Metappuccino : Large Language Model-driven Reconstruction of Sequence Read Archive Metadata for Cancer Research
Résumé
Motivation : High-throughput RNA-sequencing has significantly advanced transcriptomic profiling in oncology. Millions of RNA-seq datasets have accumulated in public databases such as the Sequence Read Archive-SRA. However, fragmented, ambiguous or missing metadata can severely limit accurate cohort selection, introduce bias and delay discoveries. Results : To address these issues, we introduce Metappuccino : a metadata enrichment tool based on a fine-tuned Mistral-7B-Instruct large language model with low-rank-adaptation (LoRA). Metappuccino can extract or infer 19 key metadata classes (e.g. organ, disease, cell type) from unstructured text. Fine-tuning was conducted with careful partitioning and training design to preserve the model's generalisation capacity, reduce data leakage, and ensure robust, context-aware inference rather than memorisation. When possible, the inferred outputs are mapped to standardised ontologies, such as Cellosaurus, Disease Ontology and Uberon, to produce consistent metadata. As a result, the fine-tuned model achieves significantly improved class prediction accuracy over the base model, performing at least as well as recent large open-source models. Furthermore, it reduces inference time by up to at least two compared to the baseline models. As a pipeline, Metappuccino complements the LLM with well-established Natural Language Processing techniques from the literature to further improve performance. By enriching the metadata of under-annotated sequences, Metappuccino creates greater value from public RNA-seq datasets, with potential applications extending beyond oncology transcriptomics.
Domaines
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| licence |
