deep learning for enzyme design and protein biophysics
gelnesr.github.io
Gina El Nesr
Loading...
Zero-shot design of a de novo metalloenzyme https://www.biorxiv.org/content/10.64898/2026.04.23.720277v1
Another day, another paper describing an ML method for generating conformational ensembles without comparison to any experimental data 😩
6/ First, dEVA proposes mutations and sidechain orientations (LigandMPNN). Then it predicts the probability and location of catalytic metal ion(s) (Metal3D-Cat)
Mutate, predict, and repeat until at the Pareto front; then chose the optimal design at the knee point. No filtering necessary!!
4/ desB was designed zero-shot. No structure prediction, no pre-defined motif, no reaction-intermediates.
Our design strategy? dEVA!
2/ Remarkably, desB also hydrolyzes phosphodiesters. A harder reaction class: different transition state, different charge, uncatalyzed half-lives >13 million years.
Rate enhancements up to 10^12 and efficiencies ALSO comparable to many natural phosphatases.
3/ desB has a binuclear zinc site, like many of nature’s most efficient hydrolases. No sequence homologs, no structural neighbors. XAS structurally confirms the binuclear active site, distances, and phosphate binding
...also consistent with alanine mutants, pH profiling, EDTA-cleavage, + inhibition
1/ Phosphomonoester hydrolysis has uncatalyzed half-lives >500,000 years.
desB hydrolyzes model substrates with kcat/KM up to1500 /Ms and rate enhancements (RE) up to 10^13. The highest reported RE of any de novo designed hydrolase & catalytic efficiencies comparable to natural phosphatases.
