Generative AI for Enzyme Design and Biocatalysis

TL;DR

Generative AI offers a new paradigm for enzyme design by learning from diverse protein sequences and structures to design or optimize enzymes for industrial biocatalysis. The paper classifies models by modality (sequence- and backbone-generating) and discusses experiments validating designs across substitutions, family expansion, and de novo scaffolds, highlighting ProteinMPNN, RFdiffusion, and related approaches. It emphasizes that experimental feedback loops, steering, and integration of physico-chemical priors are essential to align designs with real-world performance, while recognizing current end-to-end automation for new-to-nature activities is not yet achieved. The work argues that mature generative AI models can accelerate biocatalyst development when embedded in iterative, data-informed workflows.

Abstract

Sparked by innovations in generative artificial intelligence (AI), the field of protein design has undergone a paradigm shift with an explosion of new models for optimizing existing enzymes or creating them from scratch. After more than one decade of low success rates for computationally designed enzymes, generative AI models are now frequently used for designing proficient enzymes. Here, we provide a comprehensive overview and classification of generative AI models for enzyme design, highlighting models with experimental validation relevant to real-world settings and outlining their respective limitations. We argue that generative AI models now have the maturity to create and optimize enzymes for industrial applications. Wider adoption of generative AI models with experimental feedback loops can speed up the development of biocatalysts and serve as a community assessment to inform the next generation of models.

Figures (2)

• Figure 1: Classes of Generative AI models For Enzyme Design We classify generative AI models for enzyme design based on the modality they sample from. Sequence-generating models include substitution models to optimize existing enzymes, family expansion models to generate diverse novel sequences within a family, and structure-conditioned models that optimize sequences for fixed backbones. Backbone-generating models generate protein structures from noise. These are further distinguished by conditioning granularity, including fragment-conditioned, side-chain-conditioned, and atom-conditioned approaches for active site scaffolding
• Figure 2: Enzymes Designed With Generative AI. Structures of the best performing best performing enzymes designed with generative AI models. The structures were retrieved from the original publication, the AlphaFold database (for designs only containing few substitutions), or predicted with ESMFold. In cases where multiple designs for the same enzyme exist, we display the most active one. Abbreviations: Alcohol Dehydrogenase (ADH), Dihydrofolate Reductase (DHFR), Halide Methyltransferase (HMT), Lactate Dehydrogenase (LDH), Malate Dehydrogenase (MDH)