Fig. 3

Global direction at varying scaffold size and across feature attribution methods. a Global direction and b, weighted global direction values are reported at different thresholds of minimum shared MACS among testing pairs (%). In b, global direction is weighted by the number of pairs per each target. Results are shown for three loss functions, i.e. \({\mathcal {L}}_{\text{MSE}}\) (left panel), \({\mathcal {L}}_{\mathrm {MSE+AC}}\) (middle panel), and \({\mathcal {L}}_{\mathrm {MSE+UCN}}\) (right panel). Colors report different feature attribution methods, five for GNN models and atom masking for RF models. Since the three losses functions are only applied to GNN models, RF results are equivalent in the three panels. An additional random feature attribution line is included as a baseline