Built to be reversible: There has been some success in designing stable peptide filaments; however, mimicking the reversible assembly of many natural protein filaments is challenging. Dynamic filaments usually comprise independently folded and asymmetric proteins and using such building blocks requires the design of multiple intermonomer interfaces. Shen et al. report the design of self-assembling helical filaments based on previously designed stable repeat proteins. The filaments are micron scale, and their diameter can be tuned by varying the number of repeats in the monomer. Anchor and capping units, built from monomers that lack an interaction interface, can be used to control assembly and disassembly. Science , this issue p. 705

Enzymes find broad use as biocatalysts in industry and medicine owing to their exquisite selectivity, efficiency and mild reaction conditions. Custom-designed enzymes can produce tailor-made biocatalysts with potential applications that extend beyond natural reactions. However, current design methods require testing a large number of designs and mostly produce de novo enzymes with low catalytic activities ^1–3 . As a result, they require costly experimental optimization and high-throughput screening to be industrially viable ^4,5 . Here we present rotamer inverted fragment finder–diffusion (Riff-Diff), a hybrid machine learning and atomistic modelling strategy for scaffolding catalytic arrays in de novo proteins. We highlight the general applicability of Riff-Diff by designing enzymes for two mechanistically distinct chemical transformations, the retro-aldol reaction and the Morita–Baylis–Hillman reaction. We show that in both cases, it is possible to generate catalysts that exhibit activities rivalling those optimized by in vitro evolution, along with exquisite stereoselectivity. High-resolution structures of six of the designs revealed near-atomic active site design precision. The design strategy can, in principle, be applied to any catalytically competent amino acid array. These findings lay the basis for practical applicability of de novo protein catalysts in synthesis and describe fundamental principles of protein design and enzyme catalysis.

Berberine bridge enzyme-like (BBE-like) proteins form a multigene family (pfam 08031), which is present in plants, fungi and bacteria. They adopt the vanillyl alcohol-oxidase fold and predominantly show bi-covalent tethering of the FAD cofactor to a cysteine and histidine residue, respectively. The Arabidopsis thaliana genome was recently shown to contain genes coding for 28 BBE-like proteins, while featuring four distinct active site compositions. We determined the structure of a member of the AtBBE-like protein family (termed AtBBE-like 28), which has an active site composition that has not been structurally and biochemically characterized thus far. The most salient and distinguishing features of the active site found in AtBBE-like 28 are a mono-covalent linkage of a histidine to the 8α-position of the flavin-isoalloxazine ring and the lack of a second covalent linkage to the 6-position, owing to the replacement of a cysteine with a histidine. In addition, the structure reveals the interaction of a glutamic acid (Glu426) with an aspartic acid (Asp369) at the active site, which appear to share a proton. This arrangement leads to the delocalization of a negative charge at the active site that may be exploited for catalysis. The structure also indicates a shift of the position of the isoalloxazine ring in comparison to other members of the BBE-like family. The dioxygen surrogate chloride was found near the C(4a) position of the isoalloxazine ring in the oxygen pocket, pointing to a rapid reoxidation of reduced enzyme by dioxygen. A T-DNA insertional mutant line for AtBBE-like 28 results in a phenotype, that is characterized by reduced biomass and lower salt stress tolerance. Multiple sequence analysis showed that the active site composition found in AtBBE-like 28 is only present in the Brassicaceae, suggesting that it plays a specific role in the metabolism of this plant family.

Machine learning (ML) is rapidly turning into a key technology for biocatalysis. By learning patterns in amino acid sequences, protein structures, and functional data, ML models can help navigate complex fitness landscapes, uncover new enzymes in databases, and even design biocatalysts de novo. Along with advances in DNA synthesis and sequencing, laboratory automation, and high-throughput screening, ML is increasing the speed and efficiency of enzyme development. In this Outlook, we highlight recent applications of ML in the fields of enzyme discovery, design, and engineering, with a focus on current challenges and emerging solutions. Furthermore, we discuss barriers that impede a broader and faster adoption of ML-based workflows in the biocatalysis community. We conclude by suggesting best practices for fostering effective collaborations in this interdisciplinary field.