Fig. 4: Attribution analysis of the Bayesian optimization model.

a Descriptor ranking by SHAP (SHapley Additive exPlanations) values based on the Gaussian process regression (GPR) model for optimizing P_m values (P_m, probability of meso linkages) of the entire dataset of Al complexes. The top rank indicates the most significant effects across all the predictions. The positive and negative SHAP value refers to positive and negative correlation, respectively, between the measured P_m value and the corresponding feature. The larger the absolute value, the stronger the correlation is. The color coding indicates normalized high (red) to low (blue) feature values. b Correlation between P_m and %V_bur of the Al complex with the P_m value. %V_bur, percent of buried volume. c–e Multivariant regression model highlighting important descriptors impacting c the P_m values for all Al complexes, d the P_r values (probability of racemic dyad formation) in Al complexes with A_mC₁B₂ and A_mC₁B₃ ligands, and e the P_m values in Al complexes with A₃C_pB_n ligands. E_HOMO(A), HOMO energy of the A_m fragment (HOMO, the highest occupied molecular orbital); EN(A), electronegativity of the A_m fragment; V_bur, %V_bur of the whole ligand; V_bur-max(A), maximum %V_bur of the A_m fragment; Freq_min(A), minimum frequency of the A_m fragment; E_HOMO(BC), HOMO energy of the B_nC_p fragment; EN(BC), electronegativity of the B_nC_p fragment; V_bur-min(BC), minimum %V_bur of the B_nC_p fragment; Freq_min(BC), minimum frequency of the B_nC_p fragment; MAE mean absolute error. In the equations in c–e, the descriptors of the A_m fragment are highlighted in purple, and those of the B_nC_p fragment in green, and residue numbers in pink.