Fig. 3: Free energy surfaces projected along the host–guest distance.

For each of the six ligands, we compute the free energy along the s_z variable using a standard umbrella-sampling-like reweighting formula to recover the unbiased distribution¹³. The shaded areas indicate the errors, whose calculation is detailed in the Supplementary Methods. To ensure that the results do not depend on a specific realisation of the Deep-LDA CV, we repeat the training three times by using different initial weights of the NN. The resulting CVs are denoted as \({s}_{{\rm{w}}}^{a}\), \({s}_{{\rm{w}}}^{b}\), \({s}_{{\rm{w}}}^{c}\), and the corresponding FES are indicated, respectively, by dashed, dotted, and dash-dotted lines. For clarity, curves related to the same ligand but with different CVs are shifted by 1 kcal mol⁻¹, while the shift between different ligand curves is 5 kcal mol⁻¹.