Machine Learning methods for Small Wet-Lab Data Challenges in Enzyme Engineering

Abstract

Enzyme engineering is fundamentally constrained by the scarcity of high-quality sequence-function data. The reliance of traditional machine learning models on massive labeled data further exacerbates this dilemma. This review systematically surveys machine learning methods under small-sample wet-lab data conditions, covering three technical routes: zero-shot prediction of mutation effects, few-shot transfer learning (parameter-efficient fine-tuning, meta-learning, and in-context learning), and active learning-guided directed evolution. On this basis, we summarize the applications of these methods in fitness, enzyme kinetic parameters, thermostability, selectivity and substrate specificity, and other tasks, categorized by enzyme properties, and evaluate their effectiveness from the perspectives of data efficiency, optimization magnitude, and experimental validation scale. We further discuss the challenges of collaboration, tool accessibility, and computational cost in practical workflows, and finally outline future directions. This review provides a methodological reference for enzyme engineering researchers working under limited experimental data and offers a clear and comprehensive theoretical framework for methodology developers to build more powerful solutions.