PromptBio releases a preprint on PromptGenie, a system that records the steps of biomedical analysis
PromptBio releases a preprint on PromptGenie, a system that records the steps of biomedical analysis
On September 6, the PromptBio team posted a preprint on bioRxiv describing PromptGenie, a system for research with biological data. The authors explain how it turns a research question into a plan, code, program execution, and records that allow researchers to check the analysis.
Reaching a conclusion from a biomedical question requires a series of decisions: which data to use, which method to use to test the hypothesis, how to perform the analysis, and how to interpret the results. When these decisions are scattered across files, services, and correspondence, a finished table or graph makes it difficult to trace the data and choices that produced it.
In the preprint, the authors describe PromptGenie as a system that takes a research objective and the available context, builds a plan of interdependent steps, and adjusts subsequent steps in response to intermediate results or errors. Before a consequential decision, the researcher can review and revise the plan. The system then calls an existing method or writes code, runs the analysis, and passes the result to the next step.
The authors describe a record that retains the plan and its revisions, sources and data links, parameters, code, execution environment details, execution logs, and output files. This record allows researchers to trace a conclusion to the input data and computations, inspect the analysis, or rerun it. The researcher remains responsible for judging the reliability of sources and deciding whether experimental validation is needed.
The supplement to the paper illustrates this sequence with a task involving DNA regions associated with cerebral cortex development. The system received the coordinates of these regions and genomic marks, features that researchers use to identify regulatory regions. PromptGenie proposed testing whether these marks occurred in the regions more often than expected by chance: it would compare the observations with ten thousand sets of regions of the same lengths, randomly repositioned within the same chromosomes.
Please confirm this design before I run the full analysis.
After confirmation, PromptGenie ran the analysis and returned code, tables, and plots. The researcher approves the method for testing the hypothesis, while the program performs the computations and records the steps. The procedure described by the authors links the scientific decision to the computational work that produced the result.