In all science, inquiry proceeds based on observation and experimentation, exercising informed judgement and developing hypotheses to guide the design of experiments and disambiguate the theories. Artificial intelligence (AI) has dramatically improved state-of-the-art scientific research by helping scientists to formulate hypotheses, design experiments to test them, and collect and interpret data. Fundamental advances over the past decade include self-supervised learning methods that train models on broad data at scale without pre-defined labels, geometric deep learning that leverages structure and geometry informed by scientific knowledge, and generative AI methods that create action plans for experiments and produce new designs such as small molecule drugs and proteins from a diversity of data obtained from experiments, including images and sequences. Among such advances, one of the most commonly shared characteristics is learning the AI/ML model with weak forms of supervision.

To fulfil such goals, we develop a variety of learning methods on a range of structured data representations. We start by working on point clouds; we developed a universal selfsupervised pre-training method for neural feature encoders called “OcCo” and devised a quantum computing-based method named “qKC” for registration. Both methods require no labels for training and improve the robustness of the model when meeting data noise. We next focus on medical CT and CXR images, where data are usually isolated across multiple centres; therefore, we develop a federated learning framework to jointly exploit isolated data usage to improve clinical models’ performance. We next developed “GraphMVP” and “MolGraphEval” to advance the SOTA of self-supervised graph learning on molecules and provide an understanding of what structural information is captured in these methods.