University of Waterloo

Published purpose

Scientific discovery, by nature, is a lengthy and arduous journey largely due to the vastness of its exploratory scope. To illustrate, the design of a 100 amino-acid protein sequence entails the exploration of mind-boggling 20^100 possibilities. Historical approaches to scientific discovery in realms such as protein, molecular, and material design have relied heavily on the artful blend of intuition, experience, and a dash of serendipity. These empirical-driven methods, while valuable, are markedly time-intensive and laden with numerous unsuccessful attempts. The advent of artificial intelligence (AI) promised a new velocity in the discovery process. Yet, prior efforts have generally centered on small-scale models with constrained generalizability, resulting in suboptimal outcomes. The emergence of large language models (LLMs), with their remarkable reasoning and generalization prowess, holds the potential to catalyze a seismic shift in this field, thereby dramatically streamlining the scientific discovery journey. The project aims to accelerate scientific discovery with big data. By utilizing the currently popular large foundation models, it will be possible to unleash their generalization capabilities to help researchers significantly shrink the search space. By leveraging AI techniques in the process, it can make the search process more explainable and systematic.