Flow-Based Fragment Identification via Contrastive Learning of Binding Site-Specific Latent Representations

Fragment-based drug design is a promising strategy leveraging the binding of individual fragments, potentially yielding ligands with multiple key interactions, surpassing the efficiency of full ligand screening. The initial step of fragment identification remains challenging, as fragments often bind weakly and non-specifically. We propose a protein-fragment encoder, a new contrastive learning approach that captures protein-fragment interactions. Its latent space allows to perform virtual screening as well as generative design. In the latter case, fragment embeddings are generated conditioned on the protein surface. Our method locates protein-fragment interactions with high sensitivity and can be directly applied to virtual screening for which we observed competitive fragment recovery rates. The generative method outperforms common methods such as virtual screening providing a valuable starting point for fragment hit discovery. Together, these approaches contribute to advancing fragment identification and could provide valuable tools for fragment-based drug discovery. All the code and data can be found on https://figshare.com/s/c28e5d7f9d318305614e.