The recoverable resolution of virtual-cell prediction — Analysis code

Abstract

This record contains the analysis code and figure-level source tables for “The recoverable resolution of virtual-cell prediction.” The study shows that molecular measurement detail can exceed the predictive detail supported by the information available at prediction time. It establishes recoverable resolution as a design principle: predictive detail should match the biological target, the available information and the required reliability. Analyses cover human CD4+ T-cell perturbations, official State predictions, Tahoe-100M, paired lymphoblastoid cell-line responses and patient-derived colorectal-cancer organoids. LARRY lineage tracing and CellTag-multi measurements examine constraints on realized outcomes and the specificity of calibrated fate predictions. The archive includes: Scientific implementations for response prediction, experimental-support analysis, response projection and baseline representation comparisons. LARRY paired-error calculations and CellTag-multi fate-prediction and calibration analyses. Numerical source tables, pathway weights and panel mappings for five main figures, three Extended Data figures and four supplementary figures. Figure-generation scripts and schematic assets. Software requirements, dataset accessions, input specifications, implementation tests and file checksums. Figure rendering uses the CODE archive. Analyses using the included processed measurements require the companion DATA archive. Extract both archives into the same directory; they share a common CGC/ folder. Required inputs, computing resources and execution commands are documented in REPRODUCIBILITY.md. The code repository is available at GitHub. CGC-authored code is released under the MIT license; third-party resources retain their original licenses and attribution.