Background Despite rapid advances in genomic sequencing, most rare coding variants remain insufficiently characterized for clinical use, limiting the potential of personalized medicine. When classifying whether a variant is pathogenic, clinical labs adhere to diagnostic guidelines that integrate many forms of evidence, including case data, computational predictions, and functional screening data. While a great deal of clinical evidence has been curated for many variants, the majority still cannot be definitively classified as ‘pathogenic’ or ‘benign’, and thus persist as ‘Variants of Uncertain Significance’ (VUS). Variant Curation Expert Panels (VCEPs) are tasked with analyzing the available evidence for each variant to reach a classification. Methods To make use of previously curated evidence, we processed over 2.3 million free-text variant summaries from ClinVar, employing sentence-level classification to restrict to sentences that contain different forms of evidence, and removing uninformative or similar summaries. We then used labeled text summaries to train ClinVar-BERT, a model that can discern evidence of pathogenicity or benignity within variant text summaries. Results We validated ClinVar-BERT model predictions for variant summaries that are classified as uncertain using variants curated by expert panels, orthogonal functional screening data, and computational predictions. ClinVar-BERT model predictions of VUS had significantly different estimates of functional impact in clinically actionable genes, including BRCA1 (p = <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1.90 \times 10^{-20}\)</EquationSource> </InlineEquation>), TP53 (p = <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(1.14 \times 10^{-47}\)</EquationSource> </InlineEquation>), and PTEN (p = <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(3.82 \times 10^{-7}\)</EquationSource> </InlineEquation>) with an AUROC = 0.927 when classifying whether variants are damaging or are expected to retain function. Similarly, ClinVar-BERT model predictions of VUS had significantly different AlphaMissense computational scores: BRCA1 (p = <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(1.21 \times 10^{-21}\)</EquationSource> </InlineEquation>), TP53 (p = <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(3.54 \times 10^{-9}\)</EquationSource> </InlineEquation>), and PTEN (p = <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(2.60 \times 10^{-76}\)</EquationSource> </InlineEquation>). In genes screened for secondary findings or included on ClinGen expert panels, ClinVar-BERT prioritizes 7,644 variants for expert review, where 2 or more clinical summaries related to the same VUS were model-predicted to contain pathogenic evidence, and 7,042 variants with 2 or more summaries predicted to contain benign evidence. This would result in the average VCEP having 143 variants prioritized for review, ranging from 8 to 907 variants across VCEPs. Conclusions These findings suggest that ClinVar-BERT can discern evidence from diagnostic reports, useful for prioritizing variants for re-assessment by expert curation panels.

From text to translation: using language models to prioritize variants for clinical review

期刊论文
出版模式：

开放获取
发表日期：

2026年5月19日

查看全文

Weijiang Li,
Xiaomin Li,
Ethan Lavallee,
Alice Saparov,
Marinka Zitnik,
Christopher Cassa

摘要

Background

Despite rapid advances in genomic sequencing, most rare coding variants remain insufficiently characterized for clinical use, limiting the potential of personalized medicine. When classifying whether a variant is pathogenic, clinical labs adhere to diagnostic guidelines that integrate many forms of evidence, including case data, computational predictions, and functional screening data. While a great deal of clinical evidence has been curated for many variants, the majority still cannot be definitively classified as ‘pathogenic’ or ‘benign’, and thus persist as ‘Variants of Uncertain Significance’ (VUS). Variant Curation Expert Panels (VCEPs) are tasked with analyzing the available evidence for each variant to reach a classification.

Methods

To make use of previously curated evidence, we processed over 2.3 million free-text variant summaries from ClinVar, employing sentence-level classification to restrict to sentences that contain different forms of evidence, and removing uninformative or similar summaries. We then used labeled text summaries to train ClinVar-BERT, a model that can discern evidence of pathogenicity or benignity within variant text summaries.

Results

We validated ClinVar-BERT model predictions for variant summaries that are classified as uncertain using variants curated by expert panels, orthogonal functional screening data, and computational predictions. ClinVar-BERT model predictions of VUS had significantly different estimates of functional impact in clinically actionable genes, including BRCA1 (p = \(1.90 \times 10^{-20}\) ), TP53 (p = \(1.14 \times 10^{-47}\) ), and PTEN (p = \(3.82 \times 10^{-7}\) ) with an AUROC = 0.927 when classifying whether variants are damaging or are expected to retain function. Similarly, ClinVar-BERT model predictions of VUS had significantly different AlphaMissense computational scores: BRCA1 (p = \(1.21 \times 10^{-21}\) ), TP53 (p = \(3.54 \times 10^{-9}\) ), and PTEN (p = \(2.60 \times 10^{-76}\) ). In genes screened for secondary findings or included on ClinGen expert panels, ClinVar-BERT prioritizes 7,644 variants for expert review, where 2 or more clinical summaries related to the same VUS were model-predicted to contain pathogenic evidence, and 7,042 variants with 2 or more summaries predicted to contain benign evidence. This would result in the average VCEP having 143 variants prioritized for review, ranging from 8 to 907 variants across VCEPs.

Conclusions

These findings suggest that ClinVar-BERT can discern evidence from diagnostic reports, useful for prioritizing variants for re-assessment by expert curation panels.