A confound-diagnostic toolkit for in silico perturbation with single-cell foundation models

The authors note that deleting a gene token from a cell's input sequence is a convenient native way to simulate knockout effects, but the resulting embedding delta can be confounded by gene identity, universal responsiveness, tokenization coverage gaps, library-size contamination, or circular state scoring. To address this, they present a framework combining held-out increment testing, responsiveness adjustment, coverage gating, and library-size correction. The toolkit aims to help researchers distinguish genuine biological perturbation signals from technical noise, improving the reliability of hypothesis generation in single-cell genomics. As a preprint, this work is not yet peer-reviewed, but it highlights a growing concern in the use of foundation models for in silico screens.