Author rebuts leakage concern: each cell line's contribution capped at 10%

kjaganatha · x · 2026-11-20

Responding to Anshul Kundaje's leakage concern, the author explains the test set held out entire classes of cell lines (no possible matches), and each cell line's contribution to predictions is capped at 10%, so no training cell line can dominate regardless of similarity.

Related event: Stanford professor challenges Arc virtual cell challenge winner: data privilege or methodological edge(18 posts)→

Original post →