Fig. 12

Results of the pairwise XGB-FN models and the assay relatedness analysis. A, C pairwise FN models for the Ames dataset (A) and Tox21 dataset (C). For each assay, pairwise FN models were trained with each of the remaining assays and, for each pair, 20 independent runs of the model were conducted using different random seeds and evaluated on the test set. To obtain the heatmap, the median MCC score was computed for each pair and the median MCC of the single task model for respective target assay was subtracted. The diagonals represent the differences in MCC score between the full FN model and the single task XGB model as a reference. B, D For each target assay-auxiliary assay pair the difference between MCC of pairwise FN model and single task model are plotted against the MI-entropy ratio of the two assays for the Ames dataset (B) and the Tox21 dataset (D).