Adding a substrate-aware objective to dEVA during metalloenzyme design.
This example starts from a designed metalloenzyme with a buried active site and a bound substrate analog (inputs/B1.pdb). The metal site is already there. What we add is a geometric score that scores the occlusion of the ligand by the protein heavy atoms.
The first two scores are the sequence and catalytic-metal terms described in the manuscript. The third is an additional score called pocket shape to allow dEVA to trade-off enclosure against likelihood and metal probability instead of burying that trade-off in a weight.
Pocket shape is a prototype geometric filter, not a binding free energy. Calibrate
target_occon structures you trust.
| score | desired property |
|---|---|
| p(seq) | Does this sequence look like it belongs on this backbone and ligand? (LigandMPNN) |
| p(catalytic metal) | Is there still a catalytic metal at the site? (Metal3D-Cat) |
| pocket shape | Does the pocket enclose the fixed ligand pose, without clashing or sealing it? (geometry) |
Starting structure: inputs/B1.pdb. 10 generations, 5 individuals,
1 mutation per child. Full settings:
configs/substrate_example.yml.
python run.py -c configs/substrate_example.yml \
--models seq_model metal3d_model pocket_shapeThree steps, same shape as the physics example.
- Start from a posed site. The metal ligands and the substrate analog are already in the PDB. Ligand-only coordinates live in
inputs/B1_ligand.pdb(same frame as the complex). - Pick the scores. Anything that returns a number works as a score. Here we use: sequence likelihood, catalytic-metal probability, and pocket shape.
- Design with dEVA. Mutate, score, keep the non-dominated set.
Pink is a dEVA design, blue is the starting structure. The metal site and substrate pose stay put; the pocket side chains around them move.
The ligand pose is fixed. The score asks how many protein heavy atoms sit near each ligand atom (occlusion), then subtracts clashes. The score is then normalized to be between 0 and 1.
fitness = window(occlusion) − w_clash × overlap
target_occ is the occlusion you want (here 105, taken from the
starting complex). The window is not “more burial is better”: packing
the site completely would block solvent, so overshooting the target
is penalized. Higher is better; 1.0 means occlusion hit the target
and there is no clash.
| file | role |
|---|---|
configs/substrate_example.yml |
this example’s run |
models/pocket_shape.py |
ligand-pocket geometry |
models/metal3d_model.py |
catalytic-metal probability |
inputs/B1.pdb |
starting complex |
inputs/B1_ligand.pdb |
ligand pose (same frame) |