Adding a substrate-aware objective to dEVA.
dEVA is not limited to metal-coordination objectives. Here we add a third, substrate-aware objective and show that the designs organize into a Pareto front — trading substrate enclosure against sequence likelihood and catalytic-metal probability. We provide this as a simple example of optimization of a traditionally generated enzyme design.
- p(seq) — how likely the sequence is, given the backbone and ligand (LigandMPNN)
- p(catalytic metal) — probability of a catalytic metal at the site (Metal3D-Cat)
- pocket shape — how well the pocket encloses a fixed substrate pose (geometric term)
The first two are the objectives from the paper. The third represents a
geometric score that is provided alongside the other dEVA objectives (../models/pocket_shape.py).
Here we use one of the less-active enzyme designs Kim et al. from design campaign 1 B1: a metalloenzyme with a buried active site and a bound substrate analog.
Here we run dEVA using three objectives, 5 individuals, 10 generations, where we fix the catalytic metal residues. To find the full yaml file, please check ../configs/substrate_example.yml.
python run.py --config configs/substrate_example.yml \
--models seq_model metal3d_model pocket_shapeThe Pareto front represents a real trade-off: front designs raise metal probability while giving up a little sequence likelihood and pocket enclosure, and one improves pocket shape past the input. This is what a multi-objective optimizer should do — satisfy the objectives jointly and make the trade-offs explicit.
Original B1: p(seq) 0.499 · p(metal) 0.927 · pocket 0.969
Pareto front (n = 5):
| p(seq) | p(cat metal) | |
|---|---|---|
| 0.481 | 0.954 | 0.920 |
| 0.480 | 0.951 | 0.929 |
| 0.471 | 0.954 | 0.944 |
| 0.487 | 0.943 | 0.957 |
| 0.460 | 0.944 | 0.981 |
Below we compare in cyan (the new dEVA design) and orange (the original B1). We note the difference in charge and pocket shape between the original design and dEVA design, while maintaining the active site.
