Predictive models for tyrosinase inhibitors: Challenges from heterogeneous activity data determined by different experimental protocols.

Quantitative Structure-Activity Relationship (QSAR) models of tyrosinase inhibitors were built using Random Forest (RF) algorithm and evaluated by the out-of-bag estimation (R2 OOB ) and 10-fold cross validation (Q2 CV ). We found that the performances of QSAR models were closely correlated with the systematic errors of inhibitory activities of tyrosinase inhibitors arising from the different measuring protocols. By defining ERRsys , outliers with larger errors can be efficiently identified and removed from heterogeneous activity data. A reasonable QSAR model (R2 OOB of 0.74 and Q2 CV of 0.80) was obtained by the exclusion of 13 outliers with larger systematic errors. It is a clear example of the challenge for QSAR model that can overwhelm heterogeneous data from different experimental protocols.