📥 Download Datasets

Public datasets generated by the VarSem end-to-end standardization pipeline

📄

pubmed_varsem_normalized_abstract_variants.jsonl

LLM-based filtering of PubMed abstracts identified abstracts containing variant descriptions. After end-to-end standardization with VarSem, 87,330 PubMed articles with 146,127 standardizable variants were retained, each mapped to standardized genomic coordinates, transcripts, and protein-level annotations.

Format: JSON Lines (.jsonl)

PMIDs: 87,330 articles

Variants: 146,127 standardizable variants

Key Columns: pmid, title, abstract, formatted_prediction

⬇ Download

Table 1. Column structure

Field	Description
pmid	PubMed article identifier
title	Article title
abstract	Full abstract text
formatted_prediction	Standardized variants (Gene, Mutation, Standard_Mutation with Transcript / Chr / Start / End / Ref / Alt / Protein / Strand, evidence span)

87,330 PMIDs 146,127 standardizable variants

📑

pubmed_llm_human_annotated_abstract_variants.jsonl

A curated benchmark of 384 PubMed abstracts jointly annotated by LLM and human experts. Each abstract includes its variant descriptions and the corresponding VarSem-standardized information, providing a gold-standard reference for evaluating variant standardization performance.

Format: JSON Lines (.jsonl)

Records: 384 annotated abstracts

Key Columns: pmid, abstract, answer, variants

⬇ Download

Table 2. End-to-end standardization performance on the ClinVar-Based standard. (Values in %)

Method	Precision	Recall (exact match)	F1 Score
tmVar 3.0 + TransVar	65.29	57.26	61.01
DeepSeek-v3.2 (no VarLingua) + TransVar	73.83	71.92	72.86
DeepSeek-v3.2 (no VarLingua) + HGVS2STD	91.56	78.71	84.65
Full VarSem, pipeline mode (DeepSeek-v3.2)	99.21	98.74	98.97
Full VarSem, pipeline mode (Gemma-4-31B)	98.10	89.43	93.56
Full VarSem, pipeline mode (Qwen3-8B fine-tuned)	98.20	94.48	96.30
Full VarSem, Agent mode (DeepSeek-v3.2)	98.33	93.06	95.62

384 LLM + human annotated abstracts 684 variants

🗂️

pubmed_clinvar_llm_filtered_10000_variants.jsonl

10,000 variants from 7,213 ClinVar-related PubMed abstracts, filtered by LLM-based judgment. Each record includes the variant description, its ClinVar-derived standard form, and the VarSem-standardized information, supporting real-world literature evaluation.

Format: JSON Lines (.jsonl)

Records: 7,213 abstracts

Variants: 10,000

Key Columns: pmid, title, abstract, variants

⬇ Download

Table 3. Comparative evaluation of VarSem and tmVar+TransVar on real-world biomedical literature (%)

Mutation Type	Ratio	tmVar+TransVar Accuracy	VarSem Accuracy	Difference
Single Nucleotide Variant	88.82	92.63	99.76	+7.13
Deletion	7.98	90.60	98.37	+7.77
Duplication	2.22	86.94	99.10	+12.16
Deletion-Insertion	0.56	51.79	91.07	+39.28
Insertion	0.42	83.33	100.00	+16.67
Overall	100.00	92.08	99.59	+7.51

7,213 ClinVar abstracts 10,000 variants

Note: All performance metrics reported above were evaluated on a unified hardware system to ensure consistent and comparable benchmarking across methods.