Papex

Open problem · Chemistry · Open

De novo enzyme design

Can highly active enzymes be designed from scratch for any desired reaction?

Summary

Computational design has produced new enzymes for a few reactions, such as retro-aldol, Kemp elimination and Diels–Alder reactions. Designing highly active catalysts on demand for any chosen reaction has not been achieved.

Why it is open

Designed enzymes exist, and some are fast. What is missing is reliability and reach. In a 2022 review, Sarah Lovelock, David Baker, Donald Hilvert, Anthony Green and colleagues write that "the activities of designed catalysts have been quite low" and that "extensive laboratory evolution has been used to generate efficient enzymes" from them [7]. In 2026 Dina Listov and Sarel Fleishman write that until recent years design methods "yielded enzymes with low catalytic efficiencies even for simple reactions", and that the strategies which emerged since have produced catalysts "with efficiencies approaching natural enzymes", while "achieving high catalytic rates and complex, multistep mechanisms remains challenging" [13].

The problem is several problems: building a protein that folds as designed, placing the catalytic groups precisely enough, reaching natural rates, and doing all of this for a mechanism with more than one step [11, 13]. Judith Ahr, Adrian Bunzel and colleagues add in 2026 that catalytic efficiencies (kcat/Km) can be notable while turnover numbers (kcat) "often remain limited, suggesting that current approaches typically yield tight substrate binders rather than enzymes that selectively stabilize the chemical transition state" [14].

Recent design work relies heavily on machine-learning tools [14], and the published results concern particular designed enzymes for stated substrates. The reviews that assess the field say that reliably creating highly efficient enzymes "remains out of reach" [14].

No body sets a bar for a resolution. The sources point to two tests. Markus Braun and colleagues write in 2026 that "the current paradigm for generating efficient biocatalysts is to compensate for low initial efficiencies with high-throughput screening and directed evolution", which is "not suited to quickly create novel, proficient catalysts for chemical reactions that are not found in nature or are exceedingly difficult to access through high-throughput screening", so that a one-shot design capability is needed [12]. The second test is generality: a method that works for a reaction it was not built around [13, 14].

What is known

The first designed enzymes (2008 to 2010). In 2008 Daniela Röthlisberger, Olga Khersonsky and colleagues reported the computational design of eight enzymes for the Kemp elimination, a model reaction for proton transfer from carbon, with measured rate enhancements of up to 10⁵ and multiple turnovers; in vitro evolution of the designs then gave a more than 200-fold increase in catalytic efficiency, to kcat/Km of 2,600 M⁻¹ s⁻¹ [1]. In the same year Lin Jiang, Eric Althoff and colleagues designed retro-aldolases, which break a carbon-carbon bond; of 72 designs characterised experimentally, 32 had detectable activity, with "rate accelerations of up to four orders of magnitude and multiple turnovers", and crystal structures of active designs were nearly superimposable on the design models [2]. In 2010 Justin Siegel and colleagues designed enzymes for a bimolecular Diels-Alder reaction, for which no naturally occurring enzyme had been shown to catalyse the bimolecular case, with high stereoselectivity and substrate specificity confirmed by crystallography [3].

What directed evolution did to the designs (2013 to 2017). Laboratory evolution raised these activities by large factors, but it also changed the designed mechanism. Lars Giger and colleagues boosted the specific activity of a designed retro-aldolase more than 4,400-fold by random mutagenesis and screening, reaching "catalytic efficiencies approaching those of natural enzymes", and found that the designed catalytic apparatus "was unexpectedly abandoned in favor of a new lysine residue in a substrate-binding pocket created during the optimization process" [4]. Rebecca Blomberg, Hilvert and colleagues evolved a designed Kemp eliminase into an enzyme that accelerates the reaction 6 × 10⁸-fold, "approaching the exceptional efficiency of highly optimized natural enzymes such as triosephosphate isomerase", and a 1.09 Å crystal structure showed shape complementarity and precisely placed catalytic groups [5]. Richard Obexer and colleagues improved an artificial aldolase by a further factor of 30, to a rate enhancement above 10⁹, and traced the gain to a lysine, tyrosine, asparagine and tyrosine tetrad that "emerged adjacent to a computationally designed hydrophobic pocket during directed evolution" [6].

Deep-learning methods (2022 to 2023). Andy Yeh, Christoffer Norn and colleagues used a "family-wide hallucination" approach to generate idealised scaffolds with varied pockets and designed artificial luciferases for two synthetic substrates; the most active was a 13.9 kDa, thermostable enzyme with a catalytic efficiency on diphenylterazine of kcat/Km = 10⁶ M⁻¹ s⁻¹, comparable, they report, to native luciferases but with much higher substrate specificity [8]. Joseph Watson, David Juergens and colleagues then reported RFdiffusion, a generative model of protein backbones whose design tasks include enzyme active-site scaffolding [9]. Listov, Casper Goverde, Bruno Correia and Fleishman wrote in 2024 that de novo design was "still limited mostly to α-helix bundles" [10].

Recent claims (2025 and 2026). In April 2025 Anna Lauko, Samuel Pellock, Baker and colleagues published designed serine hydrolases, built by combining RFdiffusion with a method that models each state along the reaction, reporting catalytic efficiencies up to 2.2 × 10⁵ M⁻¹ s⁻¹, crystal structures within 1 Å of the design models, and catalysts in five folds not previously seen in this enzyme family [11]. Earlier attempts of theirs to place the same active site in a different fold gave activated serines but no catalytic turnover on ester substrates [11]. In Nature in 2026 Braun, Adrian Tripp and colleagues reported Riff-Diff, which scaffolds a catalytic array in a new protein; their retro-aldolases RAD35 and RAD29 gave kcat values of 3.6 × 10⁻² s⁻¹ and 3.1 × 10⁻² s⁻¹, "a roughly 5 × 10⁶-fold rate acceleration over the uncatalysed reaction", with 895 and 1,000 turnovers and, for RAD35, 99% enantiomeric excess; for the Morita-Baylis-Hillman reaction, where they report "no known natural enzymes that catalyse this transformation efficiently", their design MBH48 had a higher catalytic constant than a variant that had been through eight rounds of directed evolution [12].

Main approaches

Placing a theozyme in an existing scaffold. A minimal active site is computed from a model of the transition state and introduced into a natural protein that can hold that geometry [12]. Strongest evidence: this produced the first designed enzymes for several reactions, with crystal structures close to the design models [1, 2, 3]. Main difficulty: activities were low, and of the interactions programmed into the retro-aldolases only the buried catalytic lysine contributed significantly to catalysis, while inactive designs showed unproductive conformations [12].

Directed evolution of a design. Random mutagenesis and screening turn a weak design into an efficient enzyme [7]. Strongest evidence: rate accelerations of 6 × 10⁸ and above 10⁹ were reached this way [5, 6]. Main difficulty: evolution can discard the designed machinery, so the result does not validate the design [4]; it needs a high-throughput screen, which many target reactions do not have [12].

Generative deep learning with reaction-path filters. A network generates backbones around an active-site description, and designs are filtered by predicted compatibility with each step of the reaction [11]. Strongest evidence: serine hydrolases in new folds with efficiencies up to 2.2 × 10⁵ M⁻¹ s⁻¹ [11], and designed luciferases near native efficiency for their substrate [8]. Main difficulty: turnover numbers stay low in many designs, which points to binding rather than transition-state stabilisation [14], and multistep mechanisms remain the hard case [11, 13].

Machine learning combined with atomistic modelling. Catalytic arrays are scaffolded by a hybrid pipeline rather than by a network alone [12]. Strongest evidence: one-shot designs with activities that rival evolved variants, and six crystal structures with near-atomic active-site precision [12]. Main difficulty: both arrays used were taken from enzymes already optimised by evolution, and the authors note that "high activity cannot be inferred from precise reproduction of the apo-state alone" [12].

Open sub-questions

  1. Turnover against binding. For designs reported with high kcat/Km but low kcat, which active-site changes raise kcat without improving substrate affinity, as the transition-state account predicts [11, 14]?
  2. A reaction the method was not built around. Does a catalytic array taken from no existing enzyme, for a transformation with no natural catalyst, yield active designs at the rates reported for the retro-aldol and Morita-Baylis-Hillman arrays [12]?
  3. Predicting activity before synthesis. Which computed quantity ranks designs by measured rate better than apo-state side-chain deviation, which did not predict activity in the 2026 designs [12]?
  4. Multistep mechanisms. For a design whose mechanism has a covalent intermediate, does modelling every state of the reaction raise the fraction of active designs as it did for serine hydrolases [11]?
  5. Evolution as a test of design. When a 2025 or 2026 design is put through directed evolution, is the designed catalytic machinery kept, unlike the designed retro-aldolase that was remodelled [4, 6]?
  6. Designed catalysts in synthesis. Can a de novo enzyme perform a selective step on a complex substrate well enough to replace a protection and deprotection sequence, the question asked on the Papex page Artificial enzymes in place of protecting groups [12]?

Work on this in Papex

If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.

Key references

  1. Röthlisberger, D., Khersonsky, O., Wollacott, A. M., Jiang, L., DeChancie, J., Betker, J. et al. (2008). Kemp elimination catalysts by computational enzyme design. Nature 453(7192), 190–195. 10.1038/nature06879 · Checked 16 Oct 2026
  2. Jiang, L., Althoff, E. A., Clemente, F. R., Doyle, L., Röthlisberger, D., Zanghellini, A. et al. (2008). De novo computational design of retro-aldol enzymes. Science 319(5868), 1387–1391. 10.1126/science.1152692 · Checked 16 Oct 2026
  3. Siegel, J. B., Zanghellini, A., Lovick, H. M., Kiss, G., Lambert, A. R., St Clair, J. L. et al. (2010). Computational design of an enzyme catalyst for a stereoselective bimolecular Diels-Alder reaction. Science 329(5989), 309–313. 10.1126/science.1190239 · Checked 16 Oct 2026
  4. Giger, L., Caner, S., Obexer, R., Kast, P., Baker, D., Ban, N. & Hilvert, D. (2013). Evolution of a designed retro-aldolase leads to complete active site remodeling. Nature Chemical Biology 9(8), 494–498. 10.1038/nchembio.1276 · Checked 16 Oct 2026
  5. Blomberg, R., Kries, H., Pinkas, D. M., Mittl, P. R. E., Grütter, M. G., Privett, H. K., Mayo, S. L. & Hilvert, D. (2013). Precision is essential for efficient catalysis in an evolved Kemp eliminase. Nature 503(7476), 418–421. 10.1038/nature12623 · Checked 16 Oct 2026
  6. Obexer, R., Godina, A., Garrabou, X., Mittl, P. R. E., Baker, D., Griffiths, A. D. & Hilvert, D. (2017). Emergence of a catalytic tetrad during evolution of a highly active artificial aldolase. Nature Chemistry 9(1), 50–56 (published online 2016). 10.1038/nchem.2596 · Checked 16 Oct 2026
  7. Lovelock, S. L., Crawshaw, R., Basler, S., Levy, C., Baker, D., Hilvert, D. & Green, A. P. (2022). The road to fully programmable protein catalysis. Nature 606(7912), 49–58. 10.1038/s41586-022-04456-z · Checked 16 Oct 2026
  8. Yeh, A. H.-W., Norn, C., Kipnis, Y., Tischer, D., Pellock, S. J., Evans, D. et al. (2023). De novo design of luciferases using deep learning. Nature 614(7949), 774–780. 10.1038/s41586-023-05696-3 · Checked 16 Oct 2026
  9. Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E. et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature 620(7976), 1089–1100. 10.1038/s41586-023-06415-8 · Checked 16 Oct 2026
  10. Listov, D., Goverde, C. A., Correia, B. E. & Fleishman, S. J. (2024). Opportunities and challenges in design and optimization of protein function. Nature Reviews Molecular Cell Biology 25(8), 639–653. 10.1038/s41580-024-00718-y · Checked 16 Oct 2026
  11. Lauko, A., Pellock, S. J., Sumida, K. H., Anishchenko, I., Juergens, D., Ahern, W. et al. (2025). Computational design of serine hydrolases. Science 388(6744), eadu2454. 10.1126/science.adu2454 · Checked 16 Oct 2026
  12. Braun, M., Tripp, A., Chakatok, M., Kaltenbrunner, S., Fischer, C., Stoll, D. et al. (2026). Computational enzyme design by catalytic motif scaffolding. Nature 649(8095), 237–245 (published online 3 December 2025). 10.1038/s41586-025-09747-9 · Checked 16 Oct 2026
  13. Listov, D. & Fleishman, S. J. (2026). De novo enzyme design: controlling structure to design function. Current Opinion in Structural Biology 98, 103252. 10.1016/j.sbi.2026.103252 · Checked 16 Oct 2026
  14. Ahr, J. A., de Waal, M. M., Cotet, T.-S., Howell, D. C. & Bunzel, H. A. (2026). What deep learning can and cannot (yet) do for enzyme design. Current Opinion in Chemical Biology 95, 102770 (issue dated December 2026; available online and indexed on 2026-10-16). 10.1016/j.cbpa.2026.102770 · Checked 16 Oct 2026

Page reviewed 16 Oct 2026

Source

Open, as listed by Wikipedia · Unsolved problems in chemistry. Checked 08 Oct 2026.

The status is the source’s. Papex does not decide whether a problem is solved.

Records on this problem

No public record addresses this problem yet.

More in Chemistry