TL;DR: A Caltech–Google DeepMind–Pittsburgh team reported on Oct. 2, 2026 a design-evolution combination that builds de novo haemproteins for selective C–H amination to make chiral piperidines.
New preprint pairs protein design with evolution
A collaboration spanning the California Institute of Technology in Pasadena, Google DeepMind in London, and the University of Pittsburgh reported a synergistic design-evolution strategy for new-to-nature enzyme chemistry. The report appeared as a preprint on bioRxiv, posted October 02, 2026, under the title Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry. According to that bioRxiv abstract, the work uses engineered de novo haemproteins to synthesize chiral piperidines through regio- and enantioselective C–H amination.
The bioRxiv listing identifies Zi-Qi Li of the Division of Chemistry and Chemical Engineering at the California Institute of Technology as a lead contributor, alongside Ruijie K. Zhang of Google DeepMind and Pei-Pei Xie of the Department of Chemistry at the University of Pittsburgh. Additional Caltech contributors named on the page include Ziyang Qin, Théophile Lambert, Orna Mukhopadhyay and Deirdre Hanley. Google DeepMind contributors named include Jacob Kelly, Amy E. Danson, Josh Abramson, Sukhdeep Singh, Zachary Wu, Pushmeet Kohli and Jue Wang. Pittsburgh contributors named include Yu Zhang and Peng Liu. Frances H. Arnold of Caltech is identified as corresponding author.
The same bioRxiv page states that the authors declared no competing interest. It lists funder information as the United States Department of Energy, identifier DE-SC0021141. It classifies the posting under Enzyme Catalysis and Immobilization, Protein Structure and Dynamics, and Asymmetric Synthesis and Catalysis.
The central claim, as worded in paraphrase of the abstract, is that motif scaffolding plus computational re-design plus directed evolution can deliver demanding chemistry. According to Design-Evolution Synergy Enables New-to-Nature Enzyme Chemistry on bioRxiv, the integration demonstrates that de novo enzyme design can support high-energy-barrier transformations with programmable regio- and enantioselectivity, high catalytic efficiency, excellent stability and scaffold diversity.
What the team says it built
The abstract describes the products as chiral piperidines. It calls chiral piperidines a privileged pharmacophore in drug discovery. The catalysts are described as engineered de novo haemproteins. The reactions are described as new-to-nature regio- and enantioselective C–H amination reactions.
Regioselectivity, in the language of the report, means control over where amination occurs. Enantioselectivity means control over which mirror-image form is produced. Programmable selectivity means the design-evolution route can adjust those outcomes rather than accepting a single fixed outcome. The abstract presents those adjustable outcomes together with catalytic efficiency, enzyme stability and scaffold diversity as results of the combined approach.
The method section summarized in the abstract names three components. The first is motif scaffolding. The second is computational re-design. The third is directed evolution. The report frames the three as an integrated workflow, not as separate alternatives.
The broader framing in the abstract contrasts two histories. One history is directed evolution and machine-learning-assisted evolution, described as having transformed biocatalysis by enabling rapid optimisation of catalytic performance within compatible protein scaffolds. The other history is deep-learning-driven de novo protein design, described as having expanded accessible protein architectural space and enabled functional catalysts for increasingly challenging transformations. The new report positions its synergy at the intersection of those histories.
Why C–H amination is framed as challenging
The bioRxiv abstract calls the chemistry challenging, high-energy-barrier enzyme chemistry. That phrasing signals that direct conversion of a carbon–hydrogen bond into a carbon–nitrogen bond inside a designed active site requires precise placement of reactive intermediates. The report does not present the transformation as an extension of a natural metabolic pathway. It presents it as new-to-nature chemistry encoded in a designed haemoprotein.
Context for that difficulty appears in earlier literature cited by the preprint. The reference list includes work on enzyme-controlled nitrogen-atom transfer for regiodivergent C–H amination published in the Journal of the American Chemical Society in 2014 by Hyster, Farwell, Buller, McIntosh and Arnold. It also includes work on enantioselective intramolecular C–H amination catalyzed by engineered cytochrome P450 enzymes in vitro and in vivo, reported by McIntosh and colleagues in Angewandte Chemie International Edition in 2013. A further cited entry describes intramolecular C(sp3)–H amination of arylsulfonyl azides with engineered and artificial myoglobin-based catalysts, reported by Bordeaux, Singh and Fasan in Bioorganic & Medicinal Chemistry in 2014.
The reference list also points to transition-metal-catalyzed C–H amination literature. It includes a Chemical Reviews article from 2017 by Park, Kim and Chang on transition metal-catalyzed C–H amination scope, mechanism and applications. It includes work on iron-catalyzed direct C–H bond amination for complex N-heterocycle synthesis, reported by Hennessy and Betley in Science in 2013. It includes work on catalytic synthesis of N-heterocycles via direct C(sp3)–H amination using an air-stable iron(III) species with a redox-active ligand, reported by Bagh and colleagues in the Journal of the American Chemical Society in 2017.
The design lineage is also documented through cited computational design papers. Examples listed include de novo computational design of retro-aldol enzymes reported by Jiang and colleagues in Science in 2008, Kemp elimination catalysts by computational enzyme design reported by Röthlisberger and colleagues in Nature in 2008, and computational design of enone-binding proteins with catalytic activity for the Morita–Baylis–Hillman reaction reported by Bjelic and colleagues in ACS Chemical Biology in 2013. An iterative approach to computational enzyme design reported by Privett and colleagues in Proceedings of the National Academy of Sciences in 2012 is likewise listed. A 2016 Nature review by Huang, Boyken and Baker on the coming of age of de novo protein design is included as well.
A second 2026 example of design plus screening
Independent work posted in the same week illustrates the wider move toward combining computation with focused screening. According to Iterative Multimutant Design Creates a Family-Like Enzyme Ensemble, a ChemRxiv preprint published Oct. 1, 2026, researchers at Bielefeld University started from weak azole N-methylation by human nicotinamide N-methyltransferase and screened only 1,600 clones over three rounds.
That ChemRxiv abstract states that the generated catalysts carried up to 11 active-site substitutions. It reports mean initial-rate gains equivalent to approximately 11-fold per round, with overall gains exceeding 3,000-fold for individual enzyme-substrate pairs. It reports regioselectivity exceeding 99:1 and preparative methylation affording 87% isolated yield. It further states that all thermally characterized variants were more thermostable than the wild type.
The Bielefeld authors listed on that record are Lucas Bocquin, Om Desai, Kateryna Zelenska, Matthias Garbrecht, Tom Hoenig, Alina Klein, Friederike Reese, Kai H. Schülke and Stephan Christian Hammer. The abstract frames the outcome as transformation of one promiscuous enzyme into a family-like ensemble of complementary catalysts. It states that broad activity across diverse azoles emerged collectively from sequence-diverse specialist enzymes. It also states that mapping all mutational routes between two high-activity variants showed that each passed through a less-active intermediate, revealing why stepwise local optimization can miss high-performing variants.
The parallel is conceptual rather than chemical. The Bielefeld work addresses azole N-methylation, while the Caltech–DeepMind–Pittsburgh work addresses C–H amination to chiral piperidines. Both emphasize exploring sequence space through energetically compatible multimutant designs rather than optimizing one enzyme for one substrate at a time.
Data-efficient evolution for C–H functionalization
Another preprint from Oct. 1, 2026 adds context on extracting more information from evolution campaigns. According to Extracting Latent Sequence-Function Information from Directed Evolution to Design Improved C-H Azidases, researchers developed a strategy to recover latent information by grouping screened variants into activity-defined bins, pooling them and applying long-read sequencing.
That ChemRxiv record states that conventional workflows use only a fraction of information generated during screening, with only top-performing variants sequenced and propagated while sequence-function information from the remainder is discarded. The proposed alternative trains ordinal machine-learning models on binned data to predict beneficial mutations.
The demonstration system named in that abstract is SadX, described as an engineered Fe(II)- and alpha-ketoglutarate-dependent azidase previously evolved for site-selective C–H azidation. The abstract states that experimental validation showed several predicted single mutations enhanced azidation activity by up to 3-fold relative to the parent enzyme. Authors listed include Ayantika Ghosh of Indiana University, Sameer D’Costa of the University of Wisconsin–Madison, Christian A. Gomez, Bryce Johnson, Prajwal Prabhu, Anthony Gitter, Philip A. Romero and Jared C. Lewis. Funding is listed from the National Science Foundation and the National Institutes of Health, including a collaborative program on integrating deep learning and high-throughput experimentation to navigate protein fitness landscapes for non-native catalysis.
Together, the two Oct. 1 ChemRxiv preprints show complementary tactics around the Oct. 2 haemoprotein report. One tests compatible multi-mutation designs in small screens. The other reuses discarded screening data to guide predictions. The haemoprotein report combines motif scaffolding, re-design and directed evolution in a single pipeline.
How the pieces are described to fit together
Motif scaffolding, as named in the haemoprotein abstract, refers to building a new protein architecture around a desired active-site geometry. Computational re-design refers to adjusting sequence and structure in silico for compatibility and function. Directed evolution refers to iterative mutation and selection for improved catalytic performance. The bioRxiv abstract presents rapid optimisation within compatible scaffolds as the established strength of directed evolution and machine-learning-assisted evolution, and expanded architectural space as the established strength of deep-learning-driven de novo design.
By starting from a designed haemoprotein rather than a repurposed natural scaffold, the team seeks scaffold diversity from the outset. By applying evolution afterward, the team seeks catalytic efficiency and selectivity that computation alone does not yet reliably deliver. The reported outcome categories track that division: scaffold diversity and stability reflect design choices, while programmable regio- and enantioselectivity and high catalytic efficiency reflect evolutionary refinement of a high-energy-barrier reaction.
The perspective literature from the same period describes a similar trajectory for the field. A Molecular Cell perspective published Oct. 1, 2026, titled Programming biological behavior with AI-guided de novo protein design, states that AI-driven de novo protein design can now generate diverse molecular functions and has achieved success in engineering binders against many macromolecular targets. Authors Philipp Huettemann and Tanja Kortemme of the University of California, San Francisco frame the next step as moving beyond bimolecular interactions toward bottom-up construction of protein reaction networks, with emphasis on programmable, modular and controllable components. They list current challenges including designable protein dynamics, multi-step functionalities and non-equilibrium systems.
Drug-discovery relevance without overstating
The haemoprotein abstract connects chiral piperidines to drug discovery by calling the ring system a privileged pharmacophore. It does not name a specific drug candidate, indication, dose or commercial process. It does not report manufacturing scale, isolated yield, turnover number or clinical testing. Those details are absent from the abstract record available through the cutoff, so they are not stated here as results.
The relevance language in surrounding 2026 literature is similarly scoped to building blocks and synthetic routes. A ChemCatChem paper published Oct. 1, 2026 on imine reductase-mediated regio- and stereoselective reductive amination of diamines describes N-substituted diamines as common motifs in pharmaceutical building blocks and APIs, and reports synthesis of a precursor to the antihistamine drug tripelennamine with space-time yields up to 36 g l−1 d−1. A Chemical Record review published Sept. 29, 2026 on copper-mediated C–H amination describes copper-based reactions as cost-effective, less toxic, earth-abundant, biocompatible and environmentally benign while minimizing undesired byproducts, and discusses construction of nitrogen frameworks for drug discovery. A Current Organic Synthesis review covering 2014–2026 describes the piperazine scaffold as central to medicinal chemistry and serving as the foundation of numerous FDA-approved therapeutics, while discussing catalytic and green synthetic strategies. Those examples show why selective nitrogen chemistry attracts pharmaceutical attention, but they are separate systems from the de novo haemoprotein amination report.
What remains preprint-stage
The Design-Evolution Synergy report is identified on bioRxiv as a preprint with zero citations at the time of the listing. Preprint status means the findings had been posted for community review but are presented here only as reported, not as peer-reviewed conclusions. The funder and affiliation details confirm United States participation through Caltech, Pittsburgh and the Department of Energy award, alongside United Kingdom participation through Google DeepMind.
For readers tracking United States science and technology, the verifiable news through Oct. 3, 2026 is the posting itself and its stated approach: motif scaffolding, computational re-design and directed evolution applied to de novo haemoproteins for regio- and enantioselective C–H amination to chiral piperidines, with claimed advantages in selectivity programmability, catalytic efficiency, stability and scaffold diversity. Further assessment will depend on release of full experimental details, peer review and independent reproduction, none of which are recorded in the abstract-level sources available by the cutoff.