🧬 biology

JEDEL: Zero-Shot DNA-Encoded Library Design for Early-Stage Drug Discovery

JEDEL is a novel zero-shot framework that directly translates 3D pharmacophore representations of active ligands into experimentally realizable, synthesis-ready DNA-encoded libraries by mapping interaction patterns to scalable combinatorial routes using purchasable building blocks, thereby outperforming existing baselines across multiple protein targets without requiring target-specific retraining.

Original authors: Zygimantas Jocys, Zhanxing Zhu, Henriette M. G. Willems, Katayoun Farrahi

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zygimantas Jocys, Zhanxing Zhu, Henriette M. G. Willems, Katayoun Farrahi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a master chef trying to create a new dish that perfectly matches a specific customer's taste. In the world of drug discovery, the "customer" is a disease-causing protein, and the "taste" is a specific 3D shape of chemical features (like a lock and key) that the drug needs to fit into.

For a long time, scientists have used computers to imagine millions of new "recipes" (drug molecules) that might fit this lock. But there's a huge problem: most of these computer-generated recipes are impossible to cook. They might require ingredients you can't buy, or cooking steps that don't exist in real life. Scientists spend months trying to figure out how to actually make these virtual ideas, only to realize they can't be built.

JEDEL is a new tool that solves this by flipping the script. Instead of imagining a dish and then trying to find ingredients for it, JEDEL starts with the ingredients you already have in the pantry and designs a dish that is guaranteed to be cookable.

Here is how JEDEL works, using simple analogies:

1. The "Pantry" vs. The "Dream Menu"

Old Way (Dream Menu): Computers dream up a perfect molecule. Then, chemists have to run around the world trying to find the ingredients. Often, they can't, and the dream is discarded.
JEDEL Way (Pantry First): JEDEL looks at a massive, real-world "pantry" of over 200,000 chemical building blocks that scientists can actually buy today. It only designs molecules using these specific, available blocks. If JEDEL suggests a molecule, it is guaranteed to be something a lab can build immediately.

2. The "Blueprint" (Pharmacophore)

The paper explains that many drugs work by matching a specific 3D pattern of features (like a hydrogen bond here, a hydrophobic spot there). This pattern is called a pharmacophore.

Think of the pharmacophore as a blueprint or a mold.
JEDEL takes this blueprint (derived from known active drugs) and asks: "Which combination of real, purchasable ingredients can I mix together to fill this mold perfectly?"

3. The "Smart Translator" (The AI Engine)

The core of JEDEL is a special AI architecture (called a Joint Embedding Predictive Architecture).

Imagine a translator who speaks two languages: Geometry (the 3D shape of the blueprint) and Chemistry (the list of building blocks).
Usually, one shape can be built in many different ways (one-to-many). JEDEL learns to predict the best chemical "recipe" that fits the 3D shape without getting confused. It doesn't just guess; it learns the deep connection between the shape needed and the ingredients available.

4. The "Recipe Book" (DNA-Encoded Libraries)

JEDEL doesn't just make one molecule; it designs a whole library of them.

Think of this like a DNA-encoded library. It's like a giant catalog where every single item is tagged with a unique DNA barcode.
JEDEL designs a focused catalog of millions of potential drugs. Because it only uses real ingredients and real cooking steps (reactions), every single item in this catalog is ready to be "cooked" (synthesized) and tested immediately.

What Did They Find?

The researchers tested JEDEL on 18 different protein targets (like different types of locks). They compared JEDEL against:

Random Selection: Picking ingredients blindly.
Standard Diversity Libraries: Picking a wide variety of ingredients without a specific goal.

The Results:

Better Fit: JEDEL's molecules fit the "locks" (proteins) much better than the random or standard options.
Higher Success Rate: When they looked for molecules that actually worked (hits), JEDEL found them 2.6 times more often than the standard diversity library.
Zero-Shot Power: The coolest part? JEDEL didn't need to be retrained for each new protein. It learned the general rules of "fitting a shape with real ingredients" and applied them to new targets immediately, like a chef who can cook a perfect meal for a new customer without needing a new cookbook.

The Bottom Line

JEDEL bridges the gap between computer imagination and real-world chemistry. It stops scientists from dreaming up unbuildable molecules and starts them with a list of "synthesis-ready" candidates that are guaranteed to be buildable, focused on the specific shape they need, and ready for the lab bench. It turns the drug discovery process from a "guess and check" game into a precise, targeted design strategy.

Technical Summary: JEDEL – Zero-Shot DNA-Encoded Library Design

Problem Statement
A persistent gap exists between computational generative molecular design and experimental validation. Most existing generative models produce virtual compounds that lack guarantees of synthetic accessibility, requiring extensive downstream retrosynthetic planning and often failing to reach the laboratory bench. Conversely, experimental platforms like DNA-encoded libraries (DELs) rely on fixed physical libraries that cannot be dynamically refined based on computational hypotheses. While DELs offer a scalable, bead-based screening format where every member is theoretically synthesizable, current machine learning approaches for DELs are limited to post-screening analysis (e.g., denoising sequencing data) rather than pre-screening library design. Furthermore, many therapeutically relevant targets lack accurate protein structures, rendering structure-based design unreliable. Ligand-based 3D pharmacophore representations offer a robust alternative, yet no existing method successfully translates these pharmacophore constraints directly into synthesis-ready DEL building-block sequences.

Methodology
The authors introduce JEDEL (Joint Embedding for DNA-Encoded Libraries), a framework that maps 3D pharmacophore geometries directly to synthesis-ready DEL building-block sequences. The architecture is designed to operate strictly within the space of purchasable building blocks and validated reactions, ensuring every output is experimentally realizable by construction.

Architecture: JEDEL employs a Joint Embedding Predictive Architecture (JEPA).
- Context Encoder: Uses an E(n)-equivariant Graph Neural Network (P4-EGNN) to encode 3D pharmacophore point clouds (hydrophobic, aromatic, H-bond donors/acceptors, ionizable features), respecting rotational and translational symmetries.
- Target Encoder: Uses a message-passing GNN to encode molecular topology.
- Predictive Objective: Instead of contrastive learning (which struggles with the one-to-many mapping between pharmacophores and diverse molecular scaffolds), JEDEL uses a predictive objective. A predictor MLP learns to map the pharmacophore context embedding to the target topological embedding. An Exponential Moving Average (EMA) target encoder prevents representation collapse.
- Hierarchical Decoder: To manage a vocabulary of over 200,000 purchasable reagents, the decoder uses a hierarchical Bayesian softmax. It first selects a coarse cluster (based on reaction role and pharmacophore profile) and then a specific building block within that cluster. This reduces output dimensionality from $O(|V|)$ to $O(C + K)$ , enabling scalable generation on a single GPU.
Training Data: The model is trained on 8 million synthesis routes enumerated from 217,067 commercially available Enamine building blocks compatible with seven DEL reaction types (e.g., amide coupling, Ugi condensation, Suzuki-Miyaura). Training is performed without target-specific protein information, relying solely on pharmacophore-molecule pairs.

Key Contributions

Synthesis-by-Construction Guarantee: Unlike methods that generate virtual molecules requiring post-hoc filtering, JEDEL generates molecules directly as sequences of purchasable building blocks and reaction steps. Every output corresponds to an executable DEL synthesis protocol.
Zero-Shot Generalization: The model requires no retraining for specific protein targets. It generalizes across diverse protein families (proteases, kinases, GPCRs, etc.) by conditioning on ligand-derived pharmacophores rather than protein structures.
Bridging the Gap: JEDEL is the first framework to connect pharmacophore-driven molecular design directly to combinatorial library construction, closing the loop from computational hypothesis to experimental deployment.

Results
JEDEL was evaluated across 18 diverse protein targets against two baselines: random DEL enumeration and a curated Enamine diversity library (100k compounds).

Binding Affinity: JEDEL achieved the lowest mean docking energy (−7.49 kcal mol⁻¹) across all targets, outperforming Enamine (−7.39) and Random sampling (−6.28). It identified the single best-scoring molecule on 8 of 18 targets (44%).
Pharmacophore Recovery: JEDEL demonstrated superior fidelity in recovering reference pharmacophore features, achieving a mean P4 recovery of 89.3% and a Jaccard similarity of 0.761, compared to 88.1%/0.751 for Enamine and 78.1%/0.694 for Random.
Sample Efficiency: JEDEL showed a 2.6-fold advantage in sample efficiency (median hit rate of 14.0%) over the Enamine diversity baseline (5.4%). The improvement was most pronounced on "difficult" targets where conventional screening fails; for targets with low Enamine hit rates (≤6.3%), JEDEL achieved a mean hit rate of 2.9% versus 0.3% for Enamine.
Chemical Quality: Generated molecules exhibited high chemical validity (100%), uniqueness (99.3%), and novelty (99.9%), with drug-like properties (mean QED 0.64) and physicochemical profiles consistent with the training distribution.

Significance and Claims
The paper claims that JEDEL represents a qualitative shift in the drug discovery pipeline. By embedding synthetic constraints directly into the generative model, it eliminates the disconnect between computational proposals and experimental feasibility. The authors emphasize that JEDEL enables the design of focused, hypothesis-driven libraries at the $10^3$ – $10^4$ scale typical of bead-based platforms, where every compound is selected to satisfy specific pharmacophoric requirements.

The authors note limitations, including reliance on docking scores as a proxy for binding affinity (which may miss entropic or water-mediated effects) and the current restriction to seven reaction types. They state that experimental validation via bead-based DEL synthesis and affinity screening is required to confirm that computational predictions translate to measured binding activity. However, the framework establishes a practical, zero-shot connection between pharmacophore-driven design and combinatorial library construction, offering a direct pathway from computational hypothesis to experimental deployment.