Zero-Shot Conditional Molecule Generation with Latent Diffusion Models from Contrastive Pre-Trained Embeddings

Accurately predicting the biological activity of molecules is crucial in drug discovery, as it helps identify compounds likely to bind to specific targets, thus accelerating development and reducing experimental costs. Although recent advances have enabled predictive models to estimate molecular activity from textual assay descriptions and molecular structures, existing approaches remain focused primarily on classification and regression tasks. There is a clear gap in generating novel, biologically active molecules tailored specifically to given bioassays.
To address this limitation, this thesis introduces a generative approach integrating contrastively learned molecular embeddings from CLAMP, a state-of-the-art activity prediction model, with a latent diffusion framework. The proposed diffusion model is trained exclusively on molecular embeddings without explicit assay-specific conditioning, enabling zero-shot inference guided solely by textual assay descriptions. By leveraging CLAMP’s powerful embeddings, our method efficiently generates chemically valid, diverse, and biologically relevant molecules tailored to novel assays.
Our experiments demonstrate that assay-guided generation significantly enhances the predicted bioactivity of generated molecules compared to unguided generation. Further, linear probing experiments validate the robustness and biological informativeness of the generated molecular embeddings. Our findings indicate a clear trade-off between chemical diversity and assay specificity, with guided generation improving biological relevance at the cost of broader chemical exploration. This work advances computational drug discovery by introducing a flexible and scalable framework for zero-shot, assay-informed molecule generation, paving the way for more targeted and data-efficient drug design.