Distilling structural representations into protein sequence models DOI Creative Commons

Jeffrey Ouyang-Zhang,

Chengyue Gong, Yue Zhao

et al.

Published: Nov. 11, 2024

Abstract Protein language models, like the popular ESM2, are widely used tools for extracting evolution-based protein representations and have achieved significant success on downstream biological tasks. Representations based sequence structure however, show performance differences depending task. A major open problem is to obtain that best capture both evolutionary structural properties of proteins in general. Here we introduce I mplicit S tructure M odel ( ISM ), a sequence-only input model with structurally-enriched outperforms state-of-the-art models several well-studied benchmarks including mutation stability assessment prediction. Our key innovations microenvironment-based autoencoder generating tokens self-supervised training objective distills these into ESM2’s pre-trained model. We made ’s structure-enriched weights easily available: integrating any application using ESM2 requires changing only single line code. code available at https://github.com/jozhang97/ISM .

Language: Английский

Engineering a Photoenzyme to Use Red Light DOI Creative Commons
J. M. Carceller, Bhumika Jayee, Claire G. Page

et al.

Published: March 18, 2024

Photoenzymatic catalysis is an emerging platform for asymmetric synthesis. In most of these reactions, the protein templates a charge transfer complex between cofactor and substrate, which absorbs in blue region electromagnetic spectrum. Here, we report engineering photoenzymatic ‘ene’-reductase to utilize red light (620 nm) radical cyclization reaction. Mechanistic studies indicate that ac-tivity achieved by introducing broadly absorbing shoulder off previously identified cyan absorption feature. Molecular dynamics simulations, docking, excited-state calculations suggest 𝜋→ 𝜋* transition from flavin while feature red-shift transition, occurs upon substrate binding. Differences excitation event help disfavor alkylation cofactor, pathway catalyst decomposition observed with but not red.

Language: Английский

Citations

5
Spectral control of photoenzyme through allosteric chromophore tuning DOI
Menghan Xu, Yuki Yamanashi, Motomu Kanai

et al.

Chem, Journal Year: 2025, Volume and Issue: unknown, P. 102430 - 102430

Published: Jan. 1, 2025

Language: Английский

Citations

0
Mechanistic investigation of repurposed photoenzymes with new-to-nature reactivity DOI
Zhengyi Zhang, Mao‐Lin Li, Huimin Zhao

et al.

Current Opinion in Green and Sustainable Chemistry, Journal Year: 2025, Volume and Issue: 52, P. 101009 - 101009

Published: Feb. 27, 2025

Language: Английский

Citations

0
Unique Electron Donor–Acceptor Complex Conformation Ensures Both the Efficiency and Enantioselectivity of Photoinduced Radical Cyclization in a Non-natural Photoenzyme DOI
Matteo Capone, Gianluca Dell’Orletta, Claire G. Page

et al.

ACS Catalysis, Journal Year: 2024, Volume and Issue: 14(21), P. 16488 - 16496

Published: Oct. 23, 2024

Non-natural photoenzymatic catalysis exploits active site tunability for stereoselective radical reactions. In flavoproteins, light absorption promotes the excitation of an electron donor–acceptor (EDA) complex formed between reduced flavin cofactor and a substrate (α-chloroacetamide in this case). This can trigger chloride mesolytic cleavage, leading to cyclization (forming γ-lactam), or revert ground state. While strategy is feasible using broad UV/visible/near-infrared spectrum, low quantum yield presents significant challenge. Using multiscale computational approach, we elucidate mechanisms light-driven initiation step catalyzed by Gluconobacter oxydans "ene"-reductase mutant (GluER-G6). The experimental stems from limited population (<10%) EDA complexes with charge transfer state competent cleavage. Accessibility requires bending positioning chlorine atom near styrenic group. A subset these reactive conformers exhibits enhanced cyan/red due optimal C–Cl bond alignment flavin. Engineering GluER variant stabilize conformation expected significantly enhance catalytic efficiency when light. identified intermediates possess correct prochirality enantioselective cyclization. Our findings show that ground-state conformational selection governs both light-activated cleavage enantioselectivity.

Language: Английский

Citations

1
Distilling structural representations into protein sequence models DOI Creative Commons

Jeffrey Ouyang-Zhang,

Chengyue Gong, Yue Zhao

et al.

Published: Nov. 11, 2024

Abstract Protein language models, like the popular ESM2, are widely used tools for extracting evolution-based protein representations and have achieved significant success on downstream biological tasks. Representations based sequence structure however, show performance differences depending task. A major open problem is to obtain that best capture both evolutionary structural properties of proteins in general. Here we introduce I mplicit S tructure M odel ( ISM ), a sequence-only input model with structurally-enriched outperforms state-of-the-art models several well-studied benchmarks including mutation stability assessment prediction. Our key innovations microenvironment-based autoencoder generating tokens self-supervised training objective distills these into ESM2’s pre-trained model. We made ’s structure-enriched weights easily available: integrating any application using ESM2 requires changing only single line code. code available at https://github.com/jozhang97/ISM .

Language: Английский

Citations

1