In protein engineering, machine learning (ML) advancements have led to significant progress, including protein structure prediction (e.g., AlphaFold), sequence representation through language models, and novel protein generation. However, the impact of data curation on ML model performance is underexplored. As more sequence and structural data become available, a data-centric approach is increasingly favored over a model-centric method. A data-centric approach prioritizes high-quality, domain-specific data, ensuring ML tools are trained on datasets that accurately reflect biological complexity and diversity. This paper introduces a novel methodology that integrates ancestral sequence reconstruction (ASR) into ML models, enhancing data-centric strategies in the field. ASR uses computational techniques to infer ancient protein sequences from modern descendants, providing diverse, stable sequences with rich evolutionary information. While multiple sequence alignments (MSAs) are commonly used in protein engineering frameworks to incorporate evolutionary information, ASR offers deeper insights into protein evolution. Unlike MSAs, ASR captures mutation rates, phylogenic relationships, evolutionary trajectories, and specific ancestral sequences, giving access to novel protein sequences beyond what is available in public databases by natural selection. We employed two statistical methods for ASR: joint Bayesian inference and maximum likelihood. Bayesian approaches infer ancestral sequences by sampling from the entire posterior distribution, accounting for epistatic interactions between multiple amino acid positions to capture the nuances and uncertainties of ancestral sequences. In contrast, maximum likelihood methods estimate the most probable amino acids at individual positions in isolation. Both methods provide extensive ancestral data, enhancing ML model performance in protein sequence generation and fitness prediction tasks. Our results demonstrate that generative ML models training on either Bayesian or maximum likelihood approaches produce highly stable and diverse protein sequences. We also fine-tuned the evolutionary scale ESM protein language model with reconstructed ancestral data to obtain evolutionary-driven protein representations, and downstream stability prediction tasks for Endolysin and Lysozyme C families. For Lysozyme C, ancestral-based representations outperformed the baseline ESM in KNN classification and matched the established InterPro method. In Endolysin, our novel ASR-Dist method performed on par with or better than the baseline and other fine-tuning approaches across various classification metrics. ASR-Dist showed consistent performance in both simple and complex classification models, suggesting the effectiveness of this data-centric approach in enhancing protein representations. This work demonstrates how evolutionary data can improve ML-driven protein engineering, presenting a novel data-centric approach that expands our exploration of protein sequence space and enhances our ability to predict and design functional proteins.
