AI-Powered Science Beyond Large Datasets

The scientific community is optimistic about AI's potential following DeepMind's Nobel Prize for AlphaFold in 2024, but experts caution that this successful model relies on exceptionally rare conditions. AlphaFold was trained on 170,000 protein structures from the Protein Data Bank, a repository built over 53 years with an estimated investment of 21 billion dollars. Replicating these conditions across other scientific disciplines would take decades. Moreover, most scientific data is inherently inconsistent—cell cultures drift, reagents contain impurities, environmental conditions vary—making it impossible to create standardized, large-scale datasets needed to train modern neural networks. The true catalyst for accelerating scientific discovery will come through AI agents rather than data-heavy approaches.