Flexible Kernels for Protein Property Prediction
Efficiently predicting protein properties—such as binding affinity and thermostability—from sparse experimental data remains challenging. This work proposes a novel sequence kernel for Gaussian process models that integrates evolutionary substitution matrices with a local linearity assumption, and innovatively incorporates structure-aware substitution matrices to embed structural priors from foundation models directly into the kernel design. By synergistically leveraging both evolutionary and structural information, the method enables effective multi-task learning and significantly outperforms existing approaches based on large-model embeddings or local supervised learning across multiple protein property prediction tasks. The approach demonstrates superior data efficiency and generalization capability, particularly in low-data regimes.