PaperlayerView exact revision on arXiv

Source-linked AI summary

Generating 3D Molecular Structures Conditional on a Receptor Binding Site with Deep Generative Models

Tomohide Masuda, Matthew Ragoza, David Ryan Koes

arXiv:2010.14442v3physics.chem-phcs.LGq-bio.BM

TL;DR

Existing molecular generators largely operate on two-dimensional representations and do not directly generate structures conditioned on a three-dimensional receptor binding pocket. This paper combines receptor-conditioned convolutional generative modeling of atomic-density grids with atom and bond fitting to produce discrete 3D molecules. The generated molecules are often valid and unique with reasonable predicted interactions, but increased novelty away from the seed is accompanied by lower predicted affinity.

  • Problem

    Existing molecular generative models mainly use two-dimensional representations, while virtual screening cannot propose compounds absent from existing libraries or exhaustively explore chemical space.

    Source: §1 · p. 1, §1.1 · p. 3

  • Method

    The model uses conditional VAEs and GANs to encode receptor and ligand atomic-density grids, decode receptor-conditioned ligand densities, and fit discrete atoms and bonds.

    Source: §1 · p. 2, §1.1 · p. 3, §2.3 · p. 3

  • Results

    Valid molecules are generated at over 80% on average at variability factor 1.0, with more than 90% chemically unique; greater variability increases novelty but lowers predicted affinity.

    Source: §3.1 · p. 5, §3.3 · p. 6

  • Takeaways & Limitations

    The work demonstrates the feasibility of conditional 3D molecular structure generation and provides a basis for future property-optimized generation.

    Source: §4 · p. 10

    Generating structures from a receptor without a reference seed ligand likely requires better affinity optimization or explicit conditioning on binding affinity or other molecular properties.

    Source: §4 · p. 10

Abstract

from arXiv · show

Deep generative models have been applied with increasing success to the generation of two dimensional molecules as SMILES strings and molecular graphs. In this work we describe for the first time a deep generative model that can generate 3D molecular structures conditioned on a three-dimensional (3D) binding pocket. Using convolutional neural networks, we encode atomic density grids into separate receptor and ligand latent spaces. The ligand latent space is variational to support sampling of new molecules. A decoder network generates atomic densities of novel ligands conditioned on the receptor. Discrete atoms are then fit to these continuous densities to create molecular structures. We show that valid and unique molecules can be readily sampled from the variational latent space defined by a reference `seed' structure and generated structures have reasonable interactions with the binding site. As structures are sampled farther in latent space from the seed structure, the novelty of the generated structures increases, but the predicted binding affinity decreases. Overall, we demonstrate the feasibility of conditional 3D molecular structure generation and provide a starting point for methods that also explicitly optimize for desired molecular properties, such as high binding affinity.

1 Introduction

The paper addresses the limits of virtual screening and mainly two-dimensional molecular generative models by introducing 3D molecule generation conditioned on receptor binding pockets. It combines atomic-density representations with conditional deep generative modeling and discrete structure fitting to produce novel candidate molecules.

1 Introduction

  • Virtual screening narrows existing chemical libraries but cannot exhaustively explore chemical space or propose missing compounds and property-improving alternatives.Generative models are motivated as a way to learn chemical-space distributions and sample novel compounds.Source: §1 · p. 1

1.1 Related Work

  • Existing molecular generative efforts mainly learn continuous latent spaces from two-dimensional SMILES strings or molecular graphs.SMILES-based methods can generate invalid strings, while graph representations better capture chemical invariances.Source: §1 · p. 1, §1.1 · p. 2

1 Introduction

  • The paper introduces a deep generative model that encodes receptor and ligand atomic-density grids into latent spaces and generates 3D ligand structures conditioned on a binding pocket.Generated grids are converted into molecular structures and evaluated for validity, uniqueness, and predicted binding affinity.Source: §1 · p. 2, §1.1 · p. 3

1.1 Related Work

  • Conditional VAEs and GANs provide complementary generative mechanisms for sampling novel molecular structures and encouraging generated densities to resemble real data.VAEs support sampling from latent random variables, while GANs use a discriminator to distinguish real and generated data.Source: §1.1 · p. 2, §1 · p. 2
  • Atomic density grids retain three-dimensional conformational information and offer continuous, differentiable, permutation-invariant representations, but must be converted into discrete molecules.The paper addresses this conversion with atom-fitting and bond-adding algorithms rather than reverting to SMILES generation.Source: §1.1 · pp. 2–3

2 Methods

The method represents receptor–ligand complexes as 3D atomic-density grids, generates ligand densities with a receptor-conditioned convolutional model, fits discrete atoms and bonds, and evaluates sampled structures after optimization. Experiments use cross-docked complexes and repeated latent-space sampling to assess molecular quality and receptor interactions.

2.1 Molecular grid representation

  • Atomic-density grids encode atom type, aromaticity, hydrophobicity, and hydrogen-bond state on 48^3 voxels spanning a 23.5Å cube.Atoms are represented as Gaussian-like densities in separate channels, with random rotations during training to mitigate missing rotation invariance.Source: §2.1 · p. 3

2.2 Fitting molecules to generated grids

  • Beam search combined with gradient descent fits atom types and coordinates to generated density grids, after which distance and valence constraints assign bonds.Long bonds up to 4Å may be allowed, and energy minimization resolves inaccurate generated geometry.Source: §2.2 · p. 3

2.3 Generative model architecture and training

  • Conditional VAEs with a GAN loss encode receptor and ligand grids separately, sample a variational ligand latent space, and decode receptor-conditioned ligand densities.The receptor latent space is deterministic, conditions the ligand prior and posterior, and is concatenated with ligand latents before decoding.Source: §2.3 · pp. 3–4

2.4 Data sets

  • Training uses 725,048 low-RMSD poses from CrossDocked2020, whose clustered splits reduce target overlap and whose cross-docking expands available structures.Only poses with RMSD below 2Å were selected for high-quality structure generation.Source: §2.4 · p. 4

2.5 Evaluation procedure

  • Evaluation samples molecules from posterior or prior ligand latents, then measures validity, uniqueness, similarity, optimized RMSD, and predicted binding affinity.Each of ten test complexes receives 1,000 generations, while variability factors scale latent variance to sample farther from the seed ligand.Source: §2.5 · pp. 4–5, Figure 3 · p. 5

3 Results

The model generates mostly valid and chemically unique molecules, with increasing variability producing greater novelty but generally weaker predicted affinity and larger optimization changes. Receptor conditioning shapes generated molecular properties and poses, while higher variability can produce unconventional scaffolds and altered binding modes.

3.1 Molecular validity, uniqueness and variability factor

  • Over 80% of samples are valid at variability factor 1.0, with more than 90% chemically unique and uniqueness approaching 100% as variability increases.Source: §3.1 · p. 5

3.2 RMSD of optimization

  • Median RMSD change rises from 1.65 to 3.27 across variability factors 1.0 through prior sampling, although receptor conditioning keeps deviations generally reasonable.The reported medians are 1.65, 1.77, 2.09, 2.53, and 3.27 for variability factors 1.0, 1.5, 3.0, 5.0, and prior sampling, respectively.Source: §3.2 · pp. 5–6

3.3 Molecular size, similarity, and affinity scores

  • Higher variability shifts molecules toward smaller size, lower seed similarity, and lower predicted affinity, although every target yields some molecules outperforming the seed at variability factor 1.0.Only complexes 3h7w and 5e19 exceed 15% of molecules with better Vina affinity, at 19% and 17%, respectively.Source: §3.3 · p. 6

3.4 Relationship between affinity score and similarity

  • Affinity and similarity need not correlate: 5e19 has Pearson correlation 0.04 while generating the best-scoring ligands.The highest reported correlation is 0.55 for 4oq3 at variability factor 1.0, but many cases show no correlation.Source: §3.4 · pp. 6, 8

3.5 Conditionality on receptor

  • Receptor-conditioned and receptor-free sampling differ in energy, RMSD, and TPSA distributions, with hydrophobic sites producing lower-TPSA molecules than solvent-exposed sites.Receptor-free structures have higher unfavorable pre-optimization energy and significantly higher TPSA, while optimization substantially reduces the energy difference.Source: Figure 4 · p. 6, §3.5 · pp. 8, 10, Figure 6 · p. 8

3.6 Example structures

  • Increasing variability produces chemically and structurally distinct poses, including scaffold loss, disrupted hydrogen bonds, and unconventional scaffolds at high variability.For 3vrj, factors above 1.5 lose the seed scaffold; for 4ps7, factors 1.0 and 1.5 retain it, whereas factors 3.0 and 5.0 increasingly disrupt or replace it.Source: Figure 8 · p. 9, §3.6 · p. 10

4 Discussion

The model demonstrates feasible conditional generation of fully three-dimensional molecules within receptor binding pockets. Its usefulness is strongest near a reference seed, while broader latent-space sampling increases structural change and reduces conformational and binding-quality outcomes.

4 Discussion

  • The model generated valid molecules at a rate greater than 80% on average, with most valid molecules unique.Sampling near a reference seed produced novel structures with better predicted binding affinities and similar size to the seed.Source: §4 · p. 10
  • Receptor-conditioned 3D generation is enabled by conditional VAEs and GANs, with receptor information propagated across spatial resolutions to the decoder.The model combines atom-density grids with atom fitting and bond-adding algorithms to generate fully three-dimensional structures within binding pockets.Source: §4 · p. 10
  • Generated structures usually require only minimal geometry correction, with changes below 2Å RMSD, although larger variability produces greater structural changes and fewer locally optimal conformations.This limitation reflects the trade-off between exploring more varied structures and retaining suitable molecular geometry.Source: §4 · p. 10
  • Future improvements should regularize the ligand latent space for drug-likeness and explicitly optimize binding affinity or other molecular properties.The authors state that these approaches are likely necessary to generate structures from receptor structure without a reference seed ligand.Source: §4 · p. 10

Supplementary Materials

Figure S1 presents how affinity scores and molecular similarity relative to seed molecules change as the variable factor changes.

Supplementary Materials

  • Figure S1 tracks changes in affinity scores and molecular similarity relative to seed molecules as the variable factor changes.The figure summarizes distributional changes rather than reporting a single aggregate result.Source: Figure S1 · p. 13

Related papers

Loading 2010.14442v3…