JEDEL

JEDEL is a generative framework for designing DNA-encoded libraries (DELs) directly from three-dimensional pharmacophore representations of active ligands. DNA-encoded libraries are a workhorse of modern early-stage drug discovery: combinatorial collections of up to millions of small molecules, each tagged with a unique DNA barcode, that can be screened against a protein target in a single pooled experiment. The hard problem is designing a focused library that is both chemically realizable and enriched for binders. JEDEL addresses this by mapping pharmacophore interaction patterns — the spatial arrangement of hydrogen-bond donors, acceptors, hydrophobic groups, and other features that define how a ligand engages a target — onto concrete synthesis instructions.

The model's central design choice is to operate entirely within the space of purchasable building blocks and validated DEL reactions. Rather than emitting virtual compounds that then require downstream retrosynthesis and sourcing, JEDEL produces combinatorial synthesis routes whose outputs are experimentally realizable by construction. According to its authors, it is the first model to map pharmacophore geometry to actionable, scalable synthesis instructions at the scale of a full DEL.

JEDEL was introduced in a June 2026 arXiv preprint by researchers at the University of Southampton's School of Electronics and Computer Science and the ALBORADA Drug Discovery Institute at the University of Cambridge. It is a genuinely pretrained generative model: trained once and then applied zero-shot across 18 protein targets without any target-specific retraining.

#Key Features

• Synthesis-ready by construction: Every generated molecule is restricted to purchasable building blocks and validated reaction chemistry, so outputs map directly to a synthesizable DEL rather than to virtual compounds requiring downstream synthesis planning.
• Pharmacophore-to-structure alignment: JEDEL learns a predictive alignment between 3D pharmacophore geometry and molecular structure, then decodes it into combinatorial synthesis routes spanning potentially millions of molecules.
• Zero-shot generalization: A single pretrained model is applied across 18 protein targets with no target-specific fine-tuning, designing a focused library for each from its pharmacophore input alone.
• Sample efficiency: Across benchmarks, the generated libraries outperform random and diversity-based baselines on predicted binding affinity, pharmacophore recovery, and sample efficiency.

#Technical Details

JEDEL adopts a Joint Embedding Predictive Architecture (JEPA)-style design. An E(n)-equivariant graph neural network encodes the 3D pharmacophore of an active ligand into a representation that respects rotational and translational symmetry, which is essential for reasoning over geometric interaction patterns. A hierarchical Transformer decoder then translates this representation into combinatorial synthesis routes — selections of building blocks and reactions — that assemble into the target library. The model references PDBbind during training and, once trained, requires no per-target optimization. Evaluation spans 18 protein targets, where the focused libraries JEDEL produces exceed random and diversity-based baselines on predicted binding affinity, pharmacophore recovery, and sample efficiency. As of the preprint, no public code repository or model weights had been released; the paper is distributed under CC BY 4.0, with no software license stated.

#Applications

JEDEL targets the hit-finding stage of small-molecule drug discovery, where medicinal chemistry teams use DNA-encoded libraries to screen vast chemical space against a protein of interest. By generating synthesis-ready, target-focused libraries from a known active ligand's pharmacophore, it offers a route to design DELs that are enriched for binders while remaining buildable with on-hand reagents and established reactions. This is most directly useful to discovery groups and DEL platform teams who want to move from a structural hypothesis about binding to a physically constructable library without a separate synthesis-planning step.

#Impact

JEDEL reframes DEL design as a generative problem grounded in 3D interaction geometry rather than 2D similarity or unconstrained virtual enumeration, and it positions pharmacophore-conditioned, synthesis-aware generation as a path from in-silico design to experimentally deployable libraries. Its zero-shot performance across 18 targets suggests the learned pharmacophore-to-synthesis mapping transfers without retraining, which is a meaningful claim for a workflow where retraining per target is costly. As a recent preprint without released code or weights, its real-world adoption and independent validation remain to be established, and reported gains rest on predicted (rather than experimentally measured) binding affinity.