Expression reconstruction is not mechanism recovery. scDrugPerturb-Bench is a virtual cell benchmark that links matched control and drug-treated single-cell RNA-seq profiles to literature-curated directional key-gene evidence, testing whether virtual-cell models recover the genes, pathways and response specificity that make drug predictions biologically actionable.
A representative prediction looks excellent by expression reconstruction, yet reverses the entire known response programme.
ABCA1, ABCG1, APOE and ACSL1 should all increase. The model predicts the opposite direction for every gene at both doses.
MFS asks which parts of the biological mechanism a prediction actually recovers.
Do key genes change in the annotated direction?
Does the predicted magnitude match the measured effect?
Are mechanism genes recovered as a coordinated response?
Is the signal concentrated on mechanism-relevant genes?
Are predicted and observed pathway profiles concordant?
Are activated and suppressed pathways assigned correctly?
Traditional reconstruction metrics remain useful: they tell us whether the predicted profile resembles the measured response. MFS adds evidence that resemblance alone cannot provide.
Correlations across the two metric families are mixed and often weak. A model can therefore score well on reconstruction without preserving the genes and pathways that define the mechanism.
MFS is an important complement, not a replacement.
