings—which isolate the specific vector difference at the variant site—offered the most robust discrimination between functional categories of variants. Residue-level delta embeddings We hypothesized that the pathogenicity signal is best captured by the vector difference between the variant and reference states, rather than the absolute representation of the variant alone. We defined the Residue-Level Delta Embedding (∆E) at the variant site target_idx as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta E = E_{{{\mathrm{mutant}}}} \left[ {target\_idx} \right] - E_{{{\mathrm{wildtype}}}} \left[ {target\_idx} \right]$$\end{document} This subtraction isolates the “structural insult” induced by the variant, removing the background noise of the protein’s sequence identity. Due to the sequence length of Fibrillin-1 exceeding standard context windows, we utilized a sliding window tiling strategy (overlapping windows of ≈1000 aa) to generate these embeddings while preserving local structural context. Mechanistic categorization via structural interaction type To investigate whether the embedding space captures specific biophysical mechanisms, variants were stratified based on the structural role of the wild-type residue within the protein fold, using annotations from the UMD-FBN1 mutations database (http://www.umd.be/FBN1/): Disulfide bond disrupting: Variants affecting cysteine residues that form covalent disulfide bridges, which are critical for stabilizing the tertiary structure of FBN1 domains. Calcium binding: Variants located within high-affinity Ca2+ binding motifs (e.g., within cbEGF domains), which facilitate domain rigidity and proper inter- domain alignment. Evolutionary fitness scoring (Zero-Shot LLR) To quantify evolutionary constraint without supervision, we calculated the Log-Likelihood Ratio (LLR) using the model’s sequence_head. The LLR compares the probability of the mutant amino acid (xmut) against the wild-type (xwt) given the surrounding sequence context:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LLR = {\mathrm{log}}P\left( {x_{mut} |{\mathrm{context}}} \right) - {\mathrm{log}}P(x_{wt} |{\mathrm{context}})$$\end{document} Negative LLR values indicate evolutionary intolerance (potential pathogenicity), while values near zero imply functional neutrality. Results Embedding results The embedding analysis establishes a robust framework for clinical variant interpretation via confidence-based stratification derived from neighborhood consistency patterns. Variants can be categorized as high-confidence (> 95% agreement among nearest neighbors), moderate-confidence (60–95% agreement), or boundary cases requiring expert review (< 60% agreement). BRCA1 analysis revealed that 2–5% of variants fell into boundary cases across models, highlighting cases that may benefit from additional clinical review or functional studies. BRCA2 analysis showed fewer boundary cases (0.4–2.6%), indicating stronger semantic clustering and higher prediction confidence. The gradient of neighborhood agreement serves as a natural quality control mechanism for clinical workflows, enabling automated processing of high-confidence predictions while flagging uncertain cases for genetic counseling and expert interpretation. This approach addresses the critical need for reliable variant classification in hereditary cancer predisposition testing, where accurate interpretation directly influences clinical management decisions. Consistent performance across neighbor counts (k = 1, 5, 20) validates the methodology’s robustness, with optimal k likely being gene-specific to balance local precision and noise reduction for immediate clinical implementation. BRCA1 semantic embedding performance The k-nearest neighbors (K-NN) evaluation of BRCA1 variants (n = 3,311) demonstrated robust classification performance across all three embedding models, with systematic analysis confirming the effectiveness of semantic representations in capturing pathogenicity-relevant features (Fig. 2). The MPNet model achieved the highest overall accuracy at 97.9% with top 20 neighbors, with strong performance for both benign/likely benign variants (96.9%) and pathogenic/likely pathogenic variants (98.4%). MedEmbed showed comparable performance at 97.7% overall accuracy, while Google embeddings achieved 97.3% overall accuracy. Neighborhood consistency analysis revealed that 94.6% of benign variants and 97.4% of pathogenic variants had 80–100% agreement among their top 20 nearest neighbors when using MPNet, indicating robust semantic clustering. The high neighborhood consensus across all models suggests that BRCA1 variant annotations contain sufficient semantic information to enable reliable automated classification. MedEmbed exhibited 92.5% consistency for benign variants achieving 80–100% neighborhood agreement, while Google embeddings achieved 92.6% consistency for benign variants. This consistent performance across different embedding approaches validates the robustness of the semantic representation framework for BRCA1 variant interpretation.Fig. 2Dimensionality reduction of semantic embeddings for 3,311 BRCA1 variants. Each panel displays a 2D projection of the high-dimensional variant embeddings using three different models (rows: all-mpnet-base-v2, google-embedding, MedEmbed-large-v0.1) and three dimensionality reduction techniques (columns: PCA, t-SNE, UMAP). The distinct separation between benign (blue) and pathogenic (red) variants across all models demonstrates that the embeddings effectively capture pathogenicity-relevant features. Non-linear methods like t-SNE and UMAP reveal more defined clusters compared to the linear PCA projection, confirming the semantic space is structured by clinical significance. Dimensionality reduction of semantic embeddings for 3,311 BRCA1 variants. Each panel displays a 2D projection of the high-dimensional variant embeddings using three different models (rows: all-mpnet-base-v2, google-embedding, MedEmbed-large-v0.1) and three dimensionality reduction techniques (columns: PCA, t-SNE, UMAP). The distinct separation between benign (blue) and pathogenic (red) variants across all models demonstrates that the embeddings effectively capture pathogenicity-relevant features. Non-linear methods like t-SNE and UMAP reveal more defined clusters compared to the linear PCA projection, confirming the semantic space is structured by clinical significance. BRCA2 semantic embedding performance Our method demonstrated exceptional classification performance on BRCA2 variants (n = 4,074) across all embedding models, with MedEmbed achieving outstanding results of 99.1% overall accuracy (99.2% benign/likely benign, 99.1% pathogenic/likely pathogenic) (Fig. 3).Fig. 3Visualization of semantic embeddings for 4,074 BRCA2 variants. Following the same structure as Fig. 2, these plots show the embedding space for BRCA2 variants. The separation between benign (blue) and pathogenic (red) classes is even more pronounced than for BRCA1, with t-SNE and UMAP revealing highly compact and well-separated clusters, suggesting a very robust semantic representation. Visualization of semantic embeddings for 4,074 BRCA2 variants. Following the same structure as Fig. 2, these plots show the embedding space for BRCA2 variants. The separation between benign (blue) and pathogenic (red) classes is even more pronounced than for BRCA1, with t-SNE and UMAP revealing highly compact and well-separated clusters, suggesting a very robust semantic representation. The domain-specific medical embedding model’s superior performance highlights the value of specialized pre-training for genetic variant analysis. MPNet achieved comparable excellence with 99.0% overall accuracy (98.3% benign/likely benign, 99.4% pathogenic/likely pathogenic), while Google embeddings maintained strong performance at 97.9% overall accuracy. Neighborhood distribution analysis revealed remarkable consistency, with MedEmbed showing 99.6% of benign variants and 98.8% of pathogenic variants achieving 80–100% agreement among top 20 neighbors. MPNet demonstrated equally impressive consistency with 97.4% of benign variants and 99.3% of pathogenic variants showing high neighborhood agreement. The larger sample size for BRCA2 variants appears to contribute to enhanced model training and more stable semantic representations, enabling the development of highly reliable prediction models for clinical applications. Google embeddings, while achieving the lowest performance among the three models, still maintained clinically relevant accuracy with 94.9% of benign variants showing strong neighborhood consensus. FBN1 semantic embedding performance The FBN1 validation results demonstrate robust performance across all three embedding models, achieving consistently high accuracy rates exceeding 96% (MPNet: 96.7%, Google: 96.5%, MedEmbed: 96.6% at k = 5) (Fig. 4). These results are particularly significant as they validate the generalizability of our semantic embedding approach beyond the extensively characterized BRCA cancer predisposition genes to a gene associated with connective tissue disorders and mostly missense instead of loss-of-function pathogenic variants. The dimensional reduction visualizations reveal distinct clustering patterns for FBN1 variants, with Google embeddings showing the most compact and well-separated clusters in t-SNE and UMAP projections, while MPNet and MedEmbed demonstrated more distributed but still discriminative patterns. Notably, the FBN1 performance closely mirrors that observed for BRCA1 (97.3% with MPNet), suggesting that the embedding approach maintains effectiveness across different genetic contexts and disease mechanisms, even for genes that may have less comprehensive functional annotation compared to the intensively studied BRCA1/2 genes. This validation supports the broader clinical utility of semantic embeddings for variant interpretation across diverse genetic conditions, particularly important for rare disease contexts where sparse training data may limit traditional computational approaches.Fig. 4Validation of the semantic embedding framework’s generalizability using 1,532 FBN1 variants. This figure demonstrates that the methodology is not limited to BRCA genes. Despite the different genetic context and disease mechanism (Marfan syndrome), the embeddings for FBN1 variants maintain a strong separation between benign (blue) and pathogenic (red) classes. This result validates the robustness and broad utility of using semantic embeddings for variant interpretation across diverse, clinically important genes. Validation of the semantic embedding framework’s generalizability using 1,532 FBN1 variants. This figure demonstrates that the methodology is not limited to BRCA genes. Despite the different genetic context and disease mechanism (Marfan syndrome), the embeddings for FBN1 variants maintain a strong separation between benign (blue) and pathogenic (red) classes. This result validates the robustness and broad utility of using semantic embeddings for variant interpretation across diverse, clinically important genes. Generalization to ATM and PALB2 To further stress-test the framework’s generalizability, we evaluated ATM (n = 4,436) and PALB2 (n = 1,632). Despite differing biological functions and variant spectrums, VUS.Life maintained high predictive accuracy. As shown in Table 5, the ATM gene achieved 98.0–98.4% accuracy across all models at k = 5, while PALB2 achieved 97.2–98.4%. The consistency of these results—comparable to the highly curated BRCA datasets—confirms that the semantic embedding approach captures universal pathogenicity signals and is not an artifact of specific gene curation histories. To provide rigorous, class-imbalance robust evaluation and to assess generalization to held-out expert-panel data, Matthews Correlation Coefficient (MCC) is adopted as the primary metric for all analyses that follow.Table 5Clustering accuracy of semantic embedding models for predicting variant pathogenicity across genes. Accuracy is measured by the percentage of k-nearest neighbors sharing the same known pathogenicity classification. Higher accuracy indicates better separation of pathogenic and benign variants.GeneModelk = 5k = 10k = 15k = 20BRCA1MedEmbed-large-v0.10.9810.9800.9780.977google-embedding0.9800.9770.9740.973all-mpnet-base-v20.9840.9830.9810.979BRCA2MedEmbed-large-v0.10.9930.9930.9920.991google-embedding0.9850.9830.9810.979all-mpnet-base-v20.9920.9910.9900.990FBN1MedEmbed-large-v0.10.9660.9650.9640.963google-embedding0.9650.9640.9610.959all-mpnet-base-v20.9670.9620.9610.959ATMMedEmbed-large-v0.10.9840.9820.9790.975google-embedding0.9800.9780.9750.975all-mpnet-base-v20.9830.9810.9810.978PALB2MedEmbed-large-v0.10.9800.9790.9790.979google-embedding0.9720.9600.9520.946all-mpnet-base-v20.9840.9840.9830.981 Table 5Clustering accuracy of semantic embedding models for predicting variant pathogenicity across genes. Accuracy is measured by the percentage of k-nearest neighbors sharing the same known pathogenicity classification. Higher accuracy indicates better separation of pathogenic and benign variants.GeneModelk = 5k = 10k = 15k = 20BRCA1MedEmbed-large-v0.10.9810.9800.9780.977google-embedding0.9800.9770.9740.973all-mpnet-base-v20.9840.9830.9810.979BRCA2MedEmbed-large-v0.10.9930.9930.9920.991google-embedding0.9850.9830.9810.979all-mpnet-base-v20.9920.9910.9900.990FBN1MedEmbed-large-v0.10.9660.9650.9640.963google-embedding0.9650.9640.9610.959all-mpnet-base-v20.9670.9620.9610.959ATMMedEmbed-large-v0.10.9840.9820.9790.975google-embedding0.9800.9780.9750.975all-mpnet-base-v20.9830.9810.9810.978PALB2MedEmbed-large-v0.10.9800.9790.9790.979google-embedding0.9720.9600.9520.946all-mpnet-base-v20.9840.9840.9830.981 Clustering accuracy of semantic embedding models for predicting variant pathogenicity across genes. Accuracy is measured by the percentage of k-nearest neighbors sharing the same known pathogenicity classification. Higher accuracy indicates better separation of pathogenic and benign variants. Comprehensive multi-gene evaluation To evaluate the generalizability of VUS.Life beyond the initial four-gene pilot, the framework was extended to eight ACMG Tier 1 actionable disease genes: BRCA1, BRCA2, FBN1, ATM, PALB2, MYH7, USH2A, and PAH. These genes were selected because each has a ClinVar training set of at least 800 two-star variants and an independent ClinGen expert panel test set. LOO-CV on the training set yielded MCC of 0.895–0.989 and F1 ≥ 0.94 across all eight genes and all three embedding models (Table 6). Accuracy exceeded 96% in all cases. These results confirm that the k-NN classifier reliably discriminates pathogenic from benign variants within the ClinVar training distribution across genes with widely varying pathogenicity compositions (11–77% pathogenic in training).Table 6LOO-CV and held-out ClinGen test-set classification performance at k = 5 across eight ACMG Tier 1 disease genes and three embedding models.GeneModelLOO MCCLOO F1Test NTest MCC∆MCCFBN1MPNet0.9730.986890.653− 0.320MedEmbed0.9670.9840.285− 0.682Gemini0.9590.9790.188− 0.770BRCA1MPNet0.9810.994590.565− 0.416MedEmbed0.9770.9920.444− 0.533Gemini0.9730.9910.656− 0.317BRCA2MPNet0.9840.995530.827− 0.158MedEmbed0.9880.9960.406− 0.581Gemini0.9890.9960.463− 0.527USH2AMPNet0.9300.984290.858− 0.072MedEmbed0.8950.9760.858− 0.037Gemini0.8910.9750.671− 0.219MYH7MPNet0.9340.9411080.770− 0.164MedEmbed0.9580.9630.792− 0.167Gemini0.9450.9510.776− 0.170PAHMPNet0.9340.9682990.516− 0.418MedEmbed0.9580.9800.538− 0.421Gemini0.9560.9790.607− 0.350MCC is the primary metric. ∆MCC = Test MCC − LOO MCC. †ATM and PALB2 test sets are statistically degenerate (ATM: 10 pathogenic/1 benign; PALB2: 1 variant) and excluded from generalization conclusions. Table 6LOO-CV and held-out ClinGen test-set classification performance at k = 5 across eight ACMG Tier 1 disease genes and three embedding models.GeneModelLOO MCCLOO F1Test NTest MCC∆MCCFBN1MPNet0.9730.986890.653− 0.320MedEmbed0.9670.9840.285− 0.682Gemini0.9590.9790.188− 0.770BRCA1MPNet0.9810.994590.565− 0.416MedEmbed0.9770.9920.444− 0.533Gemini0.9730.9910.656− 0.317BRCA2MPNet0.9840.995530.827− 0.158MedEmbed0.9880.9960.406− 0.581Gemini0.9890.9960.463− 0.527USH2AMPNet0.9300.984290.858− 0.072MedEmbed0.8950.9760.858− 0.037Gemini0.8910.9750.671− 0.219MYH7MPNet0.9340.9411080.770− 0.164MedEmbed0.9580.9630.792− 0.167Gemini0.9450.9510.776− 0.170PAHMPNet0.9340.9682990.516− 0.418MedEmbed0.9580.9800.538− 0.421Gemini0.9560.9790.607− 0.350MCC is the primary metric. ∆MCC = Test MCC − LOO MCC. †ATM and PALB2 test sets are statistically degenerate (ATM: 10 pathogenic/1 benign; PALB2: 1 variant) and excluded from generalization conclusions. LOO-CV and held-out ClinGen test-set classification performance at k = 5 across eight ACMG Tier 1 disease genes and three embedding models. MCC is the primary metric. ∆MCC = Test MCC − LOO MCC. †ATM and PALB2 test sets are statistically degenerate (ATM: 10 pathogenic/1 benign; PALB2: 1 variant) and excluded from generalization conclusions. Performance on the held-out ClinGen test set was evaluated for six genes with interpretable test data; ATM (10 pathogenic, 1 benign) and PALB2 (1 variant total) were excluded from generalization conclusions as their test sets are statistically degenerate. Three generalization tiers emerged. First, USH2A and MYH7 showed the smallest LOO-to-test gap (∆MCC = − 0.04 to − 0.22): all three models achieved test MCC ≥ 0.67 for USH2A and ≥ 0.77 for MYH7, benefiting from balanced test pathogenicity distributions (MYH7: 50%/50%; USH2A: 59%/41%). Second, BRCA2 with MPNet achieved the best single test-set result (MCC = 0.827, ∆MCC = − 0.158); Gemini performed best for BRCA1 (test MCC = 0.656), while MedEmbed and Gemini failed to generalize on BRCA2 (test MCC 0.406 and 0.463), indicating model-specific sensitivity to the distribution shift for test variants of that gene. Third, PAH and FBN1 exhibited test MCC of 0.516–0.607 and 0.188–0.653, respectively, despite high F1 scores (0.93–0.97). This divergence between F1 and MCC is diagnostic of class imbalance: the PAH test set is 94.6% pathogenic (283/299 variants) and the FBN1 test set is 93.3% pathogenic (83/89 variants). A classifier predicting pathogenic for most variants achieves high F1 while failing to discriminate the few benign examples; MCC penalizes this correctly. The FBN1 test set has been expanded from the 30-variant pilot to 89 ClinGen expert-panel variants for this analysis. Ablation study: predictor scores are not the primary driver A central concern in embedding-based classifiers trained on annotated variant databases is feature circularity: if the annotation text contains pre-computed pathogenicity scores (CADD, REVEL, AlphaMissense, ClinPred, EVE) derived from the same data used for validation, observed accuracy may simply reflect ensemble re-ranking of existing tools rather than independent biological inference. To test this possibility, we conducted a systematic ablation study by removing predictor scores from the VEP annotation text before embedding and re-evaluating LOO-CV performance (Table 7, Figs. 5, 6).Table 7Gene-level LOO-CV MCC for the full-annotation baseline and Ablation 1 (all predictor scores removed) at k = 5.GeneModelBaseline MCCAblation 1 MCC∆MCCATMMPNet0.9770.972− 0.005MedEmbed0.9770.966− 0.011Gemini0.9670.883− 0.084PALB2MPNet0.9630.964+ 0.001MedEmbed0.9580.960+ 0.001Gemini0.9680.835− 0.133FBN1MPNet0.9730.949− 0.023MedEmbed0.9670.940− 0.027Gemini0.9590.877− 0.081BRCA1MPNet0.9810.955− 0.026MedEmbed0.9770.955− 0.022Gemini0.9730.921− 0.051BRCA2MPNet0.9840.985+ 0.001MedEmbed0.9880.978− 0.010Gemini0.9890.940− 0.050USH2AMPNet0.9300.874− 0.056MedEmbed0.8950.854− 0.041Gemini0.8910.817− 0.074MYH7MPNet0.9340.950+ 0.016MedEmbed0.9580.940− 0.019Gemini0.9450.800− 0.145PAHMPNet0.9340.954+ 0.019MedEmbed0.9580.901− 0.057Gemini0.9560.799− 0.157∆MCC = Ablation 1 MCC—Baseline MCC; bold values indicate |∆MCC|> 0.10.Fig. 5Dumbbell plot of gene-level LOO MCC for the full-annotation baseline (filled circles) and Ablation 1—all predictor scores removed (open circles) at k = 5. Color encodes the embedding model. A rightward open circle (MPNet on MYH7 and PAH) indicates that score removal improved performance, confirming that predictor score fields introduce distributional noise into MPNet’s embed- ding space.Fig. 6Grouped bar charts of LOO MCC per gene for baseline (solid bars) and Ablation 1 (hatched bars) at k = 5. Each panel corresponds to one embedding model. MPNet (left) shows bars of near- equal height or reversed ordering for MYH7 and PAH, confirming that predictor scores are not beneficial for this model. Gemini (right) shows the largest drops, particularly for MYH7 and PAH. Table 7Gene-level LOO-CV MCC for the full-annotation baseline and Ablation 1 (all predictor scores removed) at k = 5.GeneModelBaseline MCCAblation 1 MCC∆MCCATMMPNet0.9770.972− 0.005MedEmbed0.9770.966− 0.011Gemini0.9670.883− 0.084PALB2MPNet0.9630.964+ 0.001MedEmbed0.9580.960+ 0.001Gemini0.9680.835− 0.133FBN1MPNet0.9730.949− 0.023MedEmbed0.9670.940− 0.027Gemini0.9590.877− 0.081BRCA1MPNet0.9810.955− 0.026MedEmbed0.9770.955− 0.022Gemini0.9730.921− 0.051BRCA2MPNet0.9840.985+ 0.001MedEmbed0.9880.978− 0.010Gemini0.9890.940− 0.050USH2AMPNet0.9300.874− 0.056MedEmbed0.8950.854− 0.041Gemini0.8910.817− 0.074MYH7MPNet0.9340.950+ 0.016MedEmbed0.9580.940− 0.019Gemini0.9450.800− 0.145PAHMPNet0.9340.954+ 0.019MedEmbed0.9580.901− 0.057Gemini0.9560.799− 0.157∆MCC = Ablation 1 MCC—Baseline MCC; bold values indicate |∆MCC|> 0.10. Gene-level LOO-CV MCC for the full-annotation baseline and Ablation 1 (all predictor scores removed) at k = 5. ∆MCC = Ablation 1 MCC—Baseline MCC; bold values indicate |∆MCC|> 0.10. Dumbbell plot of gene-level LOO MCC for the full-annotation baseline (filled circles) and Ablation 1—all predictor scores removed (open circles) at k = 5. Color encodes the embedding model. A rightward open circle (MPNet on MYH7 and PAH) indicates that score removal improved performance, confirming that predictor score fields introduce distributional noise into MPNet’s embed- ding space. Grouped bar charts of LOO MCC per gene for baseline (solid bars) and Ablation 1 (hatched bars) at k = 5. Each panel corresponds to one embedding model. MPNet (left) shows bars of near- equal height or reversed ordering for MYH7 and PAH, confirming that predictor scores are not beneficial for this model. Gemini (right) shows the largest drops, particularly for MYH7 and PAH. Ablation 1 (all predictor scores removed) reveals that the three embedding models differ substantially in their score dependence. MPNet (all-mpnet-base-v2) is fully score-independent: removing all predictor scores actually improves LOO MCC on MYH7 (∆MCC = + 0.016), PAH (+ 0.019), and BRCA2 (+ 0.001). This counter-intuitive finding indicates that numerical predictor fields introduce distributional noise into MPNet’s embedding space—the model’s native sequence-semantic representation is weakened when mixed with heterogeneous numerical fields it was not pre-trained to interpret. MedEmbed (MedEmbed-large-v0.1) shows small, consistent drops across most genes (Ablation 1 ∆MCC typically − 0.01 to − 0.06), with near-zero change for PALB2; its medical-domain pre-training allows marginal benefit from clinical score terminology, but the dependence is mild. Gemini (text-embedding-004) shows the strongest score dependence: Ablation 1 ∆MCC = − 0.145 on MYH7, − 0.157 on PAH, and − 0.133 on PALB2. Critically, even for Gemini, Ablation 1 LOO MCC remains above 0.80 for all eight genes, confirming that no model relies exclusively on pre-computed scores. These findings collectively demonstrate that VUS.Life captures genuine biological signal—consequence type, population allele frequencies, and genomic context—that is independently sufficient for high variant discrimination. MPNet and MedEmbed are therefore fully deployable in contexts where pre-computed score annotation is incomplete. Pathogenicity prediction using annotation embeddings The prediction results for 3,000 not-yet-reviewed variants in both BRCA1 and 3,000 not-yet-reviewed BRCA2 variants demonstrate the promising potential of semantic embeddings for variant pathogenicity classification (Figs. 7, 8). Across all three embedding models (MPNet, Google embeddings, and MedEmbed), the unknown variants (green diamonds) exhibit spatial distributions that closely mirror the established clustering patterns of known benign (blue) and pathogenic (red) variants.Fig. 7Visualization of pathogenicity of 3,000 not-yet-reviewed variants (highlighted as green diamonds) on BRCA1. The pathogenicity is simply mapped into binary labels: benign and pathogenic. Each panel displays dimensionality reduction (PCA, UMAP, and t-SNE) of variant embeddings. The consistent spatial distribution of unknown variants relative to known benign and pathogenic variants across models and genes suggests that embeddings effectively capture pathogenicity-related semantics.Fig. 8Visualization of pathogenicity of 3,000 not-yet-reviewed variants on BRCA2. The same pat- tern is seen as BRCA1 prediction. Visualization of pathogenicity of 3,000 not-yet-reviewed variants (highlighted as green diamonds) on BRCA1. The pathogenicity is simply mapped into binary labels: benign and pathogenic. Each panel displays dimensionality reduction (PCA, UMAP, and t-SNE) of variant embeddings. The consistent spatial distribution of unknown variants relative to known benign and pathogenic variants across models and genes suggests that embeddings effectively capture pathogenicity-related semantics. Visualization of pathogenicity of 3,000 not-yet-reviewed variants on BRCA2. The same pat- tern is seen as BRCA1 prediction. To illustrate the practical application of this approach, we examined prediction results from 10 randomly selected variants each from BRCA1 and BRCA2 datasets (Tables 8, 9). The results demonstrate strong model concordance and biologically coherent predictions across different variant types. For BRCA1 variants, all three embedding models achieved perfect agreement on clearly deleterious variants such as frameshift (NC_000013.11:g.32396971_32396972dup) and stop-gained variants (NC_000013.11:g.32339409C > G), both receiving unanimous pathogenic predictions with 1.00 confidence scores based on k-nearest neighbor counts of B:0, P:20. Intronic variants consistently received benign classifications across all models with maximum confidence, while missense variants showed more nuanced predictions with generally high confidence but occasional inter-model variation in neighbors.Table 8Predicted pathogenicity of randomly selected unknown variants in BRCA1 (NC_000017.11) using different semantic embedding models.Genomic HGVSConsequenceMedEmbedgoogleMPNetCountConfPredCountConfPredCountConfPredg.43094629_43095075delsplice_acceptorB:16, P:40.800BenignB:2, P:180.900PathogenicB:10, P:100.500Uncertaing.43102853C > Gintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43067120C > Tintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43093771_43093774delinsTCframeshiftB:0, P:201.000PathogenicB:0, P:201.000PathogenicB:0, P:201.000Pathogenicg.43061638A > Gintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43065397G > Aintron_variantB:20, P:01.000BenignB:19, P:10.950BenignB:20, P:01.000Benigng.43067430delintron_variantB:20, P:01.000BenignB:8, P:120.600PathogenicB:20, P:01.000Benigng.43074502G > TMissenseB:17, P:30.850BenignB:13, P:70.650BenignB:16, P:40.800Benigng.43094690T > CMissenseB:20, P:01.000BenignB:18, P:20.900BenignB:20, P:01.000Benigng.43115774T > GMissenseB:12, P:80.600BenignB:5, P:150.750PathogenicB:16, P:40.800BenignTable 9Predicted pathogenicity of randomly selected unknown variants in BRCA2 (NC_000013.11) using different semantic embedding models.Genomic HGVSConsequenceMedEmbedgoogleMPNetCountConf.PredCountConf.PredCountConf.Predg.32,353,730 A > Gintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32340247G > TmissenseB:18, P:20.900benignB:20, P:01.000benignB:20, P:01.000benigng.32,363,243 A > GmissenseB:19, P:10.950benignB:15, P:50.750benignB:20, P:01.000benigng.32396971_32396972dupframeshiftB:0, P:201.000pathogenicB:0, P:201.000pathogenicB:0, P:201.000pathogenicg.32,370,904 A > Tintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32,322,227 A > Gintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32385434_32385445dupintron_variantB:20, P:01.000benignB:15, P:50.750benignB:20, P:01.000benigng.32,369,761 A > Tintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32380102G > AsynonymousB:20, P:01.000benignB:18, P:20.900benignB:20, P:01.000benigng.32,339,409 C > Gstop_gainedB:0, P:201.000pathogenicB:0, P:201.000pathogenicB:0, P:201.000pathogenic Table 8Predicted pathogenicity of randomly selected unknown variants in BRCA1 (NC_000017.11) using different semantic embedding models.Genomic HGVSConsequenceMedEmbedgoogleMPNetCountConfPredCountConfPredCountConfPredg.43094629_43095075delsplice_acceptorB:16, P:40.800BenignB:2, P:180.900PathogenicB:10, P:100.500Uncertaing.43102853C > Gintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43067120C > Tintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43093771_43093774delinsTCframeshiftB:0, P:201.000PathogenicB:0, P:201.000PathogenicB:0, P:201.000Pathogenicg.43061638A > Gintron_variantB:20, P:01.000BenignB:20, P:01.000BenignB:20, P:01.000Benigng.43065397G > Aintron_variantB:20, P:01.000BenignB:19, P:10.950BenignB:20, P:01.000Benigng.43067430delintron_variantB:20, P:01.000BenignB:8, P:120.600PathogenicB:20, P:01.000Benigng.43074502G > TMissenseB:17, P:30.850BenignB:13, P:70.650BenignB:16, P:40.800Benigng.43094690T > CMissenseB:20, P:01.000BenignB:18, P:20.900BenignB:20, P:01.000Benigng.43115774T > GMissenseB:12, P:80.600BenignB:5, P:150.750PathogenicB:16, P:40.800Benign Predicted pathogenicity of randomly selected unknown variants in BRCA1 (NC_000017.11) using different semantic embedding models. Table 9Predicted pathogenicity of randomly selected unknown variants in BRCA2 (NC_000013.11) using different semantic embedding models.Genomic HGVSConsequenceMedEmbedgoogleMPNetCountConf.PredCountConf.PredCountConf.Predg.32,353,730 A > Gintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32340247G > TmissenseB:18, P:20.900benignB:20, P:01.000benignB:20, P:01.000benigng.32,363,243 A > GmissenseB:19, P:10.950benignB:15, P:50.750benignB:20, P:01.000benigng.32396971_32396972dupframeshiftB:0, P:201.000pathogenicB:0, P:201.000pathogenicB:0, P:201.000pathogenicg.32,370,904 A > Tintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32,322,227 A > Gintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32385434_32385445dupintron_variantB:20, P:01.000benignB:15, P:50.750benignB:20, P:01.000benigng.32,369,761 A > Tintron_variantB:20, P:01.000benignB:20, P:01.000benignB:20, P:01.000benigng.32380102G > AsynonymousB:20, P:01.000benignB:18, P:20.900benignB:20, P:01.000benigng.32,339,409 C > Gstop_gainedB:0, P:201.000pathogenicB:0, P:201.000pathogenicB:0, P:201.000pathogenic Predicted pathogenicity of randomly selected unknown variants in BRCA2 (NC_000013.11) using different semantic embedding models. The BRCA2 results reveal similar patterns but with some instructive disagreements that highlight the method’s sensitivity to semantic context. Most notably, a splice acceptor variant (NC_000017.11:g.43094629_43095075del) demonstrated model disagreement, with MedEmbed predicting benign (B:16, P:4, confidence 0.800), Google embeddings predicting pathogenic (B:2, P:18, confidence 0.900), and MPNet remaining uncertain (B:10, P:10, confidence 0.500). This variant exemplifies cases where the semantic interpretation of clinical annotations may vary, potentially reflecting different emphasis on splice site disruption versus other contextual factors in the training literature. BRCA1 yet-to-review variants The unknown variants distribute throughout the embedding space in a manner consistent with the trained pathogenicity classes. In the MPNet and Google embedding visualizations, yet-to-review / unknown variants segregate into regions predominantly occupied by either benign or pathogenic training variants, particularly evident in the t-SNE projections, where distinct clusters emerge. The MedEmbed results show a more dispersed but still meaningful distribution, with unknown variants positioning themselves in semantically appropriate regions of the embedding space. BRCA2 yet-to-review variants Similar patterns are observed for BRCA2, where yet-to-review or unknown variants demonstrate clear spatial alignment with their predicted pathogenicity classes. The Google embeddings show particularly strong clustering behavior, with unknown variants forming coherent groups that align with either the benign or pathogenic regions. The UMAP projections across all models reveal that unknown variants maintain the same topo- logical relationships as the training data, suggesting robust semantic capture of pathogenicity-relevant features. Additional FBN1 variants The following analysis is the validation of second independent test, which include the original 30-variant ClinGen VCEP pilot and the independent 348-variant UDM sensitivity validation, aside from the expanded 89-variant ClinGen expert-panel test set. for which the MCC-based metrics are reported in the Comprehensive Multi-Gene Evaluation section above. The 2015 ACMG/AMP guidelines for variant interpretation do not fully account for gene-specific characteristics. Since 2019, the ClinGen FBN1 Variant Curation Expert Panel (VCEP) has invested substantial effort in developing consensus recommendations tailored to the unique features of this disease gene and its pathogenic variants35. To evaluate these new guidelines, the expert panel conducted a pilot study on 60 representative, challenging variants. This effort involved collaboration among three core sites and six non-core institutions, demonstrating improved interpretive consistency compared to the 2015 ACMG/AMP framework. Of these 60 variants, 30 overlapped with the 1,532 variants used to construct the FBN1 vector embedding space. Pathogenicity prediction was therefore performed only on the remaining 30 variants in this second FBN1 validation. These 30 unresolved variants segregated into established pathogenic or benign clusters, particularly evident in the UMAP projection using MPNet embeddings (Fig. 9). All 30 variants received either pathogenic or benign classifications. For 10 variants, the FBN1 VCEP—either at core sites, non-core sites, or both—had been assigned a VUS classification under the updated guidelines. Our algorithm, using the best-performing MPNet model for FBN1, classified 7 of these as pathogenic and 3 as benign. For the remaining 20 variants, MPNet predictions agreed with VCEP classifications for 17 (85%). We examined the three discordant calls, which are all missense variants predicted as pathogenic by our algorithm with MPNet model but classified as benign by the FBN1 VCEP. Notably, ClinVar lists all three as VUS, with submissions from more than four independent laboratories each, underscoring the uncertainty surrounding their true pathogenicity.Fig. 9Visualization of predicted pathogenicity of 30 variants (highlighted as green diamonds) on FBN1, piloted by the ClinGen FBN1 variant curation expert panel (VCEP). The pathogenicity is simply mapped into binary labels: benign and pathogenic. Each panel displays dimensionality reduction (PCA, UMAP, and t-SNE) of variant embeddings. The consistent spatial distribution of unknown variants relative to known benign and pathogenic variants across models and genes suggests that embeddings effectively capture pathogenicity-related semantics. Visualization of predicted pathogenicity of 30 variants (highlighted as green diamonds) on FBN1, piloted by the ClinGen FBN1 variant curation expert panel (VCEP). The pathogenicity is simply mapped into binary labels: benign and pathogenic. Each panel displays dimensionality reduction (PCA, UMAP, and t-SNE) of variant embeddings. The consistent spatial distribution of unknown variants relative to known benign and pathogenic variants across models and genes suggests that embeddings effectively capture pathogenicity-related semantics. We further validated the framework using a subset of FBN1 variants linked to specific cardiovascular phenotypes (e.g., aortic dilation, mitral valve prolapse) retrieved from the UDM-FBN1 mutations database (http://www.umd.be/FBN1/; UDM). From an initial collection of approximately 600 variants, we identified 308 that were already present in our ClinVar training set with high-quality review status; these were excluded from the validation set to prevent data leakage. The remaining 348 variants served as an independent test set. Pathogenicity predictions were generated using k = 5 nearest neighbors (Table 10). The all-mpnet-base-v2 model achieved the highest sensitivity, correctly identifying 97.13% (n = 338) of these pathogenic variants. This significantly outperformed Google Embedding (83.62%) and MedEmbed-large-v0.1 (75.00%). The superior performance of MPNet suggests that general-purpose semantic models may be more robust than domain-specific models in capturing the nuanced textual descriptions associated with complex phenotypic presentations. The spatial distribution of these test variants is visualized in Fig. 10, and detailed results are presented in Table 10.Table 10Sensitivity of pathogenicity prediction for FBN1 variants with variable cardiovascular involvement (UDM dataset, n = 348, k = 5).all-mpnet-base-v2google-embeddingMedEmbed-large-v0.1Prediction of pathogenic (n)338291261Prediction of pathogenic (%)97.13%83.62%75.00%Fig. 10Visualization of 348 external FBN1 variants from the UDM database. The green diamonds represent the query variants (associated with cardiovascular phenotypes) projected into the semantic space of 1,532 known ClinVar variants. The query variants predominantly cluster within the established pathogenic regions (red/orange) across all three models, with the tightest clustering observed in the MPNet PCA and UMAP projections. Table 10Sensitivity of pathogenicity prediction for FBN1 variants with variable cardiovascular involvement (UDM dataset, n = 348, k = 5).all-mpnet-base-v2google-embeddingMedEmbed-large-v0.1Prediction of pathogenic (n)338291261Prediction of pathogenic (%)97.13%83.62%75.00% Sensitivity of pathogenicity prediction for FBN1 variants with variable cardiovascular involvement (UDM dataset, n = 348, k = 5). Visualization of 348 external FBN1 variants from the UDM database. The green diamonds represent the query variants (associated with cardiovascular phenotypes) projected into the semantic space of 1,532 known ClinVar variants. The query variants predominantly cluster within the established pathogenic regions (red/orange) across all three models, with the tightest clustering observed in the MPNet PCA and UMAP projections. Unsupervised disentanglement of pathogenic mechanisms via protein language models To determine whether semantic embeddings capture biologically meaningful structural disruptions rather than mere sequence heuristics, we analyzed the latent space of the ESMC-600 M protein language model applied to FBN1 (Fibrillin-1). We hypothesized that a “residue-level delta embedding”—the vector difference between the mutant and wild-type representations at the specific amino acid site—would isolate the pathogenicity signal from the background noise of protein sequence identity. Topological Separation of Variant Subtypes. Dimensionality reduction (UMAP) of these delta embeddings revealed a striking, unsupervised segregation of variants based on their underlying molecular mechanism (Fig. 11, top).Fig. 11Unsupervised disentanglement of pathogenic mechanisms in FBN1. Dimensionality reduction of residue-level delta embeddings of missense FBN1 variants (n = 216). Top: Variants segregate by molecular mechanism without supervision: variants disrupting disulfide bonds (red) form a tight “structural collapse” cluster, distinct from the diffuse cluster of calcium-binding site variants (blue). Bottom: Evolutionary fitness landscape. The structural clusters align with Zero-Shot LLR scores, where disulfide disruptions exhibit the most severe evolutionary penalties (mean LLR = − 3.77), confirming that the model captures the biological severity of specific structural insults. Unsupervised disentanglement of pathogenic mechanisms in FBN1. Dimensionality reduction of residue-level delta embeddings of missense FBN1 variants (n = 216). Top: Variants segregate by molecular mechanism without supervision: variants disrupting disulfide bonds (red) form a tight “structural collapse” cluster, distinct from the diffuse cluster of calcium-binding site variants (blue). Bottom: Evolutionary fitness landscape. The structural clusters align with Zero-Shot LLR scores, where disulfide disruptions exhibit the most severe evolutionary penalties (mean LLR = − 3.77), confirming that the model captures the biological severity of specific structural insults. Disulfide bond disruptions (cysteine loss/gain). These variants formed a tightly compacted cluster distinct from other pathogenic variants. This topological tightness suggests that the model recognizes a consistent “structural collapse” signature associated with the breaking of covalent cysteine bridges—a primary driver of protein instability in Marfan syndrome. Calcium-binding alterations. Variants affecting calcium-binding epidermal growth factor-like (cbEGF) domains formed a separate, more diffuse cluster. This spatial distinction implies that the model differentiates between the global destabilization caused by disulfide loss and the local rigidity/flexibility defects caused by calcium-binding disruptions. Crucially, this mechanistic separation was achieved without providing the model with 3D structural coordinates or functional labels, demonstrating that the protein language model has implicitly learned the physics of protein folding solely through evolutionary sequence patterns. Convergence of structural impact and evolutionary fitness. To validate these structural embeddings against evolutionary constraints, we calculated Zero-Shot Log-Likelihood Ratios (LLR) for 216 missense FBN1 variants associated with cardiovascular phenotypes. The LLR serves as a proxy for evolutionary fitness, quantifying how unexpected a variant is given the protein’s evolutionary context. We observed a high degree of concordance between the structural embedding space and evolutionary fitness scores (Fig. 11, bottom): Pathogenicity gradient. LLR scores ranged from − 10.43 (highly deleterious) to + 0.45 (neutral). Structural–evolutionary coupling. Variants within the disulfide disruption structural cluster exhibited the most severe negative LLR scores (mean LLR = − 3.77), with extreme outliers (< − 7.0) exclusively confined to this region. Clinical implications of “delta” representations Our comparison of embedding strategies revealed a critical methodological insight for clinical AI. Raw mutant embeddings were dominated by sequence identity, obscuring the pathogenic signal. By utilizing residue-level delta embeddings, we successfully subtracted the “background” biological context, leaving only the vector of the pathogenic insult. This approach transforms the “black box” of the LLM into an interpretable tool capable of distinguishing between varying mechanisms of disease—such as haploinsufficiency (protein instability) versus dominant-negative effects—in silico, before experimental functional validation. Distance ratio analysis confirms robust class separation To quantitatively validate the class separation observed in the dimensionality reduction plots, we applied our Distance Ratio metric to the embeddings of BRCA1 and BRCA2 variants. The results confirm that all three embedding models produce a highly structured and discriminative semantic space. As shown in Fig. 12, the distributions of the Distance Ratio for both pathogenic (red) and benign (blue) variants are overwhelmingly concentrated to the left of the 1.0 threshold. This visually demonstrates that for most variants, the distance to same-class neighbors is significantly smaller than the distance to different-class neighbors.Fig. 12Distance Ratio distributions confirm strong separation of benign and pathogenic variants. The figure shows histograms of the Distance Ratio for BRCA1 (top row of each panel) and BRCA2 (bottom row) variants, evaluated on three different embedding models. The ratio compares a variant’s distance to same-class neighbors versus different-class neighbors. The dashed vertical line marks the theoretical boundary of 1.0; ratios below this line indicate effective clustering. The overwhelming concentration of both pathogenic (red) and benign (blue) distributions at ratios significantly less than 1.0 provides strong quantitative evidence that the semantic embedding space effectively separates variants based on their clinical significance. Distance Ratio distributions confirm strong separation of benign and pathogenic variants. The figure shows histograms of the Distance Ratio for BRCA1 (top row of each panel) and BRCA2 (bottom row) variants, evaluated on three different embedding models. The ratio compares a variant’s distance to same-class neighbors versus different-class neighbors. The dashed vertical line marks the theoretical boundary of 1.0; ratios below this line indicate effective clustering. The overwhelming concentration of both pathogenic (red) and benign (blue) distributions at ratios significantly less than 1.0 provides strong quantitative evidence that the semantic embedding space effectively separates variants based on their clinical significance. This finding is supported by quantitative analysis (Table 11), which shows that across all models and genes, over 98% of variants have a Distance Ratio less than 1.0, indicating effective clustering. The all-mpnet-base-v2 model showed the most consistent high performance across all categories, with an overall average of 98.7% of variants correctly clustered. While all models performed exceptionally well, we observed that BRCA2 variants, which have more extensive annotation data, showed slightly better class separation than BRCA1 variants. This robust quantitative validation provides high confidence in using the semantic embedding space for downstream classification and similarity search tasks.Table 11Percentage of variants with a Distance Ratio < 1.0, calculated with k = 20 nearest neighbors.ModelBRCA1 Path (%)BRCA1 Ben (%)BRCA2 Path (%)BRCA2 Ben (%)Overall (%)all-mpnet-base-v298.297.599.399.498.7google-embedding97.799.498.598.598.4MedEmbed-large-v0.197.399.598.899.798.6Values represent the proportion of variants that are closer to their same-class neighbors than to opposite-class neighbors, indicating effective separation. Table 11Percentage of variants with a Distance Ratio < 1.0, calculated with k = 20 nearest neighbors.ModelBRCA1 Path (%)BRCA1 Ben (%)BRCA2 Path (%)BRCA2 Ben (%)Overall (%)all-mpnet-base-v298.297.599.399.498.7google-embedding97.799.498.598.598.4MedEmbed-large-v0.197.399.598.899.798.6Values represent the proportion of variants that are closer to their same-class neighbors than to opposite-class neighbors, indicating effective separation. Percentage of variants with a Distance Ratio < 1.0, calculated with k = 20 nearest neighbors. Values represent the proportion of variants that are closer to their same-class neighbors than to opposite-class neighbors, indicating effective separation. Clinical implications The consistent distribution patterns across both genes and all embedding models strongly support the hypothesis that semantic embeddings can effectively capture the linguistic and clinical nuances present in variant annotations that distinguish pathogenic from benign variants. This spatial coherence between known and unknown variants indicates that the learned representations encode biologically meaningful features rather than spurious correlations, providing confidence in the method’s ability to generalize to variants of uncertain significance (VUS). The preservation of clustering structure suggests that clinicians could potentially leverage these embeddings not only for binary classification but also for confidence estimation and identification of variants requiring additional functional studies. Discussion Feature circularity and biological independence A central concern for any classifier trained on ClinVar-annotated variants is whether observed accuracy reflects genuine biological learning or simply the re-ranking of pre- computed pathogenicity scores (CADD, REVEL, AlphaMissense) already present in the annotation text. Our ablation study directly addresses this. Removing all predictor scores (Ablation 1) leaves MPNet performance unchanged or improved across eight genes—LOO MCC increases on MYH7 (∆MCC = + 0.016), PAH (+ 0.019), and BRCA2 (+ 0.001). These results demonstrate that numerical predictor fields introduce distributional noise into MPNet’s embedding space rather than providing informative signal. MedEmbed shows mild, consistent drops (typically − 0.01 to − 0.06 MCC) attributable to marginal benefit from biomedical score terminology in its pre- training corpus. Even the most score-dependent model (Gemini, Ablation 1 ∆MCC up to − 0.157 on PAH) retains LOO MCC > 0.80 for all genes, confirming that no model relies exclusively on pre-computed scores. The embedding approach captures independent biological signal—consequence type, population allele frequencies, and genomic context—that is sufficient for high variant discrimination without reference to existing prediction tools. Dataset independence and distributional shift The training set (ClinVar, two-star or higher review status) and the held-out test set (ClinGen expert panel curations) are derived from separate databases with distinct curation processes, and no variant appears in both sets. Within a gene, variants share transcript-level annotations (gene identifier, canonical transcript) as necessary biological context, but each variant’s embedding is uniquely determined by its full per-variant annotation record: consequence type, allele frequency, co-located variant identifiers, and genomic coordinates. Two variants in the same gene with different consequences occupy different positions in the embedding space. This per-variant specificity is identical to the information used by all reference-based variant classifiers and does not constitute data leakage. The systematic LOO-to-test performance gap—most pronounced for FBN1 and BRCA1—reflects a genuine distributional shift rather than contamination: ClinVar aggregates broad laboratory consensus across well-characterized variants, whereas ClinGen expert panels specifically curate contested and uncertain cases, yielding a structurally harder test distribution. Generalisation across genes and class imbalance Evaluation across eight ACMG Tier 1 genes confirms that the semantic embedding framework is not an artefact of specific gene curation histories. USH2A and MYH7 generalize best (test MCC 0.67–0.86 across all three models), benefiting from balanced test-set pathogenicity distributions. For PAH (94.6% pathogenic test variants) and FBN1 (93.3% pathogenic), the F1–MCC divergence is diagnostic of test class imbalance rather than classifier failure: a model predicting pathogenic for nearly all variants achieves F1 of 0.93–0.97 while MCC correctly penalizes the failure to discriminate the few benign variants, which had conflicting assertions hence requiring curation from ClinGen Expert Panels. This distinction is critical for clinical deployment, where both false-positive and false-negative pathogenicity calls carry significant consequences. We recommend reporting MCC as the primary metric for all future evaluations involving imbalanced variant sets. Mechanistic interpretability via protein language modeling Beyond classification accuracy, a key ambition of computational variant interpretation is mechanistic transparency. The ESMC-600M analysis of FBN1 demonstrates that residue-level delta embeddings—vector differences between mutant and wild- type representations at the affected residue—achieve unsupervised disentanglement of distinct pathogenic mechanisms without access to three-dimensional structural coordinates or functional labels. Disulfide bond-disrupting variants form a topologically compact cluster with a mean Zero-Shot LLR of − 3.77, reflecting the severe and stereo-typed thermodynamic cost of cysteine loss in FBN1 domains. Calcium-binding site variants form a spatially distinct, more diffuse cluster, consistent with local rigidity defects rather than global structural collapse. This structural–evolutionary concordance, validated against LLR scores as an independent proxy for evolutionary fitness, establishes that protein language models have implicitly encoded the physics of protein folding from evolutionary sequence patterns alone. The approach moves variant interpretation from binary labelling toward a calibrated, mechanism-aware risk assessment, with potential downstream utility in guiding functional validation priorities and understanding genotype–phenotype correlations in Marfan syndrome. Clinical utility and confidence stratification The application of VUS.Life to 3,000 unreviewed BRCA1 and BRCA2 variants illustrates the framework’s immediate clinical potential. Across all three embedding models, unknown variants segregate into established benign and pathogenic regions of the embedding space, and the neighborhood agreement score—the fraction of k nearest known variants sharing the same class—functions as a natural confidence measure. High-confidence predictions (> 95% neighborhood agreement) are suitable for automated reporting, while boundary cases (< 60% agreement) are flagged for genetic counsellor review, providing a principled triage mechanism for clinical genomics workflows. The FBN1 VCEP pilot further demonstrates concordance with expert-panel classifications at 85% (17/20 resolved variants), and the UDM validation identifies 97.1% of phenotype-confirmed pathogenic variants with MPNet, establishing that sensitivity is maintained in gene-specific and phenotype-enriched settings. Distance Ratio analysis confirms the structural integrity of this confidence framework: over 98% of known variants have a Distance Ratio < 1.0, meaning they reside closer to same- class than opposite-class neighbors, validating the geometric assumptions underlying k-NN classification. Limitations Several limitations of the present study should be noted. First, the LOO-to-test gap is real and gene-dependent, reflecting the difficulty of the ClinGen curation queue relative to the ClinVar training distribution. Deployment accuracy will be lower than LOO-CV estimates suggest for genes where ClinGen curators have specifically targeted ambiguous cases. Second, PAH and FBN1 test sets are severely imbalanced, limiting benign discrimination and suppressing MCC independently of classifier quality. Third, Gemini’s score dependence makes it less suitable when VEP annotation completeness cannot be guaranteed. Fourth, the current evaluation covers eight ACMG Tier 1 genes; generalization to genes with sparser ClinVar records, where training set sizes are very small, remains to be established. The framework is modular and embedding models are not gene-specific, such that extension to a broader gene panel is straightforward given sufficient ClinVar evidence, which we identify as a priority for future work. Conclusion VUS.Life demonstrates that semantic embedding of VEP variant annotations into high-dimensional vector spaces enables robust, scalable pathogenicity classification across diverse ACMG Tier 1 disease genes. LOO-CV across eight genes and three embedding models yielded MCC of 0.895–0.989 and F1 ≥ 0.94, with MPNet (all- mpnet-base-v2) achieving the strongest and most consistent performance. An ablation study removing all pre-computed predictor scores confirmed that MPNet captures genuine biological signal—consequence type, population allele frequencies, and genomic context—independently of existing scoring tools, making it robust to incomplete VEP annotation. Independent validation on held-out ClinGen expert-panel curations confirmed generalization, with USH2A and MYH7 showing the smallest LOO-to-test gap. The modular architecture requires no retraining for new genes and scales naturally as ClinVar evidence grows, offering a practical path toward alleviating the VUS interpretation bottleneck in clinical genomics. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2