PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jiahao Chen, Letian Gao, Yanhao Zhu, Wenbiao Zhou, Bing Su, Zhi John Lu, Bo Huang

arXiv:2605.08767v2cs.AI

TL;DR

Existing SBDD methods commonly condition generation on static empty pockets and overlook informative filler components and conformational flexibility. This paper introduces EDMolGPT, which generates molecules from low-resolution filler-derived electron-density point clouds using calculated and experimental density. Evaluations on over 100 DUD-E targets report favorable binding-compatible conformations, bioactivity, and improved performance in 3D and 2D chemical spaces.

  • Problem

    Existing SBDD methods commonly use static empty-pocket representations, overlooking filler-derived information and limiting representation of conformational flexibility.

  • Method

    EDMolGPT is a decoder-only autoregressive framework that generates 3D molecules from low-resolution filler-derived electron-density point clouds using calculated and cryo-EM/X-ray density.

  • Results

    EDMolGPT generates molecules with favorable binding-compatible 3D conformations and bioactivity across 101 targets, with improved 3D and 2D chemical-space performance and bioactive-molecule recovery.

  • Takeaways & Limitations

    Filler-derived electron density provides a conditioning representation for de novo drug design that captures binding-environment flexibility beyond rigid pocket representations.

Abstract

from arXiv · show

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/.

1. Introduction

Existing SBDD methods often reduce holo complexes to static empty pockets, overlooking filler-derived information and protein flexibility. The paper proposes using filler electron density with EDMolGPT to generate 3D molecules under dynamic, physically grounded binding-environment conditions.

  • Motivation: Existing SBDD pipelines remove ligands and solvent from holo complexes, treating the resulting pocket as a fixed scaffold for molecule generation.This formulation assumes a single static pocket conformation and suppresses intrinsic protein flexibility.
  • Motivation: Experimental electron density provides a continuous, physics-grounded representation of ensemble-averaged spatial distributions, physicochemical environments, and interaction patterns.It avoids reliance on rigid geometric abstractions while representing binding-site flexibility.
  • Motivation: Filler electron density is typically better defined and more spatially localized than pocket density in highly flexible regions, offering a more reliable conditioning signal.It encodes conformational variability and interaction patterns without rigid geometric assumptions.
  • Contributions: The method derives low-resolution electron density directly from holo-complex fillers, considering calculated density for pre-training and cryo-EM/X-ray density for experimental integration.This bypasses intermediate pocket modeling and incorporates ligand- and solvent-derived information.
  • Contributions: EDMolGPT is a decoder-only autoregressive framework that conditions 3D drug design on low-resolution electron-density point clouds.The model reorders points spatially, represents outputs with FSMILES, and uses calculated density for pre-training before experimental fine-tuning.
  • Results: 101 targets support generation of molecules with favorable 3D conformations compatible with binding pockets and demonstrated bioactivity.The reported evaluation validates the method’s potential for de novo drug discovery.

2. Related work

SBDD uses target 3D structure to guide ligand generation, while AI-based generative methods address the expense of classical workflows. Existing electron-density-guided approaches still face trade-offs among novelty, conformational fidelity, and drug-likeness.

  • Structure-based drug design: Classical SBDD workflows include docking, scoring functions, and molecular dynamics, but are computationally expensive for large-scale virtual screening.AI-based generative modeling has been introduced to address these computational limitations.
  • Electron density-guided molecule generation: Existing electron-density-guided generation methods struggle to balance scaffold novelty, 3D conformation fidelity, and drug-likeness under binding constraints.This frames the central challenge for ED-conditioned molecular generation.

3. Method

EDMolGPT converts filler-derived electron density into ordered 3D point clouds and autoregressively generates molecular structures conditioned on them. The method combines calculated or experimental ED, pharmacophore labels, FSMILES, discretized geometry, and distinct training–inference conditioning procedures.

  • 3.1. Problem formulation: EDMolGPT formulates drug design as molecule generation conditioned on a compact point cloud extracted from filler electron density.Each point contains a type and 3D coordinates, providing a compact geometric conditioning signal for the binding context.
  • 3.2. Generating point cloud: CalED is computed from solved structures using FFT-based reciprocal-space processing, whereas ExpED is obtained directly from experimental measurements.High-frequency filtering and inverse Fourier transformation reconstruct a smooth low-resolution density map before point-cloud sampling.
  • 3.2. Generating point cloud: Randomly sampled ED points receive pharmacophore labels based on their nearest filler atoms, enriching the otherwise limited chemical information in low-resolution clouds.Labels include hydrogen-bond donor, acceptor, donor/acceptor, and Other categories.
  • 3.2. Generating point cloud: The point cloud is sorted by ascending x, y, and z coordinates to provide consistent ordering for the autoregressive model.This ordering addresses the sensitivity of autoregressive models to input sequence order.
  • 3.3. Input format of molecule: Molecules are represented with improved FSMILES plus discretized 3D coordinates, bond lengths, bond angles, and dihedral angles for autoregressive generation.The FSMILES modification avoids splitting small fragments that connect rings, while relative geometric features model local structural dependencies.
  • 3.4. EDMolGPT: EDMolGPT follows a GPT-2-style decoder-only architecture that concatenates point-cloud and molecule sequences and predicts molecular tokens sequentially.During training, ED is derived from the ligand; during inference, the conditioning ED incorporates solvent.

4. Experiment

EDMolGPT is evaluated across bioactive recovery, binding affinity, conformational stability, molecular properties, and ablations on DUD-E-derived and binding-complex data. It achieves strong recovery and binding results while maintaining competitive conformational stability and balanced molecular properties.

  • Datasets: Approximately two million curated molecules support pre-training, while 40k PDBbind binding experiments provide complexes for ExpED fine-tuning.The original public collection contains approximately eight million molecules and is filtered using QED and SAS.
  • Bioactive Molecule Recovery: 41% of DUD-E targets achieve bioactive molecule recovery, with generated compounds exceeding ECFP4 TS > 0.5 against known active compounds.The evaluation uses experimentally validated active ligands with measured affinities, making recovery a direct comparison with bioactive compounds.
  • Binding Affinity: −6.92 is EDMolGPT’s average Glide score in the min-in-place setting, and 37% of generated molecules outperform force-field-sampled conformations on Min. < Re.Lower Glide scores indicate stronger predicted binding; the relative metric compares generated conformations with classical force-field sampling.
  • Conformational Stability: 33, 69, and 194 kcal/mol are EDMolGPT’s strain energies at the 25%, 50%, and 75% quantiles, respectively, without post-processing.These values are reported as comparable to Pocket2Mol and Lingo3dMol despite ED2Mol using refinement during generation.
  • Molecular Properties: EDMolGPT balances competitive QED and SAS with molecular weights close to Reference ligands, whereas ED2Mol’s apparent property advantages accompany much smaller molecules.The paper interprets the smaller ED2Mol molecules as simpler structures with reduced coverage of realistic drug-design chemical space.
  • Molecular Properties: Across molecular-weight bins, EDMolGPT maintains more stable and favorable SAS, while ED2Mol’s SAS worsens significantly as molecular weight increases.The binned comparison also reports stronger overall recovery and binding-affinity performance for EDMolGPT, while recognizing ED2Mol’s strength on selected drug-likeness properties.
  • Results on ExpED: Under ExpED, EDMolGPT can generate bioactive ligands accommodated by flexible regions that appear sterically clashing under rigid-pocket evaluation.Figure 6 describes a newly generated molecule that overlaps with the rigid pocket yet remains active.
  • Ablation Studies: Changing ED resolution produces no notable metric changes, while increasing temperature from 0.7 to 1.2 slightly lowers diversity and modestly changes alignment and recovery.The temperature ablation indicates a trade-off between scaffold diversity and structural consistency.

5. Conclusion

EDMolGPT conditions decoder-only autoregressive 3D drug design on low-resolution electron-density point clouds sampled from existing binders rather than rigid pockets. Across over 100 DUD-E targets, it outperforms existing structure-based generative methods in chemical, binding-mode, and bioactive-recovery measures.

  • 5. Conclusion: EDMolGPT uses electron-density-derived point clouds from existing binders as conditioning signals for decoder-only autoregressive 3D drug design.The approach is described as flexibly representing the binding environment and generating chemically plausible conformations.
  • 5. Conclusion: Across over 100 DUD-E targets, EDMolGPT outperforms existing structure-based generative methods in 3D and 2D chemical spaces, binding modes, and bioactive molecule recovery.The conclusion reports improved binding modes and higher recovery rates of bioactive molecules.

Impact Statement

The paper identifies no societal consequences that it considers necessary to highlight.

  • Impact Statement: The authors state that no consequences of the work require specific highlighting.No additional impact claim is provided in the supplied statement.

Overview

The appendix broadens EDMolGPT’s practical context, explains its ED-based and FSMILES representations, and documents supplementary analyses of distributions, discretization, and generation.

  • Overview: EDMolGPT uses low-resolution electron density derived from pre-existing binders rather than precise rigid pocket geometry to guide molecular generation.The approach is presented as applicable when holo structures contain binders and when active-compound conformations can be simulated computationally.
  • Overview: ExpED and CalED have noticeably different intensity distributions, motivating pre-training on stable, abundant CalED followed by fine-tuning on experimental ExpED.ExpED is generally lower and more variable than CalED because of experimental noise and structural flexibility.
  • Overview: The refined FSMILES decomposition preserves bonds when cleavage would create fragments smaller than three atoms, reducing unnecessary fragmentation.This preserves chemically meaningful groups such as hydroxyls that original BRICS-based decomposition may split.
  • Overview: Fragment atom counts are mostly moderate and fragment molecular weights remain concentrated within a reasonable range, supporting chemically meaningful decomposition.These distributions indicate a balance between fragmentation granularity and chemical coherence.
  • Overview: Coordinate discretization uses σ = 0.1 Å, giving a maximum per-coordinate quantization error of 0.05 Å, while relative geometric features constrain generation.Bond lengths, bond angles, and dihedral angles provide local structural constraints during inference.

E. Experimental setting

The experimental setting uses DUD-E-related biological evaluation, an expanded GPT-2 architecture, FSMILES and coordinate vocabularies, and ECFP4 Tanimoto Similarity for bioactivity-related assessment.

  • E. Experimental setting: DUD-E contains 102 protein targets spanning kinases, proteases, GPCRs, nuclear receptors, and ion channels.The benchmark is designed for developing and evaluating virtual screening algorithms.
  • E. Experimental setting: EDMolGPT uses 300-token FSMILES and coordinate vocabularies, plus 200 tokens each for bond lengths, bond angles, and dihedral angles.These settings provide coverage for molecular and structural variation.
  • E. Experimental setting: The model extends GPT-2 with 24 transformer blocks, hidden dimension 1024, context length 1024, and 16 attention heads.Dropout is set to 0.1 across the reported residual, embedding, and attention components.
  • E. Experimental setting: ECFP4 Tanimoto Similarity compares generated molecules with validated active molecules to assess structural analogy to proven active scaffolds.The metric complements physical measures such as docking scores.

F.1. Distribution analysis

A nearest-neighbor analysis tests whether DUD-E evaluation structures overlap with the training data, finding limited similarity consistent with genuine generation rather than memorization.

  • F.1. Distribution analysis: 60%: maximum DICE similarity between any DUD-E active-ligand point cloud and its closest training-set point cloud never exceeds 60%.The analysis retrieves the most similar training point cloud for each DUD-E active ligand.
  • F.1. Distribution analysis: The low overlap indicates that the DUD-E test structures are not contained in the training data, supporting evaluation of generation rather than simple recall.The analysis was motivated by the possibility that substantial overlap could indicate memorization.

F.2. Visualization of generation results

Visualization analyses compare ED-conditioned generations with experimentally determined ligands and report generation-speed comparisons across SBDD models.

  • F.2. Visualization of generation results: ED-conditioned point clouds produce ligands that align well with the spatial distribution of the conditioning point cloud across three representative complexes.The complexes are 1sj0, 3lan, and 2etr, and generated ligands are compared with experimentally determined ground truth.
  • F.2. Visualization of generation results: The visualizations compare generated ligands with experimentally determined ground truth for targets 1sj0, 3lan, and 2etr.Each target uses an electron-density-derived point cloud as the conditioning input.
  • F.2. Visualization of generation results: Low docking scores and structural diversity are presented together as evidence of alternative binding-mode exploration without sacrificing physical or chemical feasibility.This conclusion is stated in the context of the reported visualization and docking analyses.
  • F.2. Visualization of generation results: 1.5 seconds per molecule: EDMolGPT’s reported average generation speed is compared with speeds from other molecular-generation models.Pocket2Mol is reported at approximately 0.45 seconds per molecule.

F.4. Ablation studies on Np and pharmacophore labels cp

The ablation studies show that electron-density detail and pharmacophore guidance jointly control the balance between structural fidelity and molecular diversity.

  • Increasing Np provides a more detailed positive electron-density description, improving the Min-in-place score while slightly reducing diversity.
  • Removing pharmacophore labels cp relaxes geometric constraints, increasing diversity but lowering Min-in-place scores.
  • The observed trade-offs confirm that both Np and pharmacophore guidance are important for generated-molecule quality and structural fidelity.
Loading 2605.08767v2…