Learned Conformational Space and Pharmacophore Into Molecular Foundational Model

The emergence of large-scale chemical pre-trained models has significantly advanced our ability to capture complex relationships between molecular structures and their functions. Despite the growing interest in molecular foundational model that provide versatile representations and support molecular optimization for downstream tasks, few efforts have integrated explicit chemical knowledge-such as conformational and pharmacophore information-into pre-training. Given the highly dynamic nature of small molecules in solution, their conformational changes upon target binding, and the critical role of pharmacophore complementarity, it is essential to incorporate these factors into molecular foundational modeling. Here, we present a molecular foundational model that integrates conformational-space and pharmacophore-similarity projections during pre-training to regularize the representation space. The model adopts an Ouroboros-like architecture, where the molecular graphs are encoded into 1D representation vectors via a graph neural network and subsequently reconstructed back into SMILES sequences through an autoregressive Transformer module. This dual-module design establishes a flexible and extensible framework for both representation learning and molecular generation within a unified latent space. Extensive experiments demonstrate that our model effectively addresses a variety of practical chemical challenges, including similarity-based virtual screening, targeted poly-pharmacology design, chemical property prediction, and directed molecular optimization.