Simulations probe shape-shifting of opium poppy's reticuline-flipping enzyme using a predicted structure

What this research was about

Opium poppy makes morphine and related compounds through a long chain of chemical steps. One of them is handled by an enzyme called REPI, which takes a molecule named (S)-reticuline and converts it to the mirror-image form, (R)-reticuline, by way of an intermediate called 1,2-dehydroreticuline. REPI is a fusion protein: two working parts, known as DRS and DRR, joined into a single chain by a flexible connector. DRS uses a heme group, an iron-containing helper molecule; DRR uses NADP, a helper that carries chemical reducing power. To understand how the intermediate travels from one part to the other, researchers need the enzyme's three-dimensional shape, and no experimentally determined structure of REPI exists.

The team therefore built a computational model of the whole enzyme and set it in motion. They placed the heme and NADP helpers into the two parts by matching them onto related proteins whose crystal structures are known, docked (S)-reticuline and the intermediate into their expected sites, and ran all-atom molecular dynamics, which calculates the forces between every atom to produce a film of the protein jiggling in water. Four systems were simulated for 400 nanoseconds each: the enzyme without a bound chemical and three complexes. They then analysed the resulting trajectories for movement, tunnels through the protein and electrical charge on its surface.

Where AI came in

The only learned model in the work is AlphaFold2, a neural network that predicts a protein's folded shape from its amino-acid sequence. The researchers downloaded its ready-made prediction of REPI from the AlphaFold Protein Structure Database and used it as the scaffold on which everything else was built, discarding the first 45 residues of the chain. The database model comes with per-residue confidence scores, called pLDDT, and a map of expected positional error between residues, called PAE; these were used to judge which stretches of the chain were modelled reliably. The researchers also ran AlphaFold2 themselves through ColabFold to generate five further models of the same sequence and compared those confidence profiles and shapes with the database entry.

In effect the prediction stood in for an experiment that has not been done: in the usual route, a structure of this kind would come from crystallography or a related laboratory technique. Everything downstream of the prediction is conventional, non-learned computation — structural searches and alignments to position the helper molecules, docking software for the ligand poses, force-field setup and the molecular dynamics runs themselves, trajectory analysis, geometric tunnel detection and an electrostatics calculation. The structural and dynamic conclusions therefore rest on the predicted scaffold. The researchers report no experimental test of their predictions and state that these are to be checked later in the laboratory.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.