varpubs is a command-line tool that automates retrieval and summarization of variant-associated biomedical literature to support faster variant interpretation in oncology.

Identifying treatment-relevant evidence for rare or poorly characterized mutations often requires reviewing large volumes of PubMed articles. varpubs streamlines this process by querying literature for variant-linked publications and generating concise, role-specific three-sentence summaries using a large language model (LLM). Optional LLM-based “judges” can score articles based on user given terms (e.g., therapy relevance) to prioritize relevant articles.

Variant-level summaries per transcript are annotated to the given input VCF file as well as the PubMed IDs of the used literature sources and mean judge scores for every given entry via --judges. A caching system stores previously generated summaries to reduce runtime in subsequent analyses.

pixi global install varpubs

conda install -c bioconda varpubs

varpubs deploy-db \
  --db-path pubmed.duckdb \
  --vcf-paths variants.vcf.gz other.vcf.gz \

varpubs summarize-variants \
  --db-path pubmed.duckdb \
  --vcf-path variants_annotated.vcf.gz \
  --llm-url https://your-llm-endpoint \
  --api-key $HF_TOKEN \
  --model medgemma-27b-it \
  --role physician \
  --judges "therapy relevance" \
  --cache cache.duckdb \
  --output summaries.vcf \
  --output-cache tmp_cache.duckdb

Each LLM call starts with a budget of --max-new-tokens (default 500) and doubles it on every --retries attempt (default 3) if the response was cut off. Responses still truncated after the last attempt are discarded. Model reasoning is disabled by default and can be enabled via --enable-thinking true.

varpubs update-cache \
  --cache tmp_cache.duckdb \
  --output cache.duckdb

• Felix Wiegand
• Felix Mölder
• Hendrik Damm
• İlmay Taş
• Johannes Köster