BioMeme: biological response memory for virtual cells

BioMeme predicts next-dose responses in a new cell type using biological response memory, achieving error within 3.4% in U2OS and improving directionality in nine preregistered comparisons.

• Why it matters: Virtual-cell neural models must be retrained and cannot cite supporting experiments, leaving a gap for updateable, inspectable prediction grounded in measured data.
• What they did: BioMeme replaces a learned state-to-response layer with a memory of named measurements, predicting as an exact weighted sum and updating via one linear solve across three screens.
• The result: Across nine preregistered comparisons, memory predicted response direction better than the learners it replaced, was compound-specific (21.2% error reduction for correct compounds), replicated in reserved MCF7 data, and ran in seconds on one CPU thread.

AI summary of the abstract · may contain errors · not medical advice · read the paper.