11/02/2026
𝐋𝐞𝐚𝐫𝐧𝐞𝐝 𝐂𝐨𝐧𝐟𝐨𝐫𝐦𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐒𝐩𝐚𝐜𝐞 𝐚𝐧𝐝 𝐏𝐡𝐚𝐫𝐦𝐚𝐜𝐨𝐩𝐡𝐨𝐫𝐞 𝐢𝐧𝐭𝐨 𝐌𝐨𝐥𝐞𝐜𝐮𝐥𝐚𝐫 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐌𝐨𝐝𝐞𝐥
A key challenge in AI-driven drug discovery is moving beyond molecular analysis to enable chemically meaningful molecular design. In a newly published study in Advanced Science, researchers led by 𝑺𝒆𝒏𝒊𝒐𝒓 𝑷𝒓𝒊𝒏𝒄𝒊𝒑𝒂𝒍 𝑰𝒏𝒗𝒆𝒔𝒕𝒊𝒈𝒂𝒕𝒐𝒓, 𝑷𝒓𝒐𝒇. 𝒀𝒂𝒏𝒈 𝒁𝒉𝒂𝒏𝒈 and collaborators from Institute of Systems Medicine, Chinese Academy of Medical Sciences, present 𝑶𝒖𝒓𝒐𝒃𝒐𝒓𝒐𝒔, a molecular foundation model that bridges this gap by integrating molecular understanding and generation within a unified framework.
Ouroboros introduces a chemically grounded representation space that captures both the dynamic three-dimensional behaviour of drug molecules and their pharmacophoric similarities—features often overlooked by existing AI models. By incorporating conformational-space pharmacophore similarity as a learning signal, the model recognises pharmacologically relevant relationships even among molecules with distinct chemical scaffolds.
Crucially, Ouroboros pioneers a reconstruction learning paradigm that allows molecular representations to be translated back into complete chemical structures. This capability enables direct molecular evolution and optimisation within the learned representation space, effectively “closing the loop” between analysis and design.
The framework demonstrates strong performance across a wide range of downstream tasks, including similarity-based virtual screening, multi-target drug design, ADMET prediction, and directed molecular optimisation. By reducing trial-and-error and improving early candidate selection, Ouroboros has the potential to shorten drug discovery timelines and lower experimental costs.
Looking ahead, the team is extending Ouroboros toward target-guided molecular generation and genome-scale drug discovery, further broadening its impact across therapeutic research.
This work was published in Advanced Science on 4th January 2026:
https://doi.org/10.1002/advs.202513556
