The prediction of functions of proteins is crucial for
understanding biological mechanisms and has significant implications
in medical diagnostics. My talk will trace our decade-long exploration
in designing novel machine learning and deep learning techniques to
predict protein functions, described by the Gene Ontology (GO). I will
discuss the evolution of these methods, highlight significant
breakthroughs that have enhanced our ability to predict protein
functions accurately and efficiently. A particular focus will be on
the integration of neuro-symbolic AI methods, which combine machine
learning with logical reasoning. These methods not only enable the
incorporation of biological background knowledge into our models but
also allow for zero-shot predictions --- predicting functions for
proteins without prior observed examples. I will demonstrate how these
advancements have provided interpretable insights into protein
function and have been used in identifying candidate genes for
Mendelian diseases and enhancing variant prioritization for rare
disease diagnosis. This approach not only broadens our understanding
of protein functions in biotechnology, health, and disease, but also
opens new avenues for protein design.