OncoGenRAG: Evidence-Grounded Retrieval and BioBERT Classification for Precision Oncology Variant Interpretation
The increasing use of tumor sequencing has intensified the need for fast, traceable interpretation of genomic variants. General-purpose large language models can produce fluent answers, but unsupported statements, weak provenance, and stale knowledge limit their suitability for clinical genomics. We developed OncoGenRAG, a research framework that combines a parameter-efficiently fine-tuned…
OncoGenRAG is a novel research framework designed for precision oncology variant interpretation. This innovative system tackles the challenges posed by the rapid adoption of tumor sequencing by offering a solution that provides fast, traceable interpretation of genomic variants. It achieves this by integrating a BioBERT classifier fine-tuned with parameter efficiency, coupled with an entity-aware retrieval system that draws from a comprehensive, multi-source oncology knowledge base.
The curated knowledge base consists of 933 harmonized records sourced from various reputable databases, including CIViC, ClinVar/dbSNP, Open Targets, UniProtKB/Swiss-Prot, Ensembl Variation, and linked PubMed literature. The BioBERT classifier is adept at categorizing variants into one of five distinct labels: Pathogenic, Likely Pathogenic, Variant of Uncertain Significance, Benign, or Oncogenic.
Meanwhile, the retrieval component is responsible for ranking evidence records, employing subword TF-IDF similarity alongside explicit matches in gene, variant, and cancer-type attributes.
To ensure the reliability of its responses, OncoGenRAG incorporates a rejection rule that prevents the system from providing answers if retrieval support falls below a predetermined threshold. This mechanism serves as a safeguard against the dissemination of unsupported statements, weak provenance, or stale knowledge, which are common pitfalls of general-purpose large language models.
In rigorous testing, the classifier demonstrated impressive accuracy, achieving a 92.40% accuracy rate, a 93.15% weighted precision, a 92.40% weighted recall, and a 92.65% weighted F1 score. Additionally, an external benchmark featuring 100 clinical-style queries yielded remarkable results, with a Precision@1 of 94.5%, Precision@3 of 96.8%, and a 100% database grounding rate. Importantly, no hallucinated answers were reported, in stark contrast to a 41.0% no-hallucination rate observed in an ungrounded baseline.
However, it is crucial to note that these findings represent internal validation rather than evidence of universal safety. This is primarily due to the absence of query construction, annotator agreement, class-specific performance, calibration, and external validation data for independent analysis. Despite being a research prototype, OncoGenRAG introduces a transparent design for evidence retrieval and abstention, positioning itself as a promising tool in the field of precision oncology.
Nonetheless, its application in treatment selection should always be under the supervision of expert review.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
