Evolutionary-Scale Enzyme Kinetics Landscape Mapping
Summary¶
Muir, Asper, Notin, et al. (2025) use a high-throughput microfluidic enzymology platform (HT-MEK) to measure Michaelis-Menten parameters (kcat, Km, kcat/Km) for hundreds of adenylate kinase (ADK) orthologs and mutants sampled across bacterial and archaeal phylogeny, spanning the coldest and hottest environments on Earth, under consistent experimental conditions. This is the first evolutionary-scale, quantitative charting of an enzyme sequence-catalysis landscape at this breadth, revealing that catalytic activity varies by up to three orders of magnitude among orthologs with conserved active sites, that thermophilic enzymes are not universally slower than mesophilic ones, and that unsupervised protein language models capture sequence-encoded structure but not catalytic activity itself.
Method¶
- High-throughput kinetics: the HT-MEK microfluidic platform (1,792 chambers) enabled parallel measurement of kcat, Km, and kcat/Km for hundreds of naturally occurring ADK orthologs and site-directed mutants under identical assay conditions, avoiding the cross-study condition variability that normally confounds comparisons of enzyme kinetic parameters drawn from different publications.
- Phylogenetic sampling: orthologs were selected to span bacterial and archaeal diversity and a wide range of native thermal environments (psychrophilic to thermophilic).
- Landscape topology analysis: stepwise point mutations and domain swaps were used to test the navigability of the resulting sequence-catalysis landscape across long evolutionary timescales.
- Protein language model evaluation: an unsupervised PLM was evaluated for how well its learned sequence representations organize ADK sequences by structure versus by catalytic activity; semisupervised models combining PLM representations with this new kinetic dataset were then trained to predict catalytic parameters.
Findings¶
- Extreme kinetic variation despite conserved structure: kcat varies by up to three orders of magnitude across ADK orthologs with conserved active sites and similar predicted structures — catalytic efficiency is not well-predicted by active-site conservation alone.
- Thermophily is not a universal speed penalty: the long-standing hypothesis that thermophilic enzymes are universally slower than mesophilic counterparts (a proposed stability-activity trade-off) does not hold across this dataset.
- A rugged, multi-neighborhood landscape: the sequence-catalysis landscape is organized into at least three distinct global neighborhoods defined by different domain architectures (e.g., "H-bond lid" vs. "Zn²⁺ lid" vs. "lidless" ADK variants), meaning high catalytic activity can be achieved through multiple, structurally distinct evolutionary solutions.
- Landscape navigability: despite this ruggedness, stepwise point mutations and domain swaps show the landscape remains navigable across long evolutionary timescales through path-dependent (history-contingent) mechanisms.
- PLMs capture structure, not catalysis: an unsupervised protein language model organizes ADK sequence space in a way that reflects structural/domain-architecture groupings, but does not by itself organize sequences by catalytic activity — direct kinetic measurement remains necessary to predict function, PLM sequence similarity is an unreliable proxy.
- Improved kcat prediction from denser, consistent data: kcat prediction models trained on this dataset (rather than solely on public databases, which suffer from sparse, heterogeneous-condition measurements) outperform prior models trained only on public databases — corroborating, from an independent experimental angle, the general finding in the DLKcat critique that the composition and consistency of training data, not just model architecture, are limiting factors for enzyme kinetic parameter prediction.
Significance¶
This paper complements the sequence-to-kcat machine learning models in Enzyme Kinetic Parameter Prediction from the opposite direction: rather than training a bigger model on existing sparse, heterogeneous public databases (BRENDA, SABIO-RK), it generates a large, internally consistent experimental dataset at evolutionary scale, and uses it both to test fundamental hypotheses in enzyme evolution and to demonstrate that data quality/consistency, not just model architecture, is a binding constraint on prediction accuracy — directly relevant to the DLKcat generalization critique's conclusion that public-database-trained models struggle to generalize.
See Also¶
- Enzyme Kinetic Parameter Prediction — field overview of sequence-based kcat/Km prediction models this dataset benchmarks against.
- DLKcat Generalization Failure and Train/Test Leakage — an independent critique reaching a compatible conclusion about the limits of public-database training data.