Malva polyA query probes and LLM evaluation

Data deposits for "Malva: ultrafast, large-scale sequence discoveryin single cells" This repository contains supplementary data files referenced in themanuscript. Contents: 1. polyA_probes.fa Polyadenylation site detection probes (Sup. Fig. 10) 2. llm_evaluation_queries.json Natural language query interface evaluation dataset (Sup. Fig. 12) 1. Polyadenylation site detection probes for Malva This file contains nucleotide probes designed for detecting alternativepolyadenylation events using Malva Index. Probes were constructed frompolyASite v3.0 (Moon et al., Nucleic Acids Res. 2025) annotations. Probes with extreme GC content (<30% or >70%) were excluded. Probeheaders encode the polyASite identifier, genomic coordinates, probetype (polyA or control), and the associated gene symbol. File format: FASTA (.fa)Reference genome: GRCh38 2. Evaluation dataset for the Malva natural language query interface This file contains 10,000 synthetic test cases used to evaluate theaccuracy of the natural language query translation engine in the Malvaplatform. Test cases were generated using a local instance ofgpt-oss:120b, ensuring independence from the production model(Llama 3.1 8B). Each entry contains: - query: the natural language input string - expected_type: the correct query category (one of: simple_gene, gene_with_filters, marker_genes, sequence_search, database_lookup, pathway_query, unsupported, over_inference_trap) - expected_genes: list of gene symbols that should be extracted - expected_filters: dictionary of metadata filters explicitly stated in the query (e.g., tissue, disease, cell_type) - expected_sequence: nucleotide sequence, if applicable The over_inference_trap category (15% of cases) tests whether the modelinappropriately adds biological context as filters when not explicitlyrequested by the user. File format: JSON (.json)