Interrogating contrastive learning embeddings for structure-based virtual screening: a case study on DrugCLIP
Year: 2026
Type: preprint
Abstract: Virtual screening has become central to early-stage drug discovery, and structure-based approaches have recently been reframed as a retrieval problem through contrastive learning methods such as DrugCLIP, which project protein pockets and ligands into a shared embedding space. However, what these abstract representations encode and how they relate to conventional notions of structural and chemical similarity remain unclear. Here we present a systematic dissection of DrugCLIP's latent space. We show that its pocket embeddings, despite not being explicitly trained for this task, achieve state-of-the-art performance in pocket similarity search while running over 10,000 times faster than existing structural descriptors. We further demonstrate that this embedding space is structurally coherent and robust to conformational variation. Ligand embeddings, by contrast, capture a pocket-aware notion of chemical similarity that extends beyond conventional fingerprint-based measures. Using a rigorous de-leakage benchmark, we show that DrugCLIP generalises beyond its fine-tuning set, retrieving the crystallographic ligand within the top 1% of a library of 48,999 compounds for 56.5% of 223 pockets with low protein sequence identity (SI ≤ 30%) and limited ligand chemical similarity (TS ≤ 0.7) to the fine-tuning set. Nevertheless, performance declines under increasingly realistic screening conditions, primarily due to differences in sidechain orientation across apo, holo and AlphaFold-predicted structures, and mismatches in pocket residue composition when using predicted pockets. These findings delineate the practical limits of DrugCLIP's applicability, highlight accurate pocket identification and representation as key factors for improving performance, and provide a transferable framework for interpreting the latent spaces of related contrastive pocket-ligand models. Together, these results offer insight to guide the refinement of existing methods and the development of future contrastive screening approaches.
Institution: University College London
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7213662584