PxFquery: A Bioinformatics Tool for Large Language Model-Assisted Functional Analysis of Large-Scale Perturbation Signatures
Background/Objectives: Genetic and chemical perturbation experiments provide a systematic approach to investigate cellular responses. Large-scale resources such as Connectivity Map (CMap) contain extensive transcriptomic perturbation signatures that support functional interpretation and perturbation retrieval. However, existing access to these resources mainly relies on structured inputs, making it challenging to connect natural-language perturbation questions with experimental evidence. Methods: To address this challenge, we developed PxFquery, an evidence-grounded tool that enables natural-language exploration of large-scale perturbation resources. PxFquery uses an LLM-assisted workflow to interpret biological questions and organize responses based on retrieved perturbation evidence. It converts over 500,000 CMap/LINCS perturbation signatures into a compact functional response space and supports bidirectional perturbation-function queries. Results: Despite the sparse perturbation coverage of CMap (5.9%), PxFquery integrated related perturbation evidence and improved access to perturbation resources. Across evaluated genetic and chemical perturbation-to-function and function-to-perturbation queries, PxFquery outputs showed closer agreement with experimental reference rankings than direct and PubMed-augmented LLM approaches (paired Wilcoxon tests; p < 0.05). Functional response representation reduced storage requirements to 0.068–0.24% of the original resources and enabled lightweight deployment through a Python package (v0.5.29), website, and AI workflow interfaces. Conclusions: PxFquery provides a natural-language interface for exploring large-scale perturbation resources while maintaining connections to experimental evidence. It lowers the barrier to accessing these resources and enables their integration into diverse AI workflows.