Scientific discovery is entering a transformative era. Advances in large language models (LLMs) have revolutionized how we process knowledge, yet their impact on real-world scientific discovery remains limited, and sometimes even counterproductive. Consider the discovery of new molecules, which is critical for progress in medicine, clean energy, and advanced materials and is central to addressing many of today’s global challenges. While current AI approaches can generate vast numbers of molecules in silico, these digital advances rarely yield tangible progress in the physical world. They often propose molecules that cannot be synthesized or are incompatible with automated chemistry workflows. As a result, experimental validation remains slow and fragmented, providing little physical feedback to the models and hindering continuous improvement. The gap between rapid computational innovations and real-world breakthroughs continues to widen.

​

At Medicas AI, we are developing a physically-grounded large scientific language model for transforming scientific discovery processes. Using drug discovery as an example, our platform enables human scientists to either:

(1) specify a set of desired functions, and the model generates a molecule that fulfills them,

(2) input an underperforming molecule, and the model “repairs” it to produce a new molecule with improved functionality. In both cases, the generated molecules are guaranteed to be physically valid and automatically synthesizable.

​

We propose four new principles to design scientific foundation models for molecule discovery:

1. “Observe”- acquire, represent, and integrate knowledge from massively multiple data modalities;

2. “Think” - think critically to reason, generate and validate hypotheses;

3. “Propose and Verify” - verify hypotheses through physical experimentation in autonomous labs; 

4. Leverage the digital-physical infrastructure to continuously/non-stop generate new and high-quality molecular data targeting never-ending minimization of the uncertainty of the model.