Flexible Kernels for Protein Property Prediction

Predicting protein properties such as binding affinity, thermostability, and fluorescence from limited experimental data remains a fundamental challenge in molecular biology and bioinformatics. Traditional methods often depend heavily on structural information or large-scale deep learning embeddings, which, while powerful, are computationally expensive and sometimes lack interpretability. This paper introduces a novel class of sequence kernels grounded in evolutionary biology, specifically leveraging substitution matrices like BLOSUM50, to build data-efficient, interpretable Gaussian process models for protein property prediction.

The core innovation lies in designing flexible kernels that incorporate biological prior knowledge through correlation matrices derived from substitution matrices, with learnable exponents that adapt to specific protein landscapes. These kernels, termed kLOCK, are combined with local linearity assumptions to capture additive effects of mutations, providing a nuanced balance between linear and non-linear modeling. Furthermore, the authors propose structure-conditioned kernels (CLOCK), which map pre-trained structural embeddings into correlation matrices, enabling the model to incorporate structural information without explicit structural data during training.

Extensive benchmarking across 21 datasets demonstrates that sequence-only kernels outperform many embedding-based models, especially in data-scarce and extrapolation scenarios. In particular, the models achieve Pearson correlation coefficients up to 0.807 and reduce mean absolute errors by over 15% compared to baseline deep learning approaches. The multi-task learning framework further enhances performance by sharing information across related properties, facilitating knowledge transfer and improving generalization.

This work has profound implications for protein engineering, drug discovery, and functional annotation. By providing a biologically interpretable, computationally efficient, and highly accurate prediction tool, it paves the way for more rapid and reliable protein design workflows. The ability to incorporate structural cues via zero-shot learning from structural embeddings broadens the applicability of the approach, especially when experimental structural data is unavailable.

Despite these advances, challenges remain in extrapolating to entirely novel mutation combinations and in optimizing hyperparameters for large-scale applications. Future research will likely focus on integrating additional biological modalities, automating hyperparameter tuning, and scaling the models to encompass broader protein families. Overall, this study marks a significant step toward more intelligent, data-efficient, and biologically grounded computational tools in molecular biology.