Scientific research is being augmented and accelerated by the increasing integration of artificial intelligence (AI). This technology enables researchers to develop hypotheses, plan experiments, gather and analyze massive amounts of data, and come to conclusions that otherwise might not have been possible.
Here, we review recent advances such as self-supervised learning, which enables the training of models on massive volumes of unlabeled data, and geometric deep learning, which makes use of the structure of scientific data to improve model accuracy and effectiveness.
By analyzing many input modalities, such as photos and sequences, generative AI algorithms can generate designs, such as proteins and small-molecule medicines. We examine the key problems that persist despite such advancements as well as how these methods can assist scientists at various stages of the scientific process.
The issues brought on by poor data quality and stewardship still need to be addressed, and both creators and users of AI tools need to have a better grasp of when such approaches need to be improved.
These problems need the development of fundamental algorithmic approaches that can either contribute to or acquire scientific knowledge on their own, making them crucial areas of attention for AI innovation.
These problems cut across scientific disciplines.