Integrative machine learning approaches for enzyme kinetic parameter prediction

The prediction of enzyme kinetic parameters, particularly $\mathit{k}_{\text{cat}}$ and $\mathit{K}_{\text{m}}$, is rapidly evolving from physics-based modeling toward data-driven artificial intelligence (AI) approaches, driven by expanding biochemical databases and advances in machine learning (ML) and deep learning (DL). We focus on AI-based enzyme kinetic parameter prediction, with an emphasis on $\mathit{k}_{\text{cat}}$ and $\mathit{K}_{\text{m}}$, while also considering related parameters, such as $\mathit{K}_{\text{i}}$ and $\mathit{k}_{\text{cat}}/\mathit{K}_{\text{m}}$. The key findings reveal that: (i) Representation learning increasingly relies on pretrained protein language models, while structural embeddings often show limited additional performance gains over sequence-based approaches; (ii) Model performance is highly data-dependent: classical ML regressors can outperform DL architectures on small datasets when using comparable feature representations, though multimodal feature integration can improve generalization; and (iii) Multitask frameworks for the joint prediction of $\mathit{k}_{\text{cat}}$ and $\mathit{K}_{\text{m}}$ may provide synergistic benefits, although the magnitude of the improvement may be limited. Future progress will require standardized benchmarks, mechanistically informed model architectures, uncertainty-aware prediction, and robust out-of-distribution validation to improve biochemical relevance, interpretability, and real-world utility.