Fig. 2: Coarse-grained DMPC lipids tunnel through membrane.
From: AI-guided transition path sampling of lipid flip-flop and membrane nanoporation

a Time-average of pore reaction coordinate, \({\xi }_{{\mbox{P}}}\), evolving during the TPS MC chain. We compare samplers starting from an intact membrane (dry path \({\varPi }_{{\mbox{T}}}\), red) and from a formed pore (wet path \({\varPi }_{{\mbox{P}}}\), blue). Dark colors show a sample average, smoothed over \(10\) TPS MC steps. b Accuracy of committor models comparing different input features to random committor assignments (“rand”). Boxes show median and \(25/75\) th percentile of \(1000\) bootstraps drawn from a total of 10000 MC steps; whiskers show \(2.5/97.5\) th percentile. c, d Distribution of committor estimates for a given feature, comparing transversal displacement \(z\) (c) with a linear combination of bead-to-bead distances to neighboring lipids, \({{{\mathbf{\Delta }}}{{\bf{r}}}}^{{\mbox{all}}}\cdot {{{\bf{v}}}}_{{{\bf{r}}}}\) (d). e Projection of the TPE onto \(z\) and \({{{\mathbf{\Delta }}}{{\bf{r}}}}^{{{\rm{all}}}}\cdot {{{\bf{v}}}}_{{{\rm{r}}}}\). Gray iso-lines show the committor averaged over \(5000\) nearest-neighbors (\(0.6\)% of all data). Representative configurations (dots) are shown in the seven side-panels I–Ⅶ. Nearest neighbors are colored in white with increasing intensity, and distances in yellow.