From Holo Pockets to Electron Density: GPT-style Drug Design with Density

Abstract: Recent advances in generative modeling have enabled significant progress in structure-based drug design (SBDD). Existing methods typically condition molecule generation on empty binding pockets from holo complexes, overlooking informative components such as the filler (ligands and solvent). Here, we leverage low-resolution electron density (ED) derived from the filler as a physically grounded condition for \textit{de novo} drug design. We consider two types of ED, calculated and cryo-EM/X-ray, obtainable from computational or experimental sources, supporting unified pre-training and experimental integration. Compared with rigid pocket representations, experimental ED naturally captures conformational flexibility and provides a more faithful description of the binding environment. Based on this, we introduce EDMolGPT, a decoder-only autoregressive framework that generates molecules from low-resolution ED point clouds. By grounding generation in physically meaningful density signals, EDMolGPT mitigates structural bias and produces molecules with 3D conformations. Evaluations on 101 biological targets verify the effectiveness. Our project page: https://jiahaochen1.github.io/EDMolGPT_Page/.

• The paper introduces EDMolGPT, using electron density conditioning to generate 3D drug molecules that reflect realistic dynamic binding pockets.
• It employs FSMILES tokenization with geometric augmentation and point cloud representation to robustly capture ligand and solvent interactions.
• Experimental results demonstrate improved bioactive molecule recovery rates and energetically favorable docking profiles compared to existing SBDD methods.

EDMolGPT: GPT-Style 3D Drug Generation Conditioned on Low-Resolution Electron Density

Introduction and Motivation

The paper "From Holo Pockets to Electron Density: GPT-style Drug Design with Density" (2605.08767) addresses fundamental limitations of current AI-driven structure-based drug design (SBDD) approaches, which predominantly reconstruct the ligand-generation environment by removing all non-protein "filler" elements from holo complexes, reducing the intricate, dynamic nature of the binding site to a rigid, geometry-centric abstraction. This is problematic since intra-pocket conformational variability, interaction with solvent, and physicochemical microenvironments are poorly encoded by empty pocket representations. Recent trends have incrementally incorporated some pocket flexibility via ensemble-based modeling and electron density (ED) representations, yet these either suffer from integration complexity or low signal quality, especially in highly flexible regions.

The authors propose a paradigm shift: conditioning de novo molecule generation directly on experimentally or computationally derived low-resolution ED point clouds of the "filler"—that is, the ligand and proximal solvent. By harnessing this physically grounded, continuous 3D signal, they aim to both capture conformational plasticity lost in rigid-pocket pipelines and substantially enrich the input space for generative modeling. Experimental ED maps, as acquired from cryo-EM or X-ray crystallography, are routinely available for bioactive complexes and inherently encode ensemble behavior, interaction fields, and spatial localization of both ligand and solvent.

Contributions and Methodology

The paper introduces EDMolGPT, a GPT-style, decoder-only autoregressive transformer framework conditioned on low-resolution ED point clouds for 3D molecule generation:

• Filler ED Conditioning: Instead of using empty pocket geometry, inputs are point clouds sampled from the spatial ED distribution of the "filler" (ligand+solvent, within 4.5Å of the ligand), which may be computed analytically (CalED) or extracted from experiment (ExpED). This decouples structural bias imposed by rigid pockets and captures solvent-mediated and flexible interactions.
• Point Cloud Representation: Each ED point is annotated with pharmacophoric labels (HBA, HBD, etc.), forming a chemically aware and spatially resolved input that closely follows the underlying experimental densities.
• FSMILES Tokenization with Geometric Augmentation: Molecules are represented as a sequence of fragment-SMILES tokens, augmented with discretized Cartesian coordinates and local geometric features (bond length, angle, dihedral), robustly encoding full 3D conformation suitable for strict autoregressive generation.
• Unified Modeling and Training: EDMolGPT is pre-trained on millions of computationally generated CalED structures for large-scale coverage and fine-tuned on limited, high-fidelity ExpED from experimental complexes, ensuring generalizability and physical validity.
• Inference and Generation: At test time, molecules are generated autoregressively, each atom's position sampled from the constrained geometric manifold defined by previously generated atoms and ED point cloud alignment, yielding plausible ligand conformations consistent with the dynamic binding environment.

Experimental Results

Evaluation is conducted on DUD-E, assessing 101 target proteins with >200 validated active ligand structures per target. The paper compares EDMolGPT with prior SBDD and ED-driven generative baselines (Pocket2Mol, Lingo3DMol, TargetDiff, MolCRAFT, ECloudGen, ED2Mol).

Key results:

• Bioactive Molecule Recovery: EDMolGPT yields the highest recovery rate—41% of targets have at least one generated compound highly similar (ECFP4 TS > 0.5) to known actives. This substantially exceeds all baseline SBDD and ED-guided methods and demonstrates that ED point cloud conditioning allows robust exploration of biologically relevant chemical space.
• Binding Affinity and Pose Quality: Generated molecules achieve the lowest docking energy in GlideSP "min-in-place" (mean -6.92 kcal/mol), and 37% show binding modes superior to redocked conformers, indicating that the conformational ensemble sampled via EDMolGPT's autoregressive process is both pocket-compatible and energetically favorable.
• Conformational Stability: Without explicit force-field minimization, strain energy distributions of generated conformers are comparable with methods using explicit postprocessing, validating that geometric constraints encoded in GSMILES plus local structure preserve physical stability.
• Drug-likeness Metrics: Evaluation on QED, Synthetic Accessibility Score (SAS), and molecular weight shows that EDMolGPT produces molecules closely mirroring the complexity and practicality of real ligands, outperforming models that bias toward simple, small structures.

The model further demonstrates the capability to generate ligands that would be classified as incompatible by rigid-pocket-based evaluation—these molecules remain unseen in classical pipelines due to conformational constraints not captured in empty-pocket geometry, but are experimentally validated as active, underscoring the flexibility and scope of the ED-based paradigm.

Ablations and Practicality

Ablation studies dissect the influence of ED resolution, the number of sampled ED points (Np), and the inclusion of pharmacophore labels:

• Point Cloud Density (Np): Sufficient sampling (Np = 199) is required to balance computational cost and geometric expressivity. Decreasing Np degrades both docking scores and recovery.
• ED Resolution: Adjusting the spatial resolution of density (dmin) has limited effect compared to Np and does not dominate the quality/diversity tradeoff.
• Autoregressive Temperature: Higher temperatures increase generation diversity at the cost of slightly reduced pocket alignment and recovery.

In terms of practical application, EDMolGPT is computationally efficient (generation time ~1.5 s/molecule), outperforming diffusion-based SBDD in throughput. The architecture is agnostic to the origin of density: experimental ED maps (ExpED) or analytically computed (CalED), enabling adaptation to both solved and predicted complexes.

Theoretical and Practical Implications

By grounding ligand generation in experimentally and physically meaningful ED fields, the method tightly integrates SBDD with the actual conformational and chemical state of protein-ligand complexes. This model closes the gap between static geometric representations and the real, dynamic, noisy, and solvent-mediated environments encountered in the cell.

Key implications include:

• Advancing Generative SBDD: The direct use of ED as conditioning input could supersede hand-crafted 3D descriptors, enabling models to generalize beyond static representations and rigid binding assumptions.
• Automated Integration of Experimental Data: The framework supports seamless pipeline integration in pharmaceutical settings where holo complex ED maps are routinely generated, allowing for automated, condition-adapted generation efforts.
• Applicability in Ligand-Based Design: When structures for protein pockets are not available, but an active ligand is known, computational ED from MD ensembles can seed novel generations, further generalizing the approach.
• Extending to Flexible and Allosteric Binding: The methodology can be coupled with conformational sampling techniques or stochastic experimental ED, potentially enabling generative exploration of allosteric or cryptic site binders not visible in canonical structures.

Limitations and Future Directions

While EDMolGPT substantially improves the physical realism and relevance of generated molecules, limitations include dependence on the availability or quality of experimental or simulated ED maps, and the inability to model binding events in the complete absence of a resolved binder or reliable simulation. The fragmentation and discretization schemes, while effective, may still impact rare or highly complex topologies.

Future research will likely focus on:

• Generalizing to higher resolutions and larger contexts, possibly by integrating local pocket dynamics from MD or multistate ED averaging;
• Improved integration with end-to-end experimental pipelines, including iterative feedback from binding assays;
• Conditioning on multi-component or allosteric ED, facilitating generation against polypharmacological or cryptic targets;
• Fine-tuning on target-specific or family-specific data, maximizing the exploitability of available experimental resources.

Conclusion

EDMolGPT fundamentally advances generative SBDD by leveraging low-resolution electron density from ligand and solvent as a continuous, physically meaningful conditioning input for 3D ligand generation. The model demonstrates strong improvements in generating bioactive, pocket-compatible, and synthetically accessible ligands, overcoming the static limitations of rigid-pocket-based methods and bridging the gap between experiment and deep learning-driven molecular design. This work substantiates the role of physics-grounded generative modeling as an emerging cornerstone of next-generation AI in drug discovery.