Fast Leave-One-Out Approximation from Fragment-Target Prevalence Vectors (molFTP) : From Dummy Masking to Key-LOO for Leakage-Free Feature Construction

A fundamental question for users of predictive models is: how good is the training data (1)? One way to approach this is by delineating the model's applicability domain. A second safeguard is to prevent data leakage (2), which motivates deduplication and proper validation protocols (3) . In practice, it is standard to use cross-validation or time-series split such as SIMPD (4) . Beyond sample-level leakage (molecules crossing folds), we must also consider feature leakage (when features inadvertently encode information about held-out molecules) (2) . We return to this point in the related Work section.