Fig. 2
From: Analysis of the benefits of imputation models over traditional QSAR models for toxicity prediction
Schematic depiction of Feature Net (FN) models. In Step 1, a single task QSAR model is trained to predict the labels of missing values for each toxicity assay in turn (y1, y2, y3). In Step 2, the models are retrained using both chemical descriptors (x) and auxiliary assay labels as features. The labels for the auxiliary assay are either experimentally determined assay labels or the predictions from Step 1 where experimental values are not available