Rebecca Buller: Biocatalysis is an approach to synthetic chemistry in which enzymes carry out chemical reactions. The process of creating an efficient biocatalyst involves the identification of a suitable enzyme with some degree of initiating activity required for the desired transformation and the optimization of the initiating scaffold, usually by directed evolution to tune the enzyme properties.
Recently, the number of available protein sequences has increased by an astonishing 20 times (2023: > 2.4 billion)5,6; 2018: ~123 million7 This is thanks to technology that enables high-quality, low-cost DNA sequencing. Using machine learning, these sequences can be functionally annotated, accelerating the discovery of enzymes with useful activities. In searching for suitable starting points, ML models can also contribute to filtering the natural diversity of protein catalysts with respect to properties such as stability and solubility. Interestingly, ML can also be applied to generate entirely new protein sequences with desired functions.
Once a protein’s starting scaffold is identified, ML models can be applied to navigate the protein’s fitness landscape. ML can help prioritize which set of mutations to test in enzyme engineering campaigns by training models on experimental data. This approach helps analyze complex relationships within large datasets and identify patterns that are difficult to detect using other methods. This is important because most enzyme engineering campaigns allow only a small portion of a protein sequence to be sampled experimentally. Furthermore, experimental engineering campaigns tend to focus on single mutation steps, which means ignoring non-additive effects of mutation accumulation. However, ML-assisted directed evolution can be used to predict the fitness of protein variants with several amino acid substitutions. Using this approach, for example, halogenases can be optimized for late-stage functionalization of the macrolide Sorafen A.2 and ketoreductase3This was designed by us to produce a precursor for the anti-cancer drug ipatasertib.
Stanislav Mazurenko: One readily available application is dealing with the large number of unannotated sequences in biological databases. Although a wide variety of biocatalysts already exist in nature, we have only scratched the surface with existing, well-characterized enzymes. Recent breakthroughs in protein structure prediction, such as AlphaFold, have made it possible to access a vast “structural universe”.8 The next big step is to accumulate enough annotated enzyme data to unravel the “functional universe.” ML should be able to provide tools that can predict enzyme activity, substrate range, cofactors, optimal environment, etc. with high accuracy. We are already increasing the use of ML-based annotations in tools such as EnzymeMiner for automated mining of soluble enzymes.9.
Another area where ML can have an impact is in the teaching of protein engineering. The complexity of this task is immense, as even a single mutation can completely compromise a protein. This is often seen in our experiments and in the data collected in the mutational effects database FireProt.D.B.10 and SoluProtMutD.B.11. Mutations that affect activity often occur far from the active site and act through allostery. Therefore, the search space for protein engineering must span the entire protein, but its combinatorial complexity typically limits rational engineering to only a handful of hotspots. Today, we already have a suite of ML-based tools to navigate the entire search space, from zero-shot predictions generated by basic protein language models to task-specific predictors and small-scale models purposefully fine-tuned to single-enzyme mutation datasets.
Second, a considerable number of exciting studies on de novo enzyme design have been published, e.g., high-affinity protein binders using diffusion models have been proposed.12 or optimized protein sequences using inverse folding methods13. Such models can also be conditioned on the specific geometry of the active site, and attempts have been made to use this conditioning to generate new enzymes. This question remains unresolved, as there are many factors other than this geometry that define enzyme activity. However, the field is progressing so rapidly that it will be interesting to see whether the future of enzyme engineering will be dominated by the discovery and optimization of existing proteins or by designing entirely new proteins from scratch. In ML, on the other hand, the boundaries between these two approaches are blurring, such as in the case of de novo designs leveraging ML models trained on natural sequences.
Finally, AI is increasingly being used in laboratories at various levels, including hardware control, signal acquisition and processing, data analysis, and design, build, test, and learn cycles. I believe we will see even more exciting AI applications in laboratory automation as they free scientists from repetitive manual tasks and help optimize experimental conditions.
Yang Yang: Due to the vast protein sequence and structure databases and the similarities between natural language processing and protein sequence/structure processing, machine learning tools, especially large language models (LLMs), have the potential to accelerate protein engineering and biocatalysis. By identifying the hidden rules of the protein fitness landscape, Protein LLM can help navigate the enzyme fitness landscape by focusing on high-fitness regions and avoiding local minimum traps. Generative machine learning (ML) models have the potential to create new enzyme sequences with high success rates. Moreover, the use of synthetic enzyme design methods can lead to entirely new enzyme functions.
